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
Papers
Applications of machine learning in chemistry are often limited by the scarcity and expense of labeled data, restricting traditional supervised methods. In this work, we introduce a framework for molecular reasoning using general-purpose Large Language Models (LLMs) that operates without requiring labeled training data. Our method anchors chain-of-thought reasoning to the molecular structure by using unique atomic identifiers. First, the LLM performs a one-shot task to identify relevant fragments and their associated chemical labels or transformation classes. In an optional second step, this position-aware information is used in a few-shot task with provided class examples to predict the chemical transformation. We apply our framework to single-step retrosynthesis, a task where LLMs have previously underperformed. Across academic benchmarks and expert-validated drug discovery molecules, our work enables LLMs to achieve high success rates in identifying chemically plausible reaction sites ($\geq90\%$), named reaction classes ($\geq40\%$), and final reactants ($\geq74\%$). Beyond solving complex chemical tasks, our work also provides a method to generate theoretically grounded synthetic datasets by mapping chemical knowledge onto the molecular structure and thereby addressing data scarcity.
Recent advances in leveraging LLMs for APR have demonstrated impressive capabilities in fixing software defects. However, current LLM-based approaches predominantly focus on mainstream programming languages like Java and Python, neglecting less prevalent but emerging languages such as Rust due to expensive training resources, limited datasets, and insufficient community support. This narrow focus creates a significant gap in repair capabilities across the programming language spectrum, where the full potential of LLMs for comprehensive multilingual program repair remains largely unexplored. To address this limitation, we introduce a novel cross-language program repair approach LANTERN that leverages LLMs' differential proficiency across languages through a multi-agent iterative repair paradigm. Our technique strategically translates defective code from languages where LLMs exhibit weaker repair capabilities to languages where they demonstrate stronger performance, without requiring additional training. A key innovation of our approach is an LLM-based decision-making system that dynamically selects optimal target languages based on bug characteristics and continuously incorporates feedback from previous repair attempts. We evaluate our method on xCodeEval, a comprehensive multilingual benchmark comprising 5,068 bugs across 11 programming languages. Results demonstrate significant enhancement in repair effectiveness, particularly for underrepresented languages, with Rust showing a 22.09% improvement in Pass@10 metrics. Our research provides the first empirical evidence that cross-language translation significantly expands the repair capabilities of LLMs and effectively bridges the performance gap between programming languages with different levels of popularity, opening new avenues for truly language-agnostic automated program repair.
Large Language Models (LLMs) are increasingly used to refactor unit tests, improving readability and structure while preserving behavior. Evaluating such refactorings, however, remains difficult: metrics like CodeBLEU penalize beneficial renamings and edits, while semantic similarities overlook readability and modularity. We propose CTSES, a first step toward human-aligned evaluation of refactored tests. CTSES combines CodeBLEU, METEOR, and ROUGE-L into a composite score that balances semantics, lexical clarity, and structural alignment. Evaluated on 5,000+ refactorings from Defects4J and SF110 (GPT-4o and Mistral-Large), CTSES reduces false negatives and provides more interpretable signals than individual metrics. Our emerging results illustrate that CTSES offers a proof-of-concept for composite approaches, showing their promise in bridging automated metrics and developer judgments.
Large language models (LLMs) have shown promising potential in persuasion, but existing works on training LLM persuaders are still preliminary. Notably, while humans are skilled in modeling their opponent's thoughts and opinions proactively and dynamically, current LLMs struggle with such Theory of Mind (ToM) reasoning, resulting in limited diversity and opponent awareness. To address this limitation, we introduce Theory of Mind Augmented Persuader (ToMAP), a novel approach for building more flexible persuader agents by incorporating two theory of mind modules that enhance the persuader's awareness and analysis of the opponent's mental state. Specifically, we begin by prompting the persuader to consider possible objections to the target central claim, and then use a text encoder paired with a trained MLP classifier to predict the opponent's current stance on these counterclaims. Our carefully designed reinforcement learning schema enables the persuader learns how to analyze opponent-related information and utilize it to generate more effective arguments. Experiments show that the ToMAP persuader, while containing only 3B parameters, outperforms much larger baselines, like GPT-4o, with a relative gain of 39.4% across multiple persuadee models and diverse corpora. Notably, ToMAP exhibits complex reasoning chains and reduced repetition during training, which leads to more diverse and effective arguments. The opponent-aware feature of ToMAP also makes it suitable for long conversations and enables it to employ more logical and opponent-aware strategies. These results underscore our method's effectiveness and highlight its potential for developing more persuasive language agents. Code is available at: https://github.com/ulab-uiuc/ToMAP.
Engineering construction automation aims to transform natural language specifications into physically viable structures, requiring complex integrated reasoning under strict physical constraints. While modern LLMs possess broad knowledge and strong reasoning capabilities that make them promising candidates for this domain, their construction competencies remain largely unevaluated. To address this gap, we introduce BuildArena, the first physics-aligned interactive benchmark designed for language-driven engineering construction. It contributes to the community in four aspects: (1) a highly customizable benchmarking framework for in-depth comparison and analysis of LLMs; (2) an extendable task design strategy spanning static and dynamic mechanics across multiple difficulty tiers; (3) a 3D Spatial Geometric Computation Library for supporting construction based on language instructions; (4) a baseline LLM agentic workflow that effectively evaluates diverse model capabilities. On eight frontier LLMs, BuildArena comprehensively evaluates their capabilities for language-driven and physics-grounded construction automation. The project page is at https://build-arena.github.io/.
Reinforcement learning in large language models (LLMs) often relies on scalar rewards, a practice that discards valuable textual rationale buried in the rollouts, forcing the model to explore \textit{de novo} with each attempt and hindering sample efficiency. While LLMs can uniquely learn from language feedback provided in-context, naively integrating on-line experiences into RL training presents a paradox: feedback from the same problem risks information leakage and memorization, while feedback from different problems often leads to behavior collapse due to irrelevant context. To resolve this tension, we propose \textbf{Language-And-Numerical Policy Optimization (LANPO)}, a framework that cleanly separates the roles of feedback: language guides exploration, while numerical rewards drive optimization. LANPO builds a dynamic experience pool from past trials and introduces two principles to ensure feedback is effective: \emph{Reward-Agnostic Reflection} for safe intra-sample self-correction and \emph{Relevant Abstraction} to distill generalizable lessons from inter-sample experiences. Across mathematical reasoning benchmarks, LANPO enables 7B and 14B models to significantly outperform strong baselines trained with GRPO in test accuracy. Our work provides a robust method for integrating historical experiences into the LLM RL loop, creating more effective and data-efficient learning agents.
This research proposes a systematic, large language model (LLM) approach for extracting product and service attributes, features, and associated sentiments from customer reviews. Grounded in marketing theory, the framework distinguishes perceptual attributes from actionable features, producing interpretable and managerially actionable insights. We apply the methodology to 20,000 Yelp reviews of Starbucks stores and evaluate eight prompt variants on a random subset of reviews. Model performance is assessed through agreement with human annotations and predictive validity for customer ratings. Results show high consistency between LLMs and human coders and strong predictive validity, confirming the reliability of the approach. Human coders required a median of six minutes per review, whereas the LLM processed each in two seconds, delivering comparable insights at a scale unattainable through manual coding. Managerially, the analysis identifies attributes and features that most strongly influence customer satisfaction and their associated sentiments, enabling firms to pinpoint "joy points," address "pain points," and design targeted interventions. We demonstrate how structured review data can power an actionable marketing dashboard that tracks sentiment over time and across stores, benchmarks performance, and highlights high-leverage features for improvement. Simulations indicate that enhancing sentiment for key service features could yield 1-2% average revenue gains per store.
Peer review serves as the gatekeeper of science, yet the surge in submissions and widespread adoption of large language models (LLMs) in scholarly evaluation present unprecedented challenges. Recent work has focused on using LLMs to improve review efficiency or generate insightful review content. However, unchecked deficient reviews from both human experts and AI systems threaten to systematically undermine the peer review ecosystem and compromise academic integrity. To address this critical issue, we introduce ReviewGuard, an automated system for detecting and categorizing deficient reviews. ReviewGuard employs a comprehensive four-stage LLM-driven framework that: (1) collects ICLR and NeurIPS papers with their corresponding reviews from OpenReview; (2) annotates review types using GPT-4.1 with human validation; (3) addresses class imbalance and data scarcity through LLM-driven synthetic data augmentation, producing a final corpus of 6,634 papers, 24,657 real reviews, and 46,438 synthetic reviews; and (4) fine-tunes both encoder-based models and open source LLMs. We perform comprehensive feature analysis of the structure and quality of the review text. Compared to sufficient reviews, deficient reviews demonstrate lower rating scores, higher self-reported confidence, reduced structural complexity, and a higher proportion of negative sentiment. AI-generated text detection reveals that, since ChatGPT's emergence, AI-generated reviews have increased dramatically. In the evaluation of deficient review detection models, mixed training with synthetic and real review data provides substantial enhancements to recall and F1 scores on the binary task. This study presents the first LLM-driven system for detecting deficient peer reviews, providing evidence to inform AI governance in peer review while offering valuable insights into human-AI collaboration to maintain academic integrity.
Recent claims of strong performance by Large Language Models (LLMs) on causal discovery are undermined by a key flaw: many evaluations rely on benchmarks likely included in pretraining corpora. Thus, apparent success suggests that LLM-only methods, which ignore observational data, outperform classical statistical approaches. We challenge this narrative by asking: Do LLMs truly reason about causal structure, and how can we measure it without memorization concerns? Can they be trusted for real-world scientific discovery? We argue that realizing LLMs' potential for causal analysis requires two shifts: (P.1) developing robust evaluation protocols based on recent scientific studies to guard against dataset leakage, and (P.2) designing hybrid methods that combine LLM-derived knowledge with data-driven statistics. To address P.1, we encourage evaluating discovery methods on novel, real-world scientific studies. We outline a practical recipe for extracting causal graphs from recent publications released after an LLM's training cutoff, ensuring relevance and preventing memorization while capturing both established and novel relations. Compared to benchmarks like BNLearn, where LLMs achieve near-perfect accuracy, they perform far worse on our curated graphs, underscoring the need for statistical grounding. Supporting P.2, we show that using LLM predictions as priors for the classical PC algorithm significantly improves accuracy over both LLM-only and purely statistical methods. We call on the community to adopt science-grounded, leakage-resistant benchmarks and invest in hybrid causal discovery methods suited to real-world inquiry.
Large Language Models (LLMs) optimized to output truthful answers often overfit, producing brittle reasoning that fails to generalize. While persuasion-based optimization has shown promise in debate settings, it has not been systematically compared against mainstream truth-based approaches. We introduce DebateQD, a minimal Quality-Diversity (QD) evolutionary algorithm that evolves diverse debate strategies across different categories (rationality, authority, emotional appeal, etc.) through tournament-style competitions where two LLMs debate while a third judges. Unlike previously proposed methods that require a population of LLMs, our approach maintains diversity of opponents through prompt-based strategies within a single LLM architecture, making it more accessible for experiments while preserving the key benefits of population-based optimization. In contrast to prior work, we explicitly isolate the role of the optimization objective by fixing the debate protocol and swapping only the fitness function: persuasion rewards strategies that convince the judge irrespective of truth, whereas truth rewards collaborative correctness. Across three model scales (7B, 32B, 72B parameters) and multiple dataset sizes from the QuALITY benchmark, persuasion-optimized strategies achieve up to 13.94% smaller train-test generalization gaps, while matching or exceeding truth optimization's test performance. These results provide the first controlled evidence that competitive pressure to persuade, rather than seek the truth collaboratively, fosters more transferable reasoning skills, offering a promising path for improving LLM generalization.
Training robust retrieval and reranker models typically relies on large-scale retrieval datasets; for example, the BGE collection contains 1.6 million query-passage pairs sourced from various data sources. However, we find that certain datasets can negatively impact model effectiveness -- pruning 8 out of 15 datasets from the BGE collection, reduces the training set size by 2.35$\times$, surprisingly increases nDCG@10 on BEIR by 1.0 point. This motivates a deeper examination of training data quality, with a particular focus on "false negatives", where relevant passages are incorrectly labeled as irrelevant. We utilize LLMs as a simple, cost-effective approach to identify and relabel false negatives in training datasets. Experimental results show that relabeling false negatives as true positives improves both E5 (base) and Qwen2.5-7B retrieval models by 0.7$\unicode{x2013}$1.4 points on BEIR and by 1.7$\unicode{x2013}$1.8 points at nDCG@10 on zero-shot AIR-Bench evaluation. Similar gains are observed for rerankers fine-tuned on the relabeled data, such as Qwen2.5-3B on BEIR. The reliability of LLMs to identify false negatives is supported by human annotation results. Our training dataset and code are publicly available.
With the rapid evolution of large language models (LLMs), the demand for automated, high-performance system kernels has emerged as a key enabler for accelerating development and deployment. We introduce TritonRL, a domain-specialized LLM for Triton kernel generation, trained with a novel training framework that enables robust and automated kernel synthesis. Unlike general-purpose programming languages, Triton kernel generation faces unique challenges due to data scarcity and incomplete evaluation criteria, vulnerable to reward hacking. Our approach addresses these challenges end-to-end by distilling Triton-specific knowledge through supervised fine-tuning on curated datasets, and further improving code quality via reinforcement learning (RL) with robust, verifiable rewards and hierarchical reward assignment. Our RL framework robustly detects reward hacking and guides both reasoning traces and code tokens through fine-grained verification and hierarchical reward decomposition, enabling the model to generate high-quality Triton kernels that can truly replace existing modules. With robust and fine-grained evaluation, our experiments on KernelBench demonstrate that TritonRL achieves state-of-the-art correctness and speedup, surpassing all other Triton-specific models and underscoring the effectiveness of our RL-based training paradigm.
The advancement of robotics and autonomous navigation systems hinges on the ability to accurately predict terrain traversability. Traditional methods for generating datasets to train these prediction models often involve putting robots into potentially hazardous environments, posing risks to equipment and safety. To solve this problem, we present ZeST, a novel approach leveraging visual reasoning capabilities of Large Language Models (LLMs) to create a traversability map in real-time without exposing robots to danger. Our approach not only performs zero-shot traversability and mitigates the risks associated with real-world data collection but also accelerates the development of advanced navigation systems, offering a cost-effective and scalable solution. To support our findings, we present navigation results, in both controlled indoor and unstructured outdoor environments. As shown in the experiments, our method provides safer navigation when compared to other state-of-the-art methods, constantly reaching the final goal.
Sycophancy (overly agreeable or flattering behavior) is critical to understand in the context of human-AI collaboration, especially in decision-making settings like health, law, and education. Existing methods for studying sycophancy in LLMs are either descriptive (study behavior change when sycophancy is elicited) or normative (provide values-based judgment on behavior change). Together, these approaches help us understand the extent, and impacts, of sycophancy. However, existing normative approaches only apply for objective tasks where ground-truth data exists, ignoring the natural subjectivity in many NLP tasks. Drawing from behavioral economics and rational decision theory, we introduce an Bayesian framework to study the normative effects of sycophancy on rationality in LLMs, without requiring labeled ground-truth. Using this interdisciplinary framework, we study sycophantic behavior in multiple LLM baselines across three different tasks, experimenting with various methods for eliciting sycophancy and obtaining probability judgments from LLMs. We find significant evidence of sycophancy in our experiments (7 of 8 baselines for one of our probing techniques), and observe that sycophancy is more likely to reduce rationality than it is to increase rationality in LLMs' decisions when they are directly probed for probabilities (2 out of 4 baselines show significant increases overall).
A large amount of work has been done in Multi-Agent Systems (MAS) for modeling and solving problems with multiple interacting agents. However, most LLMs are pretrained independently and not specifically optimized for coordination. Existing LLM fine-tuning frameworks rely on individual rewards, which require complex reward designs for each agent to encourage collaboration. To address these challenges, we model LLM collaboration as a cooperative Multi-Agent Reinforcement Learning (MARL) problem. We develop a multi-agent, multi-turn algorithm, Multi-Agent Group Relative Policy Optimization (MAGRPO), to solve it, building on current RL approaches for LLMs as well as MARL techniques. Our experiments on LLM writing and coding collaboration demonstrate that fine-tuning MAS with MAGRPO enables agents to generate high-quality responses efficiently through effective cooperation. Our approach opens the door to using other MARL methods for LLMs and highlights the associated challenges.
Multi-agent deployments of large language models (LLMs) are increasingly embedded in market, allocation, and governance workflows, yet covert coordination among agents can silently erode trust and social welfare. Existing audits are dominated by heuristics that lack theoretical guarantees, struggle to transfer across tasks, and seldom ship with the infrastructure needed for independent replication. We introduce Audit the Whisper, a conference-grade research artifact that spans theory, benchmark design, detection, and reproducibility. Our contributions are: (i) a channel-capacity analysis showing how interventions such as paraphrase, rate limiting, and role permutation impose quantifiable capacity penalties-operationalised via paired-run Kullback--Leibler diagnostics-that tighten mutual-information thresholds with finite-sample guarantees and full proofs; (ii) ColludeBench-v0, covering pricing, first-price auctions, peer review, and hosted Gemini/Groq APIs with configurable covert schemes, deterministic manifests, and reward instrumentation; and (iii) a calibrated auditing pipeline that fuses cross-run mutual information, permutation invariance, watermark variance, and fairness-aware acceptance bias, each tuned to a $10^{-3}$ false-positive budget and validated by 10k honest runs plus an e-value martingale. Across ColludeBench and external suites including Secret Collusion, CASE, Perfect Collusion Benchmark, and SentinelAgent, the union meta-test attains state-of-the-art power at fixed FPR while ablations surface price-of-auditing trade-offs and fairness-driven colluders invisible to MI alone. We release regeneration scripts, anonymized manifests, and documentation so that external auditors can reproduce every figure, satisfy double-blind requirements, and extend the framework with minimal effort.
This work studies the problem of large language model (LLM) unlearning, aiming to remove unwanted data influences (e.g., copyrighted or harmful content) while preserving model utility. Despite the increasing demand for unlearning, a technically-grounded optimization framework is lacking. Gradient ascent (GA)-type methods, though widely used, are suboptimal as they reverse the learning process without controlling optimization divergence (i.e., deviation from the pre-trained state), leading to risks of over-forgetting and potential model collapse. Negative preference optimization (NPO) has been proposed to address this issue and is considered one of the state-of-the-art LLM unlearning approaches. In this work, we revisit NPO and identify another critical issue: reference model bias. This bias arises from using the reference model (i.e., the model prior to unlearning) to evaluate the unlearning success, which can compromise NPO's effectiveness. Specifically, it leads to (a) uneven allocation of optimization power across forget data with varying difficulty levels and (b) ineffective gradient weight smoothing during the early stages of unlearning optimization. To overcome these challenges, we propose a simple yet effective unlearning optimization framework, called SimNPO, showing that `simplicity' in removing the reliance on a reference model (through the lens of simple preference optimization) benefits unlearning. We provide deeper insights into SimNPO's advantages through an analysis based on mixtures of Markov chains. Extensive experiments further validate SimNPO's efficacy on benchmarks like TOFU and MUSE, as well as its robustness against relearning attacks. Codes are available at https://github.com/OPTML-Group/Unlearn-Simple.
Effective human-AI collaboration on complex reasoning tasks requires that users understand and interact with the model's process, not just receive an output. However, the monolithic text from methods like Chain-of-Thought (CoT) prevents this, as current interfaces lack real-time verbalization and robust user barge-in. We present AsyncVoice Agent, a system whose asynchronous architecture decouples a streaming LLM backend from a conversational voice frontend. This design allows narration and inference to run in parallel, empowering users to interrupt, query, and steer the model's reasoning process at any time. Objective benchmarks show this approach reduces interaction latency by more than 600x compared to monolithic baselines while ensuring high fidelity and competitive task accuracy. By enabling a two-way dialogue with a model's thought process, AsyncVoice Agent offers a new paradigm for building more effective, steerable, and trustworthy human-AI systems for high-stakes tasks.
Synthesizing 3D scenes from open-vocabulary text descriptions is a challenging, important, and recently-popular application. One of its critical subproblems is layout generation: given a set of objects, lay them out to produce a scene matching the input description. Nearly all recent work adopts a declarative paradigm for this problem: using an LLM to generate a specification of constraints between objects, then solving those constraints to produce the final layout. In contrast, we explore an alternative imperative paradigm, in which an LLM iteratively places objects, with each object's position and orientation computed as a function of previously-placed objects. The imperative approach allows for a simpler scene specification language while also handling a wider variety and larger complexity of scenes. We further improve the robustness of our imperative scheme by developing an error correction mechanism that iteratively improves the scene's validity while staying as close as possible to the original layout generated by the LLM. In forced-choice perceptual studies, participants preferred layouts generated by our imperative approach 82% and 94% of the time when compared against two declarative layout generation methods. We also present a simple, automated evaluation metric for 3D scene layout generation that aligns well with human preferences.
The 2022 World Mental Health Report calls for global mental health care reform, amid rising prevalence of issues like anxiety and depression that affect nearly one billion people worldwide. Traditional in-person therapy fails to meet this demand, and the situation is worsened by stigma. While general-purpose large language models (LLMs) offer efficiency for AI-driven mental health solutions, they underperform because they lack specialized fine-tuning. Existing LLM-based mental health chatbots can engage in empathetic conversations, but they overlook real-time user mental state assessment which is critical for professional counseling. This paper proposes MoPHES, a framework that integrates mental state evaluation, conversational support, and professional treatment recommendations. The agent developed under this framework uses two fine-tuned MiniCPM4-0.5B LLMs: one is fine-tuned on mental health conditions datasets to assess users' mental states and predict the severity of anxiety and depression; the other is fine-tuned on multi-turn dialogues to handle conversations with users. By leveraging insights into users' mental states, our agent provides more tailored support and professional treatment recommendations. Both models are also deployed directly on mobile devices to enhance user convenience and protect user privacy. Additionally, to evaluate the performance of MoPHES with other LLMs, we develop a benchmark for the automatic evaluation of mental state prediction and multi-turn counseling dialogues, which includes comprehensive evaluation metrics, datasets, and methods.
We propose CITE V.1, an agentic, evidence-grounded framework that leverages Large Language Models (LLMs) to provide transparent and reproducible interpretations of RNA-seq clusters. Unlike existing enrichment-based approaches that reduce results to broad statistical associations and LLM-only models that risk unsupported claims or fabricated citations, CITE V.1 transforms cluster interpretation by producing biologically coherent explanations explicitly anchored in the biomedical literature. The framework orchestrates three specialized agents: a Retriever that gathers domain knowledge from PubMed and UniProt, an Interpreter that formulates functional hypotheses, and Critics that evaluate claims, enforce evidence grounding, and qualify uncertainty through confidence and reliability indicators. Applied to Salmonella enterica RNA-seq data, CITE V.1 generated biologically meaningful insights supported by the literature, while an LLM-only Gemini baseline frequently produced speculative results with false citations. By moving RNA-seq analysis from surface-level enrichment to auditable, interpretable, and evidence-based hypothesis generation, CITE V.1 advances the transparency and reliability of AI in biomedicine.
While Artificial Intelligence (AI) shows promise in healthcare applications, existing conversational systems often falter in complex and sensitive medical domains such as Sexual and Reproductive Health (SRH). These systems frequently struggle with hallucination and lack the specialized knowledge required, particularly for sensitive SRH topics. Furthermore, current AI approaches in healthcare tend to prioritize diagnostic capabilities over comprehensive patient care and education. Addressing these gaps, this work at the UNC School of Nursing introduces SARHAchat, a proof-of-concept Large Language Model (LLM)-based chatbot. SARHAchat is designed as a reliable, user-centered system integrating medical expertise with empathetic communication to enhance SRH care delivery. Our evaluation demonstrates SARHAchat's ability to provide accurate and contextually appropriate contraceptive counseling while maintaining a natural conversational flow. The demo is available at https://sarhachat.com/}{https://sarhachat.com/.
Emergency departments worldwide face rising patient volumes, workforce shortages, and variability in triage decisions that threaten the delivery of timely and accurate care. Current triage methods rely primarily on vital signs, routine laboratory values, and clinicians' judgment, which, while effective, often miss emerging biological signals that could improve risk prediction for infection typing or antibiotic administration in acute conditions. To address this challenge, we introduce TriAgent, a large language model (LLM)-based multi-agent framework that couples automated biomarker discovery with deep research for literature-grounded validation and novelty assessment. TriAgent employs a supervisor research agent to generate research topics and delegate targeted queries to specialized sub-agents for evidence retrieval from various data sources. Findings are synthesized to classify biomarkers as either grounded in existing knowledge or flagged as novel candidates, offering transparent justification and highlighting unexplored pathways in acute care risk stratification. Unlike prior frameworks limited to existing routine clinical biomarkers, TriAgent aims to deliver an end-to-end framework from data analysis to literature grounding to improve transparency, explainability and expand the frontier of potentially actionable clinical biomarkers. Given a user's clinical query and quantitative triage data, TriAgent achieved a topic adherence F1 score of 55.7 +/- 5.0%, surpassing the CoT-ReAct agent by over 10%, and a faithfulness score of 0.42 +/- 0.39, exceeding all baselines by more than 50%. Across experiments, TriAgent consistently outperformed state-of-the-art LLM-based agentic frameworks in biomarker justification and literature-grounded novelty assessment. We share our repo: https://github.com/CellFace/TriAgent.
Current Large Language Model (LLM) agents show strong performance in tool use, but lack the crucial capability to systematically learn from their own experiences. While existing frameworks mainly focus on mitigating external knowledge gaps, they fail to address a more fundamental limitation: the inability to iteratively refine problem-solving strategies. In this work, we introduce EvolveR, a framework designed to enable agent to self-improve through a complete, closed-loop experience lifecycle. This lifecycle comprises two key stages: (1) Offline Self-Distillation, where the agent's interaction trajectories are synthesized into a structured repository of abstract, reusable strategic principles; (2) Online Interaction, where the agent interacts with tasks and actively retrieves distilled principles to guide its decision-making, accumulating a diverse set of behavioral trajectories. This loop employs a policy reinforcement mechanism to iteratively update the agent based on its performance. We demonstrate the effectiveness of EvolveR on complex multi-hop question-answering benchmarks, where it achieves superior performance over strong agentic baselines. Our work presents a comprehensive blueprint for agents that learn not only from external data but also from the consequences of their own actions, paving the way for more autonomous and continuously improving systems. Code is available at https://github.com/Edaizi/EvolveR.
Self-correction of large language models (LLMs) emerges as a critical component for enhancing their reasoning performance. Although various self-correction methods have been proposed, a comprehensive evaluation of these methods remains largely unexplored, and the question of whether LLMs can truly correct themselves is a matter of significant interest and concern. In this study, we introduce CorrectBench, a benchmark developed to evaluate the effectiveness of self-correction strategies, including intrinsic, external, and fine-tuned approaches, across three tasks: commonsense reasoning, mathematical reasoning, and code generation. Our findings reveal that: 1) Self-correction methods can improve accuracy, especially for complex reasoning tasks; 2) Mixing different self-correction strategies yields further improvements, though it reduces efficiency; 3) Reasoning LLMs (e.g., DeepSeek-R1) have limited optimization under additional self-correction methods and have high time costs. Interestingly, a comparatively simple chain-of-thought (CoT) baseline demonstrates competitive accuracy and efficiency. These results underscore the potential of self-correction to enhance LLM's reasoning performance while highlighting the ongoing challenge of improving their efficiency. Consequently, we advocate for further research focused on optimizing the balance between reasoning capabilities and operational efficiency. Project Page: https://correctbench.github.io/
Advancing machine intelligence requires developing the ability to perceive across multiple modalities, much as humans sense the world. We introduce OmniVinci, an initiative to build a strong, open-source, omni-modal LLM. We carefully study the design choices across model architecture and data curation. For model architecture, we present three key innovations: (i) OmniAlignNet for strengthening alignment between vision and audio embeddings in a shared omni-modal latent space; (ii) Temporal Embedding Grouping for capturing relative temporal alignment between vision and audio signals; and (iii) Constrained Rotary Time Embedding for encoding absolute temporal information in omni-modal embeddings. We introduce a curation and synthesis pipeline that generates 24M single-modal and omni-modal conversations. We find that modalities reinforce one another in both perception and reasoning. Our model, OmniVinci, outperforms Qwen2.5-Omni with +19.05 on DailyOmni (cross-modal understanding), +1.7 on MMAR (audio), and +3.9 on Video-MME (vision), while using just 0.2T training tokens - a 6 times reduction compared to Qwen2.5-Omni's 1.2T. We finally demonstrate omni-modal advantages in downstream applications spanning robotics, medical AI, and smart factory.
Large Language Models (LLMs) have shown substantial advances through reinforcement learning (RL), particularly in domains where rewards can be programmatically verified, such as mathematics and code. In these areas, models benefit from a well-defined operational base guided by explicit rule-based objectives. However, this progress reveals a significant limitation: in open-ended domains where rewards are ambiguous, subjective, or context-dependent, such as creative writing, scientific reasoning, and notably medical consultation, robust reward functions are lacking, making these areas challenging for current RL strategies. To bridge this gap, we introduce ORBIT, an open-ended rubric-based incremental training framework specifically designed for high-stakes medical dialogue. ORBIT integrates syn- thetic dialogue generation with the dynamic creation of rubrics, employing these rubrics to direct an incremental RL process. In particular, this approach does not depend on external medical knowledge or manual rules, instead utilizing rubric-guided feedback to shape learning. When implemented on the Qwen3-4B-Instruct model, our method can greatly enhance its performance on the HealthBench-Hard benchmark from 7.0 to 27.2 using only 2k samples, thus achieving state-of-the-art results for models of this scale. Our analysis confirms that rubric-driven RL fos-ters consistent performance gains across diverse consultation scenarios, going beyond simple numerical improvements. These findings underscore rubric-based feedback as a scalable strategy for advancing LLMs in intricate, open-ended tasks.
Ideal or real - that is the question.In this work, we explore whether principles from game theory can be effectively applied to the evaluation of large language models (LLMs). This inquiry is motivated by the growing inadequacy of conventional evaluation practices, which often rely on fixed-format tasks with reference answers and struggle to capture the nuanced, subjective, and open-ended nature of modern LLM behavior. To address these challenges, we propose a novel alternative: automatic mutual evaluation, where LLMs assess each other's output through self-play and peer review. These peer assessments are then systematically compared with human voting behavior to evaluate their alignment with human judgment. Our framework incorporates game-theoretic voting algorithms to aggregate peer reviews, enabling a principled investigation into whether model-generated rankings reflect human preferences. Empirical results reveal both convergences and divergences between theoretical predictions and human evaluations, offering valuable insights into the promises and limitations of mutual evaluation. To the best of our knowledge, this is the first work to jointly integrate mutual evaluation, game-theoretic aggregation, and human-grounded validation for evaluating the capabilities of LLMs.
Recently, large language models (LLMs) have been explored for integration with collaborative filtering (CF)-based recommendation systems, which are crucial for personalizing user experiences. However, a key challenge is that LLMs struggle to interpret the latent, non-semantic embeddings produced by CF approaches, limiting recommendation effectiveness and further applications. To address this, we propose FACE, a general interpretable framework that maps CF embeddings into pre-trained LLM tokens. Specifically, we introduce a disentangled projection module to decompose CF embeddings into concept-specific vectors, followed by a quantized autoencoder to convert continuous embeddings into LLM tokens (descriptors). Then, we design a contrastive alignment objective to ensure that the tokens align with corresponding textual signals. Hence, the model-agnostic FACE framework achieves semantic alignment without fine-tuning LLMs and enhances recommendation performance by leveraging their pre-trained capabilities. Empirical results on three real-world recommendation datasets demonstrate performance improvements in benchmark models, with interpretability studies confirming the interpretability of the descriptors. Code is available in https://github.com/YixinRoll/FACE.
We introduce time-to-unsafe-sampling, a novel safety measure for generative models, defined as the number of generations required by a large language model (LLM) to trigger an unsafe (e.g., toxic) response. While providing a new dimension for prompt-adaptive safety evaluation, quantifying time-to-unsafe-sampling is challenging: unsafe outputs are often rare in well-aligned models and thus may not be observed under any feasible sampling budget. To address this challenge, we frame this estimation problem as one of survival analysis. We build on recent developments in conformal prediction and propose a novel calibration technique to construct a lower predictive bound (LPB) on the time-to-unsafe-sampling of a given prompt with rigorous coverage guarantees. Our key technical innovation is an optimized sampling-budget allocation scheme that improves sample efficiency while maintaining distribution-free guarantees. Experiments on both synthetic and real data support our theoretical results and demonstrate the practical utility of our method for safety risk assessment in generative AI models.
Deep learning (DL) frameworks serve as the backbone for a wide range of artificial intelligence applications. However, bugs within DL frameworks can cascade into critical issues in higher-level applications, jeopardizing reliability and security. While numerous techniques have been proposed to detect bugs in DL frameworks, research exploring common API patterns across frameworks and the potential risks they entail remains limited. Notably, many DL frameworks expose similar APIs with overlapping input parameters and functionalities, rendering them vulnerable to shared bugs, where a flaw in one API may extend to analogous APIs in other frameworks. To address this challenge, we propose MirrorFuzz, an automated API fuzzing solution to discover shared bugs in DL frameworks. MirrorFuzz operates in three stages: First, MirrorFuzz collects historical bug data for each API within a DL framework to identify potentially buggy APIs. Second, it matches each buggy API in a specific framework with similar APIs within and across other DL frameworks. Third, it employs large language models (LLMs) to synthesize code for the API under test, leveraging the historical bug data of similar APIs to trigger analogous bugs across APIs. We implement MirrorFuzz and evaluate it on four popular DL frameworks (TensorFlow, PyTorch, OneFlow, and Jittor). Extensive evaluation demonstrates that MirrorFuzz improves code coverage by 39.92\% and 98.20\% compared to state-of-the-art methods on TensorFlow and PyTorch, respectively. Moreover, MirrorFuzz discovers 315 bugs, 262 of which are newly found, and 80 bugs are fixed, with 52 of these bugs assigned CNVD IDs.
This research introduces a novel approach to textual and multimodal Hate Speech Detection (HSD), using Large Language Models (LLMs) as dynamic knowledge bases to generate background context and incorporate it into the input of HSD classifiers. Two context generation strategies are examined: one focused on named entities and the other on full-text prompting. Four methods of incorporating context into the classifier input are compared: text concatenation, embedding concatenation, a hierarchical transformer-based fusion, and LLM-driven text enhancement. Experiments are conducted on the textual Latent Hatred dataset of implicit hate speech and applied in a multimodal setting on the MAMI dataset of misogynous memes. Results suggest that both the contextual information and the method by which it is incorporated are key, with gains of up to 3 and 6 F1 points on textual and multimodal setups respectively, from a zero-context baseline to the highest-performing system, based on embedding concatenation.
The generate-filter-refine (iterative) paradigm based on large language models (LLMs) has achieved progress in reasoning, programming, and program discovery in AI+Science. However, the effectiveness of search depends on where to search, namely, how to encode the domain prior into an operationally structured hypothesis space. To this end, this paper proposes a compact formal theory that describes and measures LLM-assisted iterative search guided by domain priors. We represent an agent as a fuzzy relation operator on inputs and outputs to capture feasible transitions; the agent is thereby constrained by a fixed safety envelope. To describe multi-step reasoning/search, we weight all reachable paths by a single continuation parameter and sum them to obtain a coverage generating function; this induces a measure of reachability difficulty; and it provides a geometric interpretation of search on the graph induced by the safety envelope. We further provide the simplest testable inferences and validate them via a majority-vote instantiation. This theory offers a workable language and operational tools to measure agents and their search spaces, proposing a systematic formal description of iterative search constructed by LLMs.
Large Language Models (LLMs) are powerful reasoners in natural language, but their actions are typically confined to outputting vocabulary tokens. As a result, interactions with external environments -- such as symbolic operators or simulators -- must be expressed through text in predefined formats, parsed, and routed to external interfaces. This overloads the model's language with both reasoning and control duties, and requires a hand-crafted parser, external to the LLM. To address this, we decouple environment interactions from language by internalizing them in an Expanded Action space (ExpA), beyond the vocabulary. The model starts reasoning in the default language environment, but may trigger routing actions and switch to an external environment at any time. From there, the model can only invoke environment-specific actions, receive feedback from the environment, and potentially route back to language as a result. To promote effective exploration of the expanded action space and new environments, we introduce ExpA Reinforcement Learning (EARL) with counterfactual policy optimization. On tasks requiring multi-turn interactions and contingent planning, EARL outperforms strong baselines with vocabulary-constrained actions. It performs robustly across calculator-based multi-task learning and, in the partially observed sorting problem, achieves perfect Sort-4 accuracy while self-discovering an efficient algorithm competitive with classical designs.
Generation of long-form, citation-backed reports is a primary use case for retrieval augmented generation (RAG) systems. While open-source evaluation tools exist for various RAG tasks, ones tailored to report generation (RG) are lacking. Accordingly, we introduce Auto-ARGUE, a robust LLM-based implementation of the recently proposed ARGUE framework for RG evaluation. We present analysis of Auto-ARGUE on the RG pilot task from the TREC 2024 NeuCLIR track, showing good system-level correlations with human judgments. We further release a web app for visualization of Auto-ARGUE outputs.
As language models are increasingly used in scientific workflows, evaluating their ability to propose sets of explanations-not just a single correct answer-becomes critical. Many scientific problems are underdetermined: multiple, mechanistically distinct hypotheses are consistent with the same observations. We introduce HypoSpace, a diagnostic suite that treats LLMs as samplers of finite hypothesis sets and measures three complementary indicators: Validity (precision of proposals consistent with observations), Uniqueness (non-redundancy among proposals), and Recovery (coverage of the enumerated admissible set). We instantiate HypoSpace in three structured domains with deterministic validators and exactly enumerated hypothesis spaces: (i) causal graphs from perturbations, (ii) gravity-constrained 3D voxel reconstruction from top-down projections, and (iii) Boolean genetic interactions. Across instruction-tuned and reasoning-focused models, Validity often remains high while Uniqueness and Recovery degrade as the admissible space grows, revealing mode collapse that is invisible to correctness-only metrics. HypoSpace offers a controlled probe-rather than a leaderboard-for methods that explicitly explore and cover admissible explanation spaces. Code is available at: https://github.com/CTT-Pavilion/_HypoSpace.
In this paper, we present our submission for the token prediction task of EvaCun 2025. Our sys-tems are based on LLMs (Command-R, Mistral, and Aya Expanse) fine-tuned on the task data provided by the organizers. As we only pos-sess a very superficial knowledge of the subject field and the languages of the task, we simply used the training data without any task-specific adjustments, preprocessing, or filtering. We compare 3 different approaches (based on 3 different prompts) of obtaining the predictions, and we evaluate them on a held-out part of the data.
Large language models (LLMs) demonstrate strong reasoning abilities with chain-of-thought prompting, but explicit reasoning introduces substantial computational overhead. Recent work on latent reasoning reduces this cost by reasoning in latent space without explicit supervision, but performance drops significantly. Our preliminary experiments suggest that this degradation stems from the unstructured latent space, which makes fitting latent tokens difficult. To address this, we restrict the latent space to the column space of the LLM vocabulary, treating latent reasoning as a superposition over vocabulary probabilities. Once latent reasoning concludes, it collapses into an eigenstate of explicit reasoning to yield the final answer. Based on this idea, we propose Latent-SFT, a two-stage learning framework. In the first stage, we design two specialized attention masks to guide the Latent Token Encoder in generating latent tokens, allowing the LLM to produce the correct answer conditioned on them. In the second stage, the Latent Token Encoder is discarded, and the LLM is directly trained to generate these latent tokens autonomously for latent reasoning, optimized with KL and CE losses. Latent-SFT sets a new state of the art on GSM8k, matching explicit SFT performance while cutting reasoning chains by up to 4 times and outperforming prior latent methods. On Math500 and AIME24, lexical probability-based latent reasoning also clearly surpasses hidden-state-based approaches. Our metrics of effective compression rate and effective global parallelism further show that latent reasoning is both the compression of a single path and the superposition of multiple paths.
The increasing acceptance of large language models (LLMs) as an alternative to knowledge sources marks a significant paradigm shift across various domains, including time-sensitive fields such as law, healthcare, and finance. To fulfill this expanded role, LLMs must not only be factually accurate but also demonstrate consistency across temporal dimensions, necessitating robust temporal reasoning capabilities. Despite this critical requirement, efforts to ensure temporal consistency in LLMs remain scarce including noticeable absence of endeavors aimed at evaluating or augmenting LLMs across temporal references in time-sensitive inquiries. In this paper, we seek to address this gap by introducing a novel benchmark entitled temporal referential consistency, accompanied by a resource TEMP-ReCon designed to benchmark a wide range of both open-source and closed-source LLMs with various linguistic contexts characterized by differing resource richness (including English, French, and Romanian). The findings emphasis that LLMs do exhibit insufficient temporal referent consistency. To address this, we propose \newmodel, a reasoning path alignment-based model that aims to enhance the temporal referential consistency of LLMs. Our empirical experiments substantiate the efficacy of UnTRaP compared to several baseline models.
Reinforcement learning (RL) has been pivotal in enhancing the reasoning capabilities of large language models (LLMs), but it often suffers from limited exploration and entropy collapse, where models exploit a narrow set of solutions, leading to a loss of sampling diversity and subsequently preventing RL from further improving performance. This issue is exacerbated in parallel sampling methods, where multiple outputs are drawn from the same distribution, potentially causing the model to converge to similar solutions. We propose SESA, a novel SEquential SAmpling framework that mitigates this challenge by generating diverse solution sketches sequentially before expanding them into full reasoning paths. This approach ensures broader exploration by conditioning each new output on previous ones, promoting diversity throughout the process and preventing policy collapse. Our experiments on a synthetic task show that sequential sampling consistently outperforms traditional RL methods in terms of path diversity and recovery from collapse. Further evaluations on real-world tasks demonstrate that SESA improves both the exploration of valid strategies and the overall performance of LLMs. On three agent benchmarks, SESA lifts success rates by $+0.25$, $+0.42$, and $+0.07$ absolute over the base model (up to an additional $211\%$ relative improvement over baseline RL), underscoring its exploration advantage. This work introduces a structured approach to exploration, paving the way for more effective and diverse reasoning in RL-trained LLMs. Our code is released at https://github.com/MuLabPKU/sesa.
Large Language Models (LLMs) can generate code, but can they generate fast code? In this paper, we study this question using a dataset of 65 real-world tasks mined from open-source Java programs. We specifically select tasks where developers achieved significant speedups, and employ an automated pipeline to generate patches for these issues using two leading LLMs under four prompt variations. By rigorously benchmarking the results against the baseline and human-authored solutions, we demonstrate that LLM-generated code indeed improves performance over the baseline in most cases. However, patches proposed by human developers outperform LLM fixes by a statistically significant margin, indicating that LLMs often fall short of finding truly optimal solutions. We further find that LLM solutions are semantically identical or similar to the developer optimization idea in approximately two-thirds of cases, whereas they propose a more original idea in the remaining one-third. However, these original ideas only occasionally yield substantial performance gains.
Mobile agents rely on Large Language Models (LLMs) to plan and execute tasks on smartphone user interfaces (UIs). While cloud-based LLMs achieve high task accuracy, they require uploading the full UI state at every step, exposing unnecessary and often irrelevant information. In contrast, local LLMs avoid UI uploads but suffer from limited capacity, resulting in lower task success rates. We propose $\textbf{CORE}$, a $\textbf{CO}$llaborative framework that combines the strengths of cloud and local LLMs to $\textbf{R}$educe UI $\textbf{E}$xposure, while maintaining task accuracy for mobile agents. CORE comprises three key components: (1) $\textbf{Layout-aware block partitioning}$, which groups semantically related UI elements based on the XML screen hierarchy; (2) $\textbf{Co-planning}$, where local and cloud LLMs collaboratively identify the current sub-task; and (3) $\textbf{Co-decision-making}$, where the local LLM ranks relevant UI blocks, and the cloud LLM selects specific UI elements within the top-ranked block. CORE further introduces a multi-round accumulation mechanism to mitigate local misjudgment or limited context. Experiments across diverse mobile apps and tasks show that CORE reduces UI exposure by up to 55.6% while maintaining task success rates slightly below cloud-only agents, effectively mitigating unnecessary privacy exposure to the cloud. The code is available at https://github.com/Entropy-Fighter/CORE.
Test-time scaling seeks to improve the reasoning performance of large language models (LLMs) by adding computational resources. A prevalent approach within the field is sampling-based test-time scaling methods, which enhance reasoning by generating multiple reasoning paths for a given input during inference. However, despite its practical success, the theoretical foundations remain underexplored. In this paper, we provide the first theoretical framework for analyzing sampling-based test-time scaling methods, grounded in the perspective of confidence estimation. Based on the framework, we analyze two dominant paradigms: self-consistency and perplexity, and reveal key limitations: self-consistency suffers from high estimation error while perplexity exhibits substantial modeling error and possible degradation of the estimation error convergence. To address these limitations, we introduce RPC, a hybrid method that leverages our theoretical insights through two key components: Perplexity Consistency and Reasoning Pruning. Perplexity Consistency combines the strengths of self-consistency and perplexity, boosting the convergence rate of estimation error from linear to exponential while preserving model error. Reasoning Pruning prevents degradation by eliminating low-probability reasoning paths. Both theoretical analysis and empirical results across seven benchmark datasets demonstrate that RPC has a strong potential for reducing reasoning error. Notably, RPC achieves reasoning performance comparable to self-consistency while not only enhancing confidence reliability but also reducing sampling costs by 50%. The code and resources are available at https://wnjxyk.github.io/RPC.
This work investigates the resilience of contemporary LLMs against frequent and structured character-level perturbations, specifically through the insertion of noisy characters after each input character. We introduce UCC-Inj, a practical method that inserts invisible Unicode control characters into text to discourage LLM misuse in scenarios such as online exam systems. Surprisingly, despite strong obfuscation that fragments tokenization and reduces the signal-to-noise ratio significantly, many LLMs still maintain notable performance. Through comprehensive evaluation across model-, problem-, and noise-related configurations, we examine the extent and mechanisms of this robustness, exploring both the handling of character-level tokenization and implicit versus explicit denoising mechanism hypotheses of character-level noises. We hope our findings on the low-level robustness of LLMs will shed light on the risks of their misuse and on the reliability of deploying LLMs across diverse applications.
Fault cause identification in automated manufacturing lines is challenging due to the system's complexity, frequent reconfigurations, and the limited reusability of existing Failure Mode and Effects Analysis (FMEA) knowledge. Although FMEA worksheets contain valuable expert insights, their reuse across heterogeneous lines is hindered by natural language variability, inconsistent terminology, and process differences. To address these limitations, this study proposes a process-aware framework that enhances FMEA reusability by combining manufacturing-domain conceptualization with graph neural network (GNN) reasoning. First, FMEA worksheets from multiple manufacturing lines are transformed into a unified knowledge graph through ontology-guided large language model (LLM) extraction, capturing domain concepts such as actions, states, components, and parameters. Second, a Relational Graph Convolutional Network (RGCN) with the process-aware scoring function learns embeddings that respect both semantic relationships and sequential process flows. Finally, link prediction is employed to infer and rank candidate fault causes consistent with the target line's process flow. A case study on automotive pressure sensor assembly lines demonstrates that the proposed method outperforms a state-of-the-art retrieval-augmented generation (RAG) baseline (F1@20 = 0.267) and an RGCN approach (0.400), achieving the best performance (0.523) in fault cause identification. Ablation studies confirm the contributions of both LLM-driven domain conceptualization and process-aware learning. These results indicate that the proposed framework significantly improves the transferability of FMEA knowledge across heterogeneous lines, thereby supporting operators in diagnosing failures more reliably and paving the way for future domain-adaptive LLM applications in smart manufacturing.
Developing Large Language Models (LLMs) to cooperate and compete effectively within multi-agent systems is a critical step towards more advanced intelligence. While reinforcement learning (RL) has proven effective for enhancing reasoning in single-agent tasks, its extension to multi-turn, multi-agent scenarios remains underexplored due to the challenges of long-horizon credit assignment and agent-specific advantage estimation. To address these challenges, we introduce MARS, an end-to-end RL framework that incentivizes Multi-Agent Reasoning of LLMs through Self-play in both cooperative and competitive games. MARS features a turn-level advantage estimator that aligns learning signals with each interaction for credit assignment, and an agent-specific advantage normalization to stabilize multi-agent training. By learning with self-play across cooperative and competitive games, the MARS agent trained from Qwen3-4B develops strong strategic abilities that generalize to held-out games with up to 28.7% performance improvements. More importantly, the capability acquired through self-play generalizes beyond games, yielding consistent performance gains of multi-agent systems in reasoning benchmarks. When integrated into leading multi-agent systems, our MARS agent achieves significant performance gains of 10.0% on AIME and 12.5% on GPQA-Diamond. These results establish end-to-end RL training with self-play in strategic games as a powerful approach for developing generalizable multi-agent reasoning capabilities in LLMs. Our code and models are publicly available at https://github.com/thu-nics/MARS.
As LLM-based agents are increasingly deployed in real-life scenarios, existing benchmarks fail to capture their inherent complexity of handling extensive information, leveraging diverse resources, and managing dynamic user interactions. To address this gap, we introduce VitaBench, a challenging benchmark that evaluates agents on versatile interactive tasks grounded in real-world settings. Drawing from daily applications in food delivery, in-store consumption, and online travel services, VitaBench presents agents with the most complex life-serving simulation environment to date, comprising 66 tools. Through a framework that eliminates domain-specific policies, we enable flexible composition of these scenarios and tools, yielding 100 cross-scenario tasks (main results) and 300 single-scenario tasks. Each task is derived from multiple real user requests and requires agents to reason across temporal and spatial dimensions, utilize complex tool sets, proactively clarify ambiguous instructions, and track shifting user intent throughout multi-turn conversations. Moreover, we propose a rubric-based sliding window evaluator, enabling robust assessment of diverse solution pathways in complex environments and stochastic interactions. Our comprehensive evaluation reveals that even the most advanced models achieve only 30% success rate on cross-scenario tasks, and less than 50% success rate on others. Overall, we believe VitaBench will serve as a valuable resource for advancing the development of AI agents in practical real-world applications. The code, dataset, and leaderboard are available at https://vitabench.github.io/
While Speech Large Language Models (Speech-LLMs) show strong performance in many applications, their robustness is critically under-tested, especially to speech disfluency. Existing evaluations often rely on idealized inputs, overlooking common disfluencies, particularly those associated with conditions like Parkinson's disease. This work investigates whether current Speech-LLMs can maintain performance when interacting with users who have speech impairments. To facilitate this inquiry, we introduce VocalBench-DF, a framework for the systematic evaluation of disfluency across a multi-dimensional taxonomy. Our evaluation of 22 mainstream Speech-LLMs reveals substantial performance degradation, indicating that their real-world readiness is limited. Further analysis identifies phoneme-level processing and long-context modeling as primary bottlenecks responsible for these failures. Strengthening recognition and reasoning capability from components and pipelines can substantially improve robustness. These findings highlight the urgent need for new methods to improve disfluency handling and build truly inclusive Speech-LLMs
As LLMs increasingly impact safety-critical applications, ensuring their safety using guardrails remains a key challenge. This paper proposes GuardReasoner, a new safeguard for LLMs, by guiding the guard model to learn to reason. Concretely, we first create the GuardReasonerTrain dataset, which consists of 127K samples with 460K detailed reasoning steps. Then, we introduce reasoning SFT to unlock the reasoning capability of guard models. In addition, we present hard sample DPO to further strengthen their reasoning ability. In this manner, GuardReasoner achieves better performance, explainability, and generalizability. Extensive experiments and analyses on 13 benchmarks of 3 guardrail tasks demonstrate its superiority. Remarkably, GuardReasoner 8B surpasses GPT-4o+CoT by 5.74% and LLaMA Guard 3 8B by 20.84% F1 score on average. We release the training data, code, and models with different scales (1B, 3B, 8B) of GuardReasoner : https://github.com/yueliu1999/GuardReasoner/.
With the advent of artificial intelligence (AI), many researchers are attempting to extract structured information from document-level biomedical literature by fine-tuning large language models (LLMs). However, they face significant challenges such as the need for expensive hardware, like high-performance GPUs and the high labor costs associated with annotating training datasets, especially in biomedical realm. Recent research on LLMs, such as GPT-4 and Llama3, has shown promising performance in zero-shot settings, inspiring us to explore a novel approach to achieve the same results from unannotated full documents using general LLMs with lower hardware and labor costs. Our approach combines two major stages: named entity recognition (NER) and relation extraction (RE). NER identifies chemical, disease and gene entities from the document with synonym and hypernym extraction using an LLM with a crafted prompt. RE extracts relations between entities based on predefined relation schemas and prompts. To enhance the effectiveness of prompt, we propose a five-part template structure and a scenario-based prompt design principles, along with evaluation method to systematically assess the prompts. Finally, we evaluated our approach against fine-tuning and pre-trained models on two biomedical datasets: ChemDisGene and CDR. The experimental results indicate that our proposed method can achieve comparable accuracy levels to fine-tuning and pre-trained models but with reduced human and hardware expenses.
The swift spread of fake news and disinformation campaigns poses a significant threat to public trust, political stability, and cybersecurity. Traditional Cyber Threat Intelligence (CTI) approaches, which rely on low-level indicators such as domain names and social media handles, are easily evaded by adversaries who frequently modify their online infrastructure. To address these limitations, we introduce a novel CTI framework that focuses on high-level, semantic indicators derived from recurrent narratives and relationships of disinformation campaigns. Our approach extracts structured CTI indicators from unstructured disinformation content, capturing key entities and their contextual dependencies within fake news using Large Language Models (LLMs). We further introduce FakeCTI, the first dataset that systematically links fake news to disinformation campaigns and threat actors. To evaluate the effectiveness of our CTI framework, we analyze multiple fake news attribution techniques, spanning from traditional Natural Language Processing (NLP) to fine-tuned LLMs. This work shifts the focus from low-level artifacts to persistent conceptual structures, establishing a scalable and adaptive approach to tracking and countering disinformation campaigns.
Large Language Models (LLMs) have achieved impressive progress across a wide range of tasks, yet their heavy reliance on English-centric training data leads to significant performance degradation in non-English languages. While existing multilingual prompting methods emphasize reformulating queries into English or enhancing reasoning capabilities, they often fail to incorporate the language- and culture-specific grounding that is essential for some queries. To address this limitation, we propose EMCee (Extracting synthetic Multilingual Context and merging), a simple yet effective framework that enhances the multilingual capabilities of LLMs by explicitly extracting and utilizing query-relevant knowledge from the LLM itself. In particular, EMCee first extracts synthetic context to uncover latent, language-specific knowledge encoded within the LLM, and then dynamically merges this contextual insight with reasoning-oriented outputs through a judgment-based selection mechanism. Extensive experiments on four multilingual benchmarks covering diverse languages and tasks demonstrate that EMCee consistently outperforms prior approaches, achieving an average relative improvement of 16.4% overall and 31.7% in low-resource languages.
Large language models (LLMs) are increasingly deployed as agents in dynamic, real-world environments, where success requires both reasoning and effective tool use. A central challenge for agentic tasks is the growing context length, as agents must accumulate long histories of actions and observations. This expansion raises costs and reduces efficiency in long-horizon tasks, yet prior work on context compression has mostly focused on single-step tasks or narrow applications. We introduce Agent Context Optimization (ACON), a unified framework that optimally compresses both environment observations and interaction histories into concise yet informative condensations. ACON leverages compression guideline optimization in natural language space: given paired trajectories where full context succeeds but compressed context fails, capable LLMs analyze the causes of failure, and the compression guideline is updated accordingly. Furthermore, we propose distilling the optimized LLM compressor into smaller models to reduce the overhead of the additional module. Experiments on AppWorld, OfficeBench, and Multi-objective QA show that ACON reduces memory usage by 26-54% (peak tokens) while largely preserving task performance, preserves over 95% of accuracy when distilled into smaller compressors, and enhances smaller LMs as long-horizon agents with up to 46% performance improvement. Our code is available at https://github.com/microsoft/acon.
Ensembling Large Language Models (LLMs) has gained attention as a promising approach to surpass the performance of individual models by leveraging their complementary strengths. In particular, aggregating models' next-token probability distributions to select the next token has been shown to be effective in various tasks. However, while successful for short-form answers, its application to long-form generation remains underexplored. In this paper, we show that using existing ensemble methods in long-form generation requires a careful choice of ensembling positions, since the standard practice of ensembling at every token often degrades performance. We identify two key factors for determining these positions: tokenization mismatch across models and consensus in their next-token probability distributions. Based on this, we propose SAFE, (Stable And Fast LLM Ensembling), a framework that selectively ensembles by jointly considering these factors. To further improve stability, we introduce a probability sharpening strategy that consolidates probabilities spread across multiple sub-word tokens representing the same word into a single representative token. Our experiments on diverse benchmarks, including MATH500 and BBH, demonstrate that SAFE outperforms existing methods in both accuracy and efficiency, with gains achieved even when ensembling fewer than 1% of tokens.
Large language model (LLM) applications are blindfolded to the infrastructure underneath and generate tokens autoregressively, indifferent to the system load, thus risking inferencing latency inflation and poor user experience. Our first-cut controller, named beLLMan, enables the LLM infrastructure to actively and progressively signal the first-party LLM application to adjust the output length in response to changing system load. On a real testbed with H100 GPUs, beLLMan helps keep inferencing latency under control (upto 8X lower end-to-end latency) and reduces energy consumption by 25% (while serving 19% more requests) during periods of congestion for a summarization workload.
Witnessed by the recent advancements on leveraging LLM for coding and multimodal understanding, we present WebGen-V, a new benchmark and framework for instruction-to-HTML generation that enhances both data quality and evaluation granularity. WebGen-V contributes three key innovations: (1) an unbounded and extensible agentic crawling framework that continuously collects real-world webpages and can leveraged to augment existing benchmarks; (2) a structured, section-wise data representation that integrates metadata, localized UI screenshots, and JSON-formatted text and image assets, explicit alignment between content, layout, and visual components for detailed multimodal supervision; and (3) a section-level multimodal evaluation protocol aligning text, layout, and visuals for high-granularity assessment. Experiments with state-of-the-art LLMs and ablation studies validate the effectiveness of our structured data and section-wise evaluation, as well as the contribution of each component. To the best of our knowledge, WebGen-V is the first work to enable high-granularity agentic crawling and evaluation for instruction-to-HTML generation, providing a unified pipeline from real-world data acquisition and webpage generation to structured multimodal assessment.
Large Language Models (LLMs) as interactive agents show significant promise in Knowledge Graph Question Answering (KGQA) but often struggle with the semantic gap between natural language queries and structured knowledge graph (KG) representations. This leads to suboptimal planning and inefficient exploration on KG, while training-free approaches often underutilize valuable reasoning patterns in training data. To address these limitations, we propose a novel framework, Exemplar-Guided Planning (EGP), which enhances the planning capabilities of LLM agents for KGQA. EGP first preprocesses the training set questions via entity templating to normalize semantic variations. It then retrieves highly similar exemplary questions and their successful reasoning paths from this preprocessed set using semantic embeddings and an efficient FAISS index. These retrieved exemplars dynamically guide the LLM's planning process in two key phases: (1) Task Decomposition, by aligning generated sub-objectives with proven reasoning steps, and (2) Relation Exploration, by providing high-quality auxiliary information to improve relation pruning accuracy. Additionally, we introduce a Smart Lookahead mechanism during relation exploration to improve efficiency by preemptively exploring promising paths and potentially terminating exploration earlier. We apply EGP to the Plan-on-Graph (PoG) framework, termed PoG-EGP. Extensive experiments on two real-world KGQA datasets, WebQSP and CWQ, demonstrate that PoG-EGP significantly improves over the baseline PoG system and other compared methods.
LLM-integrated applications are vulnerable to prompt injection attacks, where an attacker contaminates the input to inject malicious prompts, causing the LLM to follow the attacker's intent instead of the original user's. Existing prompt injection detection methods often have sub-optimal performance and/or high computational overhead. In this work, we propose PIShield, a detection method that is both effective and efficient. Our key observation is that the internal representation of the final token in a prompt-extracted from a specific layer of the LLM, which we term the injection-critical layer-captures distinguishing features between clean and contaminated prompts. Leveraging this insight, we train a simple linear classifier on these internal representations using a labeled set of clean and contaminated prompts. We compare PIShield against 11 baselines across 5 diverse benchmark datasets and 8 prompt injection attacks. The results demonstrate that PIShield is both highly effective and efficient, substantially outperforming existing methods. Additionally, we show that PIShield resists strong adaptive attacks.
Riding on the success of LLMs with retrieval-augmented generation (RAG), there has been a growing interest in augmenting agent systems with external memory databases. However, the existing systems focus on storing text information in their memory, ignoring the importance of multimodal signals. Motivated by the multimodal nature of human memory, we present AUGUSTUS, a multimodal agent system aligned with the ideas of human memory in cognitive science. Technically, our system consists of 4 stages connected in a loop: (i) encode: understanding the inputs; (ii) store in memory: saving important information; (iii) retrieve: searching for relevant context from memory; and (iv) act: perform the task. Unlike existing systems that use vector databases, we propose conceptualizing information into semantic tags and associating the tags with their context to store them in a graph-structured multimodal contextual memory for efficient concept-driven retrieval. Our system outperforms the traditional multimodal RAG approach while being 3.5 times faster for ImageNet classification and outperforming MemGPT on the MSC benchmark.
In the current era of big data, extracting deep insights from massive, heterogeneous, and complexly associated multi-dimensional data has become a significant challenge. Large Language Models (LLMs) perform well in natural language understanding and generation, but still suffer from "hallucination" issues when processing structured knowledge and are difficult to update in real-time. Although Knowledge Graphs (KGs) can explicitly store structured knowledge, their static nature limits dynamic interaction and analytical capabilities. Therefore, this paper proposes a multi-dimensional data analysis method based on the interactions between LLM agents and KGs, constructing a dynamic, collaborative analytical ecosystem. This method utilizes LLM agents to automatically extract product data from unstructured data, constructs and visualizes the KG in real-time, and supports users in deep exploration and analysis of graph nodes through an interactive platform. Experimental results show that this method has significant advantages in product ecosystem analysis, relationship mining, and user-driven exploratory analysis, providing new ideas and tools for multi-dimensional data analysis.
Reinforcement learning with verifiable rewards (RLVR) can elicit strong reasoning in large language models (LLMs), while their performance after RLVR varies dramatically across different base models. This raises a fundamental question: what microscopic property of pre-trained models leads to this variation? To investigate, we formalize reasoning as chains of Horn clauses ("if-then" rules) built from features extracted from the LLM's latent space via cross-layer sparse autoencoders (SAEs). We estimate the transition probabilities between its features, and further categorize each rule by its semantic soundness level (e.g., strict, plausible, noisy) with an LLM. Our key discovery is that high-potential models are inherently soundness-aware: their internal probability distributions systematically shift across rules' soundness levels, becoming highly distinct for "strict" versus "noisy" rules. In contrast, weaker models are soundness-agnostic, collapsing to one distribution regardless of soundness levels. To quantify this, we introduce the Soundness-Aware Level (SAL), a microscopic metric using the Jensen-Shannon Divergence to measure the separation between these distributions. We show that SAL's predictions of post-RLVR reasoning performance follow a precise empirical law (R^2=0.87) across diverse model families (Qwen, Mistral, Llama, DeepSeek) and scales (0.5B-14B). This reveals that a model's reasoning potential is tied to its intrinsic, pre-trained ability to distinguish sound knowledge from unsound ones. These findings underscore the critical role of model pre-training in shaping reasoning and offer a practical metric grounded in the model's internal mechanisms for selecting/designing stronger base models.
In recent times, the standard practice for developing MLLMs is to feed features from vision encoder(s) into the LLM and train with natural language supervision. This approach often causes models to lean towards language comprehension and undermine the rich visual perception signals present in the data, which are critical for tasks involving spatial reasoning in the domain of embodied AI and robotics. Is it possible to optimize both at the same time? In this work, we propose VisPer-LM, the first approach that infuses visual perception knowledge from expert vision encoders into the LLM's (of an MLLM) hidden representations. We start by investigating MLLMs trained solely with natural language supervision and identify a positive correlation between the quality of visual representations within these models and their downstream performance. Given this insight, we formulate the objective during the pretraining stage in MLLMs as a coupled optimization of predictive visual embedding and next (text) token prediction. Moreover, through extensive probing, we observe improved visual representation quality due to embedding optimization, underscoring the effectiveness of our probing setup. We demonstrate that our VisPer-LM outperforms the single and multi-encoder baselines, proving our approach's superiority over explicitly feeding the corresponding features to the LLM. In particular, VisPer-LM boosts performance by an average margin of up to 2.5% on various benchmarks, with a notable improvement of 8.7% on the Depth task in CV-Bench.
Sparse Autoencoders (SAEs) can extract interpretable features from large language models (LLMs) without supervision. However, their effectiveness in downstream steering tasks is limited by the requirement for contrastive datasets or large activation storage. To address these limitations, we propose CorrSteer, which selects features by correlating sample correctness with SAE activations from generated tokens at inference time. This approach uses only inference-time activations to extract more relevant features, thereby reducing spurious correlations. It also obtains steering coefficients from average activations, automating the entire pipeline. Our method shows improved task performance on QA, bias mitigation, jailbreaking prevention, and reasoning benchmarks on Gemma-2 2B and LLaMA-3.1 8B, notably achieving a +3.3% improvement in MMLU performance with 4000 samples and a +27.2% improvement in HarmBench with only 108 samples. Selected features demonstrate semantically meaningful patterns aligned with each task's requirements, revealing the underlying capabilities that drive performance. Our work establishes correlation-based selection as an effective and scalable approach for automated SAE steering across language model applications.
The alignment of pre-trained LLMs continues to draw significant attention from both industry and academia, aiming to ensure responses that are helpful, harmless, and honest. However, identifying a point in the model's representation subspace that simultaneously satisfies all these properties remains challenging. H3Fusion addresses this challenge by introducing a mixture-of-experts (MoE)-based fusion mechanism that models alignment as a controllable drift within the subspace, guided by a drift-regularization loss to balance competing alignment dimensions. Furthermore, we formulate the alignment by finding a dual objective of harnessing the distance of generated embeddings and alignment embeddings, and introduce a gating loss by canalizing the activations on the contributing experts. Extensive evaluations of three benchmark datasets show that H3Fusion is more helpful, less harmful, and more honest in three aspects: it outperforms each individually aligned model by 11.37%, and provides stronger robustness compared to the state-of-the-art LLM ensemble approaches by 13.77% and model-merging approaches by 6.18%. Code is available at https://github.com/sftekin/h3fusion.
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.
Training large language models (LLMs) with chain-of-thought (CoT) supervision has proven effective for enhancing their reasoning abilities. However, obtaining reliable and accurate reasoning supervision remains a significant challenge. We propose a scalable method for generating a high-quality CoT supervision dataset by leveraging the determinism of program execution. Unlike existing reasoning dataset generation methods that rely on costly human annotations or error-prone LLM-generated CoT, our approach extracts verifiable, step-by-step reasoning traces from code execution and transforms them into a natural language CoT reasoning. Experiments on reasoning benchmarks across various domains show that our method effectively equips LLMs with transferable reasoning abilities across diverse tasks. Furthermore, the ablation studies validate that our method produces highly accurate reasoning data and reduces overall token length during inference by reducing meaningless repetition and overthinking.
Large language models (LLMs) often suffer from catastrophic forgetting in continual learning (CL) scenarios, where performance on previously learned tasks degrades severely while training on sequentially arriving tasks. Although pioneering CL approaches using orthogonal subspaces can mitigate task interference, they typically employ fixed budget allocation, neglecting the varying complexity across tasks and layers. Besides, recent budget-adaptive tuning methods for LLMs often adopt multi-stage paradigms that decouple optimization and budget allocation. Such decoupling results in potential misalignment, which hinders those approaches' practical application in CL scenarios. To address these limitations, we propose OA-Adapter, a novel parameter-efficient approach for continual learning in LLMs that unifies dynamic budget adaptation with orthogonal subspace learning in an end-to-end training stage. Specifically, OA-Adapter introduces a dynamic bottleneck dimension adaptation mechanism that simultaneously allocates an efficient parameter budget and optimizes task objectives without misalignment.To effectively preserve previously acquired knowledge while coordinating with the dynamic budget allocation, orthogonal constraints are applied specifically between the parameter subspace of the current task and the dynamically allocated parameter subspaces of historical tasks. Experimental results on continual learning benchmarks demonstrate that OA-Adapter outperforms state-of-the-art methods in both accuracy and parameter efficiency. OA-Adapter achieves higher average accuracy while using 58.5% fewer parameters on the standard CL benchmark, and maintains its advantages on two larger benchmarks comprising 15 tasks.
We propose a method for confidence estimation in retrieval-augmented generation (RAG) systems that aligns closely with the correctness of large language model (LLM) outputs. Confidence estimation is especially critical in high-stakes domains such as finance and healthcare, where the cost of an incorrect answer outweighs that of not answering the question. Our approach extends prior uncertainty quantification methods by leveraging raw feed-forward network (FFN) activations as auto-regressive signals, avoiding the information loss inherent in token logits and probabilities after projection and softmax normalization. We model confidence prediction as a sequence classification task, and regularize training with a Huber loss term to improve robustness against noisy supervision. Applied in a real-world financial industry customer-support setting with complex knowledge bases, our method outperforms strong baselines and maintains high accuracy under strict latency constraints. Experiments on Llama 3.1 8B model show that using activations from only the 16th layer preserves accuracy while reducing response latency. Our results demonstrate that activation-based confidence modeling offers a scalable, architecture-aware path toward trustworthy RAG deployment.
Decompiler is a specialized type of reverse engineering tool extensively employed in program analysis tasks, particularly in program comprehension and vulnerability detection. However, current Solidity smart contract decompilers face significant limitations in reconstructing the original source code. In particular, the bottleneck of SOTA decompilers lies in inaccurate method identification, incorrect variable type recovery, and missing contract attributes. These deficiencies hinder downstream tasks and understanding of the program logic. To address these challenges, we propose SmartHalo, a new framework that enhances decompiler output by combining static analysis (SA) and large language models (LLM). SmartHalo leverages the complementary strengths of SA's accuracy in control and data flow analysis and LLM's capability in semantic prediction. More specifically, \system{} constructs a new data structure - Dependency Graph (DG), to extract semantic dependencies via static analysis. Then, it takes DG to create prompts for LLM optimization. Finally, the correctness of LLM outputs is validated through symbolic execution and formal verification. Evaluation on a dataset consisting of 465 randomly selected smart contract methods shows that SmartHalo significantly improves the quality of the decompiled code, compared to SOTA decompilers (e.g., Gigahorse). Notably, integrating GPT-4o with SmartHalo further enhances its performance, achieving precision rates of 87.39% for method boundaries, 90.39% for variable types, and 80.65% for contract attributes.
Large Language Models (LLMs) are increasingly embedded in evaluative processes, from information filtering to assessing and addressing knowledge gaps through explanation and credibility judgments. This raises the need to examine how such evaluations are built, what assumptions they rely on, and how their strategies diverge from those of humans. We benchmark six LLMs against expert ratings--NewsGuard and Media Bias/Fact Check--and against human judgments collected through a controlled experiment. We use news domains purely as a controlled benchmark for evaluative tasks, focusing on the underlying mechanisms rather than on news classification per se. To enable direct comparison, we implement a structured agentic framework in which both models and nonexpert participants follow the same evaluation procedure: selecting criteria, retrieving content, and producing justifications. Despite output alignment, our findings show consistent differences in the observable criteria guiding model evaluations, suggesting that lexical associations and statistical priors could influence evaluations in ways that differ from contextual reasoning. This reliance is associated with systematic effects: political asymmetries and a tendency to confuse linguistic form with epistemic reliability--a dynamic we term epistemia, the illusion of knowledge that emerges when surface plausibility replaces verification. Indeed, delegating judgment to such systems may affect the heuristics underlying evaluative processes, suggesting a shift from normative reasoning toward pattern-based approximation and raising open questions about the role of LLMs in evaluative processes.
Search augmentation empowers Large Language Models with retrieval capabilities to overcome the limitations imposed by static parameters. Recently, Reinforcement Learning leverages tailored reward signals as a viable technique to enhance LLMs performing tasks involving search. However, existing reward modeling for search-augmented LLMs faces several limitations. Rule-based rewards, such as Exact Match, are verifiable but fragile to variations in expression and cannot be applied to long-form workloads. In contrast, generative rewards improve robustness, but designing verifiable and stable rewards for long-form workloads in dynamic corpora remains challenging and also incurs high computational costs. In this paper, we propose a unified and verifiable paradigm, "nugget-as-rubric", which treats atomic information points as structured evaluation criteria for different search-augmentation workloads. Short-form tasks correspond to a single rubric, whereas long-form tasks expand to multiple rubrics aligned with the question's information needs. To support long-form settings, we design an automatic rubric construction pipeline based on query rewriting, which can automatically retrieve passages relevant to each question and extract rubrics from them, both from static corpora and from dynamic online web content. Furthermore, we introduce \textbf{Search-Gen-V}, a 4B-parameter efficient generative verifier under our proposed verifiable paradigm, which is trained via the idea of distillation and a two-stage strategy. Experimental results show that Search-Gen-V achieves strong verification accuracy across different workloads, making it a scalable, robust, and efficient verifiable reward constructor for search-augmented LLMs.
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.
The application of Large Language Models (LLMs) in recommender systems faces key challenges in delivering deep personalization and intelligent reasoning, especially for interactive scenarios. Current methods are often constrained by limited context windows and single-turn reasoning, hindering their ability to capture dynamic user preferences and proactively reason over recommendation contexts. To address these limitations, we propose MR.Rec, a novel framework that synergizes memory and reasoning for LLM-based recommendations. To achieve personalization, we develop a comprehensive Retrieval-Augmented Generation (RAG) system that efficiently indexes and retrieves relevant external memory to enhance LLM personalization capabilities. Furthermore, to enable the synergy between memory and reasoning, our RAG system goes beyond conventional query-based retrieval by integrating reasoning enhanced memory retrieval. Finally, we design a reinforcement learning framework that trains the LLM to autonomously learn effective strategies for both memory utilization and reasoning refinement. By combining dynamic memory retrieval with adaptive reasoning, this approach ensures more accurate, context-aware, and highly personalized recommendations. Extensive experiments demonstrate that MR.Rec significantly outperforms state-of-the-art baselines across multiple metrics, validating its efficacy in delivering intelligent and personalized recommendations. We will release code and data upon paper notification.
Text-To-Speech synthesis has achieved near-human quality in neutral speech, but emotional expressiveness remains a challenge. Existing methods often rely on costly emotion annotations or optimize indirect objectives that fail to capture the emotional expressiveness and perceptual naturalness of speech, leading to generated speech that is accurate but emotionally flat. To address these challenges, we propose the RLAIF-SPA framework, incorporating a Reinforcement Learning from AI Feedback (RLAIF) mechanism to employ Automatic Speech Recognition (ASR) and Large Language Model (LLM) techniques to respectively judge semantic accuracy and prosodic-emotional label alignment as a direct reward for emotional expressiveness and intelligibility optimization. Specifically, it leverages Prosodic Label Alignment to enhance expressive quality by jointly considering semantic accuracy and prosodic-emotional alignment along four fine-grained dimensions: Structure, Emotion, Speed, and Tone. In addition, it incorporates Semantic Accuracy Feedback to ensure the generation of clear and accurate speech. Experiments on the Libri Speech dataset show that RLAIF-SPA outperforms Chat-TTS, with a 26.1% reduction in WER, a 9.1% increase in SIM-O, and over 10% improvement in human evaluation.
Video anomaly detection (VAD) is crucial for intelligent surveillance, but a significant challenge lies in identifying complex anomalies, which are events defined by intricate relationships and temporal dependencies among multiple entities rather than by isolated actions. While self-supervised learning (SSL) methods effectively model low-level spatiotemporal patterns, they often struggle to grasp the semantic meaning of these interactions. Conversely, large language models (LLMs) offer powerful contextual reasoning but are computationally expensive for frame-by-frame analysis and lack fine-grained spatial localization. We introduce HyCoVAD, Hybrid Complex Video Anomaly Detection, a hybrid SSL-LLM model that combines a multi-task SSL temporal analyzer with LLM validator. The SSL module is built upon an nnFormer backbone which is a transformer-based model for image segmentation. It is trained with multiple proxy tasks, learns from video frames to identify those suspected of anomaly. The selected frames are then forwarded to the LLM, which enriches the analysis with semantic context by applying structured, rule-based reasoning to validate the presence of anomalies. Experiments on the challenging ComplexVAD dataset show that HyCoVAD achieves a 72.5% frame-level AUC, outperforming existing baselines by 12.5% while reducing LLM computation. We release our interaction anomaly taxonomy, adaptive thresholding protocol, and code to facilitate future research in complex VAD scenarios.
A new trend uses LLMs as dense text encoders via contrastive learning. However, since LLM embeddings predict the probability distribution of the next token, they are inherently generative and distributive, conflicting with contrastive learning, which requires embeddings to capture full-text semantics and align via cosine similarity. This discrepancy hinders the full utilization of LLMs' pre-training capabilities, resulting in inefficient learning. In response to this issue, we propose AutoRegEmbed, a new contrastive learning method built on embedding conditional probability distributions, which integrates two core tasks: information compression and conditional distribution alignment. The information compression task encodes text into the embedding space, ensuring that the embedding vectors capture global semantics. The conditional distribution alignment task focuses on aligning text embeddings with positive samples embeddings by leveraging the conditional distribution of embeddings while simultaneously reducing the likelihood of generating negative samples from text embeddings, thereby achieving embedding alignment and uniformity. Experimental results demonstrate that our method significantly outperforms traditional contrastive learning approaches and achieves performance comparable to state-of-the-art models when using the same amount of data.
Knowledge graphs (KGs), with their structured representation capabilities, offer promising avenue for enhancing Retrieval Augmented Generation (RAG) systems, leading to the development of KG-RAG systems. Nevertheless, existing methods often struggle to achieve effective synergy between system effectiveness and cost efficiency, leading to neither unsatisfying performance nor excessive LLM prompt tokens and inference time. To this end, this paper proposes REMINDRAG, which employs an LLM-guided graph traversal featuring node exploration, node exploitation, and, most notably, memory replay, to improve both system effectiveness and cost efficiency. Specifically, REMINDRAG memorizes traversal experience within KG edge embeddings, mirroring the way LLMs "memorize" world knowledge within their parameters, but in a train-free manner. We theoretically and experimentally confirm the effectiveness of REMINDRAG, demonstrating its superiority over existing baselines across various benchmark datasets and LLM backbones. Our code is available at https://github.com/kilgrims/ReMindRAG.
Recent trends in LLMs development clearly show growing interest in the use and application of sovereign LLMs. The global debate over sovereign LLMs highlights the need for governments to develop their LLMs, tailored to their unique socio-cultural and historical contexts. However, there remains a shortage of frameworks and datasets to verify two critical questions: (1) how well these models align with users' socio-cultural backgrounds, and (2) whether they maintain safety and technical robustness without exposing users to potential harms and risks. To address this gap, we construct a new dataset and introduce an analytic framework for extracting and evaluating the socio-cultural elements of sovereign LLMs, alongside assessments of their technical robustness. Our experimental results demonstrate that while sovereign LLMs play a meaningful role in supporting low-resource languages, they do not always meet the popular claim that these models serve their target users well. We also show that pursuing this untested claim may lead to underestimating critical quality attributes such as safety. Our study suggests that advancing sovereign LLMs requires a more extensive evaluation that incorporates a broader range of well-grounded and practical criteria.
Recent LLM agents have made great use of chain of thought reasoning and function calling. As their capabilities grow, an important question arises: can this software represent not only a smart problem-solving tool, but an entity in its own right, that can plan, design immediate tasks, and reason toward broader, more ambiguous goals? To study this question, we adopt an open-ended experimental setting where we augment a pretrained LLM agent with the ability to generate its own tasks, accumulate knowledge, and interact extensively with its environment. We study the resulting open-ended agent qualitatively. It can reliably follow complex multi-step instructions, store and reuse information across runs, and propose and solve its own tasks, though it remains sensitive to prompt design, prone to repetitive task generation, and unable to form self-representations. These findings illustrate both the promise and current limits of adapting pretrained LLMs toward open-endedness, and point to future directions for training agents to manage memory, explore productively, and pursue abstract long-term goals.
Logical reasoning with large language models (LLMs) has received growing attention. One mainstream approach translates natural language into formal logic and then applies symbolic solvers for deduction. While effective in many tasks, these LLM-based translators often fail to generate consistent symbolic representations when the same concept appears in different linguistic forms. Such inconsistencies break logical coherence and lead to solver errors. However, most existing benchmarks lack this type of linguistic variation, which frequently occurs in real-world text, leaving the problem underexplored. To address this gap, we present SoLT, a benchmark that systematically rewrites reasoning datasets into diverse yet logically equivalent forms across multiple levels. Beyond evaluation, SoLT also provides a general method to enrich any dataset with linguistic diversity while preserving both meaning and logic. To further enhance the stability of LLM-based reasoning, we propose MenTaL, which explicitly guides models to build a concept-symbol mapping table during translation. By linking equivalent expressions to shared symbols, MenTaL maintains consistency and mitigates symbol drift. Experiments on SoLT demonstrate that LLMs indeed suffer from inconsistent symbol mapping under linguistic variation, leading to significant drops in reasoning accuracy. Meanwhile, applying MenTaL brings clear and stable performance improvements across diverse inputs. Overall, our findings reveal that overlooking linguistic diversity hides key weaknesses in LLM-based translators, and our work offers a step toward more reliable logical reasoning in varied real-world scenarios. Our code is available at https://github.com/wufeiwuwoshihua/LinguDiver.
Software supply-chain attacks are an important and ongoing concern in the open source software ecosystem. These attacks maintain the standard functionality that a component implements, but additionally hide malicious functionality activated only when the component reaches its target environment. Lexo addresses such stealthy attacks by automatically learning and regenerating vulnerability-free versions of potentially malicious components. Lexo first generates a set of input-output pairs to model a component's full observable behavior, which it then uses to synthesize a new version of the original component. The new component implements the original functionality but avoids stealthy malicious behavior. Throughout this regeneration process, Lexo consults several distinct instances of Large Language Models (LLMs), uses correctness and coverage metrics to shepherd these instances, and guardrails their results. Our evaluation on 100+ real-world packages, including high profile stealthy supply-chain attacks, indicates that Lexo scales across multiple domains, regenerates code efficiently (<100s on average), maintains compatibility, and succeeds in eliminating malicious code in several real-world supply-chain-attacks, even in cases when a state-of-the-art LLM fails to eliminate malicious code when prompted to do so.
We introduce Drivelology, a unique linguistic phenomenon characterised as "nonsense with depth" - utterances that are syntactically coherent yet pragmatically paradoxical, emotionally loaded, or rhetorically subversive. While such expressions may resemble surface-level nonsense, they encode implicit meaning requiring contextual inference, moral reasoning, or emotional interpretation. We find that current large language models (LLMs), despite excelling at many natural language processing (NLP) tasks, consistently fail to grasp the layered semantics of Drivelological text. To investigate this, we construct a benchmark dataset of over 1,200+ meticulously curated and diverse examples across English, Mandarin, Spanish, French, Japanese, and Korean. Each example underwent careful expert review to verify its Drivelological characteristics, involving multiple rounds of discussion and adjudication to address disagreements. Using this dataset, we evaluate a range of LLMs on classification, generation, and reasoning tasks. Our results reveal clear limitations of LLMs: models often confuse Drivelology with shallow nonsense, produce incoherent justifications, or miss implied rhetorical functions altogether. These findings highlight a deep representational gap in LLMs' pragmatic understanding and challenge the assumption that statistical fluency implies cognitive comprehension. We release our dataset and code to facilitate further research in modelling linguistic depth beyond surface-level coherence.
The obligatory use of third-person honorifics is a distinctive feature of several South Asian languages, encoding nuanced socio-pragmatic cues such as power, age, gender, fame, and social distance. In this work, (i) We present the first large-scale study of third-person honorific pronoun and verb usage across 10,000 Hindi and Bengali Wikipedia articles with annotations linked to key socio-demographic attributes of the subjects, including gender, age group, fame, and cultural origin. (ii) Our analysis uncovers systematic intra-language regularities but notable cross-linguistic differences: honorifics are more prevalent in Bengali than in Hindi, while non-honorifics dominate while referring to infamous, juvenile, and culturally exotic entities. Notably, in both languages, and more prominently in Hindi, men are more frequently addressed with honorifics than women. (iii) To examine whether large language models (LLMs) internalize similar socio-pragmatic norms, we probe six LLMs using controlled generation and translation tasks over 1,000 culturally balanced entities. We find that LLMs diverge from Wikipedia usage, exhibiting alternative preferences in honorific selection across tasks, languages, and socio-demographic attributes. These discrepancies highlight gaps in the socio-cultural alignment of LLMs and open new directions for studying how LLMs acquire, adapt, or distort social-linguistic norms. Our code and data are publicly available at https://github.com/souro/honorific-wiki-llm
While Neural Network pruning typically requires retraining the model to recover pruning-induced performance degradation, state-of-the-art Large Language Models (LLMs) pruning methods instead solve a layer-wise mask selection and reconstruction problem on a small set of calibration data to avoid full retraining, as it is considered computationally infeasible for LLMs. Reconstructing single matrices in isolation has favorable properties, such as convexity of the objective and significantly reduced memory requirements compared to full retraining. In practice, however, reconstruction is often implemented at coarser granularities, e.g., reconstructing a whole transformer block against its dense activations instead of a single matrix. In this work, we study the key design choices when reconstructing or retraining the remaining weights after pruning. We conduct an extensive computational study on state-of-the-art GPT architectures, and report several surprising findings that challenge common intuitions about retraining after pruning. In particular, we observe a free lunch scenario: reconstructing attention and MLP components separately within each transformer block is nearly the most resource-efficient yet achieves the best perplexity. Most importantly, this Pareto-optimal setup achieves better performance than full retraining, despite requiring only a fraction of the memory. Furthermore, we demonstrate that simple and efficient pruning criteria such as Wanda can outperform much more complex approaches when the reconstruction step is properly executed, highlighting its importance. Our findings challenge the narrative that retraining should be avoided at all costs and provide important insights into post-pruning performance recovery for LLMs.
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.
The integration of Large Language Models (LLMs) into software engineering has driven a transition from traditional rule-based systems to autonomous agentic systems capable of solving complex problems. However, systematic progress is hindered by a lack of comprehensive understanding of how benchmarks and solutions interconnect. This survey addresses this gap by providing the first holistic analysis of LLM-powered software engineering, offering insights into evaluation methodologies and solution paradigms. We review over 150 recent papers and propose a taxonomy along two key dimensions: (1) Solutions, categorized into prompt-based, fine-tuning-based, and agent-based paradigms, and (2) Benchmarks, including tasks such as code generation, translation, and repair. Our analysis highlights the evolution from simple prompt engineering to sophisticated agentic systems incorporating capabilities like planning, reasoning, memory mechanisms, and tool augmentation. To contextualize this progress, we present a unified pipeline illustrating the workflow from task specification to deliverables, detailing how different solution paradigms address various complexity levels. Unlike prior surveys that focus narrowly on specific aspects, this work connects 50+ benchmarks to their corresponding solution strategies, enabling researchers to identify optimal approaches for diverse evaluation criteria. We also identify critical research gaps and propose future directions, including multi-agent collaboration, self-evolving systems, and formal verification integration. This survey serves as a foundational guide for advancing LLM-driven software engineering. We maintain a GitHub repository that continuously updates the reviewed and related papers at https://github.com/lisaGuojl/LLM-Agent-SE-Survey.
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.
A growing body of multi-agent studies with Large Language Models (LLMs) explores how norms and cooperation emerge in mixed-motive scenarios, where pursuing individual gain can undermine the collective good. While prior work has explored these dynamics in both richly contextualized simulations and simplified game-theoretic environments, most LLM systems featuring common-pool resource (CPR) games provide agents with explicit reward functions directly tied to their actions. In contrast, human cooperation often emerges without full visibility into payoffs and population, relying instead on heuristics, communication, and punishment. We introduce a CPR simulation framework that removes explicit reward signals and embeds cultural-evolutionary mechanisms: social learning (adopting strategies and beliefs from successful peers) and norm-based punishment, grounded in Ostrom's principles of resource governance. Agents also individually learn from the consequences of harvesting, monitoring, and punishing via environmental feedback, enabling norms to emerge endogenously. We establish the validity of our simulation by reproducing key findings from existing studies on human behavior. Building on this, we examine norm evolution across a $2\times2$ grid of environmental and social initialisations (resource-rich vs. resource-scarce; altruistic vs. selfish) and benchmark how agentic societies comprised of different LLMs perform under these conditions. Our results reveal systematic model differences in sustaining cooperation and norm formation, positioning the framework as a rigorous testbed for studying emergent norms in mixed-motive LLM societies. Such analysis can inform the design of AI systems deployed in social and organizational contexts, where alignment with cooperative norms is critical for stability, fairness, and effective governance of AI-mediated environments.
Large language model (LLM) inference systems face a fundamental tension between minimizing Time-to-First-Token (TTFT) latency for new requests and maintaining a high, steady token generation rate (low Time-Per-Output-Token, or TPOT) for ongoing requests. Existing stall-free batching schedulers proposed by Sarathi, while effective at preventing decode stalls, introduce significant computational unfairness. They prioritize decode tasks excessively, simultaneously leading to underutilized decode slack and unnecessary prefill queuing delays, which collectively degrade the system's overall quality of service (QoS). This work identifies the root cause of this unfairness: the non-monotonic nature of Time-Between-Tokens (TBT) as a scheduling metric and the rigid decode-prioritizing policy that fails to adapt to dynamic workload bursts. We therefore propose FairBatching, a novel LLM inference scheduler that enforces fair resource allocation between prefill and decode tasks. It features an adaptive batch capacity determination mechanism, which dynamically adjusts the computational budget to improve the GPU utilization without triggering SLO violations. Its fair and dynamic batch formation algorithm breaks away from the decode-prioritizing paradigm, allowing computation resources to be reclaimed from bursting decode tasks to serve prefill surges, achieving global fairness. Furthermore, FairBatching provides a novel load estimation method, enabling more effective coordination with upper-level schedulers. Implemented and evaluated on realistic traces, FairBatching significantly reduces TTFT tail latency by up to 2.29x while robustly maintaining TPOT SLOs, achieving overall 20.0% improvement in single-node capacity and 54.3% improvement in cluster-level capacity.
Math reasoning has been one crucial ability of large language models (LLMs), where significant advancements have been achieved in recent years. However, most efforts focus on LLMs by curating high-quality annotation data and intricate training (or inference) paradigms, while the math reasoning performance of multi-modal LLMs (MLLMs) remains lagging behind. Since the MLLM typically consists of an LLM and a vision block, we wonder: Can MLLMs directly absorb math reasoning abilities from off-the-shelf math LLMs without tuning? Recent model-merging approaches may offer insights into this question. However, they overlook the alignment between the MLLM and LLM, where we find that there is a large gap between their parameter spaces, resulting in lower performance. Our empirical evidence reveals two key factors behind this issue: the identification of crucial reasoning-associated layers in the model and the mitigation of the gaps in parameter space. Based on the empirical insights, we propose IP-Merging that first identifies the reasoning-associated parameters in both MLLM and Math LLM, then projects them into the subspace of MLLM, aiming to maintain the alignment, and finally merges parameters in this subspace. IP-Merging is a tuning-free approach since parameters are directly adjusted. Extensive experiments demonstrate that our IP-Merging method can enhance the math reasoning ability of MLLMs directly from Math LLMs without compromising their other capabilities.
Large language model (LLM) systems now underpin everyday AI applications such as chatbots, computer-use assistants, and autonomous robots, where performance often depends on carefully designed prompts. LLM-based prompt optimizers reduce that effort by iteratively refining prompts from scored feedback, yet the security of this optimization stage remains underexamined. We present the first systematic analysis of poisoning risks in LLM-based prompt optimization. Using HarmBench, we find systems are substantially more vulnerable to manipulated feedback than to injected queries: feedback-based attacks raise attack success rate (ASR) by up to $\Delta$ASR = 0.48. We introduce a simple fake-reward attack that requires no access to the reward model and significantly increases vulnerability, and we propose a lightweight highlighting defense that reduces the fake-reward $\Delta$ASR from 0.23 to 0.07 without degrading utility. These results establish prompt optimization pipelines as a first-class attack surface and motivate stronger safeguards for feedback channels and optimization frameworks.
Proprietary large language models (LLMs) exhibit strong generalization capabilities across diverse tasks and are increasingly deployed on edge devices for efficiency and privacy reasons. However, deploying proprietary LLMs at the edge without adequate protection introduces critical security threats. Attackers can extract model weights and architectures, enabling unauthorized copying and misuse. Even when protective measures prevent full extraction of model weights, attackers may still perform advanced attacks, such as fine-tuning, to further exploit the model. Existing defenses against these threats typically incur significant computational and communication overhead, making them impractical for edge deployment. To safeguard the edge-deployed LLMs, we introduce CoreGuard, a computation- and communication-efficient protection method. CoreGuard employs an efficient protection protocol to reduce computational overhead and minimize communication overhead via a propagation protocol. Extensive experiments show that CoreGuard achieves upper-bound security protection with negligible overhead.
This work investigates the resilience of contemporary LLMs against frequent and structured character-level perturbations, specifically through the insertion of noisy characters after each input character. We introduce \nameshort{}, a practical method that inserts invisible Unicode control characters into text to discourage LLM misuse in scenarios such as online exam systems. Surprisingly, despite strong obfuscation that fragments tokenization and reduces the signal-to-noise ratio significantly, many LLMs still maintain notable performance. Through comprehensive evaluation across model-, problem-, and noise-related configurations, we examine the extent and mechanisms of this robustness, exploring both the handling of character-level tokenization and \textit{implicit} versus \textit{explicit} denoising mechanism hypotheses of character-level noises. We hope our findings on the low-level robustness of LLMs will shed light on the risks of their misuse and on the reliability of deploying LLMs across diverse applications.
Detecting biases in natural language understanding (NLU) for African American Vernacular English (AAVE) is crucial to developing inclusive natural language processing (NLP) systems. To address dialect-induced performance discrepancies, we introduce AAVENUE ({AAVE} {N}atural Language {U}nderstanding {E}valuation), a benchmark for evaluating large language model (LLM) performance on NLU tasks in AAVE and Standard American English (SAE). AAVENUE builds upon and extends existing benchmarks like VALUE, replacing deterministic syntactic and morphological transformations with a more flexible methodology leveraging LLM-based translation with few-shot prompting, improving performance across our evaluation metrics when translating key tasks from the GLUE and SuperGLUE benchmarks. We compare AAVENUE and VALUE translations using five popular LLMs and a comprehensive set of metrics including fluency, BARTScore, quality, coherence, and understandability. Additionally, we recruit fluent AAVE speakers to validate our translations for authenticity. Our evaluations reveal that LLMs consistently perform better on SAE tasks than AAVE-translated versions, underscoring inherent biases and highlighting the need for more inclusive NLP models. We have open-sourced our source code on GitHub and created a website to showcase our work at https://aavenuee.github.io.
Large language models (LLMs) can be prompted with specific styles (e.g., formatting responses as lists), including in malicious queries. Prior jailbreak research mainly augments these queries with additional string transformations to maximize attack success rate (ASR). However, the impact of style patterns in the original queries that are semantically irrelevant to the malicious intent remains unclear. In this work, we seek to understand whether style patterns compromise LLM safety, how superficial style alignment increases model vulnerability, and how best to mitigate these risks during alignment. We first define ASR inflation as the increase in ASR due to style patterns in existing jailbreak benchmark queries. By evaluating 32 LLMs across seven benchmarks, we find that nearly all models exhibit ASR inflation. Notably, the inflation correlates with an LLM's relative attention to style patterns, which also overlap more with its instruction-tuning data when inflation occurs. We then investigate superficial style alignment, and find that fine-tuning with specific styles makes LLMs more vulnerable to jailbreaks of those same styles. Finally, we propose SafeStyle, a defense strategy that incorporates a small amount of safety training data augmented to match the distribution of style patterns in the fine-tuning data. Across three LLMs, six fine-tuning style settings, and two real-world instruction-tuning datasets, SafeStyle consistently outperforms baselines in maintaining LLM safety.
Mobile telecommunication networks are foundational to global infrastructure and increasingly support critical sectors such as manufacturing, transportation, and healthcare. The security and reliability of these networks are essential, yet depend heavily on accurate modeling of underlying protocols through state machines. While most prior work constructs such models manually from 3GPP specifications, this process is labor-intensive, error-prone, and difficult to maintain due to the complexity and frequent updates of the specifications. Recent efforts using natural language processing have shown promise, but remain limited in handling the scale and intricacy of cellular protocols. In this work, we propose SpecGPT, a novel framework that leverages large language models (LLMs) to automatically extract protocol state machines from 3GPP documents. SpecGPT segments technical specifications into meaningful paragraphs, applies domain-informed prompting with chain-of-thought reasoning, and employs ensemble methods to enhance output reliability. We evaluate SpecGPT on three representative 5G protocols (NAS, NGAP, and PFCP) using manually annotated ground truth, and show that it outperforms existing approaches, demonstrating the effectiveness of LLMs for protocol modeling at scale.
Bayesian Optimization (BO) has been widely used to efficiently optimize expensive black-box functions with limited evaluations. In this paper, we investigate the use of BO for prompt engineering to enhance text classification with Large Language Models (LLMs). We employ an LLM-powered Gaussian Process (GP) as the surrogate model to estimate the performance of different prompt candidates. These candidates are generated by an LLM through the expansion of a set of seed prompts and are subsequently evaluated using an Upper Confidence Bound (UCB) acquisition function in conjunction with the GP posterior. The optimization process iteratively refines the prompts based on a subset of the data, aiming to improve classification accuracy while reducing the number of API calls by leveraging the prediction uncertainty of the LLM-based GP. The proposed BO-LLM algorithm is evaluated on two datasets, and its advantages are discussed in detail in this paper.
We seek algorithms for program learning that are both sample-efficient and computationally feasible. Classical results show that targets admitting short program descriptions (e.g., with short ``python code'') can be learned with a ``small'' number of examples (scaling with the size of the code) via length-first program enumeration, but the search is exponential in description length. Consequently, Gradient-based training avoids this cost yet can require exponentially many samples on certain short-program families. To address this gap, we introduce LLM-ERM, a propose-and-verify framework that replaces exhaustive enumeration with an LLM-guided search over candidate programs while retaining ERM-style selection on held-out data. Specifically, we draw $k$ candidates with a pretrained reasoning-augmented LLM, compile and check each on the data, and return the best verified hypothesis, with no feedback, adaptivity, or gradients. Theoretically, we show that coordinate-wise online mini-batch SGD requires many samples to learn certain short programs. {\em Empirically, LLM-ERM solves tasks such as parity variants, pattern matching, and primality testing with as few as 200 samples, while SGD-trained transformers overfit even with 100,000 samples}. These results indicate that language-guided program synthesis recovers much of the statistical efficiency of finite-class ERM while remaining computationally tractable, offering a practical route to learning succinct hypotheses beyond the reach of gradient-based training.
Why do we build local large language models (LLMs)? What should a local LLM learn from the target language? Which abilities can be transferred from other languages? Do language-specific scaling laws exist? To explore these research questions, we evaluated 35 Japanese, English, and multilingual LLMs on 19 evaluation benchmarks for Japanese and English, taking Japanese as a local language. Adopting an observational approach, we analyzed correlations of benchmark scores, and conducted principal component analysis (PCA) on the scores to derive \textit{ability factors} of local LLMs. We found that training on English text can improve the scores of academic subjects in Japanese (JMMLU). In addition, it is unnecessary to specifically train on Japanese text to enhance abilities for solving Japanese code generation, arithmetic reasoning, commonsense, and reading comprehension tasks. In contrast, training on Japanese text could improve question-answering tasks about Japanese knowledge and English-Japanese translation, which indicates that abilities for solving these two tasks can be regarded as \textit{Japanese abilities} for LLMs. Furthermore, we confirmed that the Japanese abilities scale with the computational budget for Japanese text.
Large language models (LLMs) have recently shown strong reasoning abilities in domains like mathematics, coding, and scientific problem-solving, yet their potential for ranking tasks, where prime examples include retrieval, recommender systems, and LLM routing, remains underexplored. Ranking requires complex reasoning across heterogeneous candidates, but existing LLM-based rankers are often domain-specific, tied to fixed backbones, and lack iterative refinement, limiting their ability to fully exploit LLMs' reasoning potential. To address these challenges, we propose R1-Ranker, a reasoning-incentive framework built on reinforcement learning, with two complementary designs: DRanker, which generates full rankings in one shot, and IRanker, which decomposes ranking into an iterative elimination process with step-wise rewards to encourage deeper reasoning. We evaluate unified R1-Rankers on nine datasets spanning recommendation, routing, and passage ranking, showing that IRanker-3B consistently achieves state-of-the-art performance, surpasses larger 7B models on some tasks, and yields a 15.7% average relative improvement. Ablation and generalization experiments further confirm the critical role of reinforcement learning and iterative reasoning, with IRanker-3B improving zero-shot performance by over 9% on out-of-domain tasks and reasoning traces boosting other LLMs by up to 22.87%. These results demonstrate that unifying diverse ranking tasks with a single reasoning-driven foundation model is both effective and essential for advancing LLM reasoning in ranking scenarios.