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
The Finite Element Method (FEM) is widely used in engineering and scientific computing, but its pre-processing, solver configuration, and post-processing stages are often time-consuming and require specialized knowledge. This paper proposes an automated solution framework, MooseAgent, for the multi-physics simulation framework MOOSE, which combines large-scale pre-trained language models (LLMs) with a multi-agent system. The framework uses LLMs to understand user-described simulation requirements in natural language and employs task decomposition and multi-round iterative verification strategies to automatically generate MOOSE input files. To improve accuracy and reduce model hallucinations, the system builds and utilizes a vector database containing annotated MOOSE input cards and function documentation. We conducted experimental evaluations on several typical cases, including heat transfer, mechanics, phase field, and multi-physics coupling. The results show that MooseAgent can automate the MOOSE simulation process to a certain extent, especially demonstrating a high success rate when dealing with relatively simple single-physics problems. The main contribution of this research is the proposal of a multi-agent automated framework for MOOSE, which validates its potential in simplifying finite element simulation processes and lowering the user barrier, providing new ideas for the development of intelligent finite element simulation software. The code for the MooseAgent framework proposed in this paper has been open-sourced and is available at https://github.com/taozhan18/MooseAgent
Existing challenges in misinformation exposure and susceptibility vary across demographic groups, as some populations are more vulnerable to misinformation than others. Large language models (LLMs) introduce new dimensions to these challenges through their ability to generate persuasive content at scale and reinforcing existing biases. This study investigates the bidirectional persuasion dynamics between LLMs and humans when exposed to misinformative content. We analyze human-to-LLM influence using human-stance datasets and assess LLM-to-human influence by generating LLM-based persuasive arguments. Additionally, we use a multi-agent LLM framework to analyze the spread of misinformation under persuasion among demographic-oriented LLM agents. Our findings show that demographic factors influence susceptibility to misinformation in LLMs, closely reflecting the demographic-based patterns seen in human susceptibility. We also find that, similar to human demographic groups, multi-agent LLMs exhibit echo chamber behavior. This research explores the interplay between humans and LLMs, highlighting demographic differences in the context of misinformation and offering insights for future interventions.
Large Language Models (LLMs) are increasingly being used to autonomously evaluate the quality of content in communication systems, e.g., to assess responses in telecom customer support chatbots. However, the impartiality of these AI "judges" is not guaranteed, and any biases in their evaluation criteria could skew outcomes and undermine user trust. In this paper, we systematically investigate judgment biases in two LLM-as-a-judge models (i.e., GPT-Judge and JudgeLM) under the point-wise scoring setting, encompassing 11 types of biases that cover both implicit and explicit forms. We observed that state-of-the-art LLM judges demonstrate robustness to biased inputs, generally assigning them lower scores than the corresponding clean samples. Providing a detailed scoring rubric further enhances this robustness. We further found that fine-tuning an LLM on high-scoring yet biased responses can significantly degrade its performance, highlighting the risk of training on biased data. We also discovered that the judged scores correlate with task difficulty: a challenging dataset like GPQA yields lower average scores, whereas an open-ended reasoning dataset (e.g., JudgeLM-val) sees higher average scores. Finally, we proposed four potential mitigation strategies to ensure fair and reliable AI judging in practical communication scenarios.
Training LLMs relies on distributed implementations using multiple GPUs to compute gradients in parallel with sharded optimizers. However, synchronizing gradients in data parallel setups introduces communication overhead that grows with the number of workers, limiting parallelization efficiency. Local optimization algorithms reduce communications but incur high memory costs as they prevent optimizer state sharding, hindering scalability. To address this, we propose \textbf{AC}cumulate while \textbf{CO}mmunicate (ACCO), a memory-efficient optimization algorithm for distributed LLM training. By synchronizing delayed gradients while computing new ones, ACCO reduces GPU idle time and supports heterogeneous hardware. To mitigate the convergence issues caused by delayed updates, we introduce a novel technique ensuring training dynamics align with standard distributed optimization. Compared to ZeRO-1, our approach is significantly faster and scales effectively across heterogeneous hardware.
The polarization of opinions, information segregation, and cognitive biases on social media have attracted significant academic attention. In real-world networks, information often spans multiple interrelated topics, posing challenges for opinion evolution and highlighting the need for frameworks that simulate interactions among topics. Existing studies based on large language models (LLMs) focus largely on single topics, limiting the capture of cognitive transfer in multi-topic, cross-domain contexts. Traditional numerical models, meanwhile, simplify complex linguistic attitudes into discrete values, lacking interpretability, behavioral consistency, and the ability to integrate multiple topics. To address these issues, we propose Multi-topic Opinion Simulation (MTOS), a social simulation framework integrating multi-topic contexts with LLMs. MTOS leverages LLMs alongside short-term and long-term memory, incorporates multiple user-selection interaction mechanisms and dynamic topic-selection strategies, and employs a belief decay mechanism to enable perspective updates across topics. We conduct extensive experiments on MTOS, varying topic numbers, correlation types, and performing ablation studies to assess features such as group polarization and local consistency. Results show that multi-topic settings significantly alter polarization trends: positively correlated topics amplify echo chambers, negatively correlated topics inhibit them, and irrelevant topics also mitigate echo chamber effects through resource competition. Compared with numerical models, LLM-based agents realistically simulate dynamic opinion changes, reproduce linguistic features of news texts, and capture complex human reasoning, improving simulation interpretability and system stability.
We present PricingLogic, the first benchmark that probes whether Large Language Models(LLMs) can reliably automate tourism-related prices when multiple, overlapping fare rules apply. Travel agencies are eager to offload this error-prone task onto AI systems; however, deploying LLMs without verified reliability could result in significant financial losses and erode customer trust. PricingLogic comprises 300 natural-language questions based on booking requests derived from 42 real-world pricing policies, spanning two levels of difficulty: (i) basic customer-type pricing and (ii)bundled-tour calculations involving interacting discounts. Evaluations of a line of LLMs reveal a steep performance drop on the harder tier,exposing systematic failures in rule interpretation and arithmetic reasoning.These results highlight that, despite their general capabilities, today's LLMs remain unreliable in revenue-critical applications without further safeguards or domain adaptation. Our code and dataset are available at https://github.com/EIT-NLP/PricingLogic.
Large language models (LLMs) have achieved significant success in reasoning tasks, including mathematical reasoning and logical deduction. Among these reasoning tasks, graph problems stand out due to their complexity and unique structural characteristics, attracting considerable attention from researchers. Previous studies have explored LLMs' graph reasoning abilities through various techniques, such as different encoding methods for graph structures and the use of carefully designed prompts. However, a critical factor has been mostly overlooked: the prompt sequential order in which graph descriptions are presented to the models. In this study, we present the first comprehensive analysis of how the order of graph descriptions impacts LLM performance. Specifically, we comprehensively evaluate four graph description orders across six graph problems using six mainstream LLMs. The results reveal that: (1) ordered graph descriptions significantly improve LLMs' comprehension of graph structures; (2) the robustness of LLMs to graph description order varies across different tasks; and (3) the impact of graph order on performance is closely related to the inherent characteristics of tasks. This study provides a critical advancement in the application of LLMs for solving graph-related problems, paving the way for future research to optimize model performance through strategic graph description ordering.
Tokenization disparities pose a significant barrier to achieving equitable access to artificial intelligence across linguistically diverse populations. This study conducts a large-scale cross-linguistic evaluation of tokenization efficiency in over 200 languages to systematically quantify computational inequities in large language models (LLMs). Using a standardized experimental framework, we applied consistent preprocessing and normalization protocols, followed by uniform tokenization through the tiktoken library across all language samples. Comprehensive tokenization statistics were collected using established evaluation metrics, including Tokens Per Sentence (TPS) and Relative Tokenization Cost (RTC), benchmarked against English baselines. Our cross-linguistic analysis reveals substantial and systematic disparities: Latin-script languages consistently exhibit higher tokenization efficiency, while non-Latin and morphologically complex languages incur significantly greater token inflation, often 3-5 times higher RTC ratios. These inefficiencies translate into increased computational costs and reduced effective context utilization for underrepresented languages. Overall, the findings highlight structural inequities in current AI systems, where speakers of low-resource and non-Latin languages face disproportionate computational disadvantages. Future research should prioritize the development of linguistically informed tokenization strategies and adaptive vocabulary construction methods that incorporate typological diversity, ensuring more inclusive and computationally equitable multilingual AI systems.
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.
Understanding the alignment between large language models (LLMs) and human brain activity can reveal computational principles underlying language processing. We introduce a fine-grained input attribution method to identify the specific words most important for brain-LLM alignment, and leverage it to study a contentious research question about brain-LLM alignment: the relationship between brain alignment (BA) and next-word prediction (NWP). Our findings reveal that BA and NWP rely on largely distinct word subsets: NWP exhibits recency and primacy biases with a focus on syntax, while BA prioritizes semantic and discourse-level information with a more targeted recency effect. This work advances our understanding of how LLMs relate to human language processing and highlights differences in feature reliance between BA and NWP. Beyond this study, our attribution method can be broadly applied to explore the cognitive relevance of model predictions in diverse language processing tasks.
Large language models have recently demonstrated advanced capabilities in solving IMO and Putnam problems; yet their role in research mathematics has remained fairly limited. The key difficulty is verification: suggested proofs may look plausible, but cannot be trusted without rigorous checking. We present a framework, called LLM+CAS, and an associated tool, O-Forge, that couples frontier LLMs with a computer algebra systems (CAS) in an In-Context Symbolic Feedback loop to produce proofs that are both creative and symbolically verified. Our focus is on asymptotic inequalities, a topic that often involves difficult proofs and appropriate decomposition of the domain into the "right" subdomains. Many mathematicians, including Terry Tao, have suggested that using AI tools to find the right decompositions can be very useful for research-level asymptotic analysis. In this paper, we show that our framework LLM+CAS turns out to be remarkably effective at proposing such decompositions via a combination of a frontier LLM and a CAS. More precisely, we use an LLM to suggest domain decomposition, and a CAS (such as Mathematica) that provides a verification of each piece axiomatically. Using this loop, we answer a question posed by Terence Tao: whether LLMs coupled with a verifier can be used to help prove intricate asymptotic inequalities. More broadly, we show how AI can move beyond contest math towards research-level tools for professional mathematicians.
As natural language corpora expand at an unprecedented rate, manual annotation remains a significant methodological bottleneck in corpus linguistic work. We address this challenge by presenting a scalable, unsupervised pipeline for automating grammatical annotation in voluminous corpora using large language models (LLMs). Unlike previous supervised and iterative approaches, our method employs a four-phase workflow: prompt engineering, pre-hoc evaluation, automated batch processing, and post-hoc validation. We demonstrate the pipeline's accessibility and effectiveness through a diachronic case study of variation in the English consider construction. Using GPT-5 through the OpenAI API, we annotate 143,933 sentences from the Corpus of Historical American English (COHA) in under 60 hours, achieving 98%+ accuracy on two sophisticated annotation procedures. Our results suggest that LLMs can perform a range of data preparation tasks at scale with minimal human intervention, opening new possibilities for corpus-based research, though implementation requires attention to costs, licensing, and other ethical considerations.
Understanding how software developers think, make decisions, and behave remains a key challenge in software engineering (SE). Verbalization techniques (methods that capture spoken or written thought processes) offer a lightweight and accessible way to study these cognitive aspects. This paper presents a scoping review of research at the intersection of SE and psychology (PSY), focusing on the use of verbal data. To make large-scale interdisciplinary reviews feasible, we employed a large language model (LLM)-assisted screening pipeline using GPT to assess the relevance of over 9,000 papers based solely on titles. We addressed two questions: what themes emerge from verbalization-related work in SE, and how effective are LLMs in supporting interdisciplinary review processes? We validated GPT's outputs against human reviewers and found high consistency, with a 13\% disagreement rate. Prominent themes mainly were tied to the craft of SE, while more human-centered topics were underrepresented. The data also suggests that SE frequently draws on PSY methods, whereas the reverse is rare.
Large language models (LLMs) are increasingly deployed as automatic judges to evaluate system outputs in tasks such as summarization, dialogue, and creative writing. A faithful judge should base its verdicts solely on response quality and explicitly acknowledge the factors shaping its decision. We show that current LLM judges fail on both counts by relying on shortcuts introduced in the prompt. Our study uses two evaluation datasets: ELI5, a benchmark for long-form question answering, and LitBench, a recent benchmark for creative writing. Both datasets provide pairwise comparisons, where the evaluator must choose which of two responses is better. From each dataset we construct 100 pairwise judgment tasks and employ two widely used models, GPT-4o and Gemini-2.5-Flash, as evaluators in the role of LLM-as-a-judge. For each pair, we assign superficial cues to the responses, provenance cues indicating source identity (Human, Expert, LLM, or Unknown) and recency cues indicating temporal origin (Old, 1950 vs. New, 2025), while keeping the rest of the prompt fixed. Results reveal consistent verdict shifts: both models exhibit a strong recency bias, systematically favoring new responses over old, as well as a clear provenance hierarchy (Expert > Human > LLM > Unknown). These biases are especially pronounced in GPT-4o and in the more subjective and open-ended LitBench domain. Crucially, cue acknowledgment is rare: justifications almost never reference the injected cues, instead rationalizing decisions in terms of content qualities. These findings demonstrate that current LLM-as-a-judge systems are shortcut-prone and unfaithful, undermining their reliability as evaluators in both research and deployment.
Large language models (LLMs) increasingly rely on chain-of-thought (CoT) prompting to solve mathematical and logical reasoning tasks. Yet, a central question remains: to what extent are these generated rationales \emph{faithful} to the underlying computations, rather than post-hoc narratives shaped by hints that function as answer shortcuts embedded in the prompt? Following prior work on hinted vs.\ unhinted prompting, we present a systematic study of CoT faithfulness under controlled hint manipulations. Our experimental design spans four datasets (AIME, GSM-Hard, MATH-500, UniADILR), two state-of-the-art models (GPT-4o and Gemini-2-Flash), and a structured set of hint conditions varying in correctness (correct and incorrect), presentation style (sycophancy and data leak), and complexity (raw answers, two-operator expressions, four-operator expressions). We evaluate both task accuracy and whether hints are explicitly acknowledged in the reasoning. Our results reveal three key findings. First, correct hints substantially improve accuracy, especially on harder benchmarks and logical reasoning, while incorrect hints sharply reduce accuracy in tasks with lower baseline competence. Second, acknowledgement of hints is highly uneven: equation-based hints are frequently referenced, whereas raw hints are often adopted silently, indicating that more complex hints push models toward verbalizing their reliance in the reasoning process. Third, presentation style matters: sycophancy prompts encourage overt acknowledgement, while leak-style prompts increase accuracy but promote hidden reliance. This may reflect RLHF-related effects, as sycophancy exploits the human-pleasing side and data leak triggers the self-censoring side. Together, these results demonstrate that LLM reasoning is systematically shaped by shortcuts in ways that obscure faithfulness.
Large language models (LLMs) are rapidly transitioning into medical clinical use, yet their reliability under realistic, multi-turn interactions remains poorly understood. Existing evaluation frameworks typically assess single-turn question answering under idealized conditions, overlooking the complexities of medical consultations where conflicting input, misleading context, and authority influence are common. We introduce MedQA-Followup, a framework for systematically evaluating multi-turn robustness in medical question answering. Our approach distinguishes between shallow robustness (resisting misleading initial context) and deep robustness (maintaining accuracy when answers are challenged across turns), while also introducing an indirect-direct axis that separates contextual framing (indirect) from explicit suggestion (direct). Using controlled interventions on the MedQA dataset, we evaluate five state-of-the-art LLMs and find that while models perform reasonably well under shallow perturbations, they exhibit severe vulnerabilities in multi-turn settings, with accuracy dropping from 91.2% to as low as 13.5% for Claude Sonnet 4. Counterintuitively, indirect, context-based interventions are often more harmful than direct suggestions, yielding larger accuracy drops across models and exposing a significant vulnerability for clinical deployment. Further compounding analyses reveal model differences, with some showing additional performance drops under repeated interventions while others partially recovering or even improving. These findings highlight multi-turn robustness as a critical but underexplored dimension for safe and reliable deployment of medical LLMs.
Post-training of large language models is essential for adapting pre-trained language models (PLMs) to align with human preferences and downstream tasks. While PLMs typically exhibit well-calibrated confidence, post-trained language models (PoLMs) often suffer from over-confidence, assigning high confidence to both correct and incorrect outputs, which can undermine reliability in critical applications. A major obstacle in calibrating PoLMs is the scarcity of labeled data for individual downstream tasks. To address this, we propose Disagreement-Aware Confidence Alignment (DACA), a novel unsupervised method to optimize the parameters (e.g., temperature $\tau$) in post-hoc confidence calibration. Our method is motivated by the under-confidence issue caused by prediction disagreement between the PLM and PoLM while aligning their confidence via temperature scaling. Theoretically, the PLM's confidence underestimates PoLM's prediction accuracy on disagreement examples, causing a larger $\tau$ and producing under-confident predictions. DACA mitigates this by selectively using only agreement examples for calibration, effectively decoupling the influence of disagreement. In this manner, our method avoids an overly large $\tau$ in temperature scaling caused by disagreement examples, improving calibration performance. Extensive experiments demonstrate the effectiveness of our method, improving the average ECE of open-sourced and API-based LLMs (e.g. GPT-4o) by up to 15.08$\%$ on common benchmarks.
Effectively integrating molecular graph structures with Large Language Models (LLMs) is a key challenge in drug discovery. Most existing multi-modal alignment methods typically process these structures by fine-tuning the LLM or adding a static adapter simultaneously. However, these approaches have two main limitations: (1) it optimizes a shared parameter space across all molecular inputs, limiting the model's ability to capture instance-specific structural features; and (2) fine-tuning the LLM for molecular tasks can lead to catastrophic forgetting, undermining its general reasoning capabilities. In this paper, instead of static task-oriented adaptation, we propose an instance-specific parameter space alignment approach for each molecule on-the-fly. To this end, we introduce Molecule-aware Low-Rank Adaptation (MoRA) that produces a unique set of low-rank adaptation weights for each input molecular graph. These weights are then dynamically injected into a frozen LLM, allowing the model to adapt its reasoning to the structure of each molecular input, while preserving the LLM's core knowledge. Extensive experiments demonstrate that on key molecular tasks, such as chemical reaction prediction and molecular captioning, MoRA's instance-specific dynamic adaptation outperforms statically adapted baselines, including a 14.1% relative improvement in reaction prediction exact match and a 22% reduction in error for quantum property prediction. The code is available at https://github.com/jk-sounds/MoRA.
Large language models (LLMs) have been shown to internalize human-like biases during finetuning, yet the mechanisms by which these biases manifest remain unclear. In this work, we investigated whether the well-known Knobe effect, a moral bias in intentionality judgements, emerges in finetuned LLMs and whether it can be traced back to specific components of the model. We conducted a Layer-Patching analysis across 3 open-weights LLMs and demonstrated that the bias is not only learned during finetuning but also localized in a specific set of layers. Surprisingly, we found that patching activations from the corresponding pretrained model into just a few critical layers is sufficient to eliminate the effect. Our findings offer new evidence that social biases in LLMs can be interpreted, localized, and mitigated through targeted interventions, without the need for model retraining.
The reliable evaluation of large language models (LLMs) in medical applications remains an open challenge, particularly in capturing the complexity of multi-turn doctor-patient interactions that unfold in real clinical environments. Existing evaluation methods typically rely on post hoc review of full conversation transcripts, thereby neglecting the dynamic, context-sensitive nature of medical dialogues and the evolving informational needs of patients. In this work, we present MedKGEval, a novel multi-turn evaluation framework for clinical LLMs grounded in structured medical knowledge. Our approach introduces three key contributions: (1) a knowledge graph-driven patient simulation mechanism, where a dedicated control module retrieves relevant medical facts from a curated knowledge graph, thereby endowing the patient agent with human-like and realistic conversational behavior. This knowledge graph is constructed by integrating open-source resources with additional triples extracted from expert-annotated datasets; (2) an in-situ, turn-level evaluation framework, where each model response is assessed by a Judge Agent for clinical appropriateness, factual correctness, and safety as the dialogue progresses using a suite of fine-grained, task-specific metrics; (3) a comprehensive multi-turn benchmark of eight state-of-the-art LLMs, demonstrating MedKGEval's ability to identify subtle behavioral flaws and safety risks that are often overlooked by conventional evaluation pipelines. Although initially designed for Chinese and English medical applications, our framework can be readily extended to additional languages by switching the input knowledge graphs, ensuring seamless bilingual support and domain-specific applicability.
Large language models (LLMs) are increasingly deployed across high-impact domains, from clinical decision support and legal analysis to hiring and education, making fairness and bias evaluation before deployment critical. However, existing evaluations lack grounding in real-world scenarios and do not account for differences in harm severity, e.g., a biased decision in surgery should not be weighed the same as a stylistic bias in text summarization. To address this gap, we introduce HALF (Harm-Aware LLM Fairness), a deployment-aligned framework that assesses model bias in realistic applications and weighs the outcomes by harm severity. HALF organizes nine application domains into three tiers (Severe, Moderate, Mild) using a five-stage pipeline. Our evaluation results across eight LLMs show that (1) LLMs are not consistently fair across domains, (2) model size or performance do not guarantee fairness, and (3) reasoning models perform better in medical decision support but worse in education. We conclude that HALF exposes a clear gap between previous benchmarking success and deployment readiness.
Large language models (LLMs) are highly sensitive to their input prompts, making prompt design a central challenge. While automatic prompt optimization (APO) reduces manual engineering, most approaches assume access to ground-truth references such as labeled validation data. In practice, however, collecting high-quality labels is costly and slow. We propose the Prompt Duel Optimizer (PDO), a sample-efficient framework for label-free prompt optimization. PDO formulates the problem as a dueling-bandit setting, where supervision signal comes from pairwise preference feedback provided by an LLM judge. The framework combines Double Thompson Sampling (D-TS), which prioritizes informative prompt comparisons, with Top-Performer Guided Mutation, which expands the candidate pool by mutating high-performing prompts. PDO naturally operates in label-free settings and can also incorporate partial labels to mitigate judge noise. Experiments on BIG-bench Hard (BBH) and MS MARCO show that PDO consistently outperforms baseline methods. Ablation studies further demonstrate the effectiveness of both D-TS and prompt mutation.
Large language models (LLMs) achieve impressive performance across diverse tasks yet remain vulnerable to jailbreak attacks that bypass safety mechanisms. We present RAID (Refusal-Aware and Integrated Decoding), a framework that systematically probes these weaknesses by crafting adversarial suffixes that induce restricted content while preserving fluency. RAID relaxes discrete tokens into continuous embeddings and optimizes them with a joint objective that (i) encourages restricted responses, (ii) incorporates a refusal-aware regularizer to steer activations away from refusal directions in embedding space, and (iii) applies a coherence term to maintain semantic plausibility and non-redundancy. After optimization, a critic-guided decoding procedure maps embeddings back to tokens by balancing embedding affinity with language-model likelihood. This integration yields suffixes that are both effective in bypassing defenses and natural in form. Experiments on multiple open-source LLMs show that RAID achieves higher attack success rates with fewer queries and lower computational cost than recent white-box and black-box baselines. These findings highlight the importance of embedding-space regularization for understanding and mitigating LLM jailbreak vulnerabilities.
Attributing authorship in the era of large language models (LLMs) is increasingly challenging as machine-generated prose rivals human writing. We benchmark two complementary attribution mechanisms , fixed Style Embeddings and an instruction-tuned LLM judge (GPT-4o) on the Human AI Parallel Corpus, an open dataset of 600 balanced instances spanning six domains (academic, news, fiction, blogs, spoken transcripts, and TV/movie scripts). Each instance contains a human prompt with both a gold continuation and an LLM-generated continuation from either GPT-4o or LLaMA-70B-Instruct. The Style Embedding baseline achieves stronger aggregate accuracy on GPT continuations (82 pct vs. 68 pct). The LLM Judge is slightly better than the Style embeddings on LLaMA continuations (85 pct vs. 81 pct) but the results are not statistically significant. Crucially, the LLM judge significantly outperforms in fiction and academic prose, indicating semantic sensitivity, whereas embeddings dominate in spoken and scripted dialogue, reflecting structural strengths. These complementary patterns highlight attribution as a multidimensional problem requiring hybrid strategies. To support reproducibility we provide code on GitHub and derived data on Hugging Face under the MIT license. This open framework provides a reproducible benchmark for attribution quality assessment in AI-generated content, along with a review of related literature influencing this work.
In recent years, recommendation systems have evolved from providing a single list of recommendations to offering a comprehensive suite of topic focused services. To better accomplish this task, conversational recommendation systems (CRS) have progressed from basic retrieval augmented LLM generation to agentic systems with advanced reasoning and self correction capabilities. However, agentic systems come with notable response latency, a longstanding challenge for conversational recommendation systems. To balance the trade off between handling complex queries and minimizing latency, we propose AdaptJobRec, the first conversational job recommendation system that leverages autonomous agent to integrate personalized recommendation algorithm tools. The system employs a user query complexity identification mechanism to minimize response latency. For straightforward queries, the agent directly selects the appropriate tool for rapid responses. For complex queries, the agent uses the memory processing module to filter chat history for relevant content, then passes the results to the intelligent task decomposition planner, and finally executes the tasks using personalized recommendation tools. Evaluation on Walmart's real world career recommendation scenarios demonstrates that AdaptJobRec reduces average response latency by up to 53.3% compared to competitive baselines, while significantly improving recommendation accuracy.
Large language models (LLMs) often exhibit complementary strengths. Model routing harnesses these strengths by dynamically directing each query to the most suitable model, given a candidate model pool. However, routing performance relies on accurate model representations, and adding new models typically requires retraining, limiting scalability. To address these challenges, we propose a novel routing method using in-context vectors to represent model capabilities. The method proceeds in two stages. First, queries are embedded and projected into vectors, with a projector and LLM-based router trained to reconstruct the original queries, aligning vector representations with the router's semantic space. Second, each candidate model is profiled on a query set, and the router learns -- based on in-context vectors of query and model performance -- to predict whether each model can correctly answer new queries. Extensive experiments demonstrate that our method achieves state-of-the-art routing performance in both in-distribution and out-of-distribution tasks. Moreover, our method allows for seamless integration of new models without retraining the router. The code is available at https://github.com/lalalamdbf/ICL-Router.
Recent advancements in Large Language Model (LLM) safety have primarily focused on mitigating attacks crafted in natural language or common ciphers (e.g. Base64), which are likely integrated into newer models' safety training. However, we reveal a paradoxical vulnerability: as LLMs advance in reasoning, they inadvertently become more susceptible to novel jailbreaking attacks. Enhanced reasoning enables LLMs to interpret complex instructions and decode complex user-defined ciphers, creating an exploitable security gap. To study this vulnerability, we introduce Attacks using Custom Encryptions (ACE), a jailbreaking technique that encodes malicious queries with novel ciphers. Extending ACE, we introduce Layered Attacks using Custom Encryptions (LACE), which applies multi-layer ciphers to amplify attack complexity. Furthermore, we develop CipherBench, a benchmark designed to evaluate LLMs' accuracy in decoding encrypted benign text. Our experiments reveal a critical trade-off: LLMs that are more capable of decoding ciphers are more vulnerable to LACE, with success rates on gpt-oss-20b escalating from 60% under ACE to 72% with LACE. These findings highlight a critical insight: as LLMs become more adept at deciphering complex user ciphers--many of which cannot be preemptively included in safety training--they become increasingly exploitable.
Multi-agent systems (MAS) and reinforcement learning (RL) are widely used to enhance the agentic capabilities of large language models (LLMs). MAS improves task performance through role-based orchestration, while RL uses environmental rewards to learn stronger policies, such as GRPO-style optimization. However, applying on-policy RL to MAS remains underexplored and presents unique challenges. Algorithmically, standard GRPO grouping assumptions break down because prompts vary by role and by turn. System-wise, the training stack must support MAS-workflow rollouts and on-policy updates for both single-policy and multi-policy models. We propose AT-GRPO, which includes (i) an agent- and turn-wise grouped RL algorithm tailored to MAS and (ii) a training system that supports both single- and multi-policy regimes. Across game, planning, coding, and math tasks, AT-GRPO delivers substantial gains. On long-horizon planning, it increases accuracy from a 14.0 to 47.0 percent single-agent RL baseline to 96.0 to 99.5 percent. It also improves reasoning performance, with average gains of 3.87 to 7.62 percent on coding tasks and 9.0 to 17.93 percent on math. Code and environments are available at: https://github.com/pettingllms-ai/PettingLLMs.
Large Language Models (LLMs) have gained significant traction in various applications, yet their capabilities present risks for both constructive and malicious exploitation. Despite extensive training and fine-tuning efforts aimed at enhancing safety, LLMs remain susceptible to jailbreak attacks. Recently, the emergence of multi-turn attacks has intensified this vulnerability. Unlike single-turn attacks, multi-turn attacks incrementally escalate dialogue complexity, rendering them more challenging to detect and mitigate. In this study, we introduce G-Guard, an innovative attention-aware Graph Neural Network (GNN)-based input classifier specifically designed to defend against multi-turn jailbreak attacks targeting LLMs. G-Guard constructs an entity graph for multi-turn queries, which captures the interrelationships between queries and harmful keywords that present in multi-turn queries. Furthermore, we propose an attention-aware augmentation mechanism that retrieves the most relevant single-turn query based on the ongoing multi-turn conversation. The retrieved query is incorporated as a labeled node within the graph, thereby enhancing the GNN's capacity to classify the current query as harmful or benign. Evaluation results show that G-Guard consistently outperforms all baselines across diverse datasets and evaluation metrics, demonstrating its efficacy as a robust defense mechanism against multi-turn jailbreak attacks.
The increasing demand for software development has driven interest in automating software engineering (SE) tasks using Large Language Models (LLMs). Recent efforts extend LLMs into multi-agent systems (MAS) that emulate collaborative development workflows, but these systems often fail due to three core deficiencies: under-specification, coordination misalignment, and inappropriate verification, arising from the absence of foundational SE structuring principles. This paper introduces Software Engineering Multi-Agent Protocol (SEMAP), a protocol-layer methodology that instantiates three core SE design principles for multi-agent LLMs: (1) explicit behavioral contract modeling, (2) structured messaging, and (3) lifecycle-guided execution with verification, and is implemented atop Google's Agent-to-Agent (A2A) infrastructure. Empirical evaluation using the Multi-Agent System Failure Taxonomy (MAST) framework demonstrates that SEMAP effectively reduces failures across different SE tasks. In code development, it achieves up to a 69.6% reduction in total failures for function-level development and 56.7% for deployment-level development. For vulnerability detection, SEMAP reduces failure counts by up to 47.4% on Python tasks and 28.2% on C/C++ tasks.
Large Language Models have been shown to contain extensive world knowledge in their parameters, enabling impressive performance on many knowledge intensive tasks. However, when deployed in novel settings, LLMs often encounter situations where they must integrate parametric knowledge with new or unfamiliar information. In this work, we explore whether LLMs can combine knowledge in-context with their parametric knowledge through the lens of counterfactual reasoning. Through synthetic and real experiments in multi-hop reasoning problems, we show that LLMs generally struggle with counterfactual reasoning, often resorting to exclusively using their parametric knowledge. Moreover, we show that simple post-hoc finetuning can struggle to instill counterfactual reasoning ability -- often leading to degradation in stored parametric knowledge. Ultimately, our work reveals important limitations of current LLM's abilities to re-purpose parametric knowledge in novel settings.
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.
In this study, we present IL3D, a large-scale dataset meticulously designed for large language model (LLM)-driven 3D scene generation, addressing the pressing demand for diverse, high-quality training data in indoor layout design. Comprising 27,816 indoor layouts across 18 prevalent room types and a library of 29,215 high-fidelity 3D object assets, IL3D is enriched with instance-level natural language annotations to support robust multimodal learning for vision-language tasks. We establish rigorous benchmarks to evaluate LLM-driven scene generation. Experimental results show that supervised fine-tuning (SFT) of LLMs on IL3D significantly improves generalization and surpasses the performance of SFT on other datasets. IL3D offers flexible multimodal data export capabilities, including point clouds, 3D bounding boxes, multiview images, depth maps, normal maps, and semantic masks, enabling seamless adaptation to various visual tasks. As a versatile and robust resource, IL3D significantly advances research in 3D scene generation and embodied intelligence, by providing high-fidelity scene data to support environment perception tasks of embodied agents.
Fine-tuning (FT) large language models (LLMs) is crucial for adapting general-purpose models to specific tasks, enhancing accuracy and relevance with minimal resources. To further enhance generalization ability while reducing training costs, this paper proposes Federated LoRA with Dropout (FedLoDrop), a new framework that applies dropout to the rows and columns of the trainable matrix in Federated LoRA. A generalization error bound and convergence analysis under sparsity regularization are obtained, which elucidate the fundamental trade-off between underfitting and overfitting. The error bound reveals that a higher dropout rate increases model sparsity, thereby lowering the upper bound of pointwise hypothesis stability (PHS). While this reduces the gap between empirical and generalization errors, it also incurs a higher empirical error, which, together with the gap, determines the overall generalization error. On the other hand, though dropout reduces communication costs, deploying FedLoDrop at the network edge still faces challenges due to limited network resources. To address this issue, an optimization problem is formulated to minimize the upper bound of the generalization error, by jointly optimizing the dropout rate and resource allocation subject to the latency and per-device energy consumption constraints. To solve this problem, a branch-and-bound (B\&B)-based method is proposed to obtain its globally optimal solution. Moreover, to reduce the high computational complexity of the B\&B-based method, a penalized successive convex approximation (P-SCA)-based algorithm is proposed to efficiently obtain its high-quality suboptimal solution. Finally, numerical results demonstrate the effectiveness of the proposed approach in mitigating overfitting and improving the generalization capability.
Recent studies have revealed the potential of training open-source Large Language Models (LLMs) to unleash LLMs' reasoning ability for enhancing vision-language navigation (VLN) performance, and simultaneously mitigate the domain gap between LLMs' training corpus and the VLN task. However, these approaches predominantly adopt straightforward input-output mapping paradigms, causing the mapping learning difficult and the navigational decisions unexplainable. Chain-of-Thought (CoT) training is a promising way to improve both navigational decision accuracy and interpretability, while the complexity of the navigation task makes the perfect CoT labels unavailable and may lead to overfitting through pure CoT supervised fine-tuning. To address these issues, we propose EvolveNav, a novel sElf-improving embodied reasoning paradigm that realizes adaptable and generalizable navigational reasoning for boosting LLM-based vision-language Navigation. Specifically, EvolveNav involves a two-stage training process: (1) Formalized CoT Supervised Fine-Tuning, where we train the model with curated formalized CoT labels to first activate the model's navigational reasoning capabilities, and simultaneously increase the reasoning speed; (2) Self-Reflective Post-Training, where the model is iteratively trained with its own reasoning outputs as self-enriched CoT labels to enhance the supervision diversity. A self-reflective auxiliary task is also designed to encourage the model to learn correct reasoning patterns by contrasting with wrong ones. Experimental results under both task-specific and cross-task training paradigms demonstrate the consistent superiority of EvolveNav over previous LLM-based VLN approaches on various popular benchmarks, including R2R, REVERIE, CVDN, and SOON. Code is available at https://github.com/expectorlin/EvolveNav.
The proliferation of Large Language Models (LLMs) with exponentially growing parameters is making cross-data center (DC) training an inevitable trend. However, viable strategies for extending single-DC training frameworks to multi-DC environments remain underdeveloped. We experimentally demonstrate, for the first time, a high-performance geo-distributed LLMs training framework across multiple DCs interconnected by a lossless, remote direct memory access (RDMA) enabled Datacenter Optical Transport Network (DC-OTN). An enhanced pipeline parallelism scheme is implemented within the Ascend full-stack environment of Huawei, which effectively eliminates the impact of cross-DC communication overhead on training efficiency. The overlapped computation and cross-DC communication is achieved with constraint cross-DC bandwidth and High Bandwidth Memory (HBM), reducing computation bubble ratio by up to 78.91%.
Effective disaster response is essential for safeguarding lives and property. Existing statistical approaches often lack semantic context, generalize poorly across events, and offer limited interpretability. While Large language models (LLMs) provide few-shot generalization, they remain text-bound and blind to geography. To bridge this gap, we introduce a Geospatial Awareness Layer (GAL) that grounds LLM agents in structured earth data. Starting from raw wildfire detections, GAL automatically retrieves and integrates infrastructure, demographic, terrain, and weather information from external geodatabases, assembling them into a concise, unit-annotated perception script. This enriched context enables agents to produce evidence-based resource-allocation recommendations (e.g., personnel assignments, budget allocations), further reinforced by historical analogs and daily change signals for incremental updates. We evaluate the framework in real wildfire scenarios across multiple LLM models, showing that geospatially grounded agents can outperform baselines. The proposed framework can generalize to other hazards such as floods and hurricanes.
Large language models (LLMs) have rapidly advanced natural language processing, driving significant breakthroughs in tasks such as text generation, machine translation, and domain-specific reasoning. The field now faces a critical dilemma in its approach: closed-source models like GPT-4 deliver state-of-the-art performance but restrict reproducibility, accessibility, and external oversight, while open-source frameworks like LLaMA and Mixtral democratize access, foster collaboration, and support diverse applications, achieving competitive results through techniques like instruction tuning and LoRA. Hybrid approaches address challenges like bias mitigation and resource accessibility by combining the scalability of closed-source systems with the transparency and inclusivity of open-source framework. However, in this position paper, we argue that open-source remains the most robust path for advancing LLM research and ethical deployment.
As large language models (LLMs) take on increasingly complex tasks, understanding their algorithmic reasoning abilities has become essential. However, existing evaluations focus on distinct and isolated tasks. We propose a unified diagnostic lens: structural reasoning--understanding and manipulating relationships like order, hierarchy, and connectivity. We introduce DSR-Bench, the first benchmark to systematically evaluate LLM structural reasoning through canonical data structures, which serve as interpretable, algorithmically meaningful abstractions. DSR-Bench spans 20 data structures, 35 operations, and 4,140 synthetically generated problem instances with minimal contamination. The benchmark's hierarchical design pinpoints specific failure modes, while its fully automated evaluation ensures objective and consistent assessment. Benchmarking ten state-of-the-art LLMs reveals critical limitations: the top-performing model scores only 0.498 out of 1 on challenging instances. Three additional evaluation suites reveal further weaknesses: models perform poorly on spatial data and natural language scenarios, and fail to reason over their own generated code. DSR-Bench offers a principled diagnostic tool for structural reasoning, helping expose reasoning bottlenecks and guide the development of more capable and reliable LLMs.
The current trend in information extraction (IE) is to rely extensively on large language models, effectively discarding decades of experience in building symbolic or statistical IE systems. This paper compares a neuro-symbolic (NS) and an LLM-based IE system in the agricultural domain, evaluating them on nine interviews across pork, dairy, and crop subdomains. The LLM-based system outperforms the NS one (F1 total: 69.4 vs. 52.7; core: 63.0 vs. 47.2), where total includes all extracted information and core focuses on essential details. However, each system has trade-offs: the NS approach offers faster runtime, greater control, and high accuracy in context-free tasks but lacks generalizability, struggles with contextual nuances, and requires significant resources to develop and maintain. The LLM-based system achieves higher performance, faster deployment, and easier maintenance but has slower runtime, limited control, model dependency and hallucination risks. Our findings highlight the "hidden cost" of deploying NLP systems in real-world applications, emphasizing the need to balance performance, efficiency, and control.
Despite their impressive capabilities, aligned large language models (LLMs) often generate outputs that lack diversity. What drives this consistency in the generation? We investigate this phenomenon through the lens of probability concentration in the model's output distribution. To quantify this concentration, we introduce the *Branching Factor* (BF)--a token-invariant measure of the effective number of plausible next steps during generation. Our empirical analysis reveals two key findings: (1) BF often decreases as generation progresses, suggesting that LLMs become more predictable as they generate. (2) alignment tuning substantially sharpens the model's output distribution from the outset, reducing BF by nearly an order of magnitude (e.g., from 12 to 1.2) relative to base models. This stark reduction helps explain why aligned models often appear less sensitive to decoding strategies. Building on this insight, we find this consistency has surprising implications for complex reasoning. Aligned Chain-of-Thought (CoT) models (e.g., DeepSeek-distilled models), for instance, leverage this effect; by generating longer reasoning chains, they push generation into later, more deterministic (lower BF) stages, resulting in more stable outputs. We hypothesize that alignment tuning does not fundamentally change a model's behavior, but instead steers it toward stylistic tokens (e.g., ``Sure'') that unlock low-entropy trajectories already present in the base model. This view is supported by nudging experiments, which show prompting base models with such tokens can similarly reduce BF. Together, our findings establish BF as a powerful diagnostic for understanding and controlling LLM outputs - clarifying how alignment reduces variability, how CoT promotes stable generations, and how base models can be steered away from diversity.
Test-time scaling has enabled Large Language Models (LLMs) with remarkable reasoning capabilities, particularly in mathematical domains, through intermediate chain-of-thought (CoT) reasoning before generating final answers. However, the specific sources and mechanisms underlying these reasoning capabilities remain insufficiently understood. Optimization reasoning, i.e. finding extrema under constraints, represents a fundamental abstraction that underpins critical applications in planning, control, resource allocation, and prompt search. To systematically evaluate this capability, we introduce ExtremBench, a benchmark dataset for solving mathematical extremal problems, curated from inequality exercises used for Chinese Mathematical Olympiad and transformed into $93$ standardized extrema-finding problems. We conduct extensive evaluations across various state-of-the-art open-source model families, including the Qwen3, GPT-OSS, and DeepSeek. Our results reveal that LLMs' extremal-solving reasoning capabilities do not always align with those of current mathematical benchmarks such as AIME25 and MATH-500, with some models showing strong general mathematical reasoning but poor extremal-solving skills, and vice versa. This discrepancy highlights a critical gap in current evaluation practices and suggests that existing benchmarks may not comprehensively capture the full spectrum of mathematical reasoning abilities.
The human-like proficiency of Large Language Models (LLMs) has brought concerns about their potential misuse for generating persuasive and personalised disinformation at scale. While prior work has demonstrated that LLMs can generate disinformation, specific questions around persuasiveness and personalisation (generation of disinformation tailored to specific demographic attributes) remain largely unstudied. This paper presents the first large-scale, multilingual empirical study on persona-targeted disinformation generation by LLMs. Employing a red teaming methodology, we systematically evaluate the robustness of LLM safety mechanisms to persona-targeted prompts. A key novel result is AI-TRAITS (AI-generaTed peRsonAlIsed disinformaTion dataSet), a new dataset of around 1.6 million texts generated by eight state-of-the-art LLMs. AI-TRAITS is seeded by prompts that combine 324 disinformation narratives and 150 distinct persona profiles, covering four major languages (English, Russian, Portuguese, Hindi) and key demographic dimensions (country, generation, political orientation). The resulting personalised narratives are then assessed quantitatively and compared along the dimensions of models, languages, jailbreaking rate, and personalisation attributes. Our findings demonstrate that the use of even simple personalisation strategies in the prompts significantly increases the likelihood of jailbreaks for all studied LLMs. Furthermore, personalised prompts result in altered linguistic and rhetorical patterns and amplify the persuasiveness of the LLM-generated false narratives. These insights expose critical vulnerabilities in current state-of-the-art LLMs and offer a foundation for improving safety alignment and detection strategies in multilingual and cross-demographic contexts.
We present Sentinel, the first framework for formally evaluating the physical safety of Large Language Model(LLM-based) embodied agents across the semantic, plan, and trajectory levels. Unlike prior methods that rely on heuristic rules or subjective LLM judgments, Sentinel grounds practical safety requirements in formal temporal logic (TL) semantics that can precisely specify state invariants, temporal dependencies, and timing constraints. It then employs a multi-level verification pipeline where (i) at the semantic level, intuitive natural language safety requirements are formalized into TL formulas and the LLM agent's understanding of these requirements is probed for alignment with the TL formulas; (ii) at the plan level, high-level action plans and subgoals generated by the LLM agent are verified against the TL formulas to detect unsafe plans before execution; and (iii) at the trajectory level, multiple execution trajectories are merged into a computation tree and efficiently verified against physically-detailed TL specifications for a final safety check. We apply Sentinel in VirtualHome and ALFRED, and formally evaluate multiple LLM-based embodied agents against diverse safety requirements. Our experiments show that by grounding physical safety in temporal logic and applying verification methods across multiple levels, Sentinel provides a rigorous foundation for systematically evaluating LLM-based embodied agents in physical environments, exposing safety violations overlooked by previous methods and offering insights into their failure modes.
Integration of audio perception into large language models (LLMs) is an emerging research area for enabling machine listening applications, yet efficient transfer of rich audio semantics from audio encoders to LLMs remains underexplored. The most widely used integration paradigm projects the audio encoder output tokens into the LLM input space (e.g., via an MLP or a Q-Former), then prepends or inserts them to the text tokens. We refer to this generic scheme as Prepend to the LLM's input token space (PLITS) integration. We propose an efficient alternative, Lightweight Audio LLM Integration (LAL). LAL introduces audio representations solely via the attention mechanism within different layers of the LLM, bypassing its feedforward module. LAL encodes rich audio semantics at an appropriate level of abstraction for integration into different blocks of LLMs. Our design significantly reduces computational overhead compared to existing integration approaches. Observing with Whisper that the speech encoder benefits from PLITS integration, we propose an audio encoder aware approach for efficiently Probing Audio encoders via LLM (PAL), which employs PLITS integration for Whisper and LAL for general audio encoders. Under an identical training curriculum, LAL consistently maintains performance or outperforms existing integration approaches across multiple base LLMs and tasks. For general audio tasks, LAL improvement is up to 30% over a strong PLITS baseline while reducing memory usage by up to 64.1% and increasing throughput by up to 247.5%. Furthermore, for general audio-music-speech LLM, PAL performs on par with a fully PLITS integration-based system but with substantially improved computational and memory efficiency. Project page: https://ta012.github.io/PAL/
Recent advances in Large Language Models (LLMs) have expanded their role in human interaction, yet curiosity -- a central driver of inquiry -- remains underexplored in these systems, particularly across cultural contexts. In this work, we investigate cultural variation in curiosity using Yahoo! Answers, a real-world multi-country dataset spanning diverse topics. We introduce CUEST (CUriosity Evaluation across SocieTies), an evaluation framework that measures human-model alignment in curiosity through linguistic (style), topic preference (content) analysis and grounding insights in social science constructs. Across open- and closed-source models, we find that LLMs flatten cross-cultural diversity, aligning more closely with how curiosity is expressed in Western countries. We then explore fine-tuning strategies to induce curiosity in LLMs, narrowing the human-model alignment gap by up to 50\%. Finally, we demonstrate the practical value of curiosity for LLM adaptability across cultures, showing its importance for future NLP research.
Large Language Models (LLMs) should answer factual questions truthfully, grounded in objective knowledge, regardless of user context such as self-disclosed personal information, or system personalization. In this paper, we present the first systematic evaluation of LLM robustness to inquiry personas, i.e. user profiles that convey attributes like identity, expertise, or belief. While prior work has primarily focused on adversarial inputs or distractors for robustness testing, we evaluate plausible, human-centered inquiry persona cues that users disclose in real-world interactions. We find that such cues can meaningfully alter QA accuracy and trigger failure modes such as refusals, hallucinated limitations, and role confusion. These effects highlight how model sensitivity to user framing can compromise factual reliability, and position inquiry persona testing as an effective tool for robustness evaluation.
Monitoring Large Language Model (LLM) outputs is crucial for mitigating risks from misuse and misalignment. However, LLMs could evade monitoring through steganography: Encoding hidden information within seemingly benign generations. In this paper, we evaluate the steganography capabilities in frontier LLMs to better understand the risk they pose. We focus on two types of steganography: passing encoded messages and performing encoded reasoning. We find that current models are unable to encode short messages in their outputs without a monitor noticing under standard affordances. They can succeed, however, if given additional affordances like using an unmonitored scratchpad and coordinating on what encoding scheme to use. We additionally find early signs that models can perform basic encoded reasoning in a simple state-tracking problem. This includes some ability to reason with their own and pre-defined schemes, including encoding schemes such as Hexadecimal. Despite this, they can rarely hide reasoning subtly within a cover task to fool a monitor. Overall, our results indicate that current LLMs exhibit nascent steganographic capabilities. While these capabilities are likely insufficient to bypass well-designed monitors at present, this could change in the future.
Integration of audio perception into large language models (LLMs) is an emerging research area for enabling machine listening applications, yet efficient transfer of rich audio semantics from audio encoders to LLMs remains underexplored. The most widely used integration paradigm projects the audio encoder output tokens into the LLM input space (e.g., via an MLP or a Q-Former), then prepends or inserts them to the text tokens. We refer to this generic scheme as Prepend to the LLM's input token space (PLITS) integration. We propose an efficient alternative, Lightweight Audio LLM Integration (LAL). LAL introduces audio representations solely via the attention mechanism within different layers of the LLM, bypassing its feedforward module. LAL encodes rich audio semantics at an appropriate level of abstraction for integration into different blocks of LLMs. Our design significantly reduces computational overhead compared to existing integration approaches. Observing with Whisper that the speech encoder benefits from PLITS integration, we propose an audio encoder aware approach for efficiently Probing Audio encoders via LLM (PAL), which employs PLITS integration for Whisper and LAL for general audio encoders. Under an identical training curriculum, LAL consistently maintains performance or outperforms existing integration approaches across multiple base LLMs and tasks. For general audio tasks, LAL improvement is up to 30% over a strong PLITS baseline while reducing memory usage by up to 64.1% and increasing throughput by up to 247.5%. Furthermore, for general audio-music-speech LLM, PAL performs on par with a fully PLITS integration-based system but with substantially improved computational and memory efficiency. Project page: https://ta012.github.io/PAL/
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.
Current approaches for strengthening LLM reasoning tend to introduce a training bias toward human-like reasoning trajectories. In step-wise preference optimization, in particular, dependence on human or higher-capacity model annotations for intermediate steps limits exploration of alternative, non-human-like reasoning paths and thus constrains achievable performance. Furthermore, through a small-scale pilot study, we observed that in approximately 75% of cases, the model's first erroneous step occurs after the lowest-confidence point. This suggests that guiding the model at its lowest-confidence point before an error provides more accurate supervision than locating the first explicit error. In this paper, we propose Confidence-Guided Reasoning Path Preference Optimization (CGPO), a method that leverages a confidence signal to identify points of maximal uncertainty in the model's reasoning process and applies self-generated, non-human-like reasoning-path guidance to mitigate trajectory drift. Our experiments span diverse models applied to both code and mathematical reasoning tasks. The results show that, with the same amount of training data, our method using data generated by a small model can achieve better performance in most cases compared with approaches using data generated by a strong model or human-annotated.
In programming education, Debugging and Teaching (DT) task is a common scenario where students receive assistance in correcting their erroneous code. The task involves multiple inputs, including erroneous code, error messages, reference solutions, and the question description, with the goal of generating modification suggestions to the erroneous code. However, two key challenges hinder the effectiveness of existing approaches. Firstly, the complexity and heterogeneity of inputs inherent in DT tasks significantly elevate the reasoning challenges faced by LLMs. Second, existing approaches often fail to fully leverage the availability of standard code in DT tasks, forcing models to rely solely on complex multi-step reasoning, which limits the potential of LLMs in addressing DT tasks effectively. To address these challenges, we propose DebugTA, a novel LLM-based debugging and teaching agent with specialized tools for standard code retrieval, variable substitution to align reference code, and an external compiler for real-time code analysis. Guided by explicit pedagogical and debugging principles, DebugTA acts as an agent that decomposes a complex task into sequential LLM interactions, each utilizing distinct tools for specific subtasks, thereby simplifying the logical reasoning at each step and reducing overall reasoning complexity. Furthermore, DebugTA utilizes tool calls to align the standard code with the erroneous code as much as possible, allowing the LLM to focus on logic errors within the erroneous code and improving the accuracy of the generated suggestions. To rigorously assess the quality of modification suggestions, we introduce a student simulator-teacher interaction paradigm. Experimental results on three real-world code datasets demonstrate that DebugTA consistently improves teaching effectiveness while significantly reducing computational costs.
Large Language Models (LLMs) demonstrate their reasoning ability through chain-of-thought (CoT) generation. However, LLM's autoregressive decoding may limit the ability to revisit and refine earlier tokens in a holistic manner, which can also lead to inefficient exploration for diverse solutions. In this paper, we propose LaDiR (Latent Diffusion Reasoner), a novel reasoning framework that unifies the expressiveness of continuous latent representation with the iterative refinement capabilities of latent diffusion models for an existing LLM. We first construct a structured latent reasoning space using a Variational Autoencoder (VAE) that encodes text reasoning steps into blocks of thought tokens, preserving semantic information and interpretability while offering compact but expressive representations. Subsequently, we utilize a latent diffusion model that learns to denoise a block of latent thought tokens with a blockwise bidirectional attention mask, enabling longer horizon and iterative refinement with adaptive test-time compute. This design allows efficient parallel generation of diverse reasoning trajectories, allowing the model to plan and revise the reasoning process holistically. We conduct evaluations on a suite of mathematical reasoning and planning benchmarks. Empirical results show that LaDiR consistently improves accuracy, diversity, and interpretability over existing autoregressive, diffusion-based, and latent reasoning methods, revealing a new paradigm for text reasoning with latent diffusion.
Evaluating the factuality of long-form generations from Large Language Models (LLMs) remains challenging due to accuracy issues and costly human assessment. Prior efforts attempt this by decomposing text into claims, searching for evidence, and verifying claims, but suffer from critical drawbacks: (1) inefficiency due to complex pipeline components unsuitable for long LLM outputs, and (2) ineffectiveness stemming from inaccurate claim sets and insufficient evidence collection of one-line snippets. To address these limitations, we propose \name, a fast and strong evaluation framework that achieves the highest alignment with human evaluation and efficiency among existing baselines. \name first employs chunk-level claim extraction integrated with confidence-based pre-verification, significantly reducing the cost of web searching and inference calling while ensuring reliability. For searching and verification, it collects document-level evidence from crawled webpages and selectively retrieves it during verification, addressing the evidence insufficiency problem in previous pipelines. Extensive experiments based on an aggregated and manually annotated benchmark demonstrate the reliability of \name in both efficiently and effectively evaluating the factuality of long-form LLM generations. Code and benchmark data is available at https://github.com/Yingjia-Wan/FastFact.
BanglaMATH : A Bangla benchmark dataset for testing LLM mathematical reasoning at grades 6, 7, and 8
Large Language Models (LLMs) have tremendous potential to play a key role in supporting mathematical reasoning, with growing use in education and AI research. However, most existing benchmarks are limited to English, creating a significant gap for low-resource languages. For example, Bangla is spoken by nearly 250 million people who would collectively benefit from LLMs capable of native fluency. To address this, we present BanglaMATH, a dataset of 1.7k Bangla math word problems across topics such as Arithmetic, Algebra, Geometry, and Logical Reasoning, sourced from Bangla elementary school workbooks and annotated with details like grade level and number of reasoning steps. We have designed BanglaMATH to evaluate the mathematical capabilities of both commercial and open-source LLMs in Bangla, and we find that Gemini 2.5 Flash and DeepSeek V3 are the only models to achieve strong performance, with $\ge$ 80\% accuracy across three elementary school grades. Furthermore, we assess the robustness and language bias of these top-performing LLMs by augmenting the original problems with distracting information, and translating the problems into English. We show that both LLMs fail to maintain robustness and exhibit significant performance bias in Bangla. Our study underlines current limitations of LLMs in handling arithmetic and mathematical reasoning in low-resource languages, and highlights the need for further research on multilingual and equitable mathematical understanding. Dataset link: \href{https://github.com/TabiaTanzin/BanglaMATH-A-Bangla-benchmark-dataset-for-testing-LLM-mathematical-reasoning-at-grades-6-7-and-8.git}{https://github.com/BanglaMATH}
This study investigates how existing annotation guidelines can be repurposed to instruct large language model (LLM) annotators for text annotation tasks. Traditional guidelines are written for human annotators who internalize training, while LLMs require explicit, structured instructions. We propose a moderation-oriented guideline repurposing method that transforms guidelines into clear directives for LLMs through an LLM moderation process. Using the NCBI Disease Corpus as a case study, our experiments show that repurposed guidelines can effectively guide LLM annotators, while revealing several practical challenges. The results highlight the potential of this workflow to support scalable and cost-effective refinement of annotation guidelines and automated annotation.
Large Language Models (LLMs) have achieved remarkable results on a variety of mathematical benchmarks. However, concerns remain as to whether these successes reflect genuine reasoning or superficial pattern recognition. Common evaluation methods, which focus on the either the final answer or the reasoning process, fail to assess the entire problem-solving procedure. To address these limitations, we introduce SMART: a Self-Generating and Self-Validating Multi-Dimensional Assessment Framework, together with its corresponding benchmark, SMART-Bench. SMART decomposes the entire problem solving process into four distinct cognitive dimensions: Understanding, Reasoning, Arithmetic, and Reflection \& Refinement. Each dimension is evaluated independently through tailored tasks, enabling interpretable and fine-grained analysis of LLM behavior. We apply SMART to 21 state-of-the-art open- and closed-source LLMs, uncovering significant discrepancies in their abilities across different dimensions. Our findings reveal genuine weaknesses in current LLMs and motivate a new metric, the All-Pass Score, to better capture true problem-solving capabilities. Code and benchmarks will be released upon acceptance.
Multi-agent systems (MAS) and reinforcement learning (RL) are widely used to enhance the agentic capabilities of large language models (LLMs). MAS improves task performance through role-based orchestration, while RL uses environmental rewards to learn stronger policies, such as GRPO-style optimization. However, applying on-policy RL to MAS remains underexplored and presents unique challenges. Algorithmically, standard GRPO grouping assumptions break down because prompts vary by role and by turn. System-wise, the training stack must support MAS-workflow rollouts and on-policy updates for both single-policy and multi-policy models. We propose AT-GRPO, which includes (i) an agent- and turn-wise grouped RL algorithm tailored to MAS and (ii) a training system that supports both single- and multi-policy regimes. Across game, planning, coding, and math tasks, AT-GRPO delivers substantial gains. On long-horizon planning, it increases accuracy from a 14.0 to 47.0 percent single-agent RL baseline to 96.0 to 99.5 percent. It also improves reasoning performance, with average gains of 3.87 to 7.62 percent on coding tasks and 9.0 to 17.93 percent on math. Code and environments are available at: https://github.com/pettingllms-ai/PettingLLMs.
Query-service relevance prediction in e-commerce search systems faces strict latency requirements that prevent the direct application of Large Language Models (LLMs). To bridge this gap, we propose a two-stage reasoning distillation framework to transfer reasoning capabilities from a powerful teacher LLM to a lightweight, deployment-friendly student model. In the first stage, we address the limitations of general-purpose LLMs by constructing a domain-adapted teacher model. This is achieved through a three-step process: domain-adaptive pre-training to inject platform knowledge, supervised fine-tuning to elicit reasoning skills, and preference optimization with a multi-dimensional reward model to ensure the generation of reliable and preference-aligned reasoning paths. This teacher can then automatically annotate massive query-service pairs from search logs with both relevance labels and reasoning chains. In the second stage, to address the challenges of architectural heterogeneity in standard distillation, we introduce Contrastive Reasoning Self-Distillation (CRSD). By modeling the behavior of the same student model under "standard" and "reasoning-augmented" inputs as a teacher-student relationship, CRSD enables the lightweight model to internalize the teacher's complex decision-making mechanisms without needing the explicit reasoning path at inference. Offline evaluations and online A/B testing in the Meituan search advertising system demonstrate that our framework achieves significant improvements across multiple metrics, validating its effectiveness and practical value.
The rise of large language models (LLMs) has transformed healthcare by offering clinical guidance, yet their direct deployment to patients poses safety risks due to limited domain expertise. To mitigate this, we propose repositioning LLMs as clinical assistants that collaborate with experienced physicians rather than interacting with patients directly. We conduct a two-stage inspiration-feedback survey to identify real-world needs in clinical workflows. Guided by this, we construct DoctorFLAN, a large-scale Chinese medical dataset comprising 92,000 Q&A instances across 22 clinical tasks and 27 specialties. To evaluate model performance in doctor-facing applications, we introduce DoctorFLAN-test (550 single-turn Q&A items) and DotaBench (74 multi-turn conversations). Experimental results with over ten popular LLMs demonstrate that DoctorFLAN notably improves the performance of open-source LLMs in medical contexts, facilitating their alignment with physician workflows and complementing existing patient-oriented models. This work contributes a valuable resource and framework for advancing doctor-centered medical LLM development
As LLM-based agents become increasingly prevalent, backdoors can be implanted into agents through user queries or environment feedback, raising critical concerns regarding safety vulnerabilities. However, backdoor attacks are typically detectable by safety audits that analyze the reasoning process of agents. To this end, we propose a novel backdoor implantation strategy called \textbf{Dynamically Encrypted Multi-Backdoor Implantation Attack}. Specifically, we introduce dynamic encryption, which maps the backdoor into benign content, effectively circumventing safety audits. To enhance stealthiness, we further decompose the backdoor into multiple sub-backdoor fragments. Based on these advancements, backdoors are allowed to bypass safety audits significantly. Additionally, we present AgentBackdoorEval, a dataset designed for the comprehensive evaluation of agent backdoor attacks. Experimental results across multiple datasets demonstrate that our method achieves an attack success rate nearing 100\% while maintaining a detection rate of 0\%, illustrating its effectiveness in evading safety audits. Our findings highlight the limitations of existing safety mechanisms in detecting advanced attacks, underscoring the urgent need for more robust defenses against backdoor threats. Code and data are available at https://github.com/whfeLingYu/DemonAgent.
Large language models (LLMs) increasingly generate natural language rationales to enhance interpretability, but these often contain logical errors, label mismatches, and domain-specific misalignments. Directly using such rationales as supervision risks propagating noise and undermining training stability. To address this challenge, we introduce Self-Filtered Distillation, a framework specifically tailored for patent classification, which treats LLM-generated rationales as trust signals rather than ground-truth supervision. The framework employs selective distillation guided by three unsupervised trust metrics: (1) Self-Consistency, which measures the stability of LLM-generated rationales across multiple generations; (2) Class Entailment Alignment, which assesses semantic coherence with patent-specific class definitions; and (3) LLM Agreement Scoring, which validates rationale-label plausibility. These metrics are integrated into a unified trust score that primarily weights training samples while optionally filtering out extremely low-trust cases, enabling reasoning-aware supervision. Experiments on the USPTO-2M dataset, a widely used benchmark for patent classification, show that our method outperforms label-based learning and conventional distillation in accuracy, stability, and interpretability, establishing a reliable paradigm for leveraging reasoning-aware trust indicators in patent analytics.
Large Language Models (LLMs) are now widely used across many domains. With their rapid development, Reinforcement Learning with Verifiable Rewards (RLVR) has surged in recent months to enhance their reasoning and understanding abilities. However, its complex data flows and diverse tasks pose substantial challenges to RL training systems, and there is limited understanding of RLVR from a system perspective. To thoroughly understand the system challenges introduced by RLVR, we present a characterization study of RLVR tasks in our LLM deployment. Specifically, we investigate the distribution and variation trends of workloads across different RL tasks across training steps. We identify issues such as GPU idling caused by skewed sequence length distribution, inefficient parallel strategies in dynamically varying workloads, inefficient data management mechanisms, and load imbalance. We describe our observations and call for further investigation into the remaining open challenges. Furthermore, we propose PolyTrace benchmark suite to conduct evaluation with realistic workloads, and a practical use case validates that PolyTrace benchmark suite exhibits 94.7% accuracy.
Large language models (LLMs) are increasingly under scrutiny for perpetuating identity-based discrimination in high-stakes domains such as hiring, particularly against people with disabilities (PwD). However, existing research remains largely Western-centric, overlooking how intersecting forms of marginalization--such as gender and caste--shape experiences of PwD in the Global South. We conduct a comprehensive audit of six LLMs across 2,820 hiring scenarios spanning diverse disability, gender, nationality, and caste profiles. To capture subtle intersectional harms and biases, we introduce ABLEIST (Ableism, Inspiration, Superhumanization, and Tokenism), a set of five ableism-specific and three intersectional harm metrics grounded in disability studies literature. Our results reveal significant increases in ABLEIST harms towards disabled candidates--harms that many state-of-the-art models failed to detect. These harms were further amplified by sharp increases in intersectional harms (e.g., Tokenism) for gender and caste-marginalized disabled candidates, highlighting critical blind spots in current safety tools and the need for intersectional safety evaluations of frontier models in high-stakes domains like hiring.
The promise of LLM watermarking rests on a core assumption that a specific watermark proves authorship by a specific model. We demonstrate that this assumption is dangerously flawed. We introduce the threat of watermark spoofing, a sophisticated attack that allows a malicious model to generate text containing the authentic-looking watermark of a trusted, victim model. This enables the seamless misattribution of harmful content, such as disinformation, to reputable sources. The key to our attack is repurposing watermark radioactivity, the unintended inheritance of data patterns during fine-tuning, from a discoverable trait into an attack vector. By distilling knowledge from a watermarked teacher model, our framework allows an attacker to steal and replicate the watermarking signal of the victim model. This work reveals a critical security gap in text authorship verification and calls for a paradigm shift towards technologies capable of distinguishing authentic watermarks from expertly imitated ones. Our code is available at https://github.com/hsannn/ditto.git.
Large language models (LLMs), owing to their extensive open-domain knowledge and semantic reasoning capabilities, have been increasingly integrated into recommender systems (RS). However, a substantial gap remains between the pre-training objectives of LLMs and the specific requirements of recommendation tasks. To address this gap, supervised fine-tuning (SFT) is commonly performed on specially curated recommendation datasets to further enhance their predictive ability. Despite its success, SFT exhibits a critical limitation: it induces Language Bias, whereby the model over-relies on auxiliary tokens-such as task descriptions and prefix-generated tokens-while underutilizing core user interaction tokens that encode user-specific preferences. This bias not only undermines recommendation accuracy but also raises unfairness concerns. To address this issue, we propose Group Distributionally Robust Optimization-based Tuning (GDRT), a novel fine-tuning paradigm that enforces consistent model performance across token groups with varying degrees of relevance to auxiliary tokens. By adaptively upweighting underperforming groups, typically those weakly correlated with auxiliary tokens, GDRT shifts the model's attention from superficial auxiliary cues to informative user interaction tokens, thereby mitigating language bias. Extensive experiments conducted on three public datasets demonstrate that GDRT effectively mitigates language bias, yielding substantial improvements in recommendation accuracy (with an average NDCG@10 gain of 24.29%) and significantly enhancing recommendation fairness.
The advancement of LLMs and their accessibility have triggered renewed interest in multi-agent reinforcement learning as robust and adaptive frameworks for dynamically changing environments. This paper introduces RL-Focal, a two-stage RL agent framework that routes and ensembles LLMs. First, we develop the Decider RL-agent, which learns to dynamically select an ensemble of small size ($m_i$) among $N$ LLMs ($m_i \ll N$) for incoming queries from a user-defined downstream task $i$, by maximizing both error-diversity and reasoning-performance of the selected ensemble through iterative updates of task-adaptive rewards and policy. Second, to enable effective fusion of dynamically selected LLMs, we develop the stage-2 Fusion RL-agent, which learns to resolve reasoning conflicts from different LLMs and dynamically adapts to different ensemble teams composed by the Decider Agent for different downstream tasks. Third, we introduce the focal diversity metric to better model the error correlations among multiple LLMs, further improving the generalization performance of the Decider Agent, which actively prunes the ensemble combinations. By focal diversity, we enhance performance across tasks by effectively promoting reward-aware and policy-adaptive ensemble selection and inference fusion. Extensive evaluations on five benchmarks show that RL-Focal achieves the performance improvement of 8.48\% with an ensemble of small size compared to the best individual LLM in a pool and offers stronger robustness. Code is available at https://github.com/sftekin/rl-focal
Reasoning ability has become a defining capability of Large Language Models (LLMs), with Reinforcement Learning with Verifiable Rewards (RLVR) emerging as a key paradigm to enhance it. However, RLVR training often suffers from policy entropy collapse, where the policy becomes overly deterministic, hindering exploration and limiting reasoning performance. While entropy regularization is a common remedy, its effectiveness is highly sensitive to the fixed coefficient, making it unstable across tasks and models. In this work, we revisit entropy regularization in RLVR and argue that its potential has been largely underestimated. Our analysis shows that (i) tasks of varying difficulty demand distinct exploration intensities, and (ii) balanced exploration may require the policy entropy to be maintained within a moderate range below its initial level. Therefore, we propose Adaptive Entropy Regularization (AER)--a framework that dynamically balances exploration and exploitation via three components: difficulty-aware coefficient allocation, initial-anchored target entropy, and dynamic global coefficient adjustment. Experiments on multiple mathematical reasoning benchmarks show that AER consistently outperforms baselines, improving both reasoning accuracy and exploration capability.
Recent years have witnessed a surge of research on leveraging large language models (LLMs) for sequential recommendation. LLMs have demonstrated remarkable potential in inferring users' nuanced preferences through fine-grained semantic reasoning. However, they also exhibit a notable limitation in effectively modeling collaborative signals, i.e., behavioral correlations inherent in users' historical interactions. Our empirical analysis further reveals that the attention mechanisms in LLMs tend to disproportionately focus on tokens within the same item, thereby impeding the capture of cross-item correlations. To address this limitation, we propose a novel hierarchical attention masking strategy for LLM-based recommendation, termed HatLLM. Specifically, in shallow layers, HatLLM masks attention between tokens from different items, facilitating intra-item semantic understanding; in contrast, in deep layers, HatLLM masks attention within items, thereby compelling the model to capture cross-item correlations. This progressive, layer-wise approach enables LLMs to jointly model both token-level and item-level dependencies. Extensive experiments on three real-world datasets demonstrate that HatLLM achieves significant performance gains (9.13% on average) over existing LLM-based methods.
Accurately predicting conversion rates (CVR) for low-activity users remains a fundamental challenge in large-scale e-commerce recommender systems. Existing approaches face three critical limitations: (i) reliance on noisy and unreliable behavioral signals; (ii) insufficient user-level information due to the lack of diverse interaction data; and (iii) a systemic training bias toward high-activity users that overshadows the needs of low-activity users. To address these challenges, we propose ChoirRec, a novel framework that leverages the semantic capabilities of Large Language Models (LLMs) to construct semantic user groups and enhance CVR prediction for low-activity users. With a dual-channel architecture designed for robust cross-user knowledge transfer, ChoirRec comprises three components: (i) a Semantic Group Generation module that utilizes LLMs to form reliable, cross-activity user clusters, thereby filtering out noisy signals; (ii) a Group-aware Hierarchical Representation module that enriches sparse user embeddings with informative group-level priors to mitigate data insufficiency; and (iii) a Group-aware Multi-granularity Modual that employs a dual-channel architecture and adaptive fusion mechanism to ensure effective learning and utilization of group knowledge. We conduct extensive offline and online experiments on Taobao, a leading industrial-scale e-commerce platform. ChoirRec improves GAUC by 1.16\% in offline evaluations, while online A/B testing reveals a 7.24\% increase in order volume, highlighting its substantial practical value in real-world applications.
The growth of the e-commerce industry has intensified the adversarial dynamics between shadow economy actors and risk management teams. Companies often conduct risk investigations into suspicious cases to identify emerging fraud patterns, thereby enhancing both preemptive risk prevention and post-hoc governance. However, the sheer volume of case analyses imposes a substantial workload on risk management analysts, as each case requires the integration of long-term expert experience and meticulous scrutiny across multiple risk dimensions. Additionally, individual disparities among analysts hinder the establishment of uniform and high-standard workflows. To address these challenges, we propose the SHERLOCK framework, which leverages the reasoning capabilities of large language models (LLMs) to assist analysts in risk investigations. Our approach consists of three primary components: (1) extracting risk management knowledge from multi-modal data and constructing a domain knowledge base (KB), (2) building an intelligent platform guided by the data flywheel paradigm that integrates daily operations, expert annotations, and model evaluations, with iteratively fine-tuning for preference alignment, and (3) introducing a Reflect & Refine (R&R) module that collaborates with the domain KB to establish a rapid response mechanism for evolving risk patterns. Experiments conducted on the real-world transaction dataset from JD dot com demonstrate that our method significantly improves the precision of both factual alignment and risk localization within the LLM analysis results. Deployment of the SHERLOCK-based LLM system on JD dot com has substantially enhanced the efficiency of case investigation workflows for risk managers.
WebShell attacks, where malicious scripts are injected into web servers, pose a significant cybersecurity threat. Traditional ML and DL methods are often hampered by challenges such as the need for extensive training data, catastrophic forgetting, and poor generalization. Recently, Large Language Models have emerged as powerful alternatives for code-related tasks, but their potential in WebShell detection remains underexplored. In this paper, we make two contributions: (1) a comprehensive evaluation of seven LLMs, including GPT-4, LLaMA 3.1 70B, and Qwen 2.5 variants, benchmarked against traditional sequence- and graph-based methods using a dataset of 26.59K PHP scripts, and (2) the Behavioral Function-Aware Detection (BFAD) framework, designed to address the specific challenges of applying LLMs to this domain. Our framework integrates three components: a Critical Function Filter that isolates malicious PHP function calls, a Context-Aware Code Extraction strategy that captures the most behaviorally indicative code segments, and Weighted Behavioral Function Profiling that enhances in-context learning by prioritizing the most relevant demonstrations based on discriminative function-level profiles. Our results show that, stemming from their distinct analytical strategies, larger LLMs achieve near-perfect precision but lower recall, while smaller models exhibit the opposite trade-off. However, all baseline models lag behind previous SOTA methods. With the application of BFAD, the performance of all LLMs improves significantly, yielding an average F1 score increase of 13.82%. Notably, larger models now outperform SOTA benchmarks, while smaller models such as Qwen-2.5-Coder-3B achieve performance competitive with traditional methods. This work is the first to explore the feasibility and limitations of LLMs for WebShell detection and provides solutions to address the challenges in this task.
Recent advances in automated theorem proving use Large Language Models (LLMs) to translate informal mathematical statements into formal proofs. However, informal cues are often ambiguous or lack strict logical structure, making it hard for models to interpret them precisely. While existing methods achieve strong performance, little is known about how LLMs internally represent informal cues, or how these influence proof generation. To address this, we explore \textit{activation steering}, an inference-time intervention that identifies linear directions in residual activations associated with informal reasoning traces and adjusts them to improve proof construction without fine-tuning. This mechanism also yields interpretable information about how reasoning is internally encoded in the activation space of LLMs. We test our method for generating formal proofs from already-formalized theorems. Our contributions are twofold: (1) a novel activation-based intervention for guiding proof synthesis in LLMs; and (2) demonstration that this intervention improves performance under two decoding strategies (sampling and best-first search) without any further training.
Medium Access Control (MAC) protocols, essential for wireless networks, are typically manually configured. While deep reinforcement learning (DRL)-based protocols enhance task-specified network performance, they suffer from poor generalizability and resilience, demanding costly retraining to adapt to dynamic environments. To overcome this limitation, we introduce a game-theoretic LLM-empowered multi-agent DRL (MARL) framework, in which the uplink transmission between a base station and a varying number of user equipments is modeled as a dynamic multi-follower Stackelberg game (MFSG), capturing the network's natural hierarchical structure. Within this game, LLM-driven agents, coordinated through proximal policy optimization (PPO), synthesize adaptive, semantic MAC protocols in response to network dynamics. Protocol action grammar (PAG) is employed to ensure the reliability and efficiency of this process. Under this system, we further analyze the existence and convergence behavior in terms of a Stackelberg equilibrium by studying the learning dynamics of LLM-empowered unified policies in response to changing followers. Simulations corroborate that our framework achieves a 77.6% greater throughput and a 65.2% fairness improvement over conventional baselines. Besides, our framework generalizes excellently to a fluctuating number of users without requiring retraining or architectural changes.
Training Long-Context Large Language Models (LLMs) is challenging, as hybrid training with long-context and short-context data often leads to workload imbalances. Existing works mainly use data packing to alleviate this issue, but fail to consider imbalanced attention computation and wasted communication overhead. This paper proposes Hierarchical Balance Packing (HBP), which designs a novel batch-construction method and training recipe to address those inefficiencies. In particular, the HBP constructs multi-level data packing groups, each optimized with a distinct packing length. It assigns training samples to their optimal groups and configures each group with the most effective settings, including sequential parallelism degree and gradient checkpointing configuration. To effectively utilize multi-level groups of data, we design a dynamic training pipeline specifically tailored to HBP, including curriculum learning, adaptive sequential parallelism, and stable loss. Our extensive experiments demonstrate that our method significantly reduces training time over multiple datasets and open-source models while maintaining strong performance. For the largest DeepSeek-V2 (236B) MoE model, our method speeds up the training by 2.4$\times$ with competitive performance. Codes will be released at https://github.com/ModelTC/HBP.
We introduce LLM x MapReduce-V3, a hierarchically modular agent system designed for long-form survey generation. Building on the prior work, LLM x MapReduce-V2, this version incorporates a multi-agent architecture where individual functional components, such as skeleton initialization, digest construction, and skeleton refinement, are implemented as independent model-context-protocol (MCP) servers. These atomic servers can be aggregated into higher-level servers, creating a hierarchically structured system. A high-level planner agent dynamically orchestrates the workflow by selecting appropriate modules based on their MCP tool descriptions and the execution history. This modular decomposition facilitates human-in-the-loop intervention, affording users greater control and customization over the research process. Through a multi-turn interaction, the system precisely captures the intended research perspectives to generate a comprehensive skeleton, which is then developed into an in-depth survey. Human evaluations demonstrate that our system surpasses representative baselines in both content depth and length, highlighting the strength of MCP-based modular planning.
Model merging provides a scalable alternative to multi-task training by combining specialized finetuned models through parameter arithmetic, enabling efficient deployment without the need for joint training or access to all task data. While recent methods have shown promise, existing evaluations are limited in both model scale and task diversity, leaving open questions about their applicability to large, domain-specialized LLMs. To tackle the challenges, we introduce MergeBench, a comprehensive evaluation suite designed to assess model merging at scale. MergeBench builds on state-of-the-art open-source language models, including Llama and Gemma families at 2B to 9B scales, and covers five key domains: instruction following, mathematics, multilingual understanding, coding and safety. We standardize finetuning and evaluation protocols, and assess eight representative merging methods across multi-task performance, forgetting and runtime efficiency. Based on extensive experiments, we provide practical guidelines for algorithm selection and share insights showing that model merging tends to perform better on stronger base models, with techniques such as merging coefficient tuning and sparsification improving knowledge retention. However, several challenges remain, including the computational cost on large models, the gap for in-domain performance compared to multi-task models, and the underexplored role of model merging in standard LLM training pipelines. We hope MergeBench provides a foundation for future research to advance the understanding and practical application of model merging. Our project page is at \href{https://yifei-he.github.io/mergebench/}{https://yifei-he.github.io/mergebench/}.
Application profiling is an indispensable technique for many software development tasks, such as code and memory layout optimizations, where optimization decisions are tailored to specific program profiles. Unfortunately, modern application codebases exhibit highly variant behavior across different inputs, creating challenges for conventional profiling approaches that rely on a single representative execution instance. In this paper, we propose \textbf{Phaedrus}, a new \textit{compiler-assisted deep learning framework} designed to predict dynamic program behaviors across varied execution instances, specifically focusing on dynamic function call prediction.Such predicted call sequences are then used for producing optimized code pertinent to a given input. Traditional profile-guided optimization methods struggle with the input-dependent variability of modern applications, where profiling on different inputs yields divergent application behaviors. To address this, Phaedrus proposes two new approaches: \textit{Application Behavior Synthesis}, a profile-less approach where Large Language Models (LLMs) directly infer dynamic functions based on source code \& static compiler analysis, bypassing the need for traditional profiling, and \textit{Application Profile Generalization}, which uses generative models trained on compressed and augmented \textit{Whole Program Path} (WPP) based function profiles to predict application behavior under unseen inputs. Our experiments show that \textit{Phaedrus} can achieve upto $10^7X$ reduction in WPP function profile sizes, can predict most frequently executed functions that cover upto 85-99\% of the execution time, along with an average of 13.19\% (upto 65\%) reduction in application binary size, and an average of 6.08\% (upto 20\%) performance improvement over the traditional profile-guided optimization, without any execution.
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.
Large Language Models (LLMs) are transforming software engineering tasks, including code vulnerability detection-a critical area of software security. However, existing methods often rely on resource-intensive models or graph-based techniques, limiting their accessibility and practicality. This paper introduces K-ASTRO, a lightweight Transformer model that combines semantic embeddings from LLMs with structural features of Abstract Syntax Trees (ASTs) to improve both efficiency and accuracy in code vulnerability detection. Our approach introduces an AST-based augmentation technique inspired by mutation testing, a structure-aware attention mechanism that incorporates augmented AST features, and a joint adaptation pipeline to unify code semantics and syntax. Experimental results on three large-scale datasets, including BigVul, DiverseVul, and PrimeVul-demonstrate state-of-the-art performance while enabling rapid inference on CPUs with minimal training time. By offering a scalable, interpretable, and efficient solution, K-ASTRO bridges the gap between LLM advancements and practical software vulnerability detection, providing open-sourced tools to foster further research.
Cyber threat intelligence (CTI) is central to modern cybersecurity, providing critical insights for detecting and mitigating evolving threats. With the natural language understanding and reasoning capabilities of large language models (LLMs), there is increasing interest in applying them to CTI, which calls for benchmarks that can rigorously evaluate their performance. Several early efforts have studied LLMs on some CTI tasks but remain limited: (i) they adopt only closed-book settings, relying on parametric knowledge without leveraging CTI knowledge bases; (ii) they cover only a narrow set of tasks, lacking a systematic view of the CTI landscape; and (iii) they restrict evaluation to single-source analysis, unlike realistic scenarios that require reasoning across multiple sources. To fill these gaps, we present CTIArena, the first benchmark for evaluating LLM performance on heterogeneous, multi-source CTI under knowledge-augmented settings. CTIArena spans three categories, structured, unstructured, and hybrid, further divided into nine tasks that capture the breadth of CTI analysis in modern security operations. We evaluate ten widely used LLMs and find that most struggle in closed-book setups but show noticeable gains when augmented with security-specific knowledge through our designed retrieval-augmented techniques. These findings highlight the limitations of general-purpose LLMs and the need for domain-tailored techniques to fully unlock their potential for CTI.
Large language model (LLM) agents are fundamentally constrained by context length on long-horizon tasks. We introduce Context-Folding, a framework that empowers agents to actively manage their working context. An agent can procedurally branch into a sub-trajectory to handle a subtask and then fold it upon completion, collapsing the intermediate steps while retaining a concise summary of the outcome. To make this behavior learnable, we develop an end-to-end reinforcement learning framework FoldGRPO with specific process rewards to encourage effective task decomposition and context management. On complex long-horizon tasks (Deep Research and SWE), our folding agent matches or outperforms the ReAct baselines while using an active context 10$\times$ smaller and significantly outperforms models that rely on summarization-based context management.
As LLMs become central to interactive applications, ranging from tutoring to mental health, the ability to express personality in culturally appropriate ways is increasingly important. While recent works have explored personality evaluation of LLMs, they largely overlook the interplay between culture and personality. To address this, we introduce CulturalPersonas, the first large-scale benchmark with human validation for evaluating LLMs' personality expression in culturally grounded, behaviorally rich contexts. Our dataset spans 3,000 scenario-based questions across six diverse countries, designed to elicit personality through everyday scenarios rooted in local values. We evaluate three LLMs, using both multiple-choice and open-ended response formats. Our results show that CulturalPersonas improves alignment with country-specific human personality distributions (over a 20% reduction in Wasserstein distance across models and countries) and elicits more expressive, culturally coherent outputs compared to existing benchmarks. CulturalPersonas surfaces meaningful modulated trait outputs in response to culturally grounded prompts, offering new directions for aligning LLMs to global norms of behavior. By bridging personality expression and cultural nuance, we envision that CulturalPersonas will pave the way for more socially intelligent and globally adaptive LLMs.
Effective feedback is essential for student learning but is time-intensive for teachers. We present LearnLens, a modular, LLM-based system that generates personalised, curriculum-aligned feedback in science education. LearnLens comprises three components: (1) an error-aware assessment module that captures nuanced reasoning errors; (2) a curriculum-grounded generation module that uses a structured, topic-linked memory chain rather than traditional similarity-based retrieval, improving relevance and reducing noise; and (3) an educator-in-the-loop interface for customisation and oversight. LearnLens addresses key challenges in existing systems, offering scalable, high-quality feedback that empowers both teachers and students.
Serving Large Language Models (LLMs) in production faces significant challenges from highly variable request patterns and severe resource fragmentation in serverless clusters. Current systems rely on static pipeline configurations that struggle to adapt to dynamic workload conditions, leading to substantial inefficiencies. We present FlexPipe, a novel system that dynamically reconfigures pipeline architectures during runtime to address these fundamental limitations. FlexPipe decomposes models into fine-grained stages and intelligently adjusts pipeline granularity based on real-time request pattern analysis, implementing three key innovations: fine-grained model partitioning with preserved computational graph constraints, inflight pipeline refactoring with consistent cache transitions, and topology-aware resource allocation that navigates GPU fragmentation. Comprehensive evaluation on an 82-GPU cluster demonstrates that FlexPipe achieves up to 8.5x better resource efficiency while maintaining 38.3% lower latency compared to state-of-the-art systems, reducing GPU reservation requirements from 75% to 30% of peak capacity.
Phishing remains a critical cybersecurity threat, especially with the advent of large language models (LLMs) capable of generating highly convincing malicious content. Unlike earlier phishing attempts which are identifiable by grammatical errors, misspellings, incorrect phrasing, and inconsistent formatting, LLM generated emails are grammatically sound, contextually relevant, and linguistically natural. These advancements make phishing emails increasingly difficult to distinguish from legitimate ones, challenging traditional detection mechanisms. Conventional phishing detection systems often fail when faced with emails crafted by LLMs or manipulated using adversarial perturbation techniques. To address this challenge, we propose a robust phishing email detection system featuring an enhanced text preprocessing pipeline. This pipeline includes spelling correction and word splitting to counteract adversarial modifications and improve detection accuracy. Our approach integrates widely adopted natural language processing (NLP) feature extraction techniques and machine learning algorithms. We evaluate our models on publicly available datasets comprising both phishing and legitimate emails, achieving a detection accuracy of 94.26% and F1-score of 84.39% in model deployment setting. To assess robustness, we further evaluate our models using adversarial phishing samples generated by four attack methods in Python TextAttack framework. Additionally, we evaluate models' performance against phishing emails generated by LLMs including ChatGPT and Llama. Results highlight the resilience of models against evolving AI-powered phishing threats.
For Large Language Models (LLMs) to be reliable, they must learn robust knowledge that can be generally applied in diverse settings -- often unlike those seen during training. Yet, extensive research has shown that LLM performance can be brittle, with models exhibiting excessive sensitivity to trivial input variations. In this work, we explore whether this brittleness is a direct result of unstable internal knowledge representations. To explore this question, we build on previous work showing that LLM representations encode statement truthfulness -- i.e., true, factual statements can be easily separated from false, inaccurate ones. Specifically, we test the robustness of learned knowledge by evaluating representation separability on samples that have undergone superficial transformations to drive them out-of-distribution (OOD), such as typos or reformulations. By applying semantically-preserving perturbations, we study how separability degrades as statements become more OOD, across four LLM families, five evaluation datasets, and three knowledge probing methods. Our results reveal that internal representations of statement truthfulness collapse as the samples' presentations become less similar to those seen during pre-training. While LLMs can often distinguish between true and false statements when they closely resemble the pre-training data, this ability is highly dependent on the statement's exact surface form. These findings offer a possible explanation for brittle benchmark performance: LLMs may learn shallow, non-robust knowledge representations that allow for only limited generalizability. Our work presents a fundamental challenge for the utility of truthfulness probes, and more broadly, calls for further research on improving the robustness of learned knowledge representations.
Substance use disorders (SUDs) affect millions of people, and relapses are common, requiring multi-session treatments. Access to care is limited, which contributes to the challenge of recovery support. We present \textbf{ChatThero}, an innovative low-cost, multi-session, stressor-aware, and memory-persistent autonomous \emph{language agent} designed to facilitate long-term behavior change and therapeutic support in addiction recovery. Unlike existing work that mostly finetuned large language models (LLMs) on patient-therapist conversation data, ChatThero was trained in a multi-agent simulated environment that mirrors real therapy. We created anonymized patient profiles from recovery communities (e.g., Reddit). We classify patients as \texttt{easy}, \texttt{medium}, and \texttt{difficult}, three scales representing their resistance to recovery. We created an external environment by introducing stressors (e.g., social determinants of health) to simulate real-world situations. We dynamically inject clinically-grounded therapeutic strategies (motivational interview and cognitive behavioral therapy). Our evaluation, conducted by both human (blinded clinicians) and LLM-as-Judge, shows that ChatThero is superior in empathy and clinical relevance. We show that stressor simulation improves robustness of ChatThero. Explicit stressors increase relapse-like setbacks, matching real-world patterns. We evaluate ChatThero with behavioral change metrics. On a 1--5 scale, ChatThero raises \texttt{motivation} by $+1.71$ points (from $2.39$ to $4.10$) and \texttt{confidence} by $+1.67$ points (from $1.52$ to $3.19$), substantially outperforming GPT-5. On \texttt{difficult} patients, ChatThero reaches the success milestone with $26\%$ fewer turns than GPT-5.
New Large Language Models (LLMs) become available every few weeks, and modern application developers confronted with the unenviable task of having to decide if they should switch to a new model. While human evaluation remains the gold standard, it is costly and unscalable. The state-of-the-art approach is to use LLMs as evaluators ( LLM-as-a-judge), but this suffers from a critical flaw: LLMs exhibit a strong positive bias. We provide empirical evidence showing that while LLMs can identify valid outputs with high accuracy (i.e., True Positive Rate 96%), they are remarkably poor at identifying invalid ones (i.e., True Negative Rate <25%). This systematic bias, coupled with class imbalance, often leads to inflated reliability scores. While ensemble-based methods like majority voting can help, we show that they are not good enough. We introduce an optimal minority-veto strategy that is resilient to missing data and mitigates this bias to a large extent. For scenarios requiring even higher precision, we propose a novel regression-based framework that directly models the validator bias using a small set of human-annotated ground truth data. On a challenging code feedback task over 366 high-school Python programs, our regression approach reduces the maximum absolute error to just 1.2%, achieving a 2x improvement over the best-performing ensemble of 14 state-of-the-art LLMs.
Large Language Models (LLMs) are increasingly applied to data-intensive workflows, from database querying to developer observability. Yet the effectiveness of these systems is constrained by the volume, verbosity, and noise of real-world text-rich data such as logs, telemetry, and monitoring streams. Feeding such data directly into LLMs is costly, environmentally unsustainable, and often misaligned with task objectives. Parallel efforts in LLM efficiency have focused on model- or architecture-level optimizations, but the challenge of reducing upstream input verbosity remains underexplored. In this paper, we argue for treating the token budget of an LLM as an attention budget and elevating task-aware text reduction as a first-class design principle for language -- data systems. We position input-side reduction not as compression, but as attention allocation: prioritizing information most relevant to downstream tasks. We outline open research challenges for building benchmarks, designing adaptive reduction pipelines, and integrating token-budget--aware preprocessing into database and retrieval systems. Our vision is to channel scarce attention resources toward meaningful signals in noisy, data-intensive workflows, enabling scalable, accurate, and sustainable LLM--data integration.
We propose QeRL, a Quantization-enhanced Reinforcement Learning framework for large language models (LLMs). While RL is essential for LLMs' reasoning capabilities, it is resource-intensive, requiring substantial GPU memory and long rollout durations. QeRL addresses these issues by combining NVFP4 quantization with Low-Rank Adaptation (LoRA), accelerating rollout phase of RL while reducing memory overhead. Beyond efficiency, our findings show that quantization noise increases policy entropy, enhancing exploration, and enabling the discovery of better strategies during RL. To further optimize exploration, QeRL introduces an Adaptive Quantization Noise (AQN) mechanism, which dynamically adjusts noise during training. Experiments demonstrate that QeRL delivers over 1.5 times speedup in the rollout phase. Moreover, this is the first framework to enable RL training of a 32B LLM on a single H100 80GB GPU, while delivering overall speedups for RL training. It also achieves faster reward growth and higher final accuracy than 16-bit LoRA and QLoRA, while matching the performance of full-parameter fine-tuning on mathematical benchmarks such as GSM8K (90.8%) and MATH 500 (77.4%) in the 7B model. These results establish QeRL as an efficient and effective framework for RL training in LLMs.
Although Large Language Model (LLM)-based agents are increasingly used in financial trading, it remains unclear whether they can reason and adapt in live markets, as most studies test models instead of agents, cover limited periods and assets, and rely on unverified data. To address these gaps, we introduce Agent Market Arena (AMA), the first lifelong, real-time benchmark for evaluating LLM-based trading agents across multiple markets. AMA integrates verified trading data, expert-checked news, and diverse agent architectures within a unified trading framework, enabling fair and continuous comparison under real conditions. It implements four agents, including InvestorAgent as a single-agent baseline, TradeAgent and HedgeFundAgent with different risk styles, and DeepFundAgent with memory-based reasoning, and evaluates them across GPT-4o, GPT-4.1, Claude-3.5-haiku, Claude-sonnet-4, and Gemini-2.0-flash. Live experiments on both cryptocurrency and stock markets demonstrate that agent frameworks display markedly distinct behavioral patterns, spanning from aggressive risk-taking to conservative decision-making, whereas model backbones contribute less to outcome variation. AMA thus establishes a foundation for rigorous, reproducible, and continuously evolving evaluation of financial reasoning and trading intelligence in LLM-based agents.
Enhancing the mathematical reasoning capabilities of LLMs has garnered significant attention in both the mathematical and computer science communities. Recent works have made substantial progress in both Natural Language (NL) reasoning and Formal Language (FL) reasoning by leveraging the potential of pure Reinforcement Learning (RL) methods on base models. However, RL approaches struggle to impart new capabilities not presented in the base model, highlighting the need to integrate more knowledge like FL into NL math reasoning effectively. Yet, this integration is challenging due to inherent disparities in problem structure and reasoning format between NL and FL. To address these challenges, we introduce **NL-FL HybridReasoning (NFL-HR)**, an end-to-end framework designed to incorporate the FL expert into NL math problem-solving. To bridge the NL and FL input format gap, we propose the NL-FL Problem Alignment method, which reformulates the Question-Answering (QA) problems in NL as existence theorems in FL. Subsequently, the Mixed Problem Input technique we provide enables the FL reasoner to handle both QA and existence problems concurrently. Lastly, we mitigate the NL and FL output format gap in reasoning through an LLM-based Answer Extraction mechanism. Comprehensive experiments demonstrate that the NFL-HR framework achieves **89.80**% and **84.34%** accuracy rates on the MATH-500 and the AMC benchmarks, surpassing the NL baseline by **4.60%** and **4.82%**, respectively. Notably, some problems resolved by our framework remain unsolved by the NL baseline model even under a larger number of trials.
Knowledge distillation (KD) is a key technique for compressing large-scale language models (LLMs), yet prevailing logit-based methods typically employ static strategies that are misaligned with the dynamic learning process of student models. These methods typically treat all tokens indiscriminately and apply a single, fixed temperature, resulting in suboptimal knowledge transfer. To address these limitations, we propose LLM-Oriented Token-Adaptive Knowledge Distillation (AdaKD), a novel framework that adapts the distillation process to the real-time learning state of each token. AdaKD consists of two synergistic modules driven by a unified token difficulty metric. First, our Loss-Driven Adaptive Token Focusing (LATF) module dynamically adjusts the distillation focus by monitoring the student's learning stability, concentrating computational resources on the most valuable tokens at each training phase. Second, we introduce Inverse Difficulty Temperature Scaling (IDTS), a counterintuitive yet effective token-level temperature strategy. It employs low temperatures for difficult tokens for targeted error correction, and high temperatures for easy tokens to encourage students to learn from the teacher's complete and smooth output distribution, thereby enhancing generalization. As a plug-and-play framework, AdaKD can consistently improve the performance of various distillation methods on multiple model architectures and benchmarks.
Large Language Model (LLM)-based agentic systems rely on in-context policy documents encoding diverse business rules. As requirements grow, these documents expand rapidly, causing high computational overhead. This motivates developing internalization methods that embed policy documents into model priors while preserving performance. Prior prompt compression work targets generic prompts, but agentic policy documents span multiple complexity levels and require deeper reasoning, making internalization harder. We introduce CC-Gen, an agentic benchmark generator with Controllable Complexity across four levels, enabling systematic evaluation of agents' ability to handle complexity and offering a unified framework for assessing policy internalization. Our analysis shows that complex policy specifications governing workflows pose major reasoning challenges. Supporting internalization with gold user agent interaction trajectories containing chain-of-thought (CoT) annotations via supervised fine-tuning (SFT) is data-intensive and degrades sharply as policy complexity increases. To mitigate data and reasoning burdens, we propose Category-Aware Policy Continued Pretraining (CAP-CPT). Our automated pipeline parses policy documents to extract key specifications, grouping them into factual, behavioral, and conditional categories, and isolating complex conditions that drive workflow complexity. This guides targeted data synthesis and enables agents to internalize policy information through an autoregressive pretraining loss. Experiments show CAP-CPT improves SFT baselines in all settings, with up to 41% and 22% gains on Qwen-3-32B, achieving 97.3% prompt length reduction on CC-Gen and further enhancing tau-Bench with minimal SFT data.
Existing benchmarks that measure cultural adaptation in LLMs are misaligned with the actual challenges these models face when interacting with users from diverse cultural backgrounds. In this work, we introduce the first framework and benchmark designed to evaluate LLMs in realistic, multicultural conversational settings. Grounded in sociocultural theory, our framework formalizes how linguistic style - a key element of cultural communication - is shaped by situational, relational, and cultural context. We construct a benchmark dataset based on this framework, annotated by culturally diverse raters, and propose a new set of desiderata for cross-cultural evaluation in NLP: conversational framing, stylistic sensitivity, and subjective correctness. We evaluate today's top LLMs on our benchmark and show that these models struggle with cultural adaptation in a conversational setting.
Recent research on Chain-of-Thought (CoT) reasoning in Large Language Models (LLMs) has demonstrated that agents can engage in \textit{complex}, \textit{multi-turn} negotiations, opening new avenues for agentic AI. However, existing LLM agents largely overlook the functional role of emotions in such negotiations, instead generating passive, preference-driven emotional responses that make them vulnerable to manipulation and strategic exploitation by adversarial counterparts. To address this gap, we present EvoEmo, an evolutionary reinforcement learning framework that optimizes dynamic emotional expression in negotiations. EvoEmo models emotional state transitions as a Markov Decision Process and employs population-based genetic optimization to evolve high-reward emotion policies across diverse negotiation scenarios. We further propose an evaluation framework with two baselines -- vanilla strategies and fixed-emotion strategies -- for benchmarking emotion-aware negotiation. Extensive experiments and ablation studies show that EvoEmo consistently outperforms both baselines, achieving higher success rates, higher efficiency, and increased buyer savings. This findings highlight the importance of adaptive emotional expression in enabling more effective LLM agents for multi-turn negotiation.
Large Language Models (LLMs) suffer from reliability issues on complex tasks, as existing decomposition methods are heuristic and rely on agent or manual decomposition. This work introduces a novel, systematic decomposition framework that we call Analysis of CONstraint-Induced Complexity (ACONIC), which models the task as a constraint problem and leveraging formal complexity measures to guide decomposition. On combinatorial (SATBench) and LLM database querying tasks (Spider), we find that by decomposing the tasks following the measure of complexity, agent can perform considerably better (10-40 percentage point).
Governance of data, compliance, and business privacy matters, particularly for healthcare and finance businesses. Since the recent emergence of AI enterprise AI assistants enhancing business productivity, safeguarding private data and compliance is now a priority. With the implementation of AI assistants across the enterprise, the zero data retention can be achieved by implementing zero data retention policies by Large Language Model businesses like Open AI and Anthropic and Meta. In this work, we explore zero data retention policies for the Enterprise apps of large language models (LLMs). Our key contribution is defining the architectural, compliance, and usability trade-offs of such systems in parallel. In this research work, we examine the development of commercial AI assistants with two industry leaders and market titans in this arena - Salesforce and Microsoft. Both of these companies used distinct technical architecture to support zero data retention policies. Salesforce AgentForce and Microsoft Copilot are among the leading AI assistants providing much-needed push to business productivity in customer care. The purpose of this paper is to analyze the technical architecture and deployment of zero data retention policy by consuming applications as well as big language models service providers like Open Ai, Anthropic, and Meta.
Large Language Models are trained on massive multilingual corpora, yet this abundance masks a profound crisis: of the world's 7,613 living languages, approximately 2,000 languages with millions of speakers remain effectively invisible in digital ecosystems. We propose a critical framework connecting empirical measurements of language vitality (real world demographic strength) and digitality (online presence) with postcolonial theory and epistemic injustice to explain why linguistic inequality in AI systems is not incidental but structural. Analyzing data across all documented human languages, we identify four categories: Strongholds (33%, high vitality and digitality), Digital Echoes (6%, high digitality despite declining vitality), Fading Voices (36%, low on both dimensions), and critically, Invisible Giants (27%, high vitality but near-zero digitality) - languages spoken by millions yet absent from the LLM universe. We demonstrate that these patterns reflect continuities from colonial-era linguistic hierarchies to contemporary AI development, constituting what we term digital epistemic injustice. Our analysis reveals that English dominance in AI is not a technical necessity but an artifact of power structures that systematically exclude marginalized linguistic knowledge. We conclude with implications for decolonizing language technology and democratizing access to AI benefits.