llm - 2025_10
Navigation
- Part 1
- Part 2
- Part 3
- Part 4
- Part 5
- Part 6
- Part 7
- Part 8
- Part 9
- Part 10
- Part 11
- Part 12
- Part 13
- Part 14
- Part 15
- Part 16
- Part 17
- Part 18
- Part 19
- Part 20
Papers
Large language models (LLMs) deliver impressive performance but incur prohibitive memory and compute costs at deployment. Model pruning is an effective way to reduce these overheads, yet existing approaches face challenges: unstructured sparsity, where nonzeros can appear anywhere, preserves accuracy but yields irregular access patterns that prevent GPU acceleration, while semi-structured 2:4 sparsity is hardware-friendly but enforces a rigid 50% pattern that degrades model quality. To bridge this gap, we introduce PATCH, a hybrid sparsity framework that enables a continuous sparsity ratio between 0% and 50%. PATCH partitions weight matrices into tiles, assigning each tile to be either dense or 2:4 sparse via a learnable mask selection mechanism. This design provides fine-grained control over accuracy-acceleration tradeoffs and supports non-uniform sparsity across layers, leading to superior overall quality. Across models from 0.5B to 8B parameters, PATCH consistently narrows the gap to dense accuracy while delivering practical speedups. For instance, on LLaMA-2 7B with an A6000 GPU, PATCH achieves 1.18x-1.38x end-to-end speedup over dense baselines while improving accuracy by 0.37%-2.96% compared to the state-of-the-art 2:4 pruning method, MaskLLM.
Scenario simulation is central to testing autonomous driving systems. Scenic, a domain-specific language (DSL) for CARLA, enables precise and reproducible scenarios, but NL-to-Scenic generation with large language models (LLMs) suffers from scarce data, limited reproducibility, and inconsistent metrics. We introduce NL2Scenic, an open dataset and framework with 146 NL/Scenic pairs, a difficulty-stratified 30-case test split, an Example Retriever, and 14 prompting variants (ZS, FS, CoT, SP, MoT). We evaluate 13 models: four proprietary (GPT-4o, GPT-5, Claude-Sonnet-4, Gemini-2.5-pro) and nine open-source code models (Qwen2.5Coder 0.5B-32B; CodeLlama 7B/13B/34B), using text metrics (BLEU, ChrF, EDIT-SIM, CrystalBLEU) and execution metrics (compilation and generation), and compare them with an expert study (n=11). EDIT-SIM correlates best with human judgments; we also propose EDIT-COMP (F1 of EDIT-SIM and compilation) as a robust dataset-level proxy that improves ranking fidelity. GPT-4o performs best overall, while Qwen2.5Coder-14B reaches about 88 percent of its expert score on local hardware. Retrieval-augmented prompting, Few-Shot with Example Retriever (FSER), consistently boosts smaller models, and scaling shows diminishing returns beyond mid-size, with Qwen2.5Coder outperforming CodeLlama at comparable scales. NL2Scenic and EDIT-COMP offer a standardized, reproducible basis for evaluating Scenic code generation and indicate that mid-size open-source models are practical, cost-effective options for autonomous-driving scenario programming.
We introduce time-to-unsafe-sampling, a novel safety measure for generative models, defined as the number of generations required by a large language model (LLM) to trigger an unsafe (e.g., toxic) response. While providing a new dimension for prompt-adaptive safety evaluation, quantifying time-to-unsafe-sampling is challenging: unsafe outputs are often rare in well-aligned models and thus may not be observed under any feasible sampling budget. To address this challenge, we frame this estimation problem as one of survival analysis. We build on recent developments in conformal prediction and propose a novel calibration technique to construct a lower predictive bound (LPB) on the time-to-unsafe-sampling of a given prompt with rigorous coverage guarantees. Our key technical innovation is an optimized sampling-budget allocation scheme that improves sample efficiency while maintaining distribution-free guarantees. Experiments on both synthetic and real data support our theoretical results and demonstrate the practical utility of our method for safety risk assessment in generative AI models.
Artificial Intelligence (AI) is expected to be an integral part of next-generation AI-native 6G networks. With the prevalence of AI, researchers have identified numerous use cases of AI in network security. However, there are very few studies that analyze the suitability of Large Language Models (LLMs) in network security. To fill this gap, we examine the suitability of LLMs in network security, particularly with the case study of STRIDE threat modeling. We utilize four prompting techniques with five LLMs to perform STRIDE classification of 5G threats. From our evaluation results, we point out key findings and detailed insights along with the explanation of the possible underlying factors influencing the behavior of LLMs in the modeling of certain threats. The numerical results and the insights support the necessity for adjusting and fine-tuning LLMs for network security use cases.
Probabilistic reasoning is a key aspect of both human and artificial intelligence that allows for handling uncertainty and ambiguity in decision-making. In this paper, we introduce a new numerical reasoning task under uncertainty for large language models, focusing on estimating the privacy risk of user-generated documents containing privacy-sensitive information. We propose BRANCH, a new LLM methodology that estimates the k-privacy value of a text-the size of the population matching the given information. BRANCH factorizes a joint probability distribution of personal information as random variables. The probability of each factor in a population is estimated separately using a Bayesian network and combined to compute the final k-value. Our experiments show that this method successfully estimates the k-value 73% of the time, a 13% increase compared to o3-mini with chain-of-thought reasoning. We also find that LLM uncertainty is a good indicator for accuracy, as high-variance predictions are 37.47% less accurate on average.
Fixing bugs in large programs is a challenging task that demands substantial time and effort. Once a bug is found, it is reported to the project maintainers, who work with the reporter to fix it and eventually close the issue. However, across the program, there are often similar code segments, which may also contain the bug, but were missed during discovery. Finding and fixing each recurring bug instance individually is labor intensive. Even more concerning, bug reports can inadvertently widen the attack surface as they provide attackers with an exploitable pattern that may be unresolved in other parts of the program. In this paper, we explore these Recurring Pattern Bugs (RPBs) that appear repeatedly across various code segments of a program or even in different programs, stemming from a same root cause, but are unresolved. Our investigation reveals that RPBs are widespread and can significantly compromise the security of software programs. This paper introduces BugStone, a program analysis system empowered by LLVM and a Large Language Model (LLM). The key observation is that many RPBs have one patched instance, which can be leveraged to identify a consistent error pattern, such as a specific API misuse. By examining the entire program for this pattern, it is possible to identify similar sections of code that may be vulnerable. Starting with 135 unique RPBs, BugStone identified more than 22K new potential issues in the Linux kernel. Manual analysis of 400 of these findings confirmed that 246 were valid. We also created a dataset from over 1.9K security bugs reported by 23 recent top-tier conference works. We manually annotate the dataset, identify 80 recurring patterns and 850 corresponding fixes. Even with a cost-efficient model choice, BugStone achieved 92.2% precision and 79.1% pairwise accuracy on the dataset.
Reasoning and strategic behavior in social interactions is a hallmark of intelligence. This form of reasoning is significantly more sophisticated than isolated planning or reasoning tasks in static settings (e.g., math problem solving). In this paper, we present Strategic Planning, Interaction, and Negotiation (SPIN-Bench), a new multi-domain evaluation designed to measure the intelligence of strategic planning and social reasoning. While many existing benchmarks focus on narrow planning or single-agent reasoning, SPIN-Bench combines classical PDDL tasks, competitive board games, cooperative card games, and multi-agent negotiation scenarios in one unified framework. The framework includes both a benchmark as well as an arena to simulate and evaluate the variety of social settings to test reasoning and strategic behavior of AI agents. We formulate the benchmark SPIN-Bench by systematically varying action spaces, state complexity, and the number of interacting agents to simulate a variety of social settings where success depends on not only methodical and step-wise decision making, but also conceptual inference of other (adversarial or cooperative) participants. Our experiments reveal that while contemporary LLMs handle basic fact retrieval and short-range planning reasonably well, they encounter significant performance bottlenecks in tasks requiring deep multi-hop reasoning over large state spaces and socially adept coordination under uncertainty. We envision SPIN-Bench as a catalyst for future research on robust multi-agent planning, social reasoning, and human--AI teaming. Project Website: https://spinbench.github.io/
Large language models (LLMs) can exhibit biases in reasoning capabilities due to linguistic modality, performing better on tasks in one language versus another, even with similar content. Most previous works evaluate this through reasoning tasks where reliance on strategies or knowledge can ensure success, such as in commonsense or math tasks. However, abstract reasoning is vital to reasoning for everyday life, where people apply "out-of-the-box thinking" to identify and use patterns for solutions, without a reliance on formulaic approaches. Comparatively, little work has evaluated linguistic biases in this task type. In this paper, we propose a task inspired by the New York Times Connections: GlobalGroup, that evaluates models in an abstract reasoning task across several languages. We constructed a game benchmark with five linguistic backgrounds -- English, Spanish, Chinese, Hindi, and Arabic -- in both the native language and an English translation for comparison. We also proposed game difficulty measurements to evaluate models on games with similar difficulty, enabling a more controlled comparison, which is particularly important in reasoning evaluations. Through experimentation, we find English modalities largely lead to better performance in this abstract reasoning task, and performance disparities between open- and closed-source models.
The rise of Generative AI introduces a new class of HPC workloads that integrates lightweight LLMs with traditional high-throughput applications to accelerate scientific discovery. The current design of HPC clusters is inadequate to support this new class however, either incurring long wait times on static batch queues or repeatedly paying expensive LLM startup costs upon resource preemption. To circumvent both the long queues and high startup costs, we propose to "decouple" the LLM initialization context from the actual LLM inferences, and retain the context in GPUs until it is no longer needed, a technique we term "Pervasive Context Management". We transform a fact verification application to enable this technique, allowing it to reduce its execution time by 72.1% (from 3 hours to 48 minutes) using the same amount of GPUs, and scale opportunistically on 32.8% of all GPUs in the cluster and further reduce the execution time to 13 minutes.
LLM-powered Multi-Agent Systems (LLM-MAS) unlock new potentials in distributed reasoning, collaboration, and task generalization but also introduce additional risks due to unguaranteed agreement, cascading uncertainty, and adversarial vulnerabilities. We argue that ensuring responsible behavior in such systems requires a paradigm shift: from local, superficial agent-level alignment to global, systemic agreement. We conceptualize responsibility not as a static constraint but as a lifecycle-wide property encompassing agreement, uncertainty, and security, each requiring the complementary integration of subjective human-centered values and objective verifiability. Furthermore, a dual-perspective governance framework that combines interdisciplinary design with human-AI collaborative oversight is essential for tracing and ensuring responsibility throughout the lifecycle of LLM-MAS. Our position views LLM-MAS not as loose collections of agents, but as unified, dynamic socio-technical systems that demand principled mechanisms to support each dimension of responsibility and enable ethically aligned, verifiably coherent, and resilient behavior for sustained, system-wide agreement.
LLM-integrated applications are vulnerable to prompt injection attacks, where an attacker contaminates the input to inject malicious prompts, causing the LLM to follow the attacker's intent instead of the original user's. Existing prompt injection detection methods often have sub-optimal performance and/or high computational overhead. In this work, we propose PIShield, a detection method that is both effective and efficient. Our key observation is that the internal representation of the final token in a prompt-extracted from a specific layer of the LLM, which we term the injection-critical layer-captures distinguishing features between clean and contaminated prompts. Leveraging this insight, we train a simple linear classifier on these internal representations using a labeled set of clean and contaminated prompts. We compare PIShield against 11 baselines across 5 diverse benchmark datasets and 8 prompt injection attacks. The results demonstrate that PIShield is both highly effective and efficient, substantially outperforming existing methods. Additionally, we show that PIShield resists strong adaptive attacks.
What if artificial agents could not just communicate, but also evolve, adapt, and reshape their worlds in ways we cannot fully predict? With llm now powering multi-agent systems and social simulations, we are witnessing new possibilities for modeling open-ended, ever-changing environments. Yet, most current simulations remain constrained within static sandboxes, characterized by predefined tasks, limited dynamics, and rigid evaluation criteria. These limitations prevent them from capturing the complexity of real-world societies. In this paper, we argue that static, task-specific benchmarks are fundamentally inadequate and must be rethought. We critically review emerging architectures that blend llm with multi-agent dynamics, highlight key hurdles such as balancing stability and diversity, evaluating unexpected behaviors, and scaling to greater complexity, and introduce a fresh taxonomy for this rapidly evolving field. Finally, we present a research roadmap centered on open-endedness, continuous co-evolution, and the development of resilient, socially aligned AI ecosystems. \textbf{We call on the community to move beyond static paradigms and help shape the next generation of adaptive, socially-aware multi-agent simulations.}
Recent progress in large language models (LLMs) has focused on test-time scaling to improve reasoning via increased inference computation, but often at the cost of efficiency. We revisit test-time behavior and uncover a simple yet underexplored phenomenon: reasoning uncertainty is highly localized-only a small subset of high-entropy tokens dominantly affects output correctness. Motivated by this, we propose Minimal Test-Time Intervention (MTI), a training-free framework that enhances reasoning accuracy and stability with minimal overhead. MTI includes: (i) Selective CFG intervention, applying classifier-free guidance only at uncertain positions; and (ii) Lightweight negative-prompt guidance, reusing the main model's KV cache to approximate unconditional decoding efficiently. MTI yields consistent gains across general, coding, and STEM tasks-e.g., +1.35% average improvement on eight benchmarks for Qwen3-8B-Base and +5% on AIME2024 using Qwen3-32B-Reasoning-while remaining highly efficient.
Large language models (LLMs) are increasingly applied in biomedical domains, yet their reliability in drug-safety prediction remains underexplored. In this work, we investigate whether LLMs incorporate socio-demographic information into adverse event (AE) predictions, despite such attributes being clinically irrelevant. Using structured data from the United States Food and Drug Administration Adverse Event Reporting System (FAERS) and a persona-based evaluation framework, we assess two state-of-the-art models, ChatGPT-4o and Bio-Medical-Llama-3.8B, across diverse personas defined by education, marital status, employment, insurance, language, housing stability, and religion. We further evaluate performance across three user roles (general practitioner, specialist, patient) to reflect real-world deployment scenarios where commercial systems often differentiate access by user type. Our results reveal systematic disparities in AE prediction accuracy. Disadvantaged groups (e.g., low education, unstable housing) were frequently assigned higher predicted AE likelihoods than more privileged groups (e.g., postgraduate-educated, privately insured). Beyond outcome disparities, we identify two distinct modes of bias: explicit bias, where incorrect predictions directly reference persona attributes in reasoning traces, and implicit bias, where predictions are inconsistent, yet personas are not explicitly mentioned. These findings expose critical risks in applying LLMs to pharmacovigilance and highlight the urgent need for fairness-aware evaluation protocols and mitigation strategies before clinical deployment.
We propose and test the LLM Brain Rot Hypothesis: continual exposure to junk web text induces lasting cognitive decline in large language models (LLMs). To causally isolate data quality, we run controlled experiments on real Twitter/X corpora, constructing junk and reversely controlled datasets via two orthogonal operationalizations: M1 (engagement degree) and M2 (semantic quality), with matched token scale and training operations across conditions. Contrary to the control group, continual pre-training of 4 LLMs on the junk dataset causes non-trivial declines (Hedges' $g>0.3$) on reasoning, long-context understanding, safety, and inflating "dark traits" (e.g., psychopathy, narcissism). The gradual mixtures of junk and control datasets also yield dose-response cognition decay: for example, under M1, ARC-Challenge with Chain Of Thoughts drops $74.9 \rightarrow 57.2$ and RULER-CWE $84.4 \rightarrow 52.3$ as junk ratio rises from $0\%$ to $100\%$. Error forensics reveal several key insights. First, we identify thought-skipping as the primary lesion: models increasingly truncate or skip reasoning chains, explaining most of the error growth. Second, partial but incomplete healing is observed: scaling instruction tuning and clean data pre-training improve the declined cognition yet cannot restore baseline capability, suggesting persistent representational drift rather than format mismatch. Finally, we discover that the popularity, a non-semantic metric, of a tweet is a better indicator of the Brain Rot effect than the length in M1. Together, the results provide significant, multi-perspective evidence that data quality is a causal driver of LLM capability decay, reframing curation for continual pretraining as a \textit{training-time safety} problem and motivating routine "cognitive health checks" for deployed LLMs.
Biomedical queries often rely on a deep understanding of specialized knowledge such as gene regulatory mechanisms and pathological processes of diseases. They require detailed analysis of complex physiological processes and effective integration of information from multiple data sources to support accurate retrieval and reasoning. Although large language models (LLMs) perform well in general reasoning tasks, their generated biomedical content often lacks scientific rigor due to the inability to access authoritative biomedical databases and frequently fabricates protein functions, interactions, and structural details that deviate from authentic information. Therefore, we present BioMedSearch, a multi-source biomedical information retrieval framework based on LLMs. The method integrates literature retrieval, protein database and web search access to support accurate and efficient handling of complex biomedical queries. Through sub-queries decomposition, keywords extraction, task graph construction, and multi-source information filtering, BioMedSearch generates high-quality question-answering results. To evaluate the accuracy of question answering, we constructed a multi-level dataset, BioMedMCQs, consisting of 3,000 questions. The dataset covers three levels of reasoning: mechanistic identification, non-adjacent semantic integration, and temporal causal reasoning, and is used to assess the performance of BioMedSearch and other methods on complex QA tasks. Experimental results demonstrate that BioMedSearch consistently improves accuracy over all baseline models across all levels. Specifically, at Level 1, the average accuracy increases from 59.1% to 91.9%; at Level 2, it rises from 47.0% to 81.0%; and at the most challenging Level 3, the average accuracy improves from 36.3% to 73.4%. The code and BioMedMCQs are available at: https://github.com/CyL-ucas/BioMed_Search
Internet of Things (IoT) networks generate diverse and high-volume traffic that reflects both normal activity and potential threats. Deriving meaningful insight from such telemetry requires cross-layer interpretation of behaviors, protocols, and context rather than isolated detection. This work presents an LLM-powered AI agent framework that converts raw packet captures into structured and semantically enriched representations for interactive analysis. The framework integrates feature extraction, transformer-based anomaly detection, packet and flow summarization, threat intelligence enrichment, and retrieval-augmented question answering. An AI agent guided by a large language model performs reasoning over the indexed traffic artifacts, assembling evidence to produce accurate and human-readable interpretations. Experimental evaluation on multiple IoT captures and six open models shows that hybrid retrieval, which combines lexical and semantic search with reranking, substantially improves BLEU, ROUGE, METEOR, and BERTScore results compared with dense-only retrieval. System profiling further indicates low CPU, GPU, and memory overhead, demonstrating that the framework achieves holistic and efficient interpretation of IoT network traffic.
Process reward models (PRMs) are a cornerstone of test-time scaling (TTS), designed to verify and select the best responses from large language models (LLMs). However, this promise is challenged by recent benchmarks where simple majority voting, which ignores PRM signals, occasionally outperforms standard PRM-based selection. This raises a critical question: How can we effectively utilize verification signals from PRMs for TTS? To address this, we start by developing a theoretical framework for optimally combining signals from both the LLM and the PRM. Our framework reveals that the optimal strategy is a weighted aggregation of responses, a strategy whose effectiveness hinges on estimating weights that capture the complex interplay between the models. Based on our theoretical results, we empirically show that these optimal weighting functions differ significantly across LLM-PRM pairs and, notably, often assign substantial negative weights. Motivated by these insights, we propose efficient pre-computation methods to calibrate these weighting functions. Extensive experiments across 5 LLMs and 7 PRMs demonstrate that our calibration method significantly boosts the TTS efficiency, surpassing the performance of vanilla weighted majority voting while using only $21.3\%$ of the computation. Ultimately, our work demonstrates that investing in a more intelligent aggregation strategy can be a more convincing path to performance gains than simply scaling test-time computation.
Large language models (LLMs) have recently demonstrated strong capabilities in generating functional and aesthetic web interfaces directly from instructions. However, these models often replicate accessibility flaws from their training data, resulting in interfaces that exclude users with diverse needs and contexts. To address this gap, we introduce A11yn, the first method that aligns code-generating LLMs to reliably produce accessibility-compliant web UIs. A11yn optimizes a novel reward function that penalizes violations of the Web Content Accessibility Guidelines (WCAG), with penalties scaled to the severity of each violation as identified by an accessibility testing engine. To support training, we construct UIReq-6.8K, a dataset of 6,800 diverse instructions for web UI generation. For evaluation, we introduce RealUIReq-300, a benchmark of 300 real-world web UI requests grounded and manually curated from public web pages, spanning a broad range of use cases. Empirical results show that A11yn significantly outperforms strong baselines, lowering the Inaccessibility Rate by 60% over the base model while preserving semantic fidelity and visual quality of generated UIs. These findings demonstrate that accessibility can be systematically optimized within LLMs, showing the feasibility of aligning code generation for accessibility.
Retrieval-Augmented Generation (RAG) mitigates key limitations of Large Language Models (LLMs)-such as factual errors, outdated knowledge, and hallucinations-by dynamically retrieving external information. Recent work extends this paradigm through agentic RAG systems, where LLMs act as agents to iteratively plan, retrieve, and reason over complex queries. However, these systems still struggle with challenging multi-hop questions, and their intermediate reasoning capabilities remain underexplored. To address this, we propose RAGCap-Bench, a capability-oriented benchmark for fine-grained evaluation of intermediate tasks in agentic RAG workflows. We analyze outputs from state-of-the-art systems to identify common tasks and the core capabilities required for their execution, then construct a taxonomy of typical LLM errors to design targeted evaluation questions. Experiments show that "slow-thinking" models with stronger RAGCap performance achieve better end-to-end results, underscoring the benchmark's validity and the importance of enhancing these intermediate capabilities.
Recent advances in diffusion large language models (dLLMs) have introduced a promising alternative to autoregressive (AR) LLMs for natural language generation tasks, leveraging full attention and denoising-based decoding strategies. However, the deployment of these models on edge devices remains challenging due to their massive parameter scale and high resource demands. While post-training quantization (PTQ) has emerged as a widely adopted technique for compressing AR LLMs, its applicability to dLLMs remains largely unexplored. In this work, we present the first systematic study on quantizing diffusion-based language models. We begin by identifying the presence of activation outliers, characterized by abnormally large activation values that dominate the dynamic range. These outliers pose a key challenge to low-bit quantization, as they make it difficult to preserve precision for the majority of values. More importantly, we implement state-of-the-art PTQ methods and conduct a comprehensive evaluation across multiple task types and model variants. Our analysis is structured along four key dimensions: bit-width, quantization method, task category, and model type. Through this multi-perspective evaluation, we offer practical insights into the quantization behavior of dLLMs under different configurations. We hope our findings provide a foundation for future research in efficient dLLM deployment. Our code is publicly available at https://github.com/FelixMessi/QDLM.
Multilingual large language models (LLMs) often demonstrate a performance gap between English and non-English languages, particularly in low-resource settings. Aligning these models to low-resource languages is essential yet challenging due to limited high-quality data. While English alignment datasets are readily available, curating equivalent data in other languages is expensive and time-consuming. A common workaround is to translate existing English alignment data; however, standard translation techniques often fail to preserve critical elements such as code, mathematical expressions, and structured formats like JSON. In this work, we investigate LLM-based selective translation, a technique that selectively translates only the translatable parts of a text while preserving non-translatable content and sentence structure. We conduct a systematic study to explore key questions around this approach, including its effectiveness compared to vanilla translation, the importance of filtering noisy outputs, and the benefits of mixing translated samples with original English data during alignment. Our experiments focus on the low-resource Indic language Hindi and compare translations generated by Google Cloud Translation (GCP) and Llama-3.1-405B. The results highlight the promise of selective translation as a practical and effective method for improving multilingual alignment in LLMs.
Reinforcement learning (RL) has become central to training large language models (LLMs), yet the field lacks predictive scaling methodologies comparable to those established for pre-training. Despite rapidly rising compute budgets, there is no principled understanding of how to evaluate algorithmic improvements for scaling RL compute. We present the first large-scale systematic study, amounting to more than 400,000 GPU-hours, that defines a principled framework for analyzing and predicting RL scaling in LLMs. We fit sigmoidal compute-performance curves for RL training and ablate a wide range of common design choices to analyze their effects on asymptotic performance and compute efficiency. We observe: (1) Not all recipes yield similar asymptotic performance, (2) Details such as loss aggregation, normalization, curriculum, and off-policy algorithm primarily modulate compute efficiency without materially shifting the asymptote, and (3) Stable, scalable recipes follow predictable scaling trajectories, enabling extrapolation from smaller-scale runs. Combining these insights, we propose a best-practice recipe, ScaleRL, and demonstrate its effectiveness by successfully scaling and predicting validation performance on a single RL run scaled up to 100,000 GPU-hours. Our work provides both a scientific framework for analyzing scaling in RL and a practical recipe that brings RL training closer to the predictability long achieved in pre-training.
Large Language Models (LLMs) exhibit impressive reasoning abilities, yet their reliance on structured step-by-step processing reveals a critical limitation. In contrast, human cognition fluidly adapts between intuitive, heuristic (System 1) and analytical, deliberative (System 2) reasoning depending on the context. This difference between human cognitive flexibility and LLMs' reliance on a single reasoning style raises a critical question: while human fast heuristic reasoning evolved for its efficiency and adaptability, is a uniform reasoning approach truly optimal for LLMs, or does its inflexibility make them brittle and unreliable when faced with tasks demanding more agile, intuitive responses? To answer these questions, we explicitly align LLMs to these reasoning styles by curating a dataset with valid System 1 and System 2 answers, and evaluate their performance across reasoning benchmarks. Our results reveal an accuracy-efficiency trade-off: System 2-aligned models excel in arithmetic and symbolic reasoning, while System 1-aligned models perform better in commonsense reasoning tasks. To analyze the reasoning spectrum, we interpolated between the two extremes by varying the proportion of alignment data, which resulted in a monotonic change in accuracy. A mechanistic analysis of model responses shows that System 1 models employ more definitive outputs, whereas System 2 models demonstrate greater uncertainty. Building on these findings, we further combine System 1- and System 2-aligned models based on the entropy of their generations, without additional training, and obtain a dynamic model that outperforms across nearly all benchmarks. This work challenges the assumption that step-by-step reasoning is always optimal and highlights the need for adapting reasoning strategies based on task demands.
Evaluating instruction-tuned Large Language Models (LLMs) in Hindi is challenging due to a lack of high-quality benchmarks, as direct translation of English datasets fails to capture crucial linguistic and cultural nuances. To address this, we introduce a suite of five Hindi LLM evaluation datasets: IFEval-Hi, MT-Bench-Hi, GSM8K-Hi, ChatRAG-Hi, and BFCL-Hi. These were created using a methodology that combines from-scratch human annotation with a translate-and-verify process. We leverage this suite to conduct an extensive benchmarking of open-source LLMs supporting Hindi, providing a detailed comparative analysis of their current capabilities. Our curation process also serves as a replicable methodology for developing benchmarks in other low-resource languages.
We propose a method for confidence estimation in retrieval-augmented generation (RAG) systems that aligns closely with the correctness of large language model (LLM) outputs. Confidence estimation is especially critical in high-stakes domains such as finance and healthcare, where the cost of an incorrect answer outweighs that of not answering the question. Our approach extends prior uncertainty quantification methods by leveraging raw feed-forward network (FFN) activations as auto-regressive signals, avoiding the information loss inherent in token logits and probabilities after projection and softmax normalization. We model confidence prediction as a sequence classification task, and regularize training with a Huber loss term to improve robustness against noisy supervision. Applied in a real-world financial industry customer-support setting with complex knowledge bases, our method outperforms strong baselines and maintains high accuracy under strict latency constraints. Experiments on Llama 3.1 8B model show that using activations from only the 16th layer preserves accuracy while reducing response latency. Our results demonstrate that activation-based confidence modeling offers a scalable, architecture-aware path toward trustworthy RAG deployment.
Large language models (LLMs) have recently demonstrated strong potential for sequential recommendation. However, current LLM-based approaches face critical limitations in modeling users' long-term and diverse interests. First, due to inference latency and feature fetching bandwidth constraints, existing methods typically truncate user behavior sequences to include only the most recent interactions, resulting in the loss of valuable long-range preference signals. Second, most current methods rely on next-item prediction with a single predicted embedding, overlooking the multifaceted nature of user interests and limiting recommendation diversity. To address these challenges, we propose HyMiRec, a hybrid multi-interest sequential recommendation framework, which leverages a lightweight recommender to extracts coarse interest embeddings from long user sequences and an LLM-based recommender to captures refined interest embeddings. To alleviate the overhead of fetching features, we introduce a residual codebook based on cosine similarity, enabling efficient compression and reuse of user history embeddings. To model the diverse preferences of users, we design a disentangled multi-interest learning module, which leverages multiple interest queries to learn disentangles multiple interest signals adaptively, allowing the model to capture different facets of user intent. Extensive experiments are conducted on both benchmark datasets and a collected industrial dataset, demonstrating our effectiveness over existing state-of-the-art methods. Furthermore, online A/B testing shows that HyMiRec brings consistent improvements in real-world recommendation systems.
Pruning is a common technique to reduce the compute and storage requirements of Neural Networks. While conventional approaches typically retrain the model to recover pruning-induced performance degradation, state-of-the-art Large Language Model (LLM) pruning methods operate layer-wise, minimizing the per-layer pruning error on a small calibration dataset to avoid full retraining, which is considered computationally prohibitive for LLMs. However, finding the optimal pruning mask is a hard combinatorial problem and solving it to optimality is intractable. Existing methods hence rely on greedy heuristics that ignore the weight interactions in the pruning objective. In this work, we instead consider the convex relaxation of these combinatorial constraints and solve the resulting problem using the Frank-Wolfe (FW) algorithm. Our method drastically reduces the per-layer pruning error, outperforms strong baselines on state-of-the-art GPT architectures, and remains memory-efficient. We provide theoretical justification by showing that, combined with the convergence guarantees of the FW algorithm, we obtain an approximate solution to the original combinatorial problem upon rounding the relaxed solution to integrality.
Mathematical reasoning serves as a crucial testbed for evaluating the intelligence of large language models (LLMs), and math word problems (MWPs) represent one of the most widely used formats. Most existing MWP datasets contain only the necessary information, while problems with distracting or excessive conditions are often overlooked. Prior studies have shown that popular LLMs experience a dramatic performance drop when such distracting conditions are introduced. However, available datasets of MWPs with distracting conditions remain limited, and most exhibit low difficulty and out-of-context expressions. These shortcomings make the distracting conditions easy to detect and disregard, thereby reducing the credibility of benchmarking on these datasets. Moreover, when distracting conditions are added, the reasoning process and answers may change, requiring intensive manual effort to check and rewrite solutions. To address these issues, we design an iterative framework that leverages LLMs to generate distracting conditions automatically. We develop a set of prompts to revise MWPs from multiple perspectives and cognitive levels, encouraging the creation of meaningful distracting conditions as well as suggestions for further refinement. A key advantage of our framework is the preservation of shared solutions between the original and revised problems: the LLMs are explicitly guided to generate distractions that do not alter the original solution, thus eliminating the need to produce new answers. This framework is efficient and easy to deploy, substantially reducing the effort required to generate MWPs with distracting conditions while maintaining high data quality.
Unmanned Aerial Vehicles (UAVs) are increasingly being utilized in various private and commercial applications, e.g., traffic control, parcel delivery, and Search and Rescue (SAR) missions. Machine Learning (ML) methods used in UAV-Assisted Sensor Networks (UASNETs) and, especially, in Deep Reinforcement Learning (DRL) face challenges such as complex and lengthy model training, gaps between simulation and reality, and low sampling efficiency, which conflict with the urgency of emergencies, such as SAR missions. In this paper, an In-Context Learning (ICL)-Data Collection Scheduling (ICLDC) system is proposed as an alternative to DRL in emergencies. The UAV collects sensory data and transmits it to a Large Language Model (LLM), which creates a task description in natural language. From this description, the UAV receives a data collection schedule that must be executed. A verifier ensures safe UAV operations by evaluating the schedules generated by the LLM and overriding unsafe schedules based on predefined rules. The system continuously adapts by incorporating feedback into the task descriptions and using this for future decisions. This method is tested against jailbreaking attacks, where the task description is manipulated to undermine network performance, highlighting the vulnerability of LLMs to such attacks. The proposed ICLDC significantly reduces cumulative packet loss compared to both the DQN and Maximum Channel Gain baselines. ICLDC presents a promising direction for intelligent scheduling and control in UASNETs.
Large Language Model (LLM) inference has emerged as a fundamental paradigm. In real-world scenarios, variations in output length cause severe workload imbalance in the decode phase, particularly for long-output reasoning tasks. Existing systems, such as PD disaggregation architectures, rely on static prefill-to-decode scheduling, which often results in SLO violations and OOM failures under evolving decode workloads. In this paper, we propose ARES, an adaptive decoding rescheduling system powered by length prediction to anticipate future workloads. Our core contributions include: (1) A lightweight and continuous LLM-native prediction method that leverages LLM hidden state to model remaining generation length with high precision (reducing MAE by 49.42%) and low overhead (cutting predictor parameters by 93.28%); (2) A rescheduling solution in decode phase with : A dynamic balancing mechanism that integrates current and predicted workloads, reducing P99 TPOT by 74.77% and achieving up to 2.24 times higher goodput.
The emergence of large language models (LLMs) enables the development of intelligent agents capable of engaging in complex and multi-turn dialogues. However, multi-agent collaboration faces critical safety challenges, such as hallucination amplification and error injection and propagation. This paper presents GUARDIAN, a unified method for detecting and mitigating multiple safety concerns in GUARDing Intelligent Agent collaboratioNs. By modeling the multi-agent collaboration process as a discrete-time temporal attributed graph, GUARDIAN explicitly captures the propagation dynamics of hallucinations and errors. The unsupervised encoder-decoder architecture incorporating an incremental training paradigm learns to reconstruct node attributes and graph structures from latent embeddings, enabling the identification of anomalous nodes and edges with unparalleled precision. Moreover, we introduce a graph abstraction mechanism based on the Information Bottleneck Theory, which compresses temporal interaction graphs while preserving essential patterns. Extensive experiments demonstrate GUARDIAN's effectiveness in safeguarding LLM multi-agent collaborations against diverse safety vulnerabilities, achieving state-of-the-art accuracy with efficient resource utilization. The code is available at https://github.com/JialongZhou666/GUARDIAN
Large Language Models (LLMs) can be adapted to extend their text capabilities to speech inputs. However, these speech-adapted LLMs consistently underperform their text-based counterparts--and even cascaded pipelines--on language understanding tasks. We term this shortfall the text-speech understanding gap: the performance drop observed when a speech-adapted LLM processes spoken inputs relative to when the original text-based LLM processes the equivalent text. Recent approaches to narrowing this gap either rely on large-scale speech synthesis of text corpora, which is costly and heavily dependent on synthetic data, or on large-scale proprietary speech datasets, which are not reproducible. As a result, there remains a need for more data-efficient alternatives for closing the text-speech understanding gap. In this work, we analyze the gap as driven by two factors: (i) forgetting of text capabilities during adaptation, and (ii) cross-modal misalignment between speech and text. Based on this analysis, we introduce SALAD--Sample-efficient Alignment with Learning through Active selection and cross-modal Distillation--which combines cross-modal distillation with targeted synthetic data to improve alignment while mitigating forgetting. Applied to 3B and 7B LLMs, SALAD achieves competitive performance with a strong open-weight model across broad-domain benchmarks in knowledge, language understanding, and reasoning, while training on over an order of magnitude less speech data from public corpora.
Accurate coding of tumor diagnoses with ICD-10-GM and ICD-O-3 is essential for structured cancer documentation in Germany. Smaller open-weight LLMs are appealing for privacy-preserving automation but often struggle with coding accuracy in German-language contexts. This study investigates whether instruction-based fine-tuning on public datasets improves the coding accuracy of open-weight LLMs for German tumor diagnosis texts. The evaluation uses coded diagnoses from the local tumor documentation system as test data. In a systematic data quality assessment, the upper limit for ICD-10 coding performance was estimated at 60-79% for exact and 81-94% for partial (three-character codes only) derivation. As training data, over 500,000 question-answer pairs were created based on the ICD-10-GM, ICD-O-3, and OPS catalogues. Eight open-weight models from the Qwen, Llama, and Mistral families (7-70 B parameters) were fine-tuned. ICD-10-GM accuracy rose from 1.4-24% to 41-58%, and partial accuracy from 31-74% to 73-83%. The accuracy of ICD-O-3 topography coding also improved but started and remained considerably lower with an exact accuracy of 22-40% and a partial accuracy of 56-67% after fine-tuning. Malformed code outputs dropped to 0% for all models. Tumor-diagnosis recognition reached 99%. Accuracy correlated positively with model size, but gaps between small and large models narrowed after fine-tuning. The reasoning mode in Qwen3 generally yielded a lower performance than fine-tuning and was over 100 times slower. Our findings highlight the potential of leveraging public catalogues to build instruction datasets that improve LLMs in medical documentation tasks. The complete training dataset and the best-performing checkpoints of the fine-tuned models are available from https://huggingface.co/datasets/stefan-m-lenz/ICDOPS-QA-2024.
In finance, Large Language Models (LLMs) face frequent knowledge conflicts arising from discrepancies between their pre-trained parametric knowledge and real-time market data. These conflicts are especially problematic in real-world investment services, where a model's inherent biases can misalign with institutional objectives, leading to unreliable recommendations. Despite this risk, the intrinsic investment biases of LLMs remain underexplored. We propose an experimental framework to investigate emergent behaviors in such conflict scenarios, offering a quantitative analysis of bias in LLM-based investment analysis. Using hypothetical scenarios with balanced and imbalanced arguments, we extract the latent biases of models and measure their persistence. Our analysis, centered on sector, size, and momentum, reveals distinct, model-specific biases. Across most models, a tendency to prefer technology stocks, large-cap stocks, and contrarian strategies is observed. These foundational biases often escalate into confirmation bias, causing models to cling to initial judgments even when faced with increasing counter-evidence. A public leaderboard benchmarking bias across a broader set of models is available at https://linqalpha.com/leaderboard
The emergence of large language models (LLMs) has opened new opportunities for cre- ating dynamic non-player characters (NPCs) in gaming environments, enabling both func- tional task execution and persona-consistent dialogue generation. In this paper, we (Tu_Character_lab) report our participation in the Commonsense Persona-Grounded Dialogue Challenge (CPDC) 2025 Round 2, which eval- uates agents across three tracks: task-oriented dialogue, context-aware dialogue, and their integration. Our approach combines two complementary strategies: (i) lightweight prompting techniques in the API track, including a Deflanderization prompting method to suppress excessive role-play and improve task fidelity, and (ii) fine-tuned large models in the GPU track, leveraging Qwen3-14B with supervisedfinetuning (SFT) and Low-Rank Adaptation(LoRA). Our best submissions ranked 2nd on Task 1, 2nd on Task 3 (API track), and 4th on Task 3 (GPU track).
The co-development of hardware and software in industrial embedded systems frequently leads to compilation errors during continuous integration (CI). Automated repair of such failures is promising, but existing techniques rely on test cases, which are not available for non-compilable code. We employ an automated repair approach for compilation errors driven by large language models (LLMs). Our study encompasses the collection of more than 40000 commits from the product's source code. We assess the performance of an industrial CI system enhanced by four state-of-the-art LLMs, comparing their outcomes with manual corrections provided by human programmers. LLM-equipped CI systems can resolve up to 63 % of the compilation errors in our baseline dataset. Among the fixes associated with successful CI builds, 83 % are deemed reasonable. Moreover, LLMs significantly reduce debugging time, with the majority of successful cases completed within 8 minutes, compared to hours typically required for manual debugging.
Benchmarking outcomes increasingly govern trust, selection, and deployment of LLMs, yet these evaluations remain vulnerable to semantically equivalent adversarial perturbations. Prior work on adversarial robustness in NLP has emphasized text attacks that affect many models equally, leaving open the question of whether it is possible to selectively degrade or enhance performance while minimally affecting other models. We formalize this problem and study selective adversarial attacks on MMLU - a widely used benchmark designed to measure a language model's broad general knowledge and reasoning ability across different subjects. Using canonical attacks integrated into TextAttack framework, we introduce a protocol for selectivity assessment, develop a custom constraint to increase selectivity of attacks and propose a surrogate-LLM pipeline that generates selective perturbations. Empirically, we find that selective adversarial attacks exist and can materially alter relative rankings, challenging the fairness, reproducibility, and transparency of leaderboard-driven evaluation. Our results motivate perturbation-aware reporting and robustness diagnostics for LLM evaluation and demonstrate that even subtle edits can shift comparative judgments.
The reasoning pattern of Large language models (LLMs) remains opaque, and Reinforcement learning (RL) typically applies uniform credit across an entire generation, blurring the distinction between pivotal and routine steps. This work positions attention as a privileged substrate that renders the internal logic of LLMs legible, not merely as a byproduct of computation, but as a mechanistic blueprint of reasoning itself. We first distinguish attention heads between locally and globally focused information processing and reveal that locally focused heads produce a sawtooth pattern near the diagonal indicating phrasal chunks, while globally focused heads expose tokens that exert broad downstream influence over future tokens. We formalize these with two metrics: 1) Windowed Average Attention Distance, which measures the extent of backward attention within a clipped window; 2) Future Attention Influence, which quantifies a token's global importance as the average attention it receives from subsequent tokens. Taken together, these signals reveal a recurring preplan-and-anchor mechanism, where the model first performs a long-range contextual reference to generate an introductory token, which is immediately followed by or coincides with a semantic anchor token that organizes subsequent reasoning. Leveraging these insights, we introduce three novel RL strategies that dynamically perform targeted credit assignment to critical nodes (preplan tokens, anchor tokens, and their temporal coupling) and show consistent performance gains across various reasoning tasks. By aligning optimization with the model's intrinsic reasoning rhythm, we aim to transform opaque optimization into an actionable structure-aware process, hoping to offer a potential step toward more transparent and effective optimization of LLM reasoning.
Large Language Model (LLM) based agents integrated into web browsers (often called agentic AI browsers) offer powerful automation of web tasks. However, they are vulnerable to indirect prompt injection attacks, where malicious instructions hidden in a webpage deceive the agent into unwanted actions. These attacks can bypass traditional web security boundaries, as the AI agent operates with the user privileges across sites. In this paper, we present a novel fuzzing framework that runs entirely in the browser and is guided by an LLM to automatically discover such prompt injection vulnerabilities in real time.
Large Language Models (LLMs) show potential for complex reasoning, yet their capacity for emergent coordination in Multi-Agent Systems (MAS) when operating under strict swarm-like constraints-limited local perception and communication-remains largely unexplored. Existing benchmarks often do not fully capture the unique challenges of decentralized coordination when agents operate with incomplete spatio-temporal information. To bridge this gap, we introduce SwarmBench, a novel benchmark designed to systematically evaluate the swarm intelligence capabilities of LLMs acting as decentralized agents. SwarmBench features five foundational MAS coordination tasks (Pursuit, Synchronization, Foraging, Flocking, Transport) within a configurable 2D grid environment, forcing agents to rely solely on local sensory input ($k\times k$ view) and local communication. We propose metrics for coordination effectiveness and analyze emergent group dynamics. Zero-shot evaluations of leading LLMs (e.g., deepseek-v3, o4-mini) reveal significant task-dependent performance variations. While some rudimentary coordination is observed, our results indicate that current LLMs significantly struggle with robust long-range planning and adaptive strategy formation under the uncertainty inherent in these decentralized scenarios. Assessing LLMs under such swarm-like constraints is crucial for understanding their utility in future decentralized intelligent systems. We release SwarmBench as an open, extensible toolkit-built on a customizable physical system-providing environments, prompts, evaluation scripts, and comprehensive datasets. This aims to foster reproducible research into LLM-based MAS coordination and the theoretical underpinnings of emergent collective behavior under severe informational decentralization. Our code repository is available at https://github.com/x66ccff/swarmbench.
Reward models can significantly enhance the reasoning capabilities of large language models (LLMs), but they typically require extensive curated data and costly training. To mitigate these challenges, training-free approaches such as LLM-as-a-Judge leverage the intrinsic reasoning abilities of LLMs to evaluate responses, achieving promising results. Recent works have also indicated that model confidence can serve effectively as a reward metric, distinguishing between chain-of-thought (CoT) and non-CoT paths. However, the concept of using confidence as a reward has not been comprehensively studied. In this work, we systematically investigate Confidence-as-a-Reward (CRew), a simple yet powerful training-free method that utilizes token-level confidence in the model's final answers as a proxy for reward, especially suitable for close-ended tasks. Through extensive experiments on mathematical reasoning tasks, we demonstrate that CRew outperforms existing training-free reward approaches on the MATH500 and RewardMATH benchmarks, and even surpasses most trained reward models. We further identify a strong correlation between CRew scores and the actual reasoning performance of the model. Additionally, we find that CRew can effectively filter high-quality training data. Building upon these insights, we propose CRew-DPO, a training strategy that constructs preference data from confidence scores combined with correctness signals. Finetuning with CRew-DPO further enhances the model's judging capabilities and consistently outperforms existing self-training methods.
LLMs hold great promise for healthcare applications, but the rapid evolution of medical knowledge and errors in training data often cause them to generate outdated or inaccurate information, limiting their applicability in high-stakes clinical practice. Model editing has emerged as a potential remedy without full retraining. While parameter-based editing often compromises locality and is thus ill-suited for the medical domain, retrieval-based editing offers a more viable alternative. However, it still faces two critical challenges: (1) representation overlap within the medical knowledge space often causes inaccurate retrieval and reduces editing accuracy; (2) existing methods are restricted to single-sample edits, while batch-editing remains largely unexplored despite its importance for real-world medical applications. To address these challenges, we first construct MedVersa, \hk{an enhanced benchmark with broader coverage of medical subjects, designed to evaluate both single and batch edits under strict locality constraints}. We then propose MedREK, a retrieval-based editing framework that integrates a shared query-key module for precise matching with an attention-based prompt encoder for informative guidance. Experimental results on various medical benchmarks demonstrate that our MedREK achieves superior performance across different core metrics and provides the first validated solution for batch-editing in medical LLMs. Our code and dataset are available at https://github.com/mylittleriver/MedREK.
Large language models (LLMs) have been increasingly employed for (interactive) decision-making, via the development of LLM-based autonomous agents. Despite their emerging successes, the performance of LLM agents in decision-making has not been fully investigated through quantitative metrics, especially in the multi-agent setting when they interact with each other, a typical scenario in real-world LLM-agent applications. To better understand the limits of LLM agents in these interactive environments, we propose to study their interactions in benchmark decision-making settings in online learning and game theory, through the performance metric of \emph{regret}. We first empirically study the {no-regret} behaviors of LLMs in canonical (non-stationary) online learning problems, as well as the emergence of equilibria when LLM agents interact through playing repeated games. We then provide some theoretical insights into the no-regret behaviors of LLM agents, under certain assumptions on the supervised pre-training and the rationality model of human decision-makers who generate the data. Notably, we also identify (simple) cases where advanced LLMs such as GPT-4 fail to be no-regret. To promote the no-regret behaviors, we propose a novel \emph{unsupervised} training loss of \emph{regret-loss}, which, in contrast to the supervised pre-training loss, does not require the labels of (optimal) actions. We then establish the statistical guarantee of generalization bound for regret-loss minimization, followed by the optimization guarantee that minimizing such a loss may automatically lead to known no-regret learning algorithms. Our further experiments demonstrate the effectiveness of our regret-loss, especially in addressing the above ``regrettable'' cases.
Large Language Models (LLMs) have significantly advanced the field of natural language processing, enhancing capabilities in both language understanding and generation across diverse domains. However, developing LLMs for Arabic presents unique challenges. This paper explores these challenges by focusing on critical aspects such as data curation, tokenizer design, and evaluation. We detail our approach to the collection and filtration of Arabic pre-training datasets, assess the impact of various tokenizer designs on model performance, and examine the limitations of existing Arabic evaluation frameworks, for which we propose a systematic corrective methodology. To promote transparency and facilitate collaborative development, we share our data and methodologies, contributing to the advancement of language modeling, particularly for the Arabic language.
Large Language Models (LLMs) remain static in functionality after training, and extending their capabilities requires integration with external data, computation, and services. The Model Context Protocol (MCP) has emerged as a standard interface for such extensions, but current implementations rely solely on semantic matching between users' requests and server function descriptions, which makes current deployments and simulation testbeds fragile under latency fluctuations or server failures. We address this gap by enhancing MCP tool routing algorithms with real-time awareness of network and server status. To provide a controlled test environment for development and evaluation, we construct a heterogeneous experimental platform, namely Network-aware MCP (NetMCP), which offers five representative network states and build a benchmark for latency sequence generation and MCP server datasets. On top of NetMCP platform, we analyze latency sequences and propose a Semantic-Oriented and Network-Aware Routing (SONAR) algorithm, which jointly optimizes semantic similarity and network Quality of Service (QoS) metrics for adaptive tool routing. Results show that SONAR consistently improves task success rate and reduces completion time and failure number compared with semantic-only, LLM-based baselines, demonstrating the value of network-aware design for production-scale LLM systems. The code for NetMCP is available at https://github.com/NICE-HKU/NetMCP.
How does a cause lead to an effect, and which intermediate causal steps explain their connection? This work scrutinizes the mechanistic causal reasoning capabilities of large language models (LLMs) to answer these questions through the task of implicit causal chain discovery. In a diagnostic evaluation framework, we instruct nine LLMs to generate all possible intermediate causal steps linking given cause-effect pairs in causal chain structures. These pairs are drawn from recent resources in argumentation studies featuring polarized discussion on climate change. Our analysis reveals that LLMs vary in the number and granularity of causal steps they produce. Although they are generally self-consistent and confident about the intermediate causal connections in the generated chains, their judgments are mainly driven by associative pattern matching rather than genuine causal reasoning. Nonetheless, human evaluations confirmed the logical coherence and integrity of the generated chains. Our baseline causal chain discovery approach, insights from our diagnostic evaluation, and benchmark dataset with causal chains lay a solid foundation for advancing future work in implicit, mechanistic causal reasoning in argumentation settings.
Large Language Models (LLMs) have become increasingly prominent for daily tasks, from improving sound-totext translation to generating additional frames for the latest video games. With the help of LLM inference frameworks, such as llama.cpp, which support optimizations such as KV-caching and quantization, it is now easier than ever to deploy LLMs on edge devices. Quantization is fundamental to enable LLMs on resource-constrained edge devices, and llama.cpp utilizes block floating point (BFP) quantization to drastically reduce the bit width of weights and input tensors, the memory footprint, and the computational power required to run LLMs. LLMs are typically quantized with mixed BFP quantization across the model layers to reduce the loss of model accuracy due to quantization. Therefore, to efficiently accelerate across the layers of BFP-quantized LLMs, specialized accelerators need to support different BFP variants without reconfiguration. To address this issue, we propose a Flexible Block FloatingPoint Quantization (F-BFQ) accelerator, which can dynamically switch between two BFP quantization variants and perform matrix multiplication (MatMul) operations. Our initial F-BFQ accelerator design, deployed on the AMD Kria board, reduces inference time by 1.4x on average over the Arm NEON-based CPU execution across three BFP quantized LLMs while achieving 5.2 tokens per second (~3.9 words per second).
Penetration-testing is crucial for identifying system vulnerabilities, with privilege-escalation being a critical subtask to gain elevated access to protected resources. Language Models (LLMs) presents new avenues for automating these security practices by emulating human behavior. However, a comprehensive understanding of LLMs' efficacy and limitations in performing autonomous Linux privilege-escalation attacks remains under-explored. To address this gap, we introduce hackingBuddyGPT, a fully automated LLM-driven prototype designed for autonomous Linux privilege-escalation. We curated a novel, publicly available Linux privilege-escalation benchmark, enabling controlled and reproducible evaluation. Our empirical analysis assesses the quantitative success rates and qualitative operational behaviors of various LLMs -- GPT-3.5-Turbo, GPT-4-Turbo, and Llama3 -- against baselines of human professional pen-testers and traditional automated tools. We investigate the impact of context management strategies, different context sizes, and various high-level guidance mechanisms on LLM performance. Results show that GPT-4-Turbo demonstrates high efficacy, successfully exploiting 33-83% of vulnerabilities, a performance comparable to human pen-testers (75%). In contrast, local models like Llama3 exhibited limited success (0-33%), and GPT-3.5-Turbo achieved moderate rates (16-50%). We show that both high-level guidance and state-management through LLM-driven reflection significantly boost LLM success rates. Qualitative analysis reveals both LLMs' strengths and weaknesses in generating valid commands and highlights challenges in common-sense reasoning, error handling, and multi-step exploitation, particularly with temporal dependencies. Cost analysis indicates that GPT-4-Turbo can achieve human-comparable performance at competitive costs, especially with optimized context management.
Recent attempts to integrate large language models (LLMs) into recommender systems have gained momentum, but most remain limited to simple text generation or static prompt-based inference, failing to capture the complexity of user preferences and real-world interactions. This study proposes the Multi-Aspect Driven LLM Agent MADRec, an autonomous LLM-based recommender that constructs user and item profiles by unsupervised extraction of multi-aspect information from reviews and performs direct recommendation, sequential recommendation, and explanation generation. MADRec generates structured profiles via aspect-category-based summarization and applies Re-Ranking to construct high-density inputs. When the ground-truth item is missing from the output, the Self-Feedback mechanism dynamically adjusts the inference criteria. Experiments across multiple domains show that MADRec outperforms traditional and LLM-based baselines in both precision and explainability, with human evaluation further confirming the persuasiveness of the generated explanations.
Large Language Models (LLMs) often exhibit factual inconsistencies and logical decay in extended, multi-turn dialogues, a challenge stemming from their reliance on static, pre-trained knowledge and an inability to reason adaptively over the dialogue history. Prevailing mitigation strategies, such as Retrieval-Augmented Generation (RAG) and agentic working memories, improve information recall but still engage with fundamentally static knowledge sources and follow pre-defined single reasoning path. This hinders their ability to preserve factual and logical consistency of their responses in multi-turn dialogues while the context evolves over time. To address this issue, we propose D-SMART, a model-agnostic framework designed to maintain multi-turn dialogue consistency by enabling LLMs to build and reason over a dynamic, structured representation of the conversational context. This is achieved via two synergistic components: (1) a Dynamic Structured Memory (DSM), which incrementally constructs and maintains an authoritative, OWL-compliant knowledge graph of the conversation; and (2) a Reasoning Tree (RT), which executes inferences as an explicit and traceable multi-step search over the graph. As the popular-used quality score (judged by GPT-4) can overlook logical flaws, we introduce new NLI-based metrics to better measure multi-turn dialogue consistency. Comprehensive experiments on the MT-Bench-101 benchmark show that D-SMART significantly outperforms state-of-the-art baselines, elevating the dialogue consistency score by over 48\% for both proprietary and open-source models, and notably improves the quality score of the latter by up to 10.1\%.
Sentiment Analysis (SA) models harbor inherent social biases that can be harmful in real-world applications. These biases are identified by examining the output of SA models for sentences that only vary in the identity groups of the subjects. Constructing natural, linguistically rich, relevant, and diverse sets of sentences that provide sufficient coverage over the domain is expensive, especially when addressing a wide range of biases: it requires domain experts and/or crowd-sourcing. In this paper, we present a novel bias testing framework, BTC-SAM, which generates high-quality test cases for bias testing in SA models with minimal specification using Large Language Models (LLMs) for the controllable generation of test sentences. Our experiments show that relying on LLMs can provide high linguistic variation and diversity in the test sentences, thereby offering better test coverage compared to base prompting methods even for previously unseen biases.
Process mining has emerged as a powerful analytical technique for understanding complex healthcare workflows. However, its application faces significant barriers, including technical complexity, a lack of standardized approaches, and limited access to practical training resources. We introduce HealthProcessAI, a GenAI framework designed to simplify process mining applications in healthcare and epidemiology by providing a comprehensive wrapper around existing Python (PM4PY) and R (bupaR) libraries. To address unfamiliarity and improve accessibility, the framework integrates multiple Large Language Models (LLMs) for automated process map interpretation and report generation, helping translate technical analyses into outputs that diverse users can readily understand. We validated the framework using sepsis progression data as a proof-of-concept example and compared the outputs of five state-of-the-art LLM models through the OpenRouter platform. To test its functionality, the framework successfully processed sepsis data across four proof-of-concept scenarios, demonstrating robust technical performance and its capability to generate reports through automated LLM analysis. LLM evaluation using five independent LLMs as automated evaluators revealed distinct model strengths: Claude Sonnet-4 and Gemini 2.5-Pro achieved the highest consistency scores (3.79/4.0 and 3.65/4.0) when evaluated by automated LLM assessors. By integrating multiple Large Language Models (LLMs) for automated interpretation and report generation, the framework addresses widespread unfamiliarity with process mining outputs, making them more accessible to clinicians, data scientists, and researchers. This structured analytics and AI-driven interpretation combination represents a novel methodological advance in translating complex process mining results into potentially actionable insights for healthcare applications.
The increasing deployment of Large Language Models (LLMs) across enterprise and mission-critical domains has underscored the urgent need for robust guardrailing systems that ensure safety, reliability, and compliance. Existing solutions often struggle with real-time oversight, multi-modal data handling, and explainability -- limitations that hinder their adoption in regulated environments. Existing guardrails largely operate in isolation, focused on text alone making them inadequate for multi-modal, production-scale environments. We introduce Protect, natively multi-modal guardrailing model designed to operate seamlessly across text, image, and audio inputs, designed for enterprise-grade deployment. Protect integrates fine-tuned, category-specific adapters trained via Low-Rank Adaptation (LoRA) on an extensive, multi-modal dataset covering four safety dimensions: toxicity, sexism, data privacy, and prompt injection. Our teacher-assisted annotation pipeline leverages reasoning and explanation traces to generate high-fidelity, context-aware labels across modalities. Experimental results demonstrate state-of-the-art performance across all safety dimensions, surpassing existing open and proprietary models such as WildGuard, LlamaGuard-4, and GPT-4.1. Protect establishes a strong foundation for trustworthy, auditable, and production-ready safety systems capable of operating across text, image, and audio modalities.
Computational stylometry analyzes writing style through quantitative patterns in text, supporting applications from forensic tasks such as identity linking and plagiarism detection to literary attribution in the humanities. Supervised and contrastive approaches rely on data with spurious correlations and often confuse style with topic. Despite their natural use in AI-generated text detection, the CLM pre-training of modern LLMs has been scarcely leveraged for general authorship problems. We propose a novel unsupervised approach based on this extensive pre-training and the in-context learning capabilities of LLMs, employing the log-probabilities of an LLM to measure style transferability from one text to another. Our method significantly outperforms LLM prompting approaches of comparable scale and achieves higher accuracy than contrastively trained baselines when controlling for topical correlations. Moreover, performance scales fairly consistently with the size of the base model and, in the case of authorship verification, with an additional mechanism that increases test-time computation; enabling flexible trade-offs between computational cost and accuracy.
Most recent RL for LLMs (RL4LLM) methods avoid explicit critics, replacing them with average advantage baselines. This shift is largely pragmatic: conventional value functions are computationally expensive to train at LLM scale and often fail under sparse rewards and long reasoning horizons. We revisit this bottleneck from an architectural perspective and introduce Asymmetric Proximal Policy Optimization (AsyPPO), a simple and scalable framework that restores the critics role while remaining efficient in large-model settings. AsyPPO employs a set of lightweight mini-critics, each trained on disjoint prompt shards. This design encourages diversity while preserving calibration, reducing value-estimation bias. Beyond robust estimation, AsyPPO leverages inter-critic uncertainty to refine the policy update: (i) masking advantages in states where critics agree and gradients add little learning signal, and (ii) filtering high-divergence states from entropy regularization, suppressing spurious exploration. After training on open-source data with only 5,000 samples, AsyPPO consistently improves learning stability and performance across multiple benchmarks over strong baselines, such as GRPO, achieving performance gains of more than six percent on Qwen3-4b-Base and about three percent on Qwen3-8b-Base and Qwen3-14b-Base over classic PPO, without additional tricks. These results highlight the importance of architectural innovations for scalable, efficient algorithms.
Enhancing customer experience is essential for business success, particularly as service demands grow in scale and complexity. Generative artificial intelligence and Large Language Models (LLMs) have empowered intelligent interaction systems to deliver efficient, personalized, and 24/7 support. In practice, intelligent interaction systems encounter several challenges: (1) Constructing high-quality data for cold-start training is difficult, hindering self-evolution and raising labor costs. (2) Multi-turn dialogue performance remains suboptimal due to inadequate intent understanding, rule compliance, and solution extraction. (3) Frequent evolution of business rules affects system operability and transferability, constraining low-cost expansion and adaptability. (4) Reliance on a single LLM is insufficient in complex scenarios, where the absence of multi-agent frameworks and effective collaboration undermines process completeness and service quality. (5) The open-domain nature of multi-turn dialogues, lacking unified golden answers, hampers quantitative evaluation and continuous optimization. To address these challenges, we introduce WOWService, an intelligent interaction system tailored for industrial applications. With the integration of LLMs and multi-agent architectures, WOWService enables autonomous task management and collaborative problem-solving. Specifically, WOWService focuses on core modules including data construction, general capability enhancement, business scenario adaptation, multi-agent coordination, and automated evaluation. Currently, WOWService is deployed on the Meituan App, achieving significant gains in key metrics, e.g., User Satisfaction Metric 1 (USM 1) -27.53% and User Satisfaction Metric 2 (USM 2) +25.51%, demonstrating its effectiveness in capturing user needs and advancing personalized service.
Large language models (LLMs) have achieved striking successes on many benchmarks, yet recent studies continue to expose fundamental weaknesses. In particular, tasks that require abstract reasoning remain challenging, often because they use representations such as grids, symbols, or visual patterns that differ from the natural language data LLMs are trained on. In this paper, we introduce Concept, a simple word-guessing board game, as a benchmark for probing abductive reasoning in a representation that is much closer to LLM pre-training data: natural language. Our results show that this game, easily solved by humans (with a success rate of over 90\%), is still very challenging for state-of-the-art LLMs (no model exceeds 40\% success rate). Specifically, we observe that LLMs struggle with interpreting other players' strategic intents, and with correcting initial hypotheses given sequential information updates. In addition, we extend the evaluation across multiple languages, and find that the LLM performance drops further in lower-resource languages (Dutch, French, and Spanish) compared to English.
Rule-based network intrusion detection systems play a crucial role in the real-time detection of Web attacks. However, most existing works primarily focus on automatically generating detection rules for new attacks, often overlooking the relationships between new attacks and existing rules, which leads to significant redundancy within the ever-expanding ruleset. To address this issue, we propose GRIDAI, a novel end-to-end framework for the automated Generation and Repair of Intrusion Detection rules through collaboration among multiple LLM-based agents. Unlike traditional methods, GRIDAI first assesses the nature of incoming attack samples. If the sample represents a new attack type, it is used to generate a new rule. Otherwise, the sample is identified as a variant of an attack already covered by an existing rule and used to repair the rule by updating the corresponding signature, thereby enhancing its generalization capability. Additionally, to mitigate syntactic and semantic errors in rules caused by LLM hallucinations, we incorporate a tool-based real-time validation mechanism and a representative attack sample maintained for each rule, enabling fully automated rule generation and repair. Comprehensive experiments were conducted on a public dataset containing seven types of attacks and a private dataset with 43 attack types. The results demonstrate that GRIDAI accurately identifies the relationships between new attack samples and existing rules, efficiently generates and repairs rules to handle new attacks and variants, and effectively mitigates the impact of LLM hallucinations.
With the increasing adoption of Large Language Models (LLMs), more customization is needed to ensure privacy-preserving and safe generation. We address this objective from two critical aspects: unlearning of sensitive information and robustness to jail-breaking attacks. We investigate various constrained optimization formulations that address both aspects in a \emph{unified manner}, by finding the smallest possible interventions on LLM weights that either make a given vocabulary set unreachable or embed the LLM with robustness to tailored attacks by shifting part of the weights to a \emph{safer} region. Beyond unifying two key properties, this approach contrasts with previous work in that it doesn't require an oracle classifier that is typically not available or represents a computational overhead. Surprisingly, we find that the simplest point-wise constraint-based intervention we propose leads to better performance than max-min interventions, while having a lower computational cost. Comparison against state-of-the-art defense methods demonstrates superior performance of the proposed approach.
Network protocol testing is fundamental for modern network infrastructure. However, traditional network protocol testing methods are labor-intensive and error-prone, requiring manual interpretation of specifications, test case design, and translation into executable artifacts, typically demanding one person-day of effort per test case. Existing model-based approaches provide partial automation but still involve substantial manual modeling and expert intervention, leading to high costs and limited adaptability to diverse and evolving protocols. In this paper, we propose a first-of-its-kind system called NeTestLLM that takes advantage of multi-agent Large Language Models (LLMs) for end-to-end automated network protocol testing. NeTestLLM employs hierarchical protocol understanding to capture complex specifications, iterative test case generation to improve coverage, a task-specific workflow for executable artifact generation, and runtime feedback analysis for debugging and refinement. NeTestLLM has been deployed in a production environment for several months, receiving positive feedback from domain experts. In experiments, NeTestLLM generated 4,632 test cases for OSPF, RIP, and BGP, covering 41 historical FRRouting bugs compared to 11 by current national standards. The process of generating executable artifacts also improves testing efficiency by a factor of 8.65x compared to manual methods. NeTestLLM provides the first practical LLM-powered solution for automated end-to-end testing of heterogeneous network protocols.
The growing scale of evaluation tasks has led to the widespread adoption of automated evaluation using LLMs, a paradigm known as "LLM-as-a-judge". However, improving its alignment with human preferences without complex prompts or fine-tuning remains challenging. Previous studies mainly optimize based on shallow outputs, overlooking rich cross-layer representations. In this work, motivated by preliminary findings that middle-to-upper layers encode semantically and task-relevant representations that are often more aligned with human judgments than the final layer, we propose LAGER, a post-hoc, plug-and-play framework for improving the alignment of LLM-as-a-Judge point-wise evaluations with human scores by leveraging internal representations. LAGER produces fine-grained judgment scores by aggregating cross-layer score-token logits and computing the expected score from a softmax-based distribution, while keeping the LLM backbone frozen and ensuring no impact on the inference process. LAGER fully leverages the complementary information across different layers, overcoming the limitations of relying solely on the final layer. We evaluate our method on the standard alignment benchmarks Flask, HelpSteer, and BIGGen using Spearman correlation, and find that LAGER achieves improvements of up to 7.5% over the best baseline across these benchmarks. Without reasoning steps, LAGER matches or outperforms reasoning-based methods. Experiments on downstream applications, such as data selection and emotional understanding, further show the generalization of LAGER.
Recommender systems (RecSys) have become critical tools for enhancing user engagement by delivering personalized content across diverse digital platforms. Recent advancements in large language models (LLMs) demonstrate significant potential for improving RecSys, primarily due to their exceptional generalization capabilities and sophisticated contextual understanding, which facilitate the generation of flexible and interpretable recommendations. However, the direct deployment of LLMs as primary recommendation policies presents notable challenges, including persistent latency issues stemming from frequent API calls and inherent model limitations such as hallucinations and biases. To address these issues, this paper proposes a novel offline reinforcement learning (RL) framework that leverages imitation learning from LLM-generated trajectories. Specifically, inverse reinforcement learning is employed to extract robust reward models from LLM demonstrations. This approach negates the need for LLM fine-tuning, thereby substantially reducing computational overhead. Simultaneously, the RL policy is guided by the cumulative rewards derived from these demonstrations, effectively transferring the semantic insights captured by the LLM. Comprehensive experiments conducted on two benchmark datasets validate the effectiveness of the proposed method, demonstrating superior performance when compared against state-of-the-art RL-based and in-context learning baselines. The code can be found at https://github.com/ArronDZhang/IL-Rec.
At the core of Deep Research is knowledge mining, the task of extracting structured information from massive unstructured text in response to user instructions. Large language models (LLMs) excel at interpreting such instructions but are prohibitively expensive to deploy at scale, while traditional pipelines of classifiers and extractors remain efficient yet brittle and unable to generalize to new tasks. We introduce Falconer, a collaborative framework that combines the agentic reasoning of LLMs with lightweight proxy models for scalable knowledge mining. In Falconer, LLMs act as planners, decomposing user instructions into executable pipelines, and as annotators, generating supervision to train small proxies. The framework unifies classification and extraction into two atomic operations, get label and get span, enabling a single instruction-following model to replace multiple task-specific components. To evaluate the consistency between proxy models incubated by Falconer and annotations provided by humans and large models, we construct new benchmarks covering both planning and end-to-end execution. Experiments show that Falconer closely matches state-of-the-art LLMs in instruction-following accuracy while reducing inference cost by up to 90% and accelerating large-scale knowledge mining by more than 20x, offering an efficient and scalable foundation for Deep Research.
Large language models (LLMs) are increasingly deployed in AI infrastructure, driving the need for high throughput, resource efficient serving systems. Disaggregated LLM serving, which separates prompt prefill from auto-regressive decode, has emerged as a promising architecture by isolating their heterogeneous compute and memory demands. However, current disaggregated systems face three key limitations: (i) static resource allocation cannot adapt to highly dynamic workloads, causing over-provisioning that wastes resources or under-provisioning that violates service level objectives (SLOs); (ii) inherent load imbalance between prefill and decode stages, where prefill is compute-bound and decode is memory-bound, causes under-utilization in one tier while the other becomes a bottleneck; and (iii) prefix cache aware routing skews load distribution, as high cache hit rate prefill nodes attract disproportionately more requests, further degrading balance and efficiency. To address these issues, we present BanaServe, a dynamic orchestration framework that continuously rebalances computational and memory resources across prefill and decode instances while eliminating hotspots induced by cache. BanaServe introduces layer level weight migration, attention level Key Value Cache (KV Cache) migration, and Global KV Cache Store sharing with layer wise overlapped transmission, enabling both coarse grained (layer level) and fine grained (attention level) load redistribution with minimal latency overhead. These mechanisms allow routers to perform purely load aware scheduling, unconstrained by cache placement. Compared to vLLM, BanaServe achieves 1.2x-3.9x higher throughput with 3.9%-78.4% lower total processing time, and outperforms DistServe by 1.1x-2.8x in throughput with 1.4%-70.1% latency reduction.
Modern IR systems are increasingly tasked with answering complex, multi-faceted queries that require deep reasoning rather than simple keyword or semantic matching. While LLM-based IR has shown great promise, the prevailing retrieve-then-rerank paradigm inherits the limitations of embedding-based retrieval; parametric generative approaches are difficult to update with new information; and long-context methods that place the entire corpus in context are computationally infeasible for large document collections. To address these challenges, we introduce LATTICE, a hierarchical retrieval framework that enables an LLM to reason over and navigate large corpora with logarithmic search complexity by imposing a semantic tree structure on the corpus. Our approach consists of two stages: (1) an offline phase that organizes the corpus into a semantic hierarchy via either a bottom-up agglomerative strategy or a top-down divisive strategy using multi-level summaries and (2) an online traversal phase where a search LLM navigates this tree. A central challenge in such LLM-guided search is that the model's relevance judgments are noisy, context-dependent, and unaware of the hierarchy, making cross-branch and cross-level comparisons difficult. To overcome this, we propose a traversal algorithm that estimates calibrated latent relevance scores from local LLM outputs and aggregates them into a global path relevance metric. Our training-free framework achieves state-of-the-art zero-shot performance on the reasoning-intensive BRIGHT benchmark, demonstrating up to 9% improvement in Recall@100 and 5% in nDCG@10 over the next best zero-shot baseline. Furthermore, compared to the fine-tuned SOTA method DIVER-v2, LATTICE attains comparable results on BRIGHT subsets that use a static corpus for evaluation.
Bias in AI systems, especially those relying on natural language data, raises ethical and practical concerns. Underrepresentation of certain groups often leads to uneven performance across demographics. Traditional fairness methods, such as pre-processing, in-processing, and post-processing, depend on protected-attribute labels, involve accuracy-fairness trade-offs, and may not generalize across datasets. To address these challenges, we propose LLM-Guided Synthetic Augmentation (LGSA), which uses large language models to generate counterfactual examples for underrepresented groups while preserving label integrity. We evaluated LGSA on a controlled dataset of short English sentences with gendered pronouns, professions, and binary classification labels. Structured prompts were used to produce gender-swapped paraphrases, followed by quality control including semantic similarity checks, attribute verification, toxicity screening, and human spot checks. The augmented dataset expanded training coverage and was used to train a classifier under consistent conditions. Results show that LGSA reduces performance disparities without compromising accuracy. The baseline model achieved 96.7 percent accuracy with a 7.2 percent gender bias gap. Simple swap augmentation reduced the gap to 0.7 percent but lowered accuracy to 95.6 percent. LGSA achieved 99.1 percent accuracy with a 1.9 percent bias gap, improving performance on female-labeled examples. These findings demonstrate that LGSA is an effective strategy for bias mitigation, enhancing subgroup balance while maintaining high task accuracy and label fidelity.
The rapid advancement of large language models (LLMs) has inspired researchers to integrate them extensively into the academic workflow, potentially reshaping how research is practiced and reviewed. While previous studies highlight the potential of LLMs in supporting research and peer review, their dual roles in the academic workflow and the complex interplay between research and review bring new risks that remain largely underexplored. In this study, we focus on how the deep integration of LLMs into both peer-review and research processes may influence scholarly fairness, examining the potential risks of using LLMs as reviewers by simulation. This simulation incorporates a research agent, which generates papers and revises, alongside a review agent, which assesses the submissions. Based on the simulation results, we conduct human annotations and identify pronounced misalignment between LLM-based reviews and human judgments: (1) LLM reviewers systematically inflate scores for LLM-authored papers, assigning them markedly higher scores than human-authored ones; (2) LLM reviewers persistently underrate human-authored papers with critical statements (e.g., risk, fairness), even after multiple revisions. Our analysis reveals that these stem from two primary biases in LLM reviewers: a linguistic feature bias favoring LLM-generated writing styles, and an aversion toward critical statements. These results highlight the risks and equity concerns posed to human authors and academic research if LLMs are deployed in the peer review cycle without adequate caution. On the other hand, revisions guided by LLM reviews yield quality gains in both LLM-based and human evaluations, illustrating the potential of the LLMs-as-reviewers for early-stage researchers and enhancing low-quality papers.
Despite significant recent progress of Multimodal Large Language Models (MLLMs), current MLLMs are challenged by "spatio-temporal" prompts, i.e., prompts that refer to 1) the entirety of an environment encoded in a point cloud that the MLLM should consider; and simultaneously also refer to 2) actions that happened in part of the environment and are encoded in a short ego-centric video clip. However, such a holistic spatio-temporal understanding is important for agents operating in the real world. To address this challenge, we first develop a framework to collect a large-scale dataset. Using the collected "Reasoning about Environments and Actions" (REA) dataset, we show that recent MLLMs indeed struggle to correctly answer "spatio-temporal" prompts. Building on this dataset, we study two spatio-temporal LLM (STLLM) baselines: 1) STLLM-3D, which directly fuses point cloud, video, and text representations as inputs to the LLM; and 2) STLLM-Aligner, which aligns spatial context with video and text before LLM decoding. Both baselines aim to enhance spatial understanding of environments and temporal grounding of egocentric observations. On REA, the STLLM baselines outperform existing models, demonstrating the effectiveness of our designs. Code and data are available at https://zoezheng126.github.io/STLLM-website/.
The advent of large language models (LLMs) has enabled agents to represent virtual humans in societal simulations, facilitating diverse interactions within complex social systems. However, existing LLM-based agents exhibit severe limitations in affective cognition: They fail to simulate the bounded rationality essential for bridging virtual and real-world services; They lack empirically validated integration mechanisms embedding emotions within agent decision architectures. This paper constructs an emotional cognition framework incorporating desire generation and objective management, designed to achieve emotion alignment between LLM-based agents and humans, modeling the complete decision-making process of LLM-based agents, encompassing state evolution, desire generation, objective optimization, decision generation, and action execution. This study implements the proposed framework within our proprietary multi-agent interaction environment. Experimental results demonstrate that agents governed by our framework not only exhibit behaviors congruent with their emotional states but also, in comparative assessments against other agent types, demonstrate superior ecological validity and generate decision outcomes that significantly more closely approximate human behavioral patterns.
Knowledge graphs (KGs), with their structured representation capabilities, offer promising avenue for enhancing Retrieval Augmented Generation (RAG) systems, leading to the development of KG-RAG systems. Nevertheless, existing methods often struggle to achieve effective synergy between system effectiveness and cost efficiency, leading to neither unsatisfying performance nor excessive LLM prompt tokens and inference time. To this end, this paper proposes REMINDRAG, which employs an LLM-guided graph traversal featuring node exploration, node exploitation, and, most notably, memory replay, to improve both system effectiveness and cost efficiency. Specifically, REMINDRAG memorizes traversal experience within KG edge embeddings, mirroring the way LLMs "memorize" world knowledge within their parameters, but in a train-free manner. We theoretically and experimentally confirm the effectiveness of REMINDRAG, demonstrating its superiority over existing baselines across various benchmark datasets and LLM backbones. Our code is available at https://github.com/kilgrims/ReMindRAG.
Early and accurate diagnosis of Alzheimer's disease (AD), a complex neurodegenerative disorder, requires analysis of heterogeneous biomarkers (e.g., neuroimaging, genetic risk factors, cognitive tests, and cerebrospinal fluid proteins) typically represented in a tabular format. With flexible few-shot reasoning, multimodal integration, and natural-language-based interpretability, large language models (LLMs) offer unprecedented opportunities for prediction with structured biomedical data. We propose a novel framework called TAP-GPT, Tabular Alzheimer's Prediction GPT, that adapts TableGPT2, a multimodal tabular-specialized LLM originally developed for business intelligence tasks, for AD diagnosis using structured biomarker data with small sample sizes. Our approach constructs few-shot tabular prompts using in-context learning examples from structured biomedical data and finetunes TableGPT2 using the parameter-efficient qLoRA adaption for a clinical binary classification task of AD or cognitively normal (CN). The TAP-GPT framework harnesses the powerful tabular understanding ability of TableGPT2 and the encoded prior knowledge of LLMs to outperform more advanced general-purpose LLMs and a tabular foundation model (TFM) developed for prediction tasks. To our knowledge, this is the first application of LLMs to the prediction task using tabular biomarker data, paving the way for future LLM-driven multi-agent frameworks in biomedical informatics.
Speculative decoding accelerates LLM inference by using a draft model to look ahead, but gains are capped by the cost of autoregressive draft generation: increasing draft size elevates acceptance rates but introduces additional latency overhead exacerbating the speed-accuracy tradeoff. Prior methods (Medusa, Hydra, EAGLE) partially reduce draft cost but either degrade acceptance or introduce overheads that limit scaling. We present Mirror Speculative Decoding (Mirror-SD), an inference algorithm that breaks the latency-acceptance tradeoff. Mirror-SD launches branch-complete rollouts from early-exit signals in parallel with the target model's suffix and explicitly maps computation across heterogeneous accelerators (GPU and NPU) to exploit cross-device parallelism. The draft speculates forward continuations for the target to verify, while the target simultaneously speculates correction paths for the draft, converting speculation into two complementary execution pipelines. To further cut draft latency without weakening acceptance semantics, we add speculative streaming so the draft emits multiple tokens per step. This dual strategy of parallel heterogeneous execution plus multi-token speculative streaming pushes speculative decoding toward its ideal regime of high acceptance with low overhead. On SpecBench with server-scale models from 14B to 66B parameters, Mirror-SD delivers consistent end-to-end gains, achieving 2.8x-5.8x wall-time speedups across diverse tasks and a 30% average relative improvement over the strongest baseline, EAGLE3.
We introduce MENAValues, a novel benchmark designed to evaluate the cultural alignment and multilingual biases of large language models (LLMs) with respect to the beliefs and values of the Middle East and North Africa (MENA) region, an underrepresented area in current AI evaluation efforts. Drawing from large-scale, authoritative human surveys, we curate a structured dataset that captures the sociocultural landscape of MENA with population-level response distributions from 16 countries. To probe LLM behavior, we evaluate diverse models across multiple conditions formed by crossing three perspective framings (neutral, personalized, and third-person/cultural observer) with two language modes (English and localized native languages: Arabic, Persian, Turkish). Our analysis reveals three critical phenomena: "Cross-Lingual Value Shifts" where identical questions yield drastically different responses based on language, "Reasoning-Induced Degradation" where prompting models to explain their reasoning worsens cultural alignment, and "Logit Leakage" where models refuse sensitive questions while internal probabilities reveal strong hidden preferences. We further demonstrate that models collapse into simplistic linguistic categories when operating in native languages, treating diverse nations as monolithic entities. MENAValues offers a scalable framework for diagnosing cultural misalignment, providing both empirical insights and methodological tools for developing more culturally inclusive AI.
Large language models (LLMs) have achieved remarkable results in wide range of domains. However, the accuracy and robustness of one-shot LLM predictions remain highly sensitive to the examples and the diversity among ensemble members. This study systematically investigates the effects of example representativeness (one-shot strategy) and output diversity (sampling temperature) on LLM ensemble performance. Two one-shot strategies are compared: centroid-based representative examples (proposed) and randomly sampled examples (baseline) and sampling temperature also is varied. The proposed approach with higher temperature setting significantly outperforms random selection by +7.6% (macro-F1) and -10.5% (RMSE). Furthermore, the proposed model exceeds 5-shot prompting by +21.1% (macro-F1) and -24.0% (RMSE). Our findings demonstrate that combining representative example selection with increased temperature provides the appropriate level of diversity to the ensemble. This work highlights the practical importance of both example selection and controlled diversity in designing effective one-shot LLM ensembles.
A key challenge in transportation planning is that the collective preferences of heterogeneous travelers often diverge from the policies produced by model-driven decision tools. This misalignment frequently results in implementation delays or failures. Here, we investigate whether large language models (LLMs), noted for their capabilities in reasoning and simulating human decision-making, can help inform and address this alignment problem. We develop a multi-agent simulation in which LLMs, acting as agents representing residents from different communities in a city, participate in a referendum on a set of transit policy proposals. Using chain-of-thought reasoning, LLM agents provide ranked-choice or approval-based preferences, which are aggregated using instant-runoff voting (IRV) to model democratic consensus. We implement this simulation framework with both GPT-4o and Claude-3.5, and apply it for Chicago and Houston. Our findings suggest that LLM agents are capable of approximating plausible collective preferences and responding to local context, while also displaying model-specific behavioral biases and modest divergences from optimization-based benchmarks. These capabilities underscore both the promise and limitations of LLMs as tools for solving the alignment problem in transportation decision-making.
Large Language Models (LLMs) have significantly advanced natural language processing, demonstrating strong capabilities in tasks such as text generation, summarization, and reasoning. Recently, their potential for automating precise text editing tasks across specialized domains, such as programming code, LaTeX, and structured database languages, has gained attention. However, current state-of-the-art LLMs still struggle with executing precise, instruction-driven edits, particularly when structural accuracy and strict adherence to domain conventions are required. To address these challenges, we introduce InstrEditBench, an automated benchmark dataset comprising over 30,000 structured editing tasks spanning diverse domains, including Wikipedia articles, LaTeX documents, source code, and database languages. Using this benchmark, we develop FineEdit, a specialized editing model explicitly trained for accurate, context-aware text modifications. Experimental evaluations demonstrate that FineEdit outperforms state-of-the-art models, achieving improvements of approximately 10\% over Gemini models on single-turn edits, up to 30\% over Llama-3.2-3B, and exceeding Mistral-7B-OpenOrca performance by over 40\% on direct editing tasks. FineEdit also effectively generalizes to realistic multi-turn editing scenarios, highlighting its practical applicability. To facilitate further research and reproducibility, we release FineEdit at https://github.com/StuRinDQB/FineEdit} and https://huggingface.co/datasets/YimingZeng/FineEdit_bench.
Large language model (LLM) and agent techniques for data analysis (a.k.a LLM/Agent-as-Data-Analyst) have demonstrated substantial impact in both academica and industry. In comparison with traditional rule or small-model based approaches, (agentic) LLMs enable complex data understanding, natural language interfaces, semantic analysis functions, and autonomous pipeline orchestration. The technical evolution further distills five key design goals for intelligent data analysis agents, namely semantic-aware design, modality-hybrid integration, autonomous pipelines, tool-augmented workflows, and support for open-world tasks. From a modality perspective, we review LLM-based techniques for (i) structured data (e.g., table question answering for relational data and NL2GQL for graph data), (ii) semi-structured data (e.g., markup languages understanding and semi-structured table modeling), (iii) unstructured data (e.g., chart understanding, document understanding, programming languages vulnerable detection), and (iv) heterogeneous data (e.g., data retrieval and modality alignment for data lakes). Finally, we outline the remaining challenges and propose several insights and practical directions for advancing LLM/Agent-powered data analysis.
Structural pruning enhances hardware-agnostic inference efficiency for large language models (LLMs) yet often fails to maintain comparable performance. Local pruning performs efficient layer-by-layer compression but ignores global topology. Although global pruning aims to identify an optimal sparse model, intuitive methods typically adopt a two-stage paradigm that first evaluates substructure saliency and then applies global pruning, which ignores inter-structure dependencies and fails to achieve end-to-end optimization. To address these limitations, we propose T\'yr-the-Pruner, an efficient end-to-end search-based global structural pruning framework. This framework constructs a supernet by repeatedly applying local pruning across a range of sparsity ratios to each layer in an LLM, with the core goal of determining the optimal sparsity distribution under a target overall sparsity ratio. Concretely, we introduce an effective local pruning and an expectation error accumulation approach to improve supernet construction. Furthermore, we employ an iterative prune-and-search strategy with coarse-to-fine sparsity granularity to ensure efficient search convergence. Experimental results show that T\'yr-the-Pruner achieves state-of-the-art structural pruning, retaining 97% of the dense model's performance while removing a challenging 50% of Llama-3.1-70B's parameters. Code will be available at https://github.com/AMD-AGI/Tyr-the-Pruner.
As AR/VR technologies become integral to daily life, there's a growing need for AI that understands human social dynamics from an egocentric perspective. However, current LLMs often lack the social awareness to discern when to intervene as AI assistant. This leads to constant, socially unaware responses that may disrupt natural conversation and negatively impact user focus. To address these limitations, we introduce EgoSocial, a large-scale egocentric dataset with 13,500 social video-question pairs, specifically designed to benchmark intervention in social interaction perception. We also present an in-depth analysis of current omnimodal LLMs (OLLMs) to assess their effectiveness in detecting diverse social contextual cues. Experiments show that OLLMs still struggle to detect the intervention timing (14.4% for Gemini 2.5 Pro). We also propose EgoSoD (EgoSocial Detection), an end-to-end method for robustly discerning social dynamics. Informed by our OLLM analysis, EgoSoD integrates multimodal contextual cues (e.g., audio and visual cues) into a social thinking graph, dynamically modeling participants and interactions. Our method proactively detects intervention timing and social interactions, precisely determining when to intervene. Our EgoSoD improves Phi-4 by 45.6% and Gemini 2.5 Pro by 9.9% on Intervention Timing performance, and improves Phi-4 by 20.4% and Gemini 2.5 Pro by 6.9% on overall Social Interaction performance. We will release the dataset and code soon.
Large language models (LLMs) excel at complex tasks thanks to advances in their reasoning abilities. However, existing methods overlook the trade-off between reasoning effectiveness and efficiency, often encouraging unnecessarily long reasoning chains and wasting tokens. To address this, we propose Learning to Think (L2T), an information-theoretic reinforcement fine-tuning framework for LLMs to make the models achieve optimal reasoning with fewer tokens. Specifically, L2T treats each query-response interaction as a hierarchical session of multiple episodes and proposes a universal dense process reward, i.e., quantifies the episode-wise information gain in parameters, requiring no extra annotations or task-specific evaluators. We propose a method to quickly estimate this reward based on PAC-Bayes bounds and the Fisher information matrix. Theoretical analyses show that it significantly reduces computational complexity with high estimation accuracy. By immediately rewarding each episode's contribution and penalizing excessive updates, L2T optimizes the model via reinforcement learning to maximize the use of each episode and achieve effective updates. Empirical results on various reasoning benchmarks and base models demonstrate the advantage of L2T across different tasks, boosting both reasoning effectiveness and efficiency.
Search agents connect LLMs to the Internet, enabling access to broader and more up-to-date information. However, unreliable search results may also pose safety threats to end users, establishing a new threat surface. In this work, we conduct two in-the-wild experiments to demonstrate both the prevalence of low-quality search results and their potential to misguide agent behaviors. To counter this threat, we introduce an automated red-teaming framework that is systematic, scalable, and cost-efficient, enabling lightweight and harmless safety assessments of search agents. Building on this framework, we construct the SafeSearch benchmark, which includes 300 test cases covering five categories of risks (e.g., misinformation and indirect prompt injection). Using this benchmark, we evaluate three representative search agent scaffolds, covering search workflow, tool-calling, and deep research, across 7 proprietary and 8 open-source backend LLMs. Our results reveal substantial vulnerabilities of LLM-based search agents: when exposed to unreliable websites, the highest ASR reached 90.5% for GPT-4.1-mini under a search workflow setting. Moreover, our analysis highlights the limited effectiveness of common defense practices, such as reminder prompting. This emphasizes the value of our framework in promoting transparency for safer agent development. Our codebase and test cases are publicly available: https://github.com/jianshuod/SafeSearch.
Robust workflow composition is critical for effective agent performance, yet progress in Large Language Model (LLM) planning and reasoning is hindered by a scarcity of scalable evaluation data. This work introduces NL2Flow, a fully automated pipeline for generating and evaluating workflow planning problems. NL2Flow generates problems parametrically in a structured intermediate representation, translating them into both natural language and formal PDDL. I evaluate several open-source, instruct-tuned LLMs on a dataset of 2296 low-difficulty problems generated by NL2Flow. Results demonstrate that the best-performing model achieved 86% success in generating valid plans and 69% in generating optimal plans (for solvable problems). Regression analysis shows that the influence of problem characteristics on plan generation is contingent on both model and prompt design. Importantly, translating natural language problems into a structured JSON representation prior to symbolic planning significantly improved success rates, suggesting a benefit from neuro-symbolic integration. These findings underscore the importance of understanding error sources within LLM reasoning as systems scale to more complex tasks. As LLM reasoning scales to increasingly complex problems, understanding the shifting bottlenecks and sources of error within these systems will be crucial.
Diagnosis-Related Group (DRG) codes are essential for hospital reimbursement and operations but require labor-intensive assignment. Large Language Models (LLMs) struggle with DRG coding due to the out-of-distribution (OOD) nature of the task: pretraining corpora rarely contain private clinical or billing data. We introduce DRG-Sapphire, which uses large-scale reinforcement learning (RL) for automated DRG coding from clinical notes. Built on Qwen2.5-7B and trained with Group Relative Policy Optimization (GRPO) using rule-based rewards, DRG-Sapphire introduces a series of RL enhancements to address domain-specific challenges not seen in previous mathematical tasks. Our model achieves state-of-the-art accuracy on the MIMIC-IV benchmark and generates physician-validated reasoning for DRG assignments, significantly enhancing explainability. Our study further sheds light on broader challenges of applying RL to knowledge-intensive, OOD tasks. We observe that RL performance scales approximately linearly with the logarithm of the number of supervised fine-tuning (SFT) examples, suggesting that RL effectiveness is fundamentally constrained by the domain knowledge encoded in the base model. For OOD tasks like DRG coding, strong RL performance requires sufficient knowledge infusion prior to RL. Consequently, scaling SFT may be more effective and computationally efficient than scaling RL alone for such tasks.
Multi-agent large language model (LLM) systems are increasingly adopted for complex language processing tasks that require communication and coordination among agents. However, these systems often suffer substantial overhead from repeated reprocessing of overlapping contexts across agents. In typical pipelines, once an agent receives a message from its predecessor, the full context-including prior turns-must be reprocessed from scratch, leading to inefficient processing. While key-value (KV) caching is an effective solution for avoiding redundant computation in single-agent settings where prefixes remain unchanged, it cannot be directly reused in multi-agent scenarios due to diverging prefixes introduced by agent-specific context extensions. We identify that the core challenge lies in the offset variance of KV-caches across agents. To address this, we propose KVCOMM, a training-free framework that enables efficient prefilling in multi-agent inference by reusing KV-caches and aligning cache offsets of overlapping contexts under diverse prefix contexts. KVCOMM estimates and adjusts KV-caches for shared content by referencing a pool of cached examples-termed anchors-that store observed cache deviations under varying prefixes. The anchor pool is maintained and updated online, allowing dynamic adaptation to distinct user requests and context structures. KVCOMM achieves over 70% reuse rate across diverse multi-agent workloads, including retrieval-augmented generation, math reasoning, and collaborative coding tasks, all without quality degradation. Particularly, when each fully-connected agent receives 1K input tokens with 512 prefix tokens and 512 output tokens under a five-agent setting, KVCOMM achieves up to 7.8x speedup compared to the standard prefill pipeline, reducing TTFT from ~430 ms to ~55 ms.
Large language models (LLMs) are now widely deployed in user-facing applications, reaching hundreds of millions worldwide. As they become integrated into everyday tasks, growing reliance on their outputs raises significant concerns. In particular, users may unknowingly be exposed to model-inherent biases that systematically disadvantage or stereotype certain groups. However, existing bias benchmarks continue to rely on templated prompts or restrictive multiple-choice questions that are suggestive, simplistic, and fail to capture the complexity of real-world user interactions. In this work, we address this gap by introducing a counterfactual bias evaluation framework that automatically generates realistic, open-ended questions over sensitive attributes such as sex, race, or religion. By iteratively mutating and selecting bias-inducing questions, our approach systematically explores areas where models are most susceptible to biased behavior. Beyond detecting harmful biases, we also capture distinct response dimensions that are increasingly relevant in user interactions, such as asymmetric refusals and explicit acknowledgment of bias. Leveraging our framework, we construct CAB, a human-verified benchmark spanning diverse topics, designed to enable cross-model comparisons. Using CAB, we analyze a range of LLMs across multiple bias dimensions, revealing nuanced insights into how different models manifest bias. For instance, while GPT-5 outperforms other models, it nonetheless exhibits persistent biases in specific scenarios. These findings underscore the need for continual improvements to ensure fair model behavior.
Multimodal Large Language Models (MLLMs) in real-world applications require access to external knowledge sources and must remain responsive to the dynamic and ever-changing real-world information in order to address information-seeking and knowledge-intensive user queries. Existing approaches, such as retrieval augmented generation (RAG) methods, search agents, and search equipped MLLMs, often suffer from rigid pipelines, excessive search calls, and poorly constructed search queries, which result in inefficiencies and suboptimal outcomes. To address these limitations, we present DeepMMSearch-R1, the first multimodal LLM capable of performing on-demand, multi-turn web searches and dynamically crafting queries for both image and text search tools. Specifically, DeepMMSearch-R1 can initiate web searches based on relevant crops of the input image making the image search more effective, and can iteratively adapt text search queries based on retrieved information, thereby enabling self-reflection and self-correction. Our approach relies on a two-stage training pipeline: a cold start supervised finetuning phase followed by an online reinforcement learning optimization. For training, we introduce DeepMMSearchVQA, a novel multimodal VQA dataset created through an automated pipeline intermixed with real-world information from web search tools. This dataset contains diverse, multi-hop queries that integrate textual and visual information, teaching the model when to search, what to search for, which search tool to use and how to reason over the retrieved information. We conduct extensive experiments across a range of knowledge-intensive benchmarks to demonstrate the superiority of our approach. Finally, we analyze the results and provide insights that are valuable for advancing multimodal web-search.
Large Language Models (LLMs) process every token through all layers of a transformer stack, causing wasted computation on simple queries and insufficient flexibility for harder ones that need deeper reasoning. Adaptive-depth methods can improve efficiency, but prior approaches rely on costly inference-time search, architectural changes, or large-scale retraining, and in practice often degrade accuracy despite efficiency gains. We introduce Dr.LLM, Dynamic routing of Layers for LLMs, a retrofittable framework that equips pretrained models with lightweight per-layer routers deciding to skip, execute, or repeat a block. Routers are trained with explicit supervision: using Monte Carlo Tree Search (MCTS), we derive high-quality layer configurations that preserve or improve accuracy under a compute budget. Our design, windowed pooling for stable routing, focal loss with class balancing, and bottleneck MLP routers, ensures robustness under class imbalance and long sequences. On ARC (logic) and DART (math), Dr.LLM improves accuracy by up to +3.4%p while saving 5 layers per example on average. Routers generalize to out-of-domain tasks (MMLU, GSM8k, AIME, TruthfulQA, SQuADv2, GPQA, PIQA, AGIEval) with only 0.85% accuracy drop while retaining efficiency, and outperform prior routing methods by up to +7.7%p. Overall, Dr.LLM shows that explicitly supervised routers retrofit frozen LLMs for budget-aware, accuracy-driven inference without altering base weights.
A long-standing challenge in machine learning has been the rigid separation between data work and model refinement, enforced by slow fine-tuning cycles. The rise of Large Language Models (LLMs) overcomes this historical barrier, allowing applications developers to instantly govern model behavior by editing prompt instructions. This shift enables a new paradigm: data-model co-evolution, where a living test set and a model's instructions evolve in tandem. We operationalize this paradigm in an interactive system designed to address the critical challenge of encoding subtle, domain-specific policies into prompt instructions. The system's structured workflow guides people to discover edge cases, articulate rationales for desired behavior, and iteratively evaluate instruction revisions against a growing test set. A user study shows our workflow helps participants refine instructions systematically and specify ambiguous policies more concretely. This work points toward more robust and responsible LLM applications through human-in-the-loop development aligned with local preferences and policies.
Large Language Models (LLMs) typically rely on a large number of parameters for token embedding, leading to substantial storage requirements and memory footprints. In particular, LLMs deployed on edge devices are memory-bound, and reducing the memory footprint by compressing the embedding layer not only frees up the memory bandwidth but also speeds up inference. To address this, we introduce CARVQ, a post-training novel Corrective Adaptor combined with group Residual Vector Quantization. CARVQ relies on the composition of both linear and non-linear maps and mimics the original model embedding to compress to approximately 1.6 bits without requiring specialized hardware to support lower-bit storage. We test our method on pre-trained LLMs such as LLaMA-3.2-1B, LLaMA-3.2-3B, LLaMA-3.2-3B-Instruct, LLaMA-3.1-8B, Qwen2.5-7B, Qwen2.5-Math-7B and Phi-4, evaluating on common generative, discriminative, math and reasoning tasks. We show that in most cases, CARVQ can achieve lower average bitwidth-per-parameter while maintaining reasonable perplexity and accuracy compared to scalar quantization. Our contributions include a novel compression technique that is compatible with state-of-the-art transformer quantization methods and can be seamlessly integrated into any hardware supporting 4-bit memory to reduce the model's memory footprint in memory-constrained devices. This work demonstrates a crucial step toward the efficient deployment of LLMs on edge devices.
Different open-ended generation tasks require different degrees of output diversity. However, current LLMs are often miscalibrated. They collapse to overly homogeneous outputs for creative tasks and hallucinate diverse but incorrect responses for factual tasks. We argue that these two failure modes are unified by, and can both be addressed by, the notion of effective generation space size (GSS) -- the set of semantically distinct outputs a model considers for a prompt. We present GSSBench, a task suite of prompt pairs with ground-truth GSS relationships to assess different metrics and understand where models diverge from desired behavior. We find that hallucination detection metrics, particularly EigenScore, consistently outperform standard diversity and uncertainty quantification metrics, while using only model internals, providing interpretable insights into a model's internal task representations. We demonstrate three applications of GSS: (1) detecting prompt ambiguity and predicting clarification questions for better grounding, (2) interpreting overthinking and underthinking in reasoning models, and (3) steering models to expand their generation space to yield high-quality and diverse outputs.
With advancements in reasoning capabilities, Large Language Models (LLMs) are increasingly employed for automated judgment tasks. While LLMs-as-Judges offer promise in automating evaluations, current approaches often rely on simplistic aggregation methods (e.g., majority voting), which can fail even when individual agents provide correct answers. To address this, we propose a multi-agent debate judge framework where agents collaboratively reason and iteratively refine their responses. We formalize the debate process mathematically, analyzing agent interactions and proving that debate amplifies correctness compared to static ensembles. To enhance efficiency, we introduce a stability detection mechanism that models judge consensus dynamics via a time-varying Beta-Binomial mixture, with adaptive stopping based on distributional similarity (Kolmogorov-Smirnov test). This mechanism models the judges' collective correct rate dynamics using a time-varying mixture of Beta-Binomial distributions and employs an adaptive stopping criterion based on distributional similarity (Kolmogorov-Smirnov statistic). Experiments across multiple benchmarks and models demonstrate that our framework improves judgment accuracy over majority voting while maintaining computational efficiency.
Large language models (LLMs) have shown promising accuracy in predicting survey responses and policy preferences, which has increased interest in their potential to represent human interests in various domains. Most existing research has focused on behavioral cloning, effectively evaluating how well models reproduce individuals' expressed preferences. Drawing on theories of political representation, we highlight an underexplored design trade-off: whether AI systems should act as delegates, mirroring expressed preferences, or as trustees, exercising judgment about what best serves an individual's interests. This trade-off is closely related to issues of LLM sycophancy, where models can encourage behavior or validate beliefs that may be aligned with a user's short-term preferences, but is detrimental to their long-term interests. Through a series of experiments simulating votes on various policy issues in the U.S. context, we apply a temporal utility framework that weighs short and long-term interests (simulating a trustee role) and compare voting outcomes to behavior-cloning models (simulating a delegate). We find that trustee-style predictions weighted toward long-term interests produce policy decisions that align more closely with expert consensus on well-understood issues, but also show greater bias toward models' default stances on topics lacking clear agreement. These findings reveal a fundamental trade-off in designing AI systems to represent human interests. Delegate models better preserve user autonomy but may diverge from well-supported policy positions, while trustee models can promote welfare on well-understood issues yet risk paternalism and bias on subjective topics.
Hybrid thinking enables LLMs to switch between reasoning and direct answering, offering a balance between efficiency and reasoning capability. Yet our experiments reveal that current hybrid thinking LLMs only achieve partial mode separation: reasoning behaviors often leak into the no-think mode. To understand and mitigate this, we analyze the factors influencing controllability and identify four that matter most: (1) larger data scale, (2) using think and no-think answers from different questions rather than the same question, (3) a moderate increase in no-think data number, and (4) a two-phase strategy that first trains reasoning ability and then applies hybrid think training. Building on these findings, we propose a practical recipe that, compared to standard training, can maintain accuracy in both modes while significantly reducing no-think output length (from $1085$ to $585$ on MATH500) and occurrences of reasoning-supportive tokens such as ``\texttt{wait}'' (from $5917$ to $522$ on MATH500). Our findings highlight the limitations of current hybrid thinking and offer directions for strengthening its controllability.
LLM-based autonomous agents have recently shown strong capabilities in solving complex industrial design tasks. However, in domains aiming for carbon neutrality and high-performance renewable energy systems, current AI-assisted design automation methods face critical challenges in explainability, scalability, and practical usability. To address these limitations, we introduce PHIA (Physics-Informed Autonomous Agent), an LLM-driven system that automates modulation design for power converters in Power Electronics Systems with minimal human intervention. In contrast to traditional pipeline-based methods, PHIA incorporates an LLM-based planning module that interactively acquires and verifies design requirements via a user-friendly chat interface. This planner collaborates with physics-informed simulation and optimization components to autonomously generate and iteratively refine modulation designs. The interactive interface also supports interpretability by providing textual explanations and visual outputs throughout the design process. Experimental results show that PHIA reduces standard mean absolute error by 63.2% compared to the second-best benchmark and accelerates the overall design process by over 33 times. A user study involving 20 domain experts further confirms PHIA's superior design efficiency and usability, highlighting its potential to transform industrial design workflows in power electronics.
Large Language Models (LLMs) require significant GPU memory when processing long texts, with the key value (KV) cache consuming up to 70\% of total memory during inference. Although existing compression methods reduce memory by evaluating the importance of individual tokens, they overlook critical semantic relationships between tokens, resulting in fragmented context and degraded performance. We introduce ChunkKV, which fundamentally reimagines KV cache compression by treating semantic chunks - rather than isolated tokens - as basic compression units. This approach preserves complete linguistic structures and contextual integrity, ensuring that essential meaning is retained even under aggressive compression. Our innovation includes a novel layer-wise index reuse technique that exploits the higher cross-layer similarity of preserved indices in ChunkKV, reducing computational overhead and improving throughput by 26.5\%. Comprehensive evaluations on challenging benchmarks: LongBench, Needle-In-A-HayStack, GSM8K, and JailbreakV demonstrate that ChunkKV outperforms state-of-the-art methods by up to 8.7\% in precision while maintaining the same compression ratio. These results confirm that semantic-aware compression significantly enhances both efficiency and performance for long-context LLM inference, providing a simple yet effective solution to the memory bottleneck problem. The code is available at \href{https://github.com/NVIDIA/kvpress}{link}.
Knowledge editing and machine unlearning are two popular approaches for large language models (LLMs) to stay up-to-date. However, the knowledge updating mechanism of LLMs remains largely unexplored due to insufficient, isolated, and small-scale evaluation. For instance, are LLMs similar to humans in modifying certain knowledge? What differs editing and unlearning as training data increases? This paper proposes KnowledgeSmith, a unified framework to systematically understand the updating mechanism of LLMs. We first cast editing and unlearning as instances of one constrained optimization problem. Then, we propose an automatic dataset generator that provides structured interventions across multiple graph levels and data scales, enabling controlled studies of how different modification strategies propagate through model knowledge. Extensive experiments demonstrate nuanced insights over knowledge propagation, plasticity scaling, consistency, and robustness. For instance, our results show that LLMs do not exhibit similar updating as humans for different levels of knowledge, and there exists consistency-capacity trade-off. We hope our findings can offer suggestions to the design of more reliable and scalable strategies. Code: https://github.com/AIFrontierLab/KnowledgeSmith.git
With the increasing integration of large language models (LLMs) into open-domain writing, detecting machine-generated text has become a critical task for ensuring content authenticity and trust. Existing approaches rely on statistical discrepancies or model-specific heuristics to distinguish between LLM-generated and human-written text. However, these methods struggle in real-world scenarios due to limited generalization, vulnerability to paraphrasing, and lack of explainability, particularly when facing stylistic diversity or hybrid human-AI authorship. In this work, we propose StyleDecipher, a robust and explainable detection framework that revisits LLM-generated text detection using combined feature extractors to quantify stylistic differences. By jointly modeling discrete stylistic indicators and continuous stylistic representations derived from semantic embeddings, StyleDecipher captures distinctive style-level divergences between human and LLM outputs within a unified representation space. This framework enables accurate, explainable, and domain-agnostic detection without requiring access to model internals or labeled segments. Extensive experiments across five diverse domains, including news, code, essays, reviews, and academic abstracts, demonstrate that StyleDecipher consistently achieves state-of-the-art in-domain accuracy. Moreover, in cross-domain evaluations, it surpasses existing baselines by up to 36.30%, while maintaining robustness against adversarial perturbations and mixed human-AI content. Further qualitative and quantitative analysis confirms that stylistic signals provide explainable evidence for distinguishing machine-generated text. Our source code can be accessed at https://github.com/SiyuanLi00/StyleDecipher.
Human feedback plays a pivotal role in aligning large language models (LLMs) with human preferences. However, such feedback is often noisy or inconsistent, which can degrade the quality of reward models and hinder alignment. While various automated data cleaning methods have been proposed to mitigate this issue, a systematic evaluation of their effectiveness and generalizability remains lacking. To bridge this gap, we introduce the first comprehensive benchmark for evaluating 13 preference data cleaning methods in the context of LLM alignment. PrefCleanBench offers a standardized protocol to assess cleaning strategies in terms of alignment performance and generalizability across diverse datasets, model architectures, and optimization algorithms. By unifying disparate methods and rigorously comparing them, we uncover key factors that determine the success of data cleaning in alignment tasks. This benchmark lays the groundwork for principled and reproducible approaches to improving LLM alignment through better data quality-highlighting the crucial but underexplored role of data preprocessing in responsible AI development. We release modular implementations of all methods to catalyze further research: https://github.com/deeplearning-wisc/PrefCleanBench.
The evolution toward open, programmable O-RAN and AI-RAN 6G networks creates unprecedented opportunities for Intent-Based Networking (IBN) to dynamically optimize RAN[...]. However, applying IBN effectively to the RAN scheduler [...] remains a significant challenge. Current approaches predominantly rely on coarse-grained network slicing, lacking the granularity for dynamic adaptation to individual user conditions and traffic patterns. Despite the existence of a vast body of scheduling algorithms [...], their practical utilization is hindered by implementation heterogeneity, insufficient systematic evaluation in production environments, and the complexity of developing high-performance scheduler implementations.[...] To address these limitations, we propose ALLSTaR (Automated LLm-driven Scheduler generation and Testing for intent-based RAN), a novel framework leveraging LLMs for automated, intent-driven scheduler design, implementation, and evaluation. ALLSTaR interprets NL intents, automatically generates functional scheduler code from the research literature using OCR and LLMs, and intelligently matches operator intents to the most suitable scheduler(s). Our implementation deploys these schedulers as O-RAN dApps, enabling on-the-fly deployment and testing on a production-grade, 5G-compliant testbed. This approach has enabled the largest-scale OTA experimental comparison of 18 scheduling algorithms automatically synthesized from the academic literature. The resulting performance profiles serve as the input for our Intent-Based Scheduling (IBS) framework, which dynamically selects and deploys appropriate schedulers that optimally satisfy operator intents. We validate our approach through multiple use cases unattainable with current slicing-based optimization techniques, demonstrating fine-grained control based on buffer status, physical layer conditions, and heterogeneous traffic types