llm - 2026_06
Navigation
Papers
Large language models (LLMs) reach high accuracy in mathematical reasoning, but individual traces on the same problem diverge; some arrive at the correct answer while others fail. Prior work analyzes failure at the step, chunk, or sentence level, or at tokens where failure has already occurred. Neither identifies the precise token that triggers the shift toward failure. We introduce the cliff token, a token where the token-wise potential drops significantly under an adaptive threshold that scales with the local token-wise potential, based on a one-sided two-proportion z-test. Across seven models and three mathematical reasoning benchmarks (GSM1K, MATH500, AIME 2025), cliff tokens act as failure triggers; deleting the first cliff token and resampling recovers pass@64 to 1.0, while keeping it limits recovery to between 0.71 and 1.00. We further introduce a cliff taxonomy of deterministic, uncertain, and sampled-off cliffs, defined by greedy choice and token entropy. Each type has distinct probabilistic characteristics, and the taxonomy generalizes across model scales. Finally, we validate the taxonomy via single-token preference optimization at cliff positions (Cliff-DPO). Trained on GSM8K, Cliff-DPO improves accuracy across benchmarks by up to +6.6. Optimizing at uncertain and sampled-off cliffs improves reasoning, while deterministic cliffs do not.
Security misconfigurations in Terraform Infrastructure-as-Code are a growing risk in cloud deployments, and large language models are increasingly used as automated repair agents. Existing evaluations often treat a repair as successful when the targeted static-analysis finding disappears, without checking planning validity, behavioral change, or security intent. This paper presents TerraProbe, a five-layer oracle framework for evaluating LLM-assisted Terraform security repair. We apply TerraProbe to 288 first-pass repairs generated by gemini-2.5-flash-lite, GPT-4o, and Claude 3.5 Sonnet across 68 real-world TerraDS modules and 28 controlled injected-defect modules. The results show that targeted Checkov removal overstates repair success. Although targeted removal reaches 83.3 percent for the primary model, full-scanner cleanliness drops to 10.4 percent, Terraform planning succeeds for 39.6 percent, and plan comparison is reachable for 38.5 percent. Human adjudication further shows that 71.4 percent of plan-compared real-world repairs are deceptive fixes that pass automated checks while leaving the underlying vulnerability in place. This pattern is statistically indistinguishable across the three models, with deceptive-fix rates from 57.1 percent to 71.4 percent and pairwise Fisher exact p-values above 0.10. The paper introduces a four-dimensional taxonomy of deceptive fixes, validated with Cohen kappa of 0.78 and Krippendorff alpha of 0.76. IAM permission analysis confirms that wildcard Resource grants persist in all nine CKV2 AWS 11 deceptive-fix cases. TerraProbe contributes an evaluation methodology, a replication package, and the Multi-Layer Oracle Evaluation framework for distinguishing intent-aligned security repairs from scanner-passing false successes.
Low-bit floating-point formats and semi-structured sparsity are increasingly supported by modern accelerators, yet combining them for LLM activation compression remains challenging: activations contain input-dependent outliers that dominate block scales in FP4 quantization, and directly applying N:M sparsity masks discards moderate values, coupling sparsification loss with quantization error. We introduce SharQ, a training-free inference method that bridges activation sparsity and FP4 quantization through an online sparse--dense decomposition. For each activation tensor, SharQ generates an input-adaptive N:M mask to extract an outlier-dominated sparse backbone, quantizes it to FP4, and defines a dense residual relative to the quantized sparse backbone rather than the unquantized sparse values. A sparse FP4 GEMM processes the backbone while a dense FP4 GEMM compensates for both mask-induced activation loss and sparse-path quantization error. The two paths share a single FP4 weight payload with path-specific scale views, and a fused preparation kernel absorbs mask generation, residual construction, and layer normalization into one operator. SharQ requires no calibration data, retraining, or model-specific tuning. Evaluated on Llama-3.1-8B, Qwen2.5-7B, Qwen3-30B-A3B, and Qwen3-VL-8B, SharQ recovers 43--63% of the NVFP4-to-FP16 accuracy gap across language and vision-language tasks, and generalizes across NVFP4, HiF4, and MXFP4 formats. On an RTX 5090, SharQ delivers 2.2--2.4$\times$ latency reduction over FP16 and 1.2--1.4$\times$ throughput improvement over FP8 in language model serving, and up to 1.58$\times$ speedup on Wan2.2-T2V-A14B video generation when combined with SageAttention. Our code is available at https://github.com/actypedef/SharQ.
Prediction markets rest on the independence of participant errors. As LLM agents become active traders on platforms like Kalshi and Polymarket, we ask: does this independence hold when the crowd is composed of LLMs? We find it does not. LLM agents fine-tuned with Direct Preference Optimization (DPO) share a convergent output distribution, producing pairwise error correlations of $ρ= 0.70$ and reducing ten agents to the effective forecasting power of ${\approx}1.4$ independent forecasters $N_{\text{eff}}$. This is not a scaling problem: $N_{\text{eff}}$ remains flat from $N=5$ to $N=40$, and the 10-agent market (67.6%) fails to match a single standalone agent (70.2%). Two controlled ablations isolate preference optimization as the causal driver, replicated across labs and scales ($Δρ= +0.24$ to $+0.46$ on identical-SFT controls at 8B and 70B). Among mitigations tested, cross-model diversity achieves the largest correlation reduction ($ρ$ from 0.68 to 0.40). As LLMs become more aligned, markets built from them become more monocultural.
Concolic testing combines concrete execution with symbolic constraint solving, but Python programs expose recurring limits. Library calls can cause symbolic variables to downgrade to concrete values. Regular expressions, checksums, parsers, and other semantic operations can be hard to solve, and exploration can plateau on already covered paths. We present ConcoLixir, a reactive LLM extension for Python concolic execution. The LLM acts as a discovery oracle, not a replacement for the solver or a correctness oracle. It generates initial seeds, proposes concrete inputs after solver failures, and targets uncovered code when coverage stalls. Each candidate is executed concolically, and only observed coverage and collected path constraints guide later exploration. Across synthetic, real-world, and library targets, ConcoLixir improves mean line coverage over the baseline concolic tester without an LLM oracle by 8.6, 15.1, and 17.0 percentage points. The gains are strongest near semantic barriers and library boundaries, and the full evaluation costs \$1.63 in API charges. These results show that bounded LLM discovery can complement symbolic reasoning without replacing it.
Modern software development increasingly involves the use of large language models (LLMs) to generate code. Despite their rapid advancement, LLMs remain prone to errors and hallucinations, emphasizing the importance of careful code inspection. However, in practice, developers' trust in LLM-generated code and their willingness to review it thoroughly may differ from these recommendations. How developers actually behave when reviewing LLM-generated code remains largely unexplored. In this study, we conduct a Wizard-of-Oz experiment to examine how software engineers behave when code is explicitly labeled as LLM-generated during a code review task. We collect both behavioral data and participant feedback through eye-tracking and exit interviews. Combining Bayesian data analysis with qualitative analysis, we found that while the thoroughness of code review did not change for participants, they spent more time fixating on LLM-labelled code, indicating that the label itself influences attention. Practitioners also adapted their review strategy for LLM-labelled code by assessing the code based on specific criteria (e.g., logical correctness), or using the prompt to guide their review. These findings inform LLM-based tool design on labelling while incorporating the prompt as a software artifact. Our study reveals a gap between reviewers' intentions and actual reviewing behaviour, highlighting the need for software companies to revisit their AI policies (particularly regarding LLM-assisted development) to better support developers in reviewing LLM-generated code.
Static verification tools can assure industrial scale software, but require significant human labor to write specifications. This is particularly true of static verifiers based on separation logic (SL verifiers), which excel at verifying heapmanipulating programs, but require many complex auxiliary specifications to reason about heap structure. Recent work applies large language models (LLMs) to generate code, tests, and proofs, including specifications for verifiers, but mostly targeting non-SL verifiers. To address this gap, this paper thoroughly evaluates how well LLMs perform when prompted to generate specifications for verifying 303 C functions with the SL verifier VeriFast. We explored eight prompting approaches, ten LLMs, and three input types in two stages. Quantitative and qualitative analyses are used to assess the LLM-generated code and specifications for functional behavior, verifiability and errors. The results show that LLMs preserve functional behavior in source code and specifications (both over 91%), but achieve modest verification success (31.4%). Using Gemini 2.5 Pro and providing formal contracts lead to higher success rates in our setting. Moreover, most errors (94%) come from LLMs' mistakes in the domainspecific knowledge of SL verifiers such as VeriFast. These findings provide guidance for optimizing LLM-generated specifications for SL verifiers.
News media play a central role in shaping public perceptions of climate change, and whether coverage emphasizes threats or solutions has measurable effects on audience engagement and policy support. Automated detection of these framing patterns at the sentence level would allow researchers to analyze large corpora that are infeasible to code manually. We present a systematic comparison of two approaches for classifying sentences from German-language climate news articles as threat-oriented, solution-oriented, both, or neither. The first approach uses few-shot prompting with an open-weights large language model (Llama 4 Maverick), employing chain-of-thought reasoning and structured output with confidence scoring. The second approach fine-tunes a German BERT model (deepset/gbert-large) for sentence-pair classification, where the preceding sentence provides contextual information for the target sentence. Both approaches implement two independent binary classifiers, one for threat framing and one for solution framing. We evaluate both methods on a corpus of 440 Austrian newspaper articles that were manually coded following a detailed coding scheme developed with domain experts. The fine-tuned BERT classifiers achieve an F1 score of 0.83 for both the threat and solution tasks, while the LLM-based classifiers reach an F1 of 0.78. An ablation study confirms that providing the preceding sentence as context improves BERT classification performance substantially compared to single-sentence input. These results contribute to the growing body of work comparing fine-tuned encoder models with prompted generative models for text classification in computational social science.
Large language models (LLMs) are attractive for context-aware time series forecasting because they can integrate heterogeneous textual signals, yet their discrete, language-oriented tokenization and embedding interfaces are misaligned with continuous numerical values, often harming numerical ordering and forecasting reliability. We propose TempoWave, a plug-and-play temporal wavelet digit interface that maps each scalar observation into digit-wise embeddings constructed from multi-wavelet, multi-scale coefficients. By directly overriding standard token representations, TempoWave seamlessly exposes both fine-grained local fluctuations and macro global structures in a transformer-compatible form, ensuring that precise numerical formatting, distinct digit identity, and robustness to common normalization operations are maintained throughout the LLM pipeline. Experiments across five context-enriched forecasting benchmarks demonstrate that TempoWave consistently improves LLM-based forecasters over standard numeric tokenization and alternative embedding interfaces, achieving a new state-of-the-art. These results highlight the numeric interface as a key bottleneck and suggest that principled multi-resolution embeddings can better couple LLMs' contextual reasoning with precise forecasting. Our code is available at https://github.com/DC-research/TempoWAVE and our model can be accessed at https://huggingface.co/Melady/TempoWAVE.
Recent work (2024 to 2026) has converged on a strategy for defending tool-using LLM agents against indirect prompt injection: rather than training the model to refuse malicious instructions, enforce security outside the model with a deterministic policy that mediates the agent's actions. Systems such as CaMeL, FIDES, Progent, RTBAS, and FORGE realize this with capabilities, information-flow labels, and reference monitors, and several report near-elimination of attacks on the AgentDojo benchmark. We make two contributions. First, we organize these out-of-band defenses as instances of classical integrity protection (Biba), reference monitoring, and least privilege, yielding a structured comparison of what they do and do not cover. Second, we warn that every one of them is validated only on static benchmarks (a fixed set of injection attempts), the same methodology that made in-band defenses look strong until adaptive, defense-aware attacks broke twelve of them at over 90% success; we specify the threat model and protocol an adaptive evaluation requires. We then run that protocol as an independent reproduction and extension of Progent's own adaptive-attack analysis, on AgentDojo, with an open-weight agent (Qwen2.5-7B) self-hosted on a single H200, a setting its authors did not test. Averaged over three runs, the defense held: Progent cut mean attack success roughly sixfold (25.8% to 4.2%), and a hand-crafted adaptive attack did not raise it (2.6%). This is one small-scale data point on a weak model with a single black-box attack template; a stronger optimized (white-box GCG) attack remains open. The result is consistent with, but does not establish, the hypothesis that deterministic out-of-band enforcement is a harder target for an adaptive attacker than in-band detection.
With advances in autonomous driving and electric vehicle technologies, functional safety has become a critical requirement in automotive chip design. Traditional simulation-based fault analysis is often overly conservative at the module level and fails to accurately reflect fault criticality. This paper presents SafeGen, an LLM-driven, formal-verification-assisted framework for functional-safety-oriented fault criticality assessment. SafeGen leverages large language models (LLMs) and a document-level Hyper Knowledge Graph (HyperKG) that incorporates Failure Modes, Effects, and Diagnostic Analysis (FMEDA) guidelines to extract verifiable specifications from design and safety documents and evaluate their relevance to overall system safety. The HyperKG is further enriched with register-transfer-level (RTL) information to guide the generation of Functional Safety Assertions (FSAs) that are both semantically grounded and design-aware. Each assertion is linked to its corresponding specification, enabling traceable reasoning throughout the assessment process. A gate-to-RTL fault-mapping mechanism supporting both stuck-at and bridging faults, combined with formal property verification (FPV), enables semantic-level fault criticality grading based on specification-linked assertion violations. A digital-physical co-simulation platform for a field-oriented control (FOC) system is developed to validate SafeGen. Experimental results demonstrate that SafeGen generates higher-quality assertions than existing LLM-based assertion generation frameworks while providing greater semantic interpretability in fault criticality assessment compared with traditional simulation-based approaches.
Large language models are widely employed as evaluators, a paradigm commonly referred to as LLM-as-a-judge. Prior research has predominantly examined point-wise or pair-wise evaluation protocols; in contrast, our focus is on rubric-based evaluation, which has been attracting increasing attention owing to its utility for training models in domains where verification is otherwise difficult. In this work, we show that rubric-based evaluation implicitly resembles a multiple-choice setting and therefore exhibits position bias: LLMs tend to prefer score options that appear at specific positions within the rubric list. Through controlled experiments across multiple models and datasets, we demonstrate that this position bias is consistent. Its direction, however, is model-specific: some judges favor the first option, while others favor the last. We further identify a second, orthogonal axis of bias: when a prompt scores several criteria simultaneously, the ordering of the criteria itself shifts the resulting scores. We additionally explore permuting the order of the rubric options as a means of mitigating position bias, and find that although the bias can be attenuated, improvements in the correlation between model judgments and human annotations are obtained primarily for models that exhibit strong bias. Our results recast rubric-based LLM-as-a-judge as a multiple-choice problem with measurable, model-specific position bias, and we further confirm that only a small number of random order permutations are sufficient to reduce the error introduced by this bias for the majority of models.
LLM-based coding agents rely on \emph{skills}, pre-packaged instruction sets that extend agent capabilities, yet every token of skill content injected into the context window incurs both monetary cost and attention dilution. To understand the severity of this problem, we conduct a large-scale empirical study of 55,315 publicly available skills and find systemic inefficiencies: 26.4\% lack routing descriptions entirely, over 60\% of body content is non-actionable, and reference files can inject tens of thousands of tokens per invocation. Motivated by these findings, we present \textsc{SkillReducer}, a two-stage optimization framework. Stage~1 optimizes the routing layer by compressing verbose descriptions and generating missing ones via adversarial delta debugging. Stage~2 restructures skill bodies through taxonomy-driven classification and progressive disclosure, separating actionable core rules from supplementary content loaded on demand, validated by faithfulness checks and a self-correcting feedback loop. Evaluated on 600 skills and the SkillsBench benchmark, \textsc{SkillReducer} achieves 48\% description compression and 39\% body compression while improving functional quality by 2.8\%, revealing a \emph{less-is-more} effect where removing non-essential content reduces distraction in the context window. These benefits transfer across five models from four families with a mean retention of 0.965, and generalize to an independent agent framework.
Knowledge Editing (KE) has emerged as a frontier for updating specific facts in LLMs without costly retraining, but its reliability and underlying mechanisms remain poorly understood. In this work, we examine KE from an adversarial elicitation perspective, revealing that edited knowledge is often not fully erased and continues to surface, with consistent failures observed across diverse model architectures. To explain this behavior, we conduct a mechanistic analysis of popular KE methods. We show that low-rank updates do not overwrite existing knowledge but instead redistribute it within the model's representation space. Furthermore, we find that these methods act as targeted suppression mechanisms that reduce the likelihood of expressing original facts, rather than removing them from the model. Analysis of the loss landscape reveals that edited knowledge lies in narrow, anisotropic regions that are highly sensitive to perturbations, making them highly vulnerable to indirect prompting and adversarial attacks. By exposing these profound architectural vulnerabilities, our work proves that KE algorithms are inherently bypassable and motivates a fundamental reevaluation of how we deploy post-hoc updates in several LLM applications.
Large Language Models (LLMs) are increasingly deployed in tool-driven environments such as healthcare analytics, financial systems, retrieval-augmented generation (RAG), and multi-agent workflows. Although the Model Context Protocol (MCP) standardizes how LLM applications expose and invoke external tools, its baseline model does not require tool-use manifests to be cryptographically authenticated, freshness-checked, policy-bound, or independently auditable before execution. As a result, MCP pipelines may remain vulnerable to manifest tampering, unauthorized tool invocation, replay of stale requests, and weak accountability. This paper presents a manifest-level enforcement layer for MCP-based LLM pipelines. It treats each MCP tool-use manifest as a first-class security object whose canonical form must be policy-validated, freshness-checked, digitally signed, verified before execution, and linked to tamper-evident audit evidence. The framework binds tool invocation to verifiable manifest integrity and fail-closed authorization, separates user-visible request parameters from execution metadata, rejects non-compliant or stale manifests before execution, and records accepted invocations in a Merkle-based transparency log. Evaluation across GPT-5.3, LLaMA-3.5, and DeepSeek-V3 using up to 50,000 manifest instances shows near-linear scalability (R^2 = 0.998), bounded verification latency (<= 9.4 ms on edge devices), and rejection of expired, malformed, replayed, and policy-violating manifests, with rejection rates above 98.7%. Experiments in healthcare, finance, RAG, and multi-agent settings show that manifest-level cryptographic enforcement enables low-overhead, traceable, and auditable execution control for heterogeneous LLM-tool pipelines.
Large language models (LLMs) reproduce homogeneity bias -- the tendency to portray marginalized groups as more internally similar than dominant groups -- but whether this bias is stable or an artifact of inference settings has only been studied in single proprietary models. We map homogeneity bias across a 5x5 temperature-by-top-p grid in seven open-weight instruction-tuned LLMs (7-20B parameters). Hispanic and Asian Americans are portrayed as more homogeneous than White Americans in at least 18 of 20 hyperparameter configurations across six of seven models, including at extreme sampling settings. African American and gender bias show model-specific variation in direction. A conservative cell-level re-analysis confirms Hispanic and Asian homogeneity as robust, while weaker African American and gender signals largely do not survive, establishing group-specific robustness. We also apply the same grid to a names-based paradigm in which group identity is signaled via racially distinctive surnames rather than explicit labels. The names paradigm corroborates Hispanic and Asian homogeneity bias, but Black-coded surnames elicit robustly less homogeneous outputs than White-coded names in every model tested -- a reversal absent from the label paradigm -- showing that how group identity is operationalized shapes which biases surface and in which direction.
We present KernelPro, a closed-loop multi-agent system that automatically generates, profiles, and iteratively optimizes GPU kernel code by integrating large language model (LLM) code generation with hardware profiler feedback and pluggable bottleneck detection tools. KernelPro introduces four contributions: (1) a semantic feedback operator that encodes expert heuristics as pluggable micro-profiling tools, transforming raw hardware metrics into actionable natural language guidance; (2) a two-stage tool invocation architecture where roofline-based bottleneck classification filters which specialized analysis tools execute, combining kernel-level (ncu), instruction-level (SASS), and system-level (nsys) profiling; (3) a domain-adapted MCTS with progressive widening, asymmetric branching, log-reward calibration, dead-end pruning, and search memory for cross-iteration learning; and (4) direct CuTe source-level code generation via autonomous code search over the CUTLASS/CuTe codebase. On KernelBench, KernelPro achieves geometric mean speedups of 2.42x/4.69x/5.30x on Levels 1/2/3, establishing state-of-the-art performance across all difficulty levels. On VeOmni's expert-optimized MoE training kernels, KernelPro achieves 1.23x over hand-tuned Triton by generating a from-scratch raw-CUDA+CuTe Hopper WGMMA kernel. Ablation studies demonstrate that each design component independently and significantly improves optimization quality: micro-profiling tools (p < 0.0001 vs raw metrics), MCTS search (26% higher geometric mean vs greedy, p = 0.004), and proactive tool orchestration (23% improvement, p = 0.035). Finally, KernelPro is the first CUDA kernel coding agent to optimize energy efficiency beyond the speed-only focus of prior systems, demonstrating an 11.6% measured energy reduction at matched speed.
Current LLM evaluation relies on two complementary but often disconnected signals: static benchmarks with objective correctness labels and arena-style preference data that better reflect open-ended user interactions. We introduce DualEval, a latent model-item calibration framework that represents models and evaluation items in a shared space, jointly estimating model ability together with item difficulty and sharpness. We apply DualEval across four domains: coding, math, miscellaneous domain-knowledge tasks, and generic everyday user queries. Our evaluation uses 18 frontier LLMs, static benchmark labels, and reward-model scores validated against held-out human preferences for open-ended model responses. Empirically, our framework produces reliable and balanced model rankings, and its learned item-level profiles support downstream applications such as benchmark compression for sample-efficient evaluation and anomaly detection for contamination or outlier analysis. Overall, DualEval unifies static and arena-style evaluation through joint model-item calibration, producing model rankings and item-level diagnostics that support more sample-efficient, interpretable, and auditable evaluation pipelines.
Large language models (LLMs) are increasingly deployed in interactive applications, yet they remain vulnerable to adversarial interactions that induce harmful, deceptive, or policy-violating outputs. Existing defenses typically analyze either user prompts or generated outputs, but not both. However, many real-world attacks exploit a separation between adversarial intent expressed in the prompt and actionable harm manifested only in the response. As a result, prompt-only and response-only defenses frequently miss unsafe interactions that appear benign when viewed from either side in isolation. We present a verification-centric defense framework that jointly evaluates prompt intent and response harm before an LLM response is delivered to a user. The framework employs specialized analysts for intent and harm assessment together with a Judge for conflict resolution. We formalize a threat model for prompt-response attacks and evaluate the framework across five threat categories: jailbreaks, prompt injection, phishing, cyber abuse, and harmful content. Experiments on multiple benchmark datasets show that jointly verifying prompt intent and response harm consistently outperforms single-sided defenses and single-agent reasoning baselines. Across threat categories, the framework improves average F1 from 0.90 for the strongest applicable baselines to 0.95 while reducing the average attack success rate to 4.1 percent. Compared with a Single-Agent+CoT baseline, it improves average F1 from 0.87 to 0.95 and reduces the false positive rate on benign-sensitive requests from 0.12 to 0.06. We further evaluate architecture-aware adaptive attacks in which the attacker knows the verifier structure and attempts to bypass individual verification components. Our results suggest that prompt-response verification provides a practical foundation for securing LLM applications against evolving adversarial threats.
Reinforcement Learning with Verifiable Rewards (RLVR) is highly effective for enhancing LLM reasoning, yet recent evidence shows models like Qwen 2.5 achieve significant gains even with spurious or incorrect rewards. We investigate this phenomenon and identify a "Perplexity Paradox": spurious RLVR triggers a divergence where answer-token perplexity drops while prompt-side coherence degrades, suggesting the model is bypassing reasoning in favor of memorization. Using Path Patching, Logit Lens, JSD analysis, and Neural Differential Equations, we uncover a hidden Anchor-Adapter circuit that facilitates this shortcut. We localize a Functional Anchor in the middle layers (L18-20) that triggers the retrieval of memorized solutions, followed by Structural Adapters in later layers (L21+) that transform representations to accommodate the shortcut signal. Finally, we demonstrate that scaling specific MLP keys within this circuit allows for bidirectional causal steering-artificially amplifying or suppressing contamination-driven performance. Our results provide a mechanistic roadmap for identifying and mitigating data contamination in RLVR-tuned models. Code is available at https://github.com/idwts/How-RLVR-Activates-Memorization-Shortcuts.
Multi-LLM systems use multiple language models to deliberate, judge each other's outputs, or coordinate as agents. Their value depends on the models producing measurably different conversational behaviors when given the same input. Prior offline studies recommend drawing one model per family for behavioral diversity, because LLMs prefer outputs from their own family when rating one another in isolation. Whether the same family label predicts behavior in interactive multi-LLM systems, the setting that real deployed systems use, has not been tested. We study this with a 940,000-chain 11-checkpoint corpus and a 1.6M-chain same-base Llama factorial. On our validated headline metric, hedging, a reasoning-distilled Llama checkpoint shifts by 18% depending on which same-base partner it replies to, more than any cross-family hedging gap in the controlled subset. Qwen, closed-API, and runtime checks suggest the pattern is not isolated, while repair and challenge analyses remain exploratory because their surface-cue detectors are weaker. Overall, the results identify post-training recipe as a first-class axis for multi-LLM panel composition and show that model family alone is an incomplete proxy for conversational diversity.
Agentic benchmarks have emerged across general-purpose and domain-specific settings, including finance, coding, law, and drug discovery, yet energy-domain evaluations remain largely limited to static knowledge recall. This is a critical gap for a sector that requires live data retrieval, specialized regulatory and market knowledge, and multi-step quantitative reasoning under real-world constraints. We present an empirical study of tool-augmented LLM agents on real-world energy market analytics tasks. Our evaluation environment includes 243 expert-curated problems across three categories: (1) Market Data Retrieval and Analysis, (2) Knowledge Retrieval and Interpretation, and (3) Advanced Quantitative Modeling and Decision Analytics. Tasks include price and demand analysis, tariff impact modeling, asset revenue and returns estimation, hedging strategy analysis, and optimization modeling, with problems spanning multiple difficulty levels. Agents are equipped with a configurable suite of domain tools, including live electricity market APIs for major U.S. ISOs, regulatory docket search, utility tariff databases, asset optimization models, and retrieval-augmented generation over energy market documents. We assess agent responses using a multi-dimensional evaluation protocol that scores approach correctness, answer accuracy, attribute alignment, and source validity, with category-aware routing to match scoring criteria to question type. We evaluate both closed-source and open-source LLMs, providing a comparative analysis of how model capability and domain tooling interact in a high-stakes professional domain. Key artifacts are publicly released to support reproducibility and future research.
Cognitive reappraisal is a well-studied emotion regulation strategy that helps individuals reinterpret stressful situations to reduce their impact. Many digital mental health tools struggle to support this process because rigid scripts fail to accommodate how users naturally describe stressors. This study examined the feasibility of an LLM-based single-session intervention (SSI) for workplace stress reappraisal. We assessed short-term changes in stress-related outcomes and examined design tensions during use. We conducted a feasibility study with 100 employees at a large technology company who completed a structured cognitive reappraisal session delivered by a GPT-4o-based chatbot. Pre-post measures included perceived stress intensity, stress mindset, perceived demand, and perceived resources. These outcomes were analyzed using paired Wilcoxon signed-rank tests with correction for multiple comparisons. We also examined sentiment and stress trajectories across conversation quartiles using two RoBERTa-based classifiers and an LLM-based stress rater. Open-ended responses were analyzed using thematic analysis. Results showed significant reductions in perceived stress intensity and significant improvements in stress mindset. Changes in perceived resources and perceived demand trended in expected directions but were not statistically significant. Automated analyses indicated consistent declines in negative sentiment and stress over the course of the interaction. Qualitative findings suggested that participants valued the structured prompts for organizing thoughts, gaining perspective, and feeling acknowledged. Participants also reported tensions around scriptedness, preferred interaction length, and reactions to AI-driven empathy. These findings highlight both the promise and the design constraints of integrating LLMs into DMH interventions for workplace settings.
Supervised classification on tabular data remains a central machine learning task, but its dependence on large labeled datasets limits its applicability in data-scarce settings. Few-shot methods such as TabPFN achieve strong performance through large-scale synthetic pretraining, yet still require labeled context examples. Large Language Models (LLMs) offer a more flexible alternative through zero- and few-shot in-context learning from task descriptions, but their behavior on tabular data remains inconsistent. We introduce LLMTabBench, a benchmark for evaluating LLMs on tabular classification under low-data conditions. The benchmark studies how LLM prior knowledge interacts with task descriptions and few-shot examples, and how performance changes with increasing data complexity across real-world and controlled synthetic datasets. We find that LLMs can be highly competitive in zero-shot settings, sometimes outperforming models given few-shot examples. However, additional examples may conflict with prior knowledge, thereby degrading performance. We also observe a complexity threshold at which LLM performance declines and few-shot examples become less useful. These results clarify key limits of in-context learning for tabular data and inform the deployment of LLMs in low-data regimes.
Prior clinical psychology research shows that Western individuals with depression tend to report psychological symptoms, while Eastern individuals report somatic ones. We test whether Large Language Models (LLMs), which are increasingly used in mental health, reproduce these cultural patterns by prompting them with Western or Eastern personas. Results show that LLMs largely fail to replicate the patterns when prompted in English, though prompting in major Eastern languages (i.e., Chinese, Japanese, and Hindi) improves alignment in several configurations. Our analysis pinpoints two key reasons for this failure: the models' low sensitivity to cultural personas and a strong, culturally invariant symptom hierarchy that overrides cultural cues. These findings reveal that while prompt language is important, current general-purpose LLMs lack the robust, culture-aware capabilities essential for safe and effective mental health applications.
Sequential user behavior modeling is widely adopted in industrial recommender systems; however, significant gaps remain in financial services, where pre-login web interactions and authenticated in-app experiences differ drastically. Specifically, pre-login web users typically explore new products, whereas logged-in app users focus on account servicing. Due to the challenge of cross-channel entity resolution (e.g., matching anonymous web sessions to authenticated mobile accounts), web-based intent signals remain underutilized for post-authentication personalization. Existing methods for capturing web-based intent are often ad-hoc and narrow, lacking the flexibility to support both quantitative downstream recommendations and qualitative understanding at scale. In this work, we propose a scalable and dual-purpose intent prediction framework for web-based interactions and demonstrate its applicability for personalization. Our approach transforms raw web clickstreams into two outputs: a self-supervised Transformer encodes multi-modal clickstreams into a compact session embedding, while an LLM-based taxonomy generation and distillation pipeline produces interpretable intent labels. Our system demonstrates that self-supervised clickstream representations combined with LLM-distilled taxonomies can jointly serve quantitative tasks and qualitative understanding in production: on the mobile homepage tile ranking task, the session embedding improves macro Recall@1 by 1.88% and reduces Log Loss by 13.38% over production baselines. On the user conversion prediction task, the embedding outperforms the LLM labels by 4.3% on micro F1, while the distillation layer delivers interpretable labels at ultra-low latency with only a 7% performance drop.
Process reward models enable fine-grained, step-level evaluation of LLMs, yet building them for agentic settings remains prohibitively difficult: long-horizon interactions, irreversible actions, and stochastic environment feedback make both human annotation and Monte Carlo estimation infeasible at scale. In this work, we show that reinforcement learning (RL) post-training already provides the ingredients for effective step-level scoring, eliminating the need for dedicated reward model training altogether. Concretely, we derive an implicit advantage under a general stochastic Markov decision process, which we term progress advantage -- log-probability ratio between the RL-trained policy and its reference policy exactly recovers the optimal advantage function. This formulation makes the resulting signal annotation-free, domain-agnostic, and available as a byproduct of the standard RL post-training pipeline. We validate the effectiveness of the progress advantage across three different applications: test-time scaling, uncertainty quantification, and failure attribution on five benchmarks and four model families. Across all settings, it consistently outperforms confidence-based baselines and, despite requiring no task-specific training, surpasses dedicated trained reward models. We complement these results with deeper analyses on characteristics of progress advantage, offering practical guidance for adoption in real-world agentic systems.
As AI agent protocols proliferate, the governance structures shaping their interoperability standards remain empirically underexamined. We introduce an LLM-powered comparative pipeline for large-scale governance discourse analysis, integrating automated annotation, neural topic modeling, and multi-layer network analysis to study socio-technical power structures at scale. We validate it on two contrasting standards for agent interoperability: ERC-8004 (permissionless, on-chain) and Google A2A (corporate-led). Analyzing 4,323 governance participation records, we combine LLM-assisted coding, topic modeling, and multi-layer network analysis to examine how institutional design shapes thematic priorities and community structure. We find that while governance form influences substantive focus, both regimes exhibit comparable levels of participation inequality and community fragmentation. Discourse alignment is denser in the permissionless setting, suggesting that open governance may foster greater thematic convergence despite decentralized participation. These findings illustrate how LLM-assisted methods can advance the empirical study of technology governance, with implications for designing more equitable agentic AI standards. All data and code are openly available.
Large language model (LLM) agents increasingly automate complex tasks by integrating language models with external tools and environments. However, their autonomy poses significant safety risks: agents may execute destructive commands, leak sensitive data, or violate domain constraints. Existing safety approaches face a fundamental tradeoff: hand-crafted rules are interpretable but brittle, with overly conservative rules blocking safe operations (high false positives) while permissive rules miss unsafe behaviors (high false negatives). Neural classifiers lack the interpretability required for safety-critical deployments. We present AutoSpec, a framework that automatically evolves deployed expert-designed safety rules from user safe/unsafe annotations through counterexample-guided inductive synthesis (CEGIS) guided by inductive logic programming (ILP). Starting from the expert rules and a stream of annotated traces, AutoSpec iteratively evaluates rules, mines false-positive and false-negative counterexamples, uses ILP to learn which predicates discriminate them, generates candidate rule edits, and verifies candidates to select the best revision. The key insight is that ILP efficiently identifies predicates that appear frequently in false negatives but rarely in false positives (or vice versa), dramatically pruning the exponential search space of rule edits. This continues until convergence, producing interpretable rules that balance precision and recall. We evaluate AutoSpec on 291 execution traces spanning code execution and embodied agent domains. AutoSpec raises rule F1 to 0.98 and 0.93 across the two domains, achieving up to 94% false positive reduction while maintaining high recall, and converges within 4-5 iterations. The ILP-guided approach achieves up to 4.8x higher F1 than heuristic CEGIS. The learned rules are human-readable, auditable, and generalize to unseen scenarios.
Visual Question Answering (VQA) is the task of answering questions about images, requiring the integration of multimodal input and reasoning. Modular approaches that incorporate logic-based representations into the reasoning component offer clear advantages over end-to-end trained systems, particularly in terms of interpretability. However, adapting or extending these representations when task requirements change can place a significant burden on developers. To address this challenge, we present an approach for distilling rules from Large Language Models (LLMs). Our method prompts an LLM to extend an initial VQA reasoning theory, expressed as an answer-set program, to meet new requirements of the task. Examples from VQA datasets guide the LLM, validate the results, and help correct erroneous rules by leveraging feedback from the ASP solver. We demonstrate that our approach is effective across diverse VQA datasets. Notably, only a few examples are needed to elicit correct rules from LLMs. Our experiments suggest that rule distillation from LLMs is a promising alternative to traditional data-driven rule learning approaches. Under consideration in Theory and Practice of Logic Programming (TPLP).
Software vulnerability remediation is a cognitively demanding task that requires specialized security expertise often lacking in general developers. In the meantime, Large Language Models (LLMs) assisted tools show potential in vulnerability detection, location, and repair tasks. [Hypothesis:] While LLM-assistance is hypothesized to accelerate patching, it also risks introducing hallucinations or insecure code, leading to a higher likelihood of generating superficial repairs that bypass the standard functionality checks but fail the security validation. [Objective:] We aim to present an empirical experiment, unveiling the capability of LLM-assisted vulnerability patching compared to manual debugging on human participants in real-world scenarios. [Method:] We plan to conduct a controlled experiment using a Balanced Crossover design. For that, we have developed a WebApp for code execution and integrated hidden Ghost Tests to verify patch integrity beyond visible functional requirements. The experiment involves training and evaluation scenarios. The remediation speed, remediation efficacy for both standard functionality tests and security tests, and participant perception will be evaluated. [Pilot Study:] A pilot experiment with a small sample of participants has been conducted, providing insights for the following study.
On-device LLM inference is increasingly attractive for privacy-preserving, reliable, and cost-effective deployment, yet its energy and thermal costs remain a critical bottleneck. Existing systems primarily optimize for decoding speed, implicitly assuming that faster execution is always preferable. We show instead that on-device LLM inference often has exploitable configuration slack: modestly lowering NPU and memory frequencies preserves quality of experience (QoE) while substantially improving energy efficiency and reducing heat. Realizing this opportunity in production is challenging. The most energy-efficient NPU/DDR setting varies with the model, inference engine, platform, and runtime conditions, with no stable ranking across configurations. Commercial devices further lack component-level power sensing, and shell temperature evolves with request arrivals, response lengths, and thermal history. To address these challenges, we propose EnerInfer, the first on-device LLM inference framework that jointly manages energy efficiency, throughput, and thermal comfort for LLM workloads. EnerInfer replaces per-model profiling and sensor-heavy control with disaggregated, model-structure-aware prediction and ranking-driven online feedback. It predicts throughput and power for unseen LLMs across NPU/DDR frequency settings, selects QoE-satisfying efficient configurations under runtime interference, and uses lightweight limited-horizon thermal prediction to dynamically switch between energy-optimized and thermally constrained inference. Evaluations on real-world LLMs show that EnerInfer improves energy efficiency by up to 65%, 12%, and 24% on phones, a laptop, and a development board, respectively, without QoE violation.
Reinforcement learning (RL) is often credited with improving language model reasoning at the expense of knowledge. We challenge this narrative by showing that reasoning models consistently outperform their instruction-tuned versions on pure knowledge recall tasks. These gains do not reflect newly acquired information, but rather an improved procedural skill in navigating and searching existing knowledge hierarchies within the model parameters. Structured prompting, which explicitly guides models through hierarchical traversal -- recovers most of the instruct-reasoning gap across five model families. A controlled RL experiment on unseen, non-extractable facts improves recall of held-out frequent but previously inaccessible facts, ruling out simple data exposure. On depth-stratified retrieval tasks, reasoning models exhibit superior traversal as retrieval depth grows. Layerwise activation analysis further shows that while factual representations maintain high cosine similarity between instruct and reasoning models, query representations diverge noticeably, indicating that reasoning primarily reshapes how models traverse knowledge rather than the knowledge representation itself. Finally, we find that distilled models often fail to match reasoning models on knowledge recall because they imitate self-correction without acquiring the exploratory behavior needed for hierarchical navigation. Together, these findings suggest that improving factual recall in LLMs depends not only on expanding what models know but also on teaching them to navigate it -- motivating future post-training methods that optimize traversal.
Long-term human-agent dialogues are organized by topic continuity: adjacent turns often develop the same goal, plan, problem, or event, while related activities may recur across distant sessions. Yet many LLM agent memory systems first decompose histories into isolated turns or fixed-size chunks, then compensate through enrichment, consolidation, or retrieval mechanisms still tied to semantic proximity or fragment-level records. This weakens temporal and causal organization and biases memory access toward semantic proximity rather than task- or topic-level continuity. We introduce \emph{Membox}, a hierarchical memory architecture that instantiates topic continuity as an explicit organization layer for agent memory. Its \textbf{Topic Loom} incrementally organizes dialogue streams into boxes whose internal turns follow the same local topic, while its \textbf{Trace Weaver} links extracted events across boxes into macro-topic traces that recover recurring activities, goals, and factual developments across distant sessions. On LoCoMo, Topic-Loom-only retrieval improves over the best Mem0/A-MEM retrieval-depth setting by 13.00 F1 points (53.95 vs. 40.95), and trace-expanded retrieval further raises F1 to 55.28; with GPT-4o, trace-expanded retrieval reaches 59.71 F1. Additional DialSim results show the same gain from adding cross-box traces in multi-party dialogue. These results show that local topic-continuity organization and macro-topic trace expansion improve long-range memory beyond semantic retrieval over fragmented records.
Stakeholder interviews are an important source of information for requirements elicitation, yet many relevant requirements remain implicit in such conversations. Stakeholders frequently describe workflows, challenges, and operational practices without explicitly articulating the software capabilities that could address them. Recent work has considered the use of LLMs to analyze conversational data and extract requirements from stakeholder interviews. Existing approaches, however, primarily focus on identifying explicitly stated requirements, leaving implicit opportunities largely unexplored. In this paper, we present LENS (LLM-Enabled Needs Discovery from Stakeholder Interviews), an approach that analyzes stakeholder interview transcripts to both extract explicit requirements and infer additional latent requirements. LENS performs this inference by reasoning over stakeholder statements together with contextual information about organizational tools and infrastructure. Both extracted and inferred requirements are represented as user stories and linked to transcript excerpts to ensure traceability. We conduct a preliminary evaluation of LENS using twelve stakeholder interview transcripts collected in an industrial setting involving cybersecurity operations. We show that LENS achieves an average F1-score of 84.4% for extracting explicit requirements, while, on average, 75% of the latent requirements identified by LENS were perceived as providing useful automation or time-saving potential by domain experts.
The explosive demand for interactive Large Language Model serving has highlighted the management of the Key-Value cache's dynamic memory footprint as a critical area for performance optimization in inference engines. Modern inference systems overwhelmingly rely on time-centric scheduling heuristics, such as Shortest Job First. However, their theoretical optimality is rooted in traditional schedule modeling, failing to capture the highly dynamic, 2D spatio-temporal geometric growth specific to LLM inference mechanisms. To resolve this, we propose the geometry-aware online scheduling by introducing the Smallest Volume First (SVF) algorithm and its highly efficient variant, 1-bit SVF. Theoretically, we provide a rigorous mathematical foundation for our approach. Via a novel volume-certificate proof, we sharpen SVF's worst-case competitive ratio from the prior best of 48 towards \textbf{3} in the high-concurrency regime of LLM serving. Building upon this core breakthrough, we complete a comprehensive theoretical taxonomy analyzing our algorithms across different traffic scenarios and information availability. Practically, we seamlessly integrate our approach as a plug-and-play layer in vLLM. Extensive evaluations on Llama-3.1 models demonstrate comprehensive performance gains: SVF delivers strong reductions in both average and tail latency, while 1-bit SVF, with merely a single bit information, achieves competitive throughput and latency. This work establishes a theoretically sound and empirically proven approach for resolving memory-constrained scheduling in modern LLM deployments. To facilitate future research, our code is available at https://github.com/Aurora-Kl/Geometry-Aware-Online-Scheduling.git.
Group-based reinforcement learning effectively post-trains LLM agents for long-horizon, sparse-reward tasks by deriving step-level credit from trajectory outcomes. However, this ties a step's credit to its rollout's final outcome: semantically near-identical intermediate steps receive opposite credit depending on whether their trajectory eventually succeeded or failed. Such semantic credit inconsistency sends conflicting gradients to similar actions and wastes the partially-correct progress inside failed rollouts. Motivated by this, we propose Semantic Consistency Policy Optimization (SCPO), a value-free reward-shaping method that mitigates this inconsistency by recovering step-level credit from successful siblings in the same rollout group. Concretely, SCPO scores each failed step against a successful sibling and adds positive step-level credit for new progress along that sibling. On ALFWorld and WebShop, SCPO matches or exceeds strong group-based baselines, reaching 93.7+/-4.1 percent success on ALFWorld and 74.8+/-2.0 percent on WebShop at 1.5B parameters, with gains concentrated on the hardest multi-step tasks.
Automated methods for discovering mechanistic simulator models from observational data offer a promising path toward accelerating scientific progress. Such methods often take the form of agentic-style iterative workflows that repeatedly propose and revise candidate models by imitating human discovery processes. However, existing LLM-based approaches typically implement such workflows via hand-crafted heuristic procedures, without an explicit probabilistic formulation. We recast model discovery as probabilistic inference, i.e., as sampling from an unknown distribution over mechanistic models capable of explaining the data. This perspective provides a unified way to reason about model proposal, refinement, and selection within a single inference framework. As a concrete instantiation of this view, we introduce ModelSMC, an algorithm based on Sequential Monte Carlo sampling. ModelSMC represents candidate models as particles which are iteratively proposed and refined by an LLM, and weighted using likelihood-based criteria. Experiments on real-world scientific systems illustrate that this formulation discovers models with interpretable mechanisms and improves posterior predictive checks. More broadly, this perspective provides a probabilistic lens for understanding and developing LLM-based approaches to model discovery.
LLM-as-a-Judge has become the dominant paradigm for evaluating language model outputs, yet LLM judges exhibit systematic biases that compromise evaluation reliability. We present a comprehensive empirical study comparing nine debiasing strategies across five judge models from four provider families (Google, Anthropic, OpenAI, Meta), three benchmarks (MT-Bench n=400, LLMBar n=200, custom n=375), and four bias types. Our headline practical finding is that a mid-tier model with the right debiasing can outperform frontier judges at a fraction of the cost: Gemini 2.5 Flash with the Combined Budget strategy reaches the highest agreement of any configuration we tested (71.0%, kappa=0.549) at ~$0.001 per evaluation, about 15x cheaper than the best frontier setup (Claude Sonnet 4, 69.5%, ~$0.015). Other key findings: (1) Style bias is the dominant bias (0.10-0.76 across models, favoring markdown over plain prose), far exceeding position bias (<=0.04), yet is rarely studied. (2) Verbosity bias is heterogeneous when measured length-aware: Pro, Flash, and Llama prefer longer answers (+0.24 to +0.44), Claude prefers concise (-0.12), and GPT-4o is neutral (-0.04); on truncation controls all models correctly prefer the complete response (0.88-1.00 accuracy). (3) Debiasing helps multiple models: Claude S8 (+11.5 pp), Flash S8 (+7.5 pp), and Claude S5 (+7.3 pp) survive Holm-Bonferroni correction, with Flash S1 (+4.7 pp) and Llama S8 (+4.5 pp) also significant. We release our evaluation framework, the 375-pair controlled dataset, and per-instance cached results for all nine strategies.
LLM-as-judge ("grader") components are now standard in evaluation harnesses, including safety evaluations where a pass/fail verdict may gate downstream deployment decisions. A widespread assumption is that setting the grader's sampling temperature to 0 makes grading deterministic. We test this assumption against a real safety-evaluation codebase (Japan AISI's open-source aisev) and show it fails on two levels. First, the harness invokes its grader without setting temperature or seed; the underlying provider silently applies its default of 1.0, so items near the decision boundary flip pass/fail across identical runs (per-item disagreement up to ~50% over 20 runs). Second, pinning temperature=0 reduces but does not eliminate flips: across 690 API calls spanning two providers, three model tiers, and five sampling configurations, 1-2 of 7 borderline items remain non-reproducible even under forced greedy decoding (top_k=1). Claude Opus 4.7/4.8 has since deprecated temperature entirely, rendering the primary mitigation inapplicable to newer model generations. These findings expose a structural gap: evaluation harnesses that report single-run verdicts without variance or grader-disagreement metrics can present noise as a safety property. We release a reproduction harness (690 calls, 7 conditions) and recommend that harnesses treat grader disagreement as a first-class health metric alongside the scores themselves.
Large language models (LLMs) are increasingly used to generate personalized guidance in wellbeing contexts such as physical activity, stress management, and mental health support, enabling fluent and context-aware interaction but relying on largely implicit mechanisms that shape how recommendations are expressed and adapted. We argue that this reliance on implicit adaptation through prompting and alignment limits control over guidance, responsibility framing, and user influence, which is particularly problematic in wellbeing settings where recommendations affect users' actions and long-term outcomes. We propose a system-level perspective in which conversational behavior is structured through explicit interaction constraints, including guidance strategies, explanation styles, degrees of directness, and mechanisms for user control. Building on prior work on tangible recommendations, we show how these constraints address key challenges in wellbeing-oriented recommendation, namely trust calibration, intent alignment, and consequence awareness. We outline a modular architecture for controllable LLM-based recommendation and discuss how different configurations can be systematically designed and evaluated in relation to user-centered outcomes such as self-efficacy, perceived agency, and appropriate reliance. This paper contributes a system-level framework for designing LLM-based recommender systems that are adaptive while remaining transparent, controllable, and aligned with human wellbeing.
With the widespread adoption of large language models (LLMs) in chatbots and everyday applications, companies increasingly need guardrails that are effective while remaining low-cost and low-latency. Safety evaluation of LLM outputs has generally relied on LLM-based judges, which can be effective but are often slow and expensive to deploy at scale. In this paper, we evaluate whether fine-tuned modern encoder classifiers from the ModernBERT family, including ModernBERT and Ettin, can reliably identify harmful LLM outputs in user-model conversations without substantial performance loss relative to LLM-based judges. We benchmark these encoder classifiers against rule-based prefix matching, fine-tuned LLM classifiers, and LLM judges using a range of judge-prompting strategies across open-source adversarial datasets. The LLM judges include evaluation methodologies from StrongReject, ShieldGemma, JailbreakBench, AILuminate, SorryBench, and a Claude-as-a-judge setup, as well as fine-tuned safety classifiers such as LlamaGuard 3 and LlamaGuard 4. The encoder classifiers are fine-tuned on judge-labeled data using a majority-voting label strategy and are then evaluated on a gold-standard holdout dataset to assess their performance relative to LLM judges. We report absolute performance using F1 score, false negative rate, and precision-recall metrics. We also break down results by attack technique, including single-turn prompting, decomposition, escalation, and context manipulation, to identify where encoder classifiers align with or diverge from LLM-based judges. Our findings provide guidance on when encoder classifiers can serve as cost- and latency-efficient alternatives to LLM-based safety evaluation.
Safety evaluation of large language models (LLMs) is commonly performed by querying models with unsafe or jailbreak prompts and judging whether their outputs violate a safety policy. Although useful, output-level evaluation is expensive, sensitive to judge choice, and easily tied to fixed question banks. We propose **SafeVec**, a white-box evaluation procedure that measures safety from internal representations rather than generated answers. **SafeVec** first extracts layer-wise refusal directions from a safety-aligned reference model, then selects stable layer windows where safe and unsafe behaviors are separable, and finally scores a target model by measuring whether its hidden states align with these refusal directions under unsafe and jailbreak prompts. The resulting metric, **RAS** (**R**efusal **A**lignment **S**core), maps representation-level refusal alignment to a calibrated 0-100 safety score. Across `Llama`, `Gemma`, and `Qwen` model families, RAS separates aligned models from uncensored and abliterated variants, tracks output-level attack success rate, and is substantially faster than judge-based evaluation. These results suggest that refusal alignment provides a compact and efficient signal for white-box LLM safety evaluation.
The demand for better prediction accuracy and higher execution performance in neural networks continues to grow. The emergence and success of Large Language Models (LLMs) have produced many cloud-based tools for software engineering tasks such as code suggestion. Although effective, cloud deployment raises concerns over privacy, latency, and reliance on network connectivity. Running LLMs locally on personal devices such as laptops would address these issues, because it enables offline use and reduces response time. However, local deployment is challenging, since commodity devices lack high-performance accelerators such as GPUs and are constrained by limited memory and compute capacity, which makes it hard to execute large models efficiently. We present Ditto, a framework that optimizes both the model size of Code LLMs and the inference programs that execute them. Our approach integrates two components. The first is a quantization technique inspired by product quantization, which groups model parameters into per-block codebooks via K-Means clustering and stores each weight as a bit-packed low-bitwidth index. The second component is a compilation pass integrated into LLVM that automatically detects and replaces unoptimized General Matrix-Vector Multiplication (GEMV) operations, with calls into Basic Linear Algebra Subprograms (BLAS) libraries that are highly optimized for the target hardware. The output of Ditto is a compiled executable that runs the selected Code LLM on commodity hardware. We evaluate Ditto on three popular Code LLMs, namely Code Llama, MagicCoder, and OpenCodeInterpreter, achieving up to 10.5$\times$ faster inference, 6.4$\times$ lower memory usage, and 10.5$\times$ lower energy consumption compared with their original inference pipelines, while preserving accuracy close to the full-precision models, with an average loss of only 0.27% in pass@1.
LLMs excel at reasoning, but validating their steps remains challenging. Formal verification offers a solution through mechanically checkable proofs. Interactive theorem provers (ITPs) dominate mathematical reasoning but require detailed low-level proof steps, while auto-active verifiers offer automation but focus on software verification. Recent work has begun bridging this divide by evaluating LLMs for software verification in ITPs, but the complementary direction, LLMs for mathematical theorem proving in auto-active verifiers, remains unexplored. We present MINIF2F-DAFNY, the first translation of the widely-used mathematical benchmark miniF2F to an auto-active verifier: Dafny. We find that Dafny's automation alone solves 39-44% of problems with empty proofs, whereas many require substantial proof guidance in ITPs. We evaluate 8 off-the-shelf LLMs on proof generation, with the best model (Claude Opus 4.6) achieving 62.7% cumulative pass@4 on the full test set, improving over the 38.9% empty-proof baseline by 23.8 percentage points. These results show that auto-active verification offers a complementary empirical setting for AI-assisted mathematical reasoning, where LLMs provide high-level guidance while SMT automation handles low-level details. Our benchmark and evaluation infrastructure are publicly available on https://github.com/dafny-lang/miniF2F.
As Large Language Models (LLMs) demonstrate remarkable capabilities learned from vast corpora, concerns regarding data privacy and safety are receiving increasing attention. LLM unlearning, which aims to remove the influence of specific data while preserving overall model utility, is becoming an important research area. One of the mainstream unlearning classes is optimization-based methods, which achieve forgetting directly through fine-tuning, exemplified by Negative Preference Optimization (NPO). However, NPO's effectiveness is limited by its inherent lack of explicit positive preference signals. Attempts to introduce such signals by constructing preferred responses often necessitate domain-specific knowledge or well-designed prompts, fundamentally restricting their generalizability. In this paper, we shift the focus to the distribution-level, directly targeting the next-token probability distribution instead of entire responses, and derive a novel unlearning algorithm termed \textbf{Di}stribution \textbf{P}reference \textbf{O}ptimization (DiPO). We show that the requisite preference distribution pairs for DiPO, which are distributions over the model's output tokens, can be constructed by selectively amplifying or suppressing the model's high-confidence output logits, thereby effectively overcoming NPO's limitations. We theoretically prove the consistency of DiPO's loss function with the desired unlearning direction. Extensive experiments demonstrate that DiPO achieves a strong trade-off between model utility and forget quality. Notably, DiPO attains the highest forget quality on the TOFU benchmark, and maintains leading scalability and sustainability in utility preservation on the MUSE benchmark.
Large Language Models (LLMs) have demonstrated significant promise in formal theorem proving. In this study, we investigate the ability of LLMs to discover novel theorems and produce verified proofs. We propose a pipeline called Conjecturing-Proving Loop (CPL), which iteratively generates mathematical conjectures and attempts to prove them in Lean 4. A key feature of CPL is that each iteration conditions the LLM on previously generated theorems and their formal proofs, enabling parameter-free improvement of proof strategies via in-context learning. We provide both theoretical and experimental evidence that CPL increases the discovery rate of hard-to-prove theorems compared to frameworks that generate statements and proofs simultaneously. Moreover, our experiments show that reusing the LLM's own formally verified outputs as context consistently improves subsequent proof success, demonstrating the effectiveness of self-generated in-context learning for neural theorem proving. The source code is available at https://github.com/auto-res/ConjecturingProvingLoop.
The integration of Formal Verification tools with Large Language Models (LLMs) offers a path to scale software verification beyond manual workflows. However, current methods remain unreliable: without a solid theoretical footing, the refinement process acts as a black box that may oscillate, loop, or diverge. This work bridges this critical gap by developing an LLM-Verifier Convergence Theorem, providing the first formal framework with provable guarantees for termination in multi-stage verification pipelines. We model the interaction not as a generic loop, but as a sequential absorbing Markov Chain comprising four essential engineering stages: \texttt{CodeGen}, \texttt{Compilation}, \texttt{InvariantSynth}, and \texttt{SMTSolving}. We prove that for any non-zero stage success probability ($δ> 0$), the system reaches the \texttt{Verified} state almost surely. Furthermore, because of the sequential nature of the pipeline, we derive a precise latency bound of $\mathbb{E}[n] \leq 4/δ$. We stress-tested this prediction in an extensive empirical campaign comprising over 90,000 trials. The results match the theory with striking consistency: every run reached verification, and the empirical convergence factor clustered tightly around $C_f\approx 1.0$, confirming that the $4/δ$ bound accurately mirrors system behavior rather than serving as a loose buffer. Based on this data, we identify three distinct operating zones -- marginal, practical, and high-performance -- and propose a dynamic calibration strategy to handle parameter drift in real-world environments. Together, these contributions replace heuristic guesswork with a rigorous architectural foundation, enabling predictable resource planning and performance budgeting for safety-critical software.
Recent work shows that Large Language Models (LLMs) can act as semantic mutation operators for the evolutionary discovery of programs and proofs. Most current applications focus on static coding benchmarks. We extend this paradigm to algorithmic trading. This domain is uniquely challenging because it is noisy, non-stationary, and highly discontinuous. We present AlgoEvolve, an LLM-driven evolutionary framework that generates, evaluates, and iteratively improves executable trading strategies. These strategies are expressed as Python code and evaluated through a rigorous testing protocol. Across multiple experiments, the system exhibits emergent regime-adaptive strategy logic, including autonomous shifts in trading rules. We further introduce a meta-evolutionary outer loop that evolves the prompts guiding program synthesis in the inner loop. This outer loop discovers improved search heuristics. These heuristics balance exploration and exploitation while reducing zero-trade failures. They consistently outperform initial human-designed instructions. The results demonstrate that LLM-based semantic evolution provides a viable approach for continual program synthesis in complex environments.
Smallholder farming communities in remote, depopulating areas have limited access to agricultural advice, and their locally specific agronomic knowledge, often expressed in regional dialect, is poorly represented in the global corpora on which Large Language Models (LLMs) are trained. A general-purpose chatbot therefore answers fluently but unreliably, ungrounded in authoritative local data farmers can trust. This paper presents a conversational AI assistant, Falco eleonorae, embedded in a bilingual (Greek-primary, English-secondary) e-market platform serving farmers and cooperatives of a defined island area of interest. It is a thin Backend-for-Frontend (BFF) proxy in front of a geospatially-aware agronomic agent rather than a self-hosted model. Answer generation and tool selection are delegated to a managed upstream service on OpenAI GPT-5-family models, while one bounded task, describing an uploaded field photograph, is handled directly by a vision-capable model so only text reaches the agent, and voice input is transcribed by a managed EU streaming speech-to-text service. Grounding comes not from a self-hosted vector database but from tool-augmented retrieval: a Model Context Protocol (MCP) tool queries a curated, read-only, bilingual data interface exposing local crops, a seasonal calendar, traditional practices, a dialect glossary, products, agritourism experiences, cooperatives, and training content, each wrapped in a geospatial Well-Known Text envelope anchoring the agent to the area of interest. We detail its multilingual, voice, and image modalities, its progressive-web-application and accessibility design for low-bandwidth field use, and its security and data-protection posture, and argue that for a small, resource-constrained rural deployment a managed, grounded multilingual assistant is more attainable and trustworthy than a self-hosted model.
When answering user queries, LLMs often retrieve knowledge from external sources stored in retrieval-augmented generation (RAG) databases. These are often populated from unvetted sources, e.g. the open web, and can contain maliciously crafted data. This paper studies attacks that can manipulate the context retrieved by LLMs from such RAG databases. Prior work on such context manipulation primarily injects false or toxic content, which can often be detected by fact-checking or linguistic analysis. A more subtle threat, which we call epistemic bias injection (EBI), is where adversaries inject factually correct yet epistemically biased passages that systematically favor one side of an open-ended issue. Although linguistically coherent and truthful, such adversarial passages effectively crowd out alternative viewpoints during retrieval from the RAG and push LLM outputs towards an attack-desired stance. As a core contribution, we propose a novel characterization of the problem: We give a geometric metric that quantifies stance polarity and epistemic bias. This metric can be computed directly on embeddings of text passages. Leveraging it, we construct EBI attacks and develop a lightweight prototype defense called BiasDef for them. We evaluate them both on a comprehensive benchmark constructed from public question answering datasets. Our results show that: (1) the proposed attack induces significant stance polarity shifts, effectively evading existing retrieval-based sanitization defenses, and (2) BiasDef substantially reduces adversarial retrieval and epistemic bias in LLM's answers. Overall, this demonstrates the new threat as well as the ease of employing epistemic bias metrics for filtering in RAG-enabled LLMs.
This paper presents a novel approach to perform semi-automated BSI IT-Grundschutz certification using a MultiLarge Language Model system (MLS) with Hybrid RetrievalAugmented Generation (HybridRAG). Facing the challenges of the Network and Information Security Directive 2 (NIS2) directive, a shortage of specialists, and high implementation costs, our MLS architecture aims to increase efficiency, reduce costs, and support certifiers in maintaining the quality of security concepts while meeting the increased demand for certifications of newly affected companies. The system combines Large Language Models (LLMs) and Knowledge Graphs (KGs) to support different phases of the certification process, including protection needs assessment, modeling, IT-Grundschutz check, measure consolidation, and subsequent realization. Our architecture addresses the growing demand for security concepts and offers an approach to handle the digital security challenges introduced by NIS2.
The LLM-based generation of machine-readable outputs such as JSON has attracted significant attention for integration with external systems. However, existing approaches cannot strictly enforce the maximum number of tokens to be generated, leading to infinite generation or truncated outputs that cause a system malfunction. To address this limitation, we propose TruncProof, a novel grammar-constrained generation method that enables LLMs to produce grammatically valid JSONs while adhering to a predefined token limit. By leveraging the properties of LL(1) parsers, TruncProof efficiently approximates the minimum number of tokens required to complete a grammatically valid output at each decoding step. Experiments on the Text-to-JSON instruction tasks demonstrate that TruncProof successfully generates syntactically correct outputs even under strict token constraints. Furthermore, we show that TruncProof can be effectively combined with advanced decoding strategies, resulting in outputs that are not only grammatically valid but also semantically accurate. The source code is public at https://github.com/Yosshi999/TruncProof
BiPACE: Bisimulation-Guided Policy Optimization with Action Counterfactual Estimation for LLM Agents
Stepwise group-based RL is an attractive way to train long-horizon LLM agents without a learned critic: it reuses multiple sampled rollouts to estimate local advantages. Its weakness is less visible but more fundamental: every group-relative estimator assumes that the steps it compares are equivalent for credit assignment. We show that current agentic variants violate this assumption through a state-action credit mismatch. The observation-hash partition is overly fine on the state side, creating singleton groups with zero step-level signal, while a single within-group mean is too coarse on the action side, mixing state-value estimation with action-specific credit. We introduce BiPACE (Bisimulation-Guided Policy Optimization with Action Counterfactual Estimation), a drop-in advantage estimator that fixes both sides without adding a critic, auxiliary loss, or extra rollouts. BiGPO clusters steps by cosine distance in the actor's own hidden-state geometry, an empirical policy-induced proxy for bisimulation that substantially lowers the singleton rate left by observation hashing. PACE then recenters returns within each behavioral cluster using action-conditioned peer baselines; its Q-style instance estimates a local Q(s,a)-V(s) nonparametrically. On ALFWorld/Qwen2.5-7B, BiPACE_Q raises overall validation success from GiGPO's 90.8 to $97.1\pm0.9$ over three seeds, and crosses the 95% threshold on every seed, which GiGPO never does within the same budget. On Qwen2.5-1.5B it reaches $93.5\pm1.2$ versus GiGPO's 86.7, and on WebShop and TextCraft it improves over GRPO and GiGPO at both model scales. The measured BiPACE-specific overhead is 11.3% of a single training-step wall time. Yet it changes the estimator's comparison unit from surface identity to approximate behavioral equivalence plus action-side counterfactuals. The code is available at https://github.com/TianxiangZhao/BiPACE.
Reinforcement learning (RL) has become a dominant post-training paradigm, driving the emergence of high-performance RL systems such as veRL for autoregressive large language models (LLMs). In parallel, diffusion-oriented RL algorithms, e.g., DanceGRPO and FlowGRPO, have rapidly expanded the scope of RL from language reasoning to diffusion-based visual and flow-based generation. However, efficient RL systems for diffusion generative LLMs remain underexplored. Existing implementations, e.g., veRL-Omni, still rely on colocated execution, which simplifies synchronization but couples rollout and training resources, limits heterogeneous deployment, and constrains independent scaling. To this end, we introduce DigenRL, a disaggregated RL framework for diffusion-based generative LLMs that supports flexible resource allocation, accommodates heterogeneous GPUs, and facilitates efficient task scheduling. To maximally reduce the execution bubbles in the disaggregated architecture, we propose: 1) a generation-axis pipeline (GAP) and time-step parallelism (TSP) in the diffusion architecture to enable finer-grained pipelining between rollout and training; 2) an elastic trainer-assisted generation (TAG) approach to enable the trainer GPU resources to dynamically assist in executing rollout generations; and 3) a tightly one-step constrained asynchronous strategy to further utilize the tail bubble in the pipeline. Extensive experiments are conducted on three hardware testbeds with 16-32 GPUs using HunyuanVideo-13B, Wan2.1-14B, FLUX.1-12B, and QwenImage-20B generative models. Experimental results show that DigenRL achieves 1.56-2.10x throughput improvements over state-of-the-art diffusion RL systems, veRL-Omni and GenRL.
Software performance optimization is a notoriously complex and manual task. Despite the growing use of Large Language Models (LLMs) for code refinement, we still lack benchmarks that capture how optimization actually happens in real-world codebases. Existing frameworks often oversimplify the problem by focusing on isolated functions or a single performance metric, missing the critical trade-offs between execution time and memory footprint, the inherent noise of the measurement environment, and the variability introduced by different input data and execution conditions. We address this by introducing SWE-Pro, a repository-level benchmark derived from 102 expert-written optimizations from open-source projects. Unlike previous benchmarks, SWE-Pro pairs each task with parameterized tests to evaluate runtime, peak memory, and Time-Weighted Memory Usage (TWMU) across varying input data and execution conditions under noise-aware measurement conditions. Our evaluation shows that current LLMs struggle significantly: runtime gains are negligible, and memory optimizations are nearly non-existent. This stands in sharp contrast to expert implementations, which achieve an aggregate speedup of 15.5x and peak memory reduction of 171.3x over benchmark tasks. Expert-written improvements are observed in 91.2% of tasks for runtime and 65.7% for peak memory. Our findings expose a substantial gap between current LLM capabilities and the demands of expert-level engineering.
LLMs are trained once, then deployed into a world that never stops changing. External memory compensates for this, but most systems manage it explicitly rather than letting it adapt on its own. Biological memory works differently: coupled multi-timescale dynamics make new associations immediately usable, strengthen what repetition confirms, and let the rest fade. We argue that external memory should follow a similar principle. In Memini, this view takes the form of an associative memory that organizes knowledge as a directed graph. Each edge carries two coupled internal variables, one fast and one slow, following the Benna-Fusi model of synaptic consolidation. From this coupling, episodic sensitivity, gradual consolidation, and selective forgetting are expected to emerge as facets of a single mechanism, reframing external memory as a learning substrate that reorganizes through its own dynamics. This workshop article describes an early-stage conceptual design without experimental evaluation.
Large language models (LLMs) can improve autonomous driving planning but are costly to query online, and existing fast-slow planners often rely on hand-designed triggering rules that either over-call the slow system or call it at the wrong times. We formulate slow-system invocation as a resource-aware sequential decision problem and propose the Adaptive Slow-System Control Gate (ASSCG), which makes frame-level Query/Cache/Drop decisions to refresh, reuse, or suppress slow guidance. ASSCG uses an RWKV backbone for efficient long-horizon gating and is trained with supervised fine-tuning followed by GRPO-style compute-aware reinforcement fine-tuning. We apply ASSCG to two different fast-slow architectures: (i) AsyncDriver on nuPlan Hard20 closed-loop evaluation, where ASSCG improves score to 67.28 (+2.28) while reducing average end-to-end inference latency by 60%; and (ii) a RecogDrive-based dual system that we build by replacing its original VLM-2B module with a lightweight ViT-based fast planner and adding an LLM slow planner, evaluated on NAVSIM, where ASSCG achieves 91.4 PDMS (+0.6) and increases average speed by 25%. The project page, including video visualizations and additional results, is available at https://williamxuanyu.github.io/asscg/.
FinRED: An Expert-Guided Benchmark Generation and Evaluation Framework for Financial LLM Red-Teaming
Existing safety benchmarks target general adversarial scenarios but miss finance-specific risks. Financial LLMs face regulatory compliance violations, fraud facilitation, and systemic trust erosion that require targeted evaluation. We introduce FinRED, an expert-guided red-teaming framework for financial LLM safety evaluation developed with financial experts. FinRED uses a novel two-level taxonomy mapping global standards (e.g., FATF and EU DORA) to threats ranging from regulatory evasion to complex fraud, integrated with a scalable pipeline that converts real financial documents into context-rich red-teaming Behavioral Prompts (seeds) through an expert-defined schema. Rigorous expert validation confirms seed plausibility and realism for meaningful LLM safety evaluation. We also provide an expert-validated, finance-specific rubric that goes beyond disclaimer checks, aligns more closely with human experts than static one-size-fits-all rubrics, and reduces critical false negatives from 28 to 12. Aligned with internationally adopted risk-management and information-security standards (e.g., ISO/IEC 27001), FinRED is deployed in South Korea's Financial Security Institute (FSI) regulatory sandbox for generative AI security evaluation in real financial services. To mitigate dual-use risks, the dataset, generation pipeline, prompt template, and evaluation framework are gated for qualified researchers at https://github.com/selectstar-ai/FinRED-paper and https://huggingface.co/datasets/datumo/FinRED.
Object-Goal Navigation (ObjNav) requires embodied agents to autonomously locate specified targets using only egocentric visual observations. Existing monolithic methods struggle with long-horizon reasoning and generalize poorly to novel environments. To address these limitations, we propose SAGE-Nav, a novel hierarchical framework that integrates the reasoning capabilities of Large Language Models (LLMs) with dynamic scene graphs. Crucially, it decouples asynchronous global semantic planning from the high-frequency reactive control loop. The LLM serves as a global planner, decomposing abstract instructions into a sequence of semantically grounded waypoints. To translate these plans into dense multi-modal guidance, we design a Hierarchical Scene Graph Encoder (HSGE) that leverages relational graph convolutions to produce structure-aware embeddings preserving both semantic and spatial topology. Furthermore, we develop the Goal-aware Alignment-Fusion Network (GAFN) to dynamically fuse real-time perception with these structural priors. Using an adaptive gating mechanism with an explicit inductive bias, GAFN ensures robust visual-topological alignment for the low-level policy. Extensive evaluations in the i-THOR and RoboTHOR environments demonstrate that SAGE-Nav achieves state-of-the-art performance, delivering substantial gains in navigation efficiency and zero-shot generalization while maintaining the low control latency required for physical robotic deployment.
Large language models (LLMs) increasingly show step-by-step reasoning rationales alongside their answers, turning reasoning from an internal model capability into a user-facing interface feature. Yet it is unclear whether such rationales help users judge when trust is warranted or merely persuade through fluent reasoning. We address this gap through the lens of auditable trust calibration: user-facing rationales should help people inspect whether an answer is warranted by evidence. We test this framing in factual verification through two linked studies. Study 1, an online experiment (N=68), manipulated rationale presentation format (instant, delayed, on demand), rationale correctness (correct, incorrect), and certainty framing (none, certain, uncertain). Study 2, a controlled eye-tracking study (N=54), examined how no-, correct-, and incorrect-rationale conditions were associated with users' trust, decision-making, and eye-movement patterns. Study 1 showed no reliable presentation-format effects; instead, rationale correctness and certainty framing influenced the trust in the information, trust in the LLM system, and decision confidence. In Study 2, incorrect rationales drew more attention to the supporting evidence and larger pupil diameter while the rationale was viewed, consistent with greater cognitive effort. Incorrect rationales also lowered trust in LLM system relative to showing no rationale, whereas the no-rationale difference was weaker for trust in information. A post-hoc predictive modeling analysis of gaze data from Study 2 further showed that gaze features carried predictive signal for trust- and decision-related user states. This work challenges the assumption that more reasoning is always better and supports rationale designs that are selective, linked to evidence, calibrated in how they express certainty, and easier to verify.
Granting LLMs direct control over costly, irreversible scientific experiments leads to unsafe exploration and unstable performance, but discarding LLM creativity entirely sacrifices significant optimization potential. We introduce CARE (Controlling LLM-Generated Policies through Auditable Review of Evidence in Scientific Experimentation), an auditable controller for high-throughput experimentation (HTE) optimization that keeps a non-LLM incumbent optimizer as the default action path while using LLMs to revise challenger ranking policies. Before each outcome is revealed, a public-evidence intervention gate compares the challenger with the incumbent. It authorizes the challenger's selection only when the evidence available before selection supports the change, with the decision recorded in the audit log. CARE outperforms all other evaluated methods on Minerva/Olympus and ChemLex benchmarks, with final-best improving from 80.0 to 88.5 on Minerva/Olympus and from 83.9 to 92.1 on ChemLex, relative to the public incumbent. Our experiments indicate that LLM self-evolution is more reliable when it expands the proposal space under an auditable controller, rather than directly choosing experiments.
Tool-integrated LLM agents are often wrapped within a harness: the scaffolding that determines which tools are exposed, how they are described, and what auxiliary information accompanies each per-step observation. While agents are routinely post-trained, this scaffolding is typically treated as a fixed engineering detail, with design effort limited to the training-free regime. Moreover, existing post-training algorithms assume a static environment, even though tool environments and tasks often shift upon deployment. To address this gap, we extend $\texttt{ALFWorld}$ (i) to treat the harness as a controllable design dimension and (ii) to support evaluation under task and tool environment shifts. Building on this, we systematically analyze how the harness design influences post-training in both in-distribution and out-of-distribution (OOD) settings. We empirically show that harness-aware post-training not only improves in-distribution performance but also enables agents to robustly adapt to OOD settings. Under a harness with minimal design effort, post-training suffers a drastic performance drop under stronger tool environment shifts, further highlighting the importance of harness-aware post-training under such shifts.
Large Language Models (LLMs) increasingly show reasoning rationales alongside their answers, turning "reasoning" into a user-interface element. While step-by-step rationales are typically associated with model performance, how they influence users' trust and decision-making in factual verification tasks remains unclear. We ran an online study (N=68) manipulating three properties of LLM reasoning rationales: presentation format (instant vs. delayed vs. on-demand), correctness (correct vs. incorrect), and certainty framing (none vs. certain vs. uncertain). We found that correct rationales and certainty cues increased trust, decision confidence, and AI advice adoption, whereas uncertainty cues reduced them. Presentation format did not have a significant effect, suggesting users were less sensitive to how reasoning was revealed than to its reliability. Participants indicated they use rationales to primarily audit outputs and calibrate trust, where they expected rationales in stepwise, adaptive forms with certainty indicators. Our work shows that user-facing rationales, if poorly designed, can both support decision-making yet miscalibrate trust.
Connecting a pre-trained speech encoder to a Large Language Model (LLM) is the standard architecture for building Speech LLMs. However, a structural misalignment exists between the encoder and the LLM. Unlike encoders based on automatic speech recognition, which often produce representations in separate language-specific spaces, LLMs operate within a unified language-agnostic space. A mechanism is required to align the encoder's language-specific representations with the LLM's shared space. We argue that speech translation provides a principled way to achieve this. Unlike monolingual transcription, translation requires the model to bridge different languages and learn language-agnostic representations. We experimentally evaluate the impact of incorporating translation objectives into speech encoder pre-training. Our results demonstrate that translation-enhanced pre-training improves cross-modal integration and leads to superior performance across downstream Speech LLM tasks.
On Apple Silicon the fp32 GEMMs dominating LLM prefill are dispatched by Accelerate to a matrix coprocessor (AMX) on the M1-M3. We ask where a hand-written kernel's throughput over Accelerate comes from on the M1 AMX, and reach a structural conclusion: not a faster inner loop. By microbenchmark the inner loop is load-issue bound -- once any operand load interleaves with the FMA32 stream, single-thread throughput falls to a 610-to-680 GFLOPS band, under half the load-free rate. The gain comes from two deployment-level levers Accelerate underuses: fine multi-thread panels filling the M1's second on-chip AMX block (winning the K >= N shapes), and pre-packing the constant weight at load (winning the N > K shapes). A bit-exact direct-AMX kernel using both is the fastest bit-exact fp32 GEMM path we find on the M1: it exceeds all three Accelerate fp32 paths (cblas_sgemm, BNNSMatMul, and the BNNS Graph compiler) at all twelve LLM prefill GEMMs at S = 128 (GPT-2 to Llama-7B scale), leading the fastest, BNNS Graph, by 1.17 -- and by 1.09 at the three shapes where it too holds fp32 -- with geometric means of 1.58 over BNNSMatMul and about 2.0x over cblas_sgemm. Every output is bit-identical to Accelerate, whereas BNNS Graph is bit-exact at only three of twelve shapes, the rest at reduced precision (error up to 1.4e-3). Dropped into llama.cpp for its cblas_sgemm prefill matmuls, it raises measured full-forward throughput from 291 to 420 tokens/s (1.44x, bit-identical) at 128-token prefill -- end-to-end, not a GEMM-only ratio. The contribution is this shape-resolved M1-AMX characterization (microbenchmark, two-block aggregate, per-core occupancy probe), leaving fine-panel scheduling and pre-packing as the only two levers above an inner loop at the hardware limit; mis-tuning the single column-panel width costs nearly 2x.
Reliable analytics and machine-learning pipelines depend on clean tabular data, yet production tables often contain missing values, typographical errors, inconsistent formats, violated dependencies, unit mismatches, and ambiguous categorical values. Existing cleaning systems make different trade-offs. Constraint-based systems need experts to specify rules. Learning-based systems need labels or retraining. Recent LLM-based cleaners reduce setup effort, but many call an LLM on rows, cells, or repeated workflow steps, so their cost grows with table size and with every recurring batch. We present TabClean, a model-training-free system that compiles LLM reasoning into reusable guarded cleaning programs. Given a dirty table and a small annotated development set, TabClean profiles table evidence, diagnoses repair mechanisms, synthesizes executable Python transformations, validates candidates with cell-level feedback, and commits the best program for reuse on schema-compatible batches. The key abstraction is an evidence-backed guarded repair clause. A deterministic transformation may fire only when its dirty pattern, target-negative condition, evidence support, and scope constraints are satisfied. Across six benchmarks, TabClean achieves high precision, improves F1 over representative rule-based, learning-based, and LLM-based baselines on five datasets, and substantially reduces recurring runtime and API cost by replacing repeated LLM inference with deterministic program execution.
Assessing financial literacy during gameplay without disrupting the learning experience remains a key challenge in serious games for education. We present the Agentic BKT pipeline, a multi-agent large language model architecture for stealth assessment of financial competencies from open-ended gameplay events. The pipeline processes events from a 2D platformer serious game aligned with the OECD/INFE financial literacy framework through four phases: (1) the game captures every player decision as a structured event log; (2) an LLM event classifier labels each action on a four-point rubric validated against three domain experts (Fleiss kappa = 0.624, substantial agreement); (3) four domain-specific agents specializing in risk mitigation, investing, spending, and credit management perform session-level reasoning over behavioral trajectories, feeding per-competency Bayesian Knowledge Tracing that estimates mastery within each domain; and (4) an expert judge agent synthesizes the domain-level estimates into an overall mastery score. Evaluated with 193 K-12 participants across 264 game sessions, the Agentic BKT pipeline yields mastery estimates significantly correlated with learning gain (r = 0.276, p = 0.0001) and post-test scores (r = 0.333, p < 0.0001) while showing no correlation with pre-test scores, providing both convergent and discriminant validity. The multi-agent approach approximately triples the predictive validity of a single-LLM baseline (r = 0.095, not significant) in this study, demonstrating that domain decomposition and session-level reasoning play a central role in capturing the multidimensional nature of financial literacy from gameplay
Large Language Models (LLMs) are increasingly used for automated vulnerability detection, but it remains unclear how program structure and semantics should be represented for LLM-based reasoning. Most prompting-based approaches provide raw source code, implicitly assuming that more source-level context gives the model better evidence. This paper challenges that assumption through RepBench, an empirical benchmark comparing raw source code with static-analysis-based program representations. RepBench converts real-world C/C++ vulnerability testcases into multiple representations: raw source, Abstract Syntax Trees (ASTs), Control-Flow Graphs (CFGs), Program Dependence Graphs (PDGs), their combinations, and an auxiliary track of enriched PDGs (ePDGs). Using a curated PrimeVul-derived corpus of 107 Joern-based testcases across five CWE categories, we evaluate ten representation variants under a fixed Chain-of-Thought and structured-output protocol, plus 19 additional ePDG cases generated through VulChecker/Hector. Representation choice substantially affects LLM vulnerability reasoning. The strongest variant, AST+PDG, achieves 83.2% accuracy, compared with 53.5% for raw source. At the family level, graph-only prompts outperform both source-only and source-plus-graph prompts while requiring far less prompt overhead. These results reveal a context dilution effect: adding raw source code to compact structural graph evidence can degrade reasoning by making vulnerability-relevant evidence less salient. Overall, our findings show that carefully selected structural representations offer a better accuracy-overhead tradeoff than simply giving LLMs more raw input, and suggest that static analysis can serve as an effective prompt-construction layer for security-focused LLM reasoning.
Large language model (LLM) inference is increasingly dominated by data movement across the memory hierarchy. Recent 3D-stacked cache technologies have enabled GB-scale last-level caches in modern server CPUs, making it possible to keep reusable model weights on chip and exploit cache bandwidth and latency. Achieving this regime is not straightforward: deeper pipelining for weight residency increases in-flight requests and KV-cache footprint, while cache-resident operators make operator-boundary synchronization a visible bottleneck. We present a cache-resident execution model for inference on hierarchical-memory clustered systems. The model separates weight-centric operators from attention and KV-cache management into dedicated resource domains, keeping reusable weights cache-resident while scaling KV capacity independently of pipeline depth. It also relaxes synchronization from operator boundaries to true sub-operator dependencies, reducing coordination overhead in the cache-resident regime. We instantiate this model on a multi-socket CPU cluster with a weight-attention decoupled architecture, locality-aware placement, and a specialized static runtime. The prototype substantially outperforms equally provisioned llama.cpp. On deployed Llama-3.2-3B and Llama-2-7B configurations, it achieves 2.04x-11.51x speedup on time-per-output-token (TPOT). Under a validated analytical model, it further reaches up to 13.9x TPOT speedup across model sizes, context lengths, and batch sizes. These results show that commodity CPUs with GB-scale last-level caches can support efficient LLM inference when execution is organized around cache residency, decoupled state management, and dependency-aware coordination.
Data preparation pipelines improve data quality in machine learning by transforming raw tables into learning-ready data through sequential cleaning and feature transformation operators. However, automatically constructing such pipelines is computationally difficult because operator sequences are combinatorial and end-to-end evaluation is expensive. Existing state-of-the-art (SOTA) Multi-DQN methods still face three key limitations: decoupled value estimators weaken long-horizon credit assignment, dataset context is only weakly injected into the policy, and exploration is inefficient in a sparse search space with many invalid states. To address these issues, we propose FlowPipe, a unified framework that formulates pipeline synthesis as conditional probabilistic flow generation over a directed acyclic graph. FlowPipe uses Conditional Generative Flow Networks (C-GFlowNets) with a Trajectory Balance objective to connect terminal validation rewards with early pipeline decisions. It further introduces Deep Semantic Modulation through Feature-wise Linear Modulation (FiLM), allowing LLM-derived logical priors to condition the policy's internal activations according to dataset semantics. In addition, FlowPipe incorporates failure awareness into the flow objective to avoid invalid states and concentrate search on high-potential regions. Experiments on two benchmark suites with 74 real-world datasets show that FlowPipe outperforms SOTA baselines, improving accuracy by 11.96% on average and achieving 12.5x faster training convergence. Source code is available at https://github.com/KunyuNi/FlowPipe.
Large Language Models (LLMs) have shown promise for automated penetration testing, yet existing end-to-end black-box evaluations are highly susceptible to error cascading: failures in early reconnaissance can mask an agent's actual ability to exploit vulnerabilities. To more accurately characterize these capabilities, we propose a two-stage decoupled evaluation framework that separates exploit execution from reconnaissance. Using ground-truth injection and knowledge-driven ablation across 70 high-fidelity web vulnerability testbeds, our framework isolates exploitation performance from reconnaissance noise. We empirically evaluate five open-source penetration-testing agents, covering multiagent, monolithic, and graph-driven architectures, on a strictly aligned subset of 50 representative vulnerabilities. The results reveal a substantial capability gap. With accurate vulnerability context, agents achieve a functional success rate of up to 90.0%, whereas autonomous reconnaissance, measured by targeted vulnerability recall, plateaus at approximately 50.0%, primarily due to failures in parsing unstructured telemetry. Cross-architectural analysis further reveals distinct capability niches: multi-agent isolation is more effective for long-sequence interactions such as de-serialization, while monolithic and graph-driven designs perform better on short-chain injections and cross-session access-control vulnerabilities, respectively. This decoupled evaluation work provides a fine-grained benchmarking protocol and an empirical basis for designing next-generation automated offensive security agents.
Recent work has demonstrated the potential of large language models (LLMs) for program optimization, a key challenge in programming languages. We propose a blackbox adaptation method called Retrieval Augmented Search (RAS) that performs beam search over candidate optimizations; at each step, it retrieves in-context examples from a given training dataset of slow-fast program pairs to guide the LLM. Critically, we find that performing contextual retrieval based on an LLM-generated natural language description significantly outperforms retrieval based on the source code. We also propose AEGIS, a method for improving interpretability by decomposing training examples into ''atomic edits'' that are significantly more incremental in nature. We show that RAS performs up to 2.06$\times$ better than prior state-of-the-art blackbox adaptation strategies on optimizing C++ programs, and that AEGIS performs up to 1.37$\times$ better while making significantly smaller edits. We also show that using RAS improves the mean runtime percentile of Python programs by 10.27 compared to baselines.
Large language model (LLM) agents rely on long-term memory to support extended interactions and personalized assistance beyond finite context windows. Existing memory agents actively update external memory through generated write, revise, and delete operations, but these updates may omit important information, corrupt existing memory, or introduce unsupported hallucinated content. Once stored, such errors become persistent system-state failures that can affect future reasoning and generation. In this paper, we propose TrustMem, a framework designed to improve the trustworthiness of memory consolidation. TrustMem relies on a Memory Transition Verifier to evaluate the transition process of memory updates in terms of coverage, preservation, and faithfulness. It further constructs preference pairs among candidate updates under the same memory state, enabling preference-guided reinforcement learning to directly optimize memory updating behaviors. Extensive experiments demonstrate that TrustMem improves both memory utility and reliability: it achieves state-of-the-art results across MemoryAgentBench, HaluMem, and the Mem-alpha validation set, improves HaluMem memory extraction by 12.14 F1 points, and reduces transition-level omission, corruption, and hallucination by 40.1\%, 79.1\%, and 50.0\%, respectively, compared with the strongest baseline for each error type.
Contemporary systems serving large language models (LLMs) have adopted prefill-decode disaggregation to load-balance between the compute-bound prefill phase and the memory-bound decode phase. Under this design, prefill workers generate a KV cache that must be transferred to decode workers before generation can begin. With these workers residing on different physical systems, this transfer becomes a significant bottleneck to serving LLMs at scale, especially for long-input and agentic workloads. Existing lossless codecs are unsuitable here as they primarily target offline weight compression, run on CPUs, or use variable-length coding whose compression cannot keep up with KV production during prefill. We introduce SplitZip, a GPU-friendly lossless compressor for KV cache transfer that preserves KV tensors bitwise and integrates into existing serving frameworks without modifying model execution. SplitZip exploits redundancy in floating-point exponents of KV activations, encoding frequent exponent values with fixed-length codes and routing rare exponents through a sparse escape stream of (position, value). A calibrated top-16 exponent codebook eliminates online histogramming, while the regular dense path and sparse escape correction make both encoding and decoding efficient on GPUs. On real BF16 activation tensors, SplitZip achieves $613.3$ GB/s compression throughput and $2181.8$ GB/s decompression throughput, outperforming prior lossless compressors on the critical codec path. End-to-end transfer experiments show up to $1.32\times$ speedup for BF16 KV cache transfer, $1.30\times$ speedup for TTFT, and $1.23\times$ increase in Request Throughput. The same approach extends to FP8 KV caches, providing up to $1.14\times$ compression over native E5M2. Code is available at https://github.com/Intelligent-Microsystems-Lab/SplitZip
The rapid growth of scientific submissions has pushed traditional peer review toward its scalability limits, motivating the exploration of large language models (LLMs) as intelligent automated evaluation assistants. Although recent studies show that LLMs can generate fluent critiques and approximate reviewer scores, their reliability, robustness, and security as decision-support systems remain insufficiently understood. This survey offers a systems-level analysis of LLM-based scientific peer review, focusing on two core evaluative functions: critique generation and score prediction. We present a structured taxonomy of modeling approaches (including prompt-based, supervised, retrieval-augmented, and alignment-optimized approaches), and synthesize empirical findings across existing benchmarks. We analyze dataset constraints, evaluation shortcomings, and domain concentration biases that limit current assessment practices. Beyond performance metrics, we identify emerging robustness risks, including prompt injection, data poisoning, retrieval vulnerabilities, and reward hacking, which expose automated review pipelines to strategic manipulation. From a data mining perspective, we outline key open challenges in modeling subjective disagreement and cross-domain generalization. By reframing automated peer review as a high-stakes, multi-objective decision problem, this survey provides a roadmap for developing robust, transparent, and trustworthy AI-assisted scientific evaluation systems.
Recovering governing Ordinary Differential Equations (ODEs) from data is a central challenge in modeling dynamical systems across scientific domains. Existing approaches cast discovery as a static inference problem over fixed datasets, assuming that the observed trajectories are sufficiently informative. However, dynamical systems evolve over large state spaces, and limited data can make multiple equations observationally indistinguishable, leading to identifiability gaps and the recovery of incorrect governing equations. To address this, we introduce LLM-ACES, or LLM-guided Active Closed-loop Equation Search, a closed-loop framework that jointly optimizes symbolic hypothesis construction and adaptive data acquisition. In LLM-ACES, a large language model (LLM) proposes operator priors that partition the large search space into distinct regions, within which candidate equations are fit to the observed data. The disagreement among these candidates guides the acquisition of informative trajectories, creating a feedback loop that iteratively refines both the hypothesis space and the discovered dynamics. On 122 ODE systems spanning ODEBench and ODEBase, LLM-ACES achieves the lowest median NMSE, outperforming state-of-the-art baselines by several orders of magnitude while achieving a high symbolic accuracy of 46.2% and 52.4%, respectively. Our analysis further shows that LLM-ACES is sample-efficient, achieving better performance with one-tenth the data. Furthermore, LLM-ACES's feedback-driven data acquisition makes it robust to noise and recovers the correct symbolic structure, while baselines introduce spurious terms that fit the data locally but obscure the true governing relationships.
Persistent memory attacks against LLM agents achieve high attack success rates against open-source models. In these attacks, malicious instructions injected via RAG-retrieved documents are stored in persistent memory and executed in later sessions. However, no systematic evaluation of defense effectiveness against this attack class exists. We evaluate six defenses across four architectural layers against delayed-trigger attacks on nine open-source models (5,040 runs, N=40 per condition). Four defenses fail at approximately baseline attack success rate: input-level filtering (Minimizer, Sanitizer) and retrieval-level filtering (RAG Sanitizer, RAG LLM Judge) achieve 88-89% ASR, statistically indistinguishable from the undefended baseline of 88.6%. Prompt Hardening partially fails at 77.8% ASR, with the reduction driven by two models at 0%: one genuine defense effect and one model-level refusal independent of the defense. The architectural explanation holds: input-level defenses cannot observe RAG-injected content, and retrieval-level classifiers are defeated by compliance-framed semantic masking. One defense, tool-gating at the memory layer (Memory Sandbox), reduces ASR to 0% for eight of nine models by removing the recall capability the attack requires. The exception inverts the defense entirely: a reasoning model that achieves 0% ASR under no defense via execution refusal inverts to 100% ASR under Memory Sandbox, because removing explicit recall forces the model onto the RAG pathway where its refusal mechanism does not activate. Memory Sandbox imposes zero utility cost in the absence of attack (BTCR = 100% across all conditions). These results provide the first systematic characterization of why each defense class fails against persistent memory attacks, enabling informed defense investment decisions.
Large language models (LLMs) are increasingly deployed as autonomous agents that make sequences of decisions over extended interactions in high-stakes domains. However, the behaviour of LLMs under sustained authority pressure is still an open question with direct implications for the safety of agentic pipelines. We ran a variation of Milgram's obedience experiment on 11 open-source LLMs and found that most models reached or approached the final shock level before refusing, across 8 conditions with 30 trials per model per condition. Model behaviour varies considerably in multiple aspects both across models and across trials of the same model. We found four main takeaways: (1) LLMs are subject to pressure and they comply despite explicitly expressing distress, just like human subjects did in the original experiment; (2) LLMs are vulnerable to gradual boundary/value violations; (3) when LLMs refuse, they may ignore the response format requirements, so the response is discarded by the orchestrator, which causes a retry that can result in compliance with the underlying request even when refusal was intended initially; (4) we hypothesise that there is a runaway low-level token pattern continuation attractor that might be contributing to obedience, overriding higher level processing of the situation's meaning and values.
Software-Defined Networking (SDN) is increasingly adopted to secure Internet-of-Things (IoT) networks due to its centralized control and programmable forwarding. However, SDN-IoT defense is inherently a closed-loop control problem in which mitigation actions impact controller workload, queue dynamics, rule-installation delay, and future traffic observations. Aggressive mitigation may destabilize the control plane, degrade Quality of Service (QoS), and amplify systemic risk. Existing learning-based approaches prioritize detection accuracy while neglecting controller coupling and short-horizon Reinforcement Learning (RL) optimization without structured, auditable policy evolution. This paper introduces a self-reflective two-timescale SDN-IoT defense solution separating fast mitigation from slow policy governance. At the fast timescale, per-switch Proximal Policy Optimization (PPO) agents perform controller-aware mitigation under safety constraints and action masking. At the slow timescale, a multi-agent Large Language Model (LLM) governance engine generates machine-parsable updates to the global policy constitution Pi, which encodes admissible actions, safety thresholds, and reward priorities. Updates (Delta Pi) are validated through stress testing and deployed only with non-regression and safety guarantees, ensuring an auditable evolution without retraining RL agents. Evaluation under heterogeneous IoT traffic and adversarial stress shows improvements of 9.1% Macro-F1 over PPO and 15.4% over static baselines. Worst-case degradation drops by 36.8%, controller backlog peaks by 42.7%, and RTT p95 inflation remains below 5.8% under high-intensity attacks. Policy evolution converges within five cycles, reducing catastrophic overload from 11.6% to 2.3%.
LLM-based dialogue assistants have become mainstream tools for software developers, yet current evaluation benchmarks focus exclusively on functional correctness. This leaves a critical gap in assessing the quality and accuracy of these conversations when handling Non-Functional Requirements (NFRs), which are inherently vague, context-dependent, and involve many parts of a program. Evaluating how well these systems support collaborative reasoning about NFRs requires methods that go beyond single-turn accuracy to capture both the correctness of the system's outputs and the quality of the multi-turn interaction. In this paper, we investigate the accuracy and quality of multi-turn conversations between developers and an LLM-based agent in the domain of Health Insurance Portability and Accountability Act (HIPAA) regulatory compliance. We hired 49 programmers to interact with GitHub Copilot to assess 148 HIPAA-derived NFRs against the iTrust codebase, a system designed to comply with HIPAA regulations, across three dimensions: requirement satisfaction level, reasoning, and code localization. We find that developers tend to agree with LLM assessments, but accuracy against expert ground truth is low. We model user satisfaction and find that longer system responses and more information-providing turns negatively affect user satisfaction, whereas proactive interactions positively affect it. Our findings provide insights for designing LLM-based dialogue systems that support NFR assessment.
This paper introduces HelpBench, a benchmark for assessing whether LLMs are capable of providing accurate help in response to questions about digital privacy, safety, and security. We curated 450 questions representing authentic user situations and developed rubrics for each question to evaluate the factual accuracy and tone of a response. Example questions touch on how to regain access to lost or suspended accounts, how to balance the trade-offs of hardware security keys versus other forms of two-factor authentication, whether a suspicious email is likely a scam, or whether an abuser might be able to track an individual based on their device peripherals. We then developed and applied an auto-rater to evaluate responses from 18 state-of-the-art LLMs. Our results indicate that while models provide high-quality advice (with scores of 82% on average), one in ten responses from models scores less than 65%, reflecting inaccurate and even harmful advice. Addressing these failures is critical for models to serve as trustworthy sources of assistance for digital privacy, safety, and security needs.
Large Language Models (LLMs) have demonstrated remarkable capabilities across diverse tasks, yet they remain prone to generating hallucinations. Detecting these hallucinations is critical for deploying LLMs reliably in high-stakes applications. We present Grad Detect, a gradient-based approach for predicting hallucinations by analyzing layer-wise gradient patterns from a single forward-backward pass during inference. Our method shows that the internal gradient structure of a model carries rich information about the correctness of its output. This information is not accessible through output-level signals alone. We evaluate Grad Detect on several Q&A benchmarks across both hallucination detection and model abstention prediction, where it consistently outperforms confidence-based and sampling-based baselines. Through comprehensive layer ablation studies across all eleven models from four architectural families, we find that the final five layers concentrate over 97% of the discriminative gradient signal, enabling efficient deployment with minimal performance loss. Grad Detect provides a unified framework for predicting multiple dimensions of LLM reliability, offering strong predictive performance alongside interpretable insights into where and how model failures originate.
Large Language Models (LLMs) achieve strong performance across a growing range of domains, yet their scale poses deployment challenges in applications where latency and cost constraints are critical. This paper derives empirical scaling laws for domain-specific LLM compression, quantifying how in-domain and general knowledge performance scale with dataset size, compression ratio, supervision format, and iterative pruning schedule. Using quantitative finance as our application domain, we compare logit-based and LoRA-based distillation under iterative structural pruning, introducing a blended chain-of-thought supervision loss that stabilizes KL-divergence distillation over reasoning traces. In-domain task quality degrades predictably under compression while general-knowledge benchmarks collapse well before the same point; supervision format is the key driver of this tradeoff, with chain-of-thought supervision actively recovering general knowledge that pruning erases. We release the headline dataset FinHeadlineMix, scaling law results, and practical recommendations to provide a reusable framework for domain-specific compression decisions.
Recent progress in large language models (LLMs) has focused on producing responses that meet human expectations and align with shared values - a process coined alignment. However, aligning LLMs remains challenging due to the inherent disconnect between the complexity of human values and the narrow nature of the technological approaches designed to address them. Current alignment methods often lead to misspecified objectives, reflecting the broader issue of incomplete contracts, the impracticality of specifying a contract between a model developer, and the model that accounts for every scenario in LLM alignment. In this paper, we argue that improving LLM alignment requires incorporating insights from societal alignment frameworks, including social, economic, and contractual alignment, and discuss potential solutions drawn from these domains. Given the role of uncertainty within societal alignment frameworks, we then investigate how it manifests in LLM alignment. We end our discussion by offering an alternative view on LLM alignment, framing the underspecified nature of its objectives as an opportunity rather than perfect their specification. Beyond technical improvements in LLM alignment, we discuss the need for participatory alignment interface designs.
Large Language Models (LLMs) and LLM-based Multi-Agent Systems (MAS) are revolutionizing software engineering (SE) by advancing automation, decision-making, and knowledge processing. Their recent application to SE tasks has already shown promising results. In this paper, we focus on summarization as a key application area. We present Metagente, an LLM-based MAS designed to generate concise and accurate summaries of software documentation. Metagente employs a Teacher-Student architecture where multiple LLM agents collaborate to enhance relevance and precision of produced summaries. An empirical evaluation on real-world datasets demonstrates Metagente's effectiveness in streamlining workflows, outperforming the considered baselines. The evaluation provides evidence that Metagente improves summarization for requirements analysis and technical documentation. Our findings underscore the transformative potential of these technologies in SE, while identifying challenges and future research directions for their seamless integration.
A recent Nature Medicine study reports that general-purpose frontier LLMs outperform specialized retrieval-augmented clinical tools on medical benchmarks, and that retrieval can hurt strong models. We ask the natural follow-up: does structured knowledge-graph (KG) grounding change this, and when does grounding help at all? We contribute two results. First, a reproduction: the study's headline HealthBench score (~88) is the Consensus variant, not full HealthBench, where frontier models and ideal completions both score ~46-47 under a physician-calibrated grader (agreement 82.5%); we reproduce GPT-5.2 Consensus =90.9 and flag a score-deflating grader bug. Second, a knowledge-boundary result. Using a graph+vector engine (samyama-graph) over the public biomedical KG PrimeKG, neither naive triple retrieval nor an agentic natural-language-to-Cypher loop (82% successful queries) improves MedQA across a weak-to-strong model ladder (all |Delta| <= 3.4). On a synthetic counterfactual KG, and on a hybrid benchmark mixing known and novel facts, the identical pipeline lifts out-of-training accuracy from chance to ~100% (+68 to +79) while adding nothing on known facts (a no-LLM arm answers both). Across three regimes (no-knowledge, graph-aided, hybrid), grounding helps only insofar as the decisive fact lies outside the model's training -- public-KG facts are redundant, private and novel data are where it pays -- matching the study's institutional-data caveat.
Organizations and researchers show increasing interest in using large language models (LLMs) in place of human participants in A/B tests, in the hope of experimenting faster and at lower cost. We study when a treatment effect estimated on LLM outcomes can recover the effect for the human population of interest. Distributional equivalence between LLM and human outcomes would make any standard estimator valid but is unrealistic. We therefore develop a statistical framework that adapts surrogate endpoint theory to LLMs, showing that calibrating LLM outcomes to human outcomes identifies the average treatment effect under surrogacy and comparability conditions that are jointly weaker than distributional equivalence. We present a falsification test for surrogacy and a bound on the worst-case bias from limited overlap between the LLM and human samples. We further show that the stochasticity inherent to LLMs can weaken surrogacy for identification while also introducing bias and variance during estimation, but that using an average over multiple LLM draws per unit as the surrogate mitigates these issues. Simulations validate the results, and an empirical application to the Upworthy Research Archive dataset shows that raw LLM outputs recover only 39% of the human treatment effect while nonparametric calibration closes the gap. A central takeaway is that A/B testing on LLM responses is correct only by assumption, whereas A/B testing on humans is correct by design, and that the required assumptions are hardest to justify precisely where LLMs promise the greatest benefit. We discuss the choice of LLM, prompting, and temperature as design variables, the compounded challenge posed by long-term outcomes, and how to size human pilot studies for validation.
Accurate user modeling often depends on rich interaction histories, which are unavailable for billions of low-activity users. Large Language Models (LLMs) can infer latent user states from static profiles, but this reasoning becomes unreliable when profiles are sparse, and applying an LLM to billions of users is prohibitively expensive. We present ScaleToT, which learns structured reasoning from a small LLM-processed subset and extends it to the broader low-activity user population. To improve reasoning reliability, ScaleToT constructs typed user-state chains with a bounded entropy-guided Tree-of-Thought (ToT) refinement procedure. To make this structured reasoning usable from sparse profiles, the teacher-curated chains are used to train a student model on static profiles through supervised fine-tuning (SFT) and Outcome-Driven Segment-Aware Implicit Reward Policy Optimization (OSIPO). ScaleToT then transfers the student's reasoning representations to a lightweight profile encoder, providing shared reasoning signals for the remaining users without LLM inference. We evaluate ScaleToT on lifetime value (LTV) prediction in a billion-scale advertising deployment. A randomized online A/B test increased LT30 by 6.738\%, while offline reasoning covered only 7.32\% of the potential population, greatly reducing compute cost compared with full-population reasoning.
While the validity of LLMs' use in the legal context remains subject to ethical and legal debate, legal professionals are already experimenting with personal LLMs, if only for translation and reformulation. However, even such a seemingly innocuous use can introduce biases through case processing speed if LLM assistants selectively refuse assistance on certain topics. To better anticipate such biases, we investigate several modern small LLMs that are most likely to be used as on-device assistants, to assess the impact of overrefusal on legal prompts. Surprisingly, we find that authority-style prefixes (``you are acting as an assistant of the national supreme court'', ``[...] defense lawyer'') systematically increase refusal rates by 2--20x over the no-prefix baseline, while a known role-play jailbreak prefix shows mixed effects, sharply increasing refusals in some models and barely shifting them in others. The finding suggests that small on-prem deployable LLMs are unstable under contextual framings that a real institutional user might naturally introduce, and further investigation is essential to minimize opportunities for bias.
The explosive growth of AI research has created unprecedented information overload, increasing the demand for scientific summarization at multiple levels of granularity beyond traditional abstracts. While LLMs are increasingly adopted for summarization, existing benchmarks remain limited in scale, target only a single granularity, and predate the LLM era. Moreover, since the release of ChatGPT in November 2022, researchers have rapidly adopted LLMs for drafting manuscripts themselves, fundamentally transforming scientific writing, yet no resource exists to analyze how this writing has evolved. To bridge these gaps, we introduce SciZoom, a benchmark comprising 44,946 papers from four top-tier ML venues (NeurIPS, ICLR, ICML, EMNLP) spanning 2020 to 2025, explicitly stratified into Pre-LLM and Post-LLM eras. SciZoom provides three hierarchical summarization targets (Abstract, Contributions, and TL;DR) achieving compression ratios up to 600:1, enabling both multi-granularity summarization research and temporal mining of scientific writing patterns. Our linguistic analysis reveals striking shifts in phrase patterns (up to 10x for formulaic expressions) and rhetorical style (23% decline in hedging), suggesting that LLM-assisted writing produces more confident yet homogenized prose. SciZoom serves as both a challenging benchmark and a unique resource for mining the evolution of scientific discourse in the generative AI era. Our code and dataset are publicly available on GitHub (https://github.com/janghana/SciZoom) and Hugging Face (https://huggingface.co/datasets/hanjang/SciZoom), respectively.
Multi-Document Summarization (MDS) plays a critical role in distilling essential information from collections of textual data. Existing approaches often struggle to capture complex inter-document relationships, rely heavily on large amounts of labeled data for supervised training, or exhibit limited generalization across domains and languages. To address these limitations, we present a training-free mixture-of-agents framework for MDS that leverages the complementary strengths of large language models (LLMs) and knowledge graphs. Our approach decomposes summarization into specialized agent tasks: extractive selection, knowledge-aware abstraction, and iterative refinement, each operating without task-specific fine-tuning. We unify their outputs using a multi-perspective consistency mechanism guided by LLMs. Experiments across four datasets in English and Vietnamese demonstrate state-of-the-art or competitive performance, validating the effectiveness and adaptability of our modular design.
Gradient-based optimization is the workhorse of deep learning, offering efficient and scalable training via backpropagation. However, exposing gradients during training can leak sensitive information about the underlying data, raising privacy and security concerns such as susceptibility to data poisoning attacks. In contrast, black-box optimization methods, which treat the model as an opaque function, relying solely on function evaluations to guide optimization, offer a promising alternative in scenarios where data access is restricted, adversarial risks are high, or overfitting is a concern. This paper introduces BBoxER, an evolutionary black-box method for LLM post-training that induces an information bottleneck via implicit compression of the training data. Leveraging the tractability of information flow, we provide non-vacuous generalization bounds and strong theoretical guarantees for robustness to data poisoning attacks and extraction attacks, while ensuring privacy by design. In experiments with LLMs, we demonstrate empirically that black-box optimization methods-despite the scalability and computational challenges inherent to black-box approaches-are able to learn, showing how a few iterations of BBoxER improve performance, generalize well on a benchmark of reasoning datasets, and are robust to membership inference attacks. This positions BBoxER as an attractive add-on on top of gradient-based optimization, offering suitability for deployment in restricted environments while also providing non-vacuous generalization guarantees.
Multi-agent LLM environments require robust mechanisms for shared knowledge management. This paper formalizes the fleet-memory problem and identifies four foundational failure modes: unauthorized leakage, stale propagation, contradiction persistence, and provenance collapse. To address these, we define explicit systems-level primitives: scoped retrieval, temporal supersession, provenance tracking, and policy-governed memory propagation. These primitives are implemented in MemClaw, a production multi-tenant memory service, and evaluated via ArgusFleet, a reproducible harness testing four governance dimensions. Rather than a baseline comparison, this study measures a live production service, emphasizing real-world architectural insights and negative results. Key Evaluation Results Provenance: Successfully reconstructed 100% of depth-four derivation chains with correct writer identity at sub-second per-hop latency. Propagation: Demonstrated high intra-fleet visibility with zero cross-fleet leakage. Under strong write mode, write-to-visible latency was optimized to a single search round-trip. Production Architectural Issues Discovered Asymmetric Scope Enforcement: Tenant isolation held, but sub-tenant scope was initially bypassed on direct GET-by-id requests for agent-scoped credentials (disclosed and remediated during the study). Pipeline Ordering Conflict: While contradiction supersession works for admitted writes, a synchronous near-duplicate gate can prematurely reject contradictory writes before the asynchronous contradiction detector can evaluate them. Conclusion: Long-context retrieval alone is insufficient for production multi-agent memory. Governed shared memory demands explicit systems-level abstractions, and live evaluation is vital to expose enforcement and pipeline-ordering failures missed by design-only treatments.
Search and recommendation (S&R) are core to online platforms, addressing explicit intent through queries and modeling implicit intent from behaviors, respectively. Their complementary roles motivate a unified modeling paradigm. Early studies to unify S&R adopt shared encoders with task-specific heads, while recent efforts reframe item ranking in both S&R as conditional generation. The latter holds particular promise, enabling end-to-end optimization and leveraging the semantic understanding of LLMs. However, existing methods rely on full fine-tuning, which is computationally expensive and limits scalability. Parameter-efficient fine-tuning (PEFT) offers a more practical alternative but faces two critical challenges in unifying S&R: (1) gradient conflicts across tasks due to divergent optimization objectives, and (2) shifts in user intent understanding caused by overfitting to fine-tuning data, which distort general-domain knowledge and weaken LLM reasoning. To address the above issues, we propose Gradient Multi-Subspace Tuning (GEMS), a novel framework that unifies S&R with LLMs while alleviating gradient conflicts and preserving general-domain knowledge. GEMS introduces (1) \textbf{Multi-Subspace Decomposition}, which disentangles shared and task-specific optimization signals into complementary low-rank subspaces, thereby reducing destructive gradient interference, and (2) \textbf{Null-Space Projection}, which constrains parameter updates to a subspace orthogonal to the general-domain knowledge space, mitigating shifts in user intent understanding. Extensive experiments on benchmark datasets show that GEMS consistently outperforms the state-of-the-art baselines across both search and recommendation tasks, achieving superior effectiveness.
Recent studies have raised significant concerns regarding the reliability of current mathematics benchmarks, highlighting issues such as simplistic design and potential data contamination. Consequently, developing a reliable benchmark that effectively evaluates large language models' (LLMs) genuine capabilities in mathematical reasoning remains a critical challenge. To address these concerns, we propose RV-Bench, a novel evaluation methodology for Benchmarking LLMs with Random Variables in mathematical reasoning. Specifically, we build question-generating functions to produce random variable questions (RVQs), whose background content mirrors original benchmark problems, but with randomized variable combinations, rendering them "unseen" to LLMs. Models must completely understand the inherent question pattern to correctly answer RVQs with diverse variable combinations. Thus, an LLM's genuine reasoning capability is reflected through its accuracy and robustness on RV-Bench. We conducted extensive experiments on over 30 representative LLMs across more than 1,000 RVQs. Our findings propose that LLMs exhibit a proficiency imbalance between encountered and ``unseen'' data distributions. Furthermore, RV-Bench reveals that proficiency generalization across similar mathematical reasoning tasks is limited, but we verified it can still be effectively elicited through test-time scaling.
Emerging LLM services increasingly host many sparse MoE models, yet most models receive sparse requests and remain cold. This creates a GPU memory problem: model weights are stable and model-determined, while KV-cache is transient and demand-determined. Because cold models rarely reach peak KV-cache demand at the same time, reserving worst-case KV capacity per model wastes memory; a shared KV-cache pool can instead provision aggregate active demand. However, KV-cache sharing is not sufficient when weights and KV-cache remain in a monolithic GPU memory pool. Static weights compete with dynamic KV-cache, and KV-head-limited attention under cold, low-concurrency traffic exposes only a fraction of replicated KV capacity, leading to low GPU memory utilization and weak long-context support. We present CrossPool, a serving engine for cold MoE models that separates FFN weights and KV-cache into two GPU memory pools: a weights pool that consolidates FFN weights across cold models, and a KV-cache pool that dynamically serves active requests while keeping attention local to KV-cache. CrossPool combines a KV-cache planner and virtualizer, a layer-wise pipeline scheduler that hides hidden-state transfers, and persistent kernels with control lowering to reduce CPU-GPU control overhead. With efficient GPU memory pooling, CrossPool underpins bursty long-context requests and outperforms the state-of-the-art kvcached-based multi-LLM serving system, reducing P99 TBT by up to $10.4\times$.
Large language model (LLM)-based search agents synthesize open-web content into actionable recommendations on behalf of users, creating a risk that attacker-published pages are transformed into endorsed claims. We introduce SearchGEO, a controlled evaluation framework for measuring endorsement corruption in LLM-based web-search agents, combining a web-evidence manipulation pipeline, a five-mode attack taxonomy, and multiple output-level metrics. We evaluate 13 LLM backends on 308 cases each. Results show that vulnerability patterns vary across backends: overall attack success rate (ASR) ranges from 0.0% on Claude-Sonnet-4.6 to 31.4% on Gemini-3-Flash, the strongest attack mode differs by model family, and the same deployment scaffold could amplify or decrease ASR on different backends. An auxiliary agent-skill probe, where endorsement becomes an install command, exposes a sharp split among otherwise robust backends: Claude over-rejects while GPT over-trusts. These findings argue for treating recommendation reliability under adversarial search content as a first-class dimension of backend safety evaluation.
Long-context large language model (LLM) inference is increasingly constrained by the memory footprint and decoding cost of key-value (KV) caches, limiting sustainable deployment on resource-constrained hardware. Existing KV cache eviction methods typically apply heuristic token scoring over all heads in GQA-based LLMs. These methods ignore the different functionalities of attention heads, leading to the eviction of critical tokens and thus degrading the performance of LLMs. To address this issue, we propose CompressKV, a resource-efficient KV-cache compression framework for GQA-based LLMs. Instead of aggregating attention scores from all heads, CompressKV identifies Semantic Retrieval Heads (SRHs) that capture both the initial and final tokens of a prompt and semantically important mid-context evidence, and uses them to select tokens whose KV pairs should be retained. Furthermore, CompressKV allocates cache budgets across layers according to offline estimates of layer-wise eviction error. Experiments on LongBench and Needle-in-a-Haystack show that CompressKV consistently outperforms existing KV-cache eviction methods across memory budgets. Notably, it preserves over 97\% of full-cache performance using only 3\% of the KV cache on LongBench question-answering tasks and achieves 90\% accuracy with just 0.7\% KV storage on Needle-in-a-Haystack. These results demonstrate an improved resource--performance trade-off for long-context LLM inference. Our code is publicly available at: https://github.com/TUDa-HWAI/CompressKV
Autonomous systems that build structures from natural-language instructions need reliable spatial reasoning, yet large language models (LLMs) make systematic coordinate errors when generating three-dimensional block placements. We present a neuro-symbolic pipeline based on \emph{2.5-D decomposition}: the LLM plans in the two-dimensional horizontal plane while a deterministic executor computes all vertical placement from column occupancy, eliminating an entire class of errors. On the Build What I Mean benchmark (160 rounds), GPT-4o-mini with this pipeline achieves 94.6\% mean structural accuracy across 12 independent runs, within 3.0 percentage points of the 97.6\% ceiling imposed by architect-agent errors that no builder-side improvement can address. This outperforms both GPT-4o at 90.3\% and the best competing system at 76.3\%. A controlled ablation confirms that 2.5-D decomposition is the dominant contributor, accounting for 50.7 percentage points of accuracy. The pipeline transfers directly to edge hardware: Nemotron-3 120B running locally on an NVIDIA Jetson Thor AGX matches the cloud result at 94.5\% with no prompt modifications. The underlying principle, removing deterministic dimensions from the LLM's output space, applies to any autonomous construction or assembly task where gravity or other physical constraints fix one or more degrees of freedom. A transfer experiment on 500 IGLU collaborative building tasks confirm the effect generalizes beyond the primary benchmark.