llm - 2025_12
Navigation
Papers
By incorporating visual inputs, Multimodal Large Language Models (MLLMs) extend LLMs to support visual reasoning. However, this integration also introduces new vulnerabilities, making MLLMs susceptible to multimodal jailbreak attacks and hindering their safe deployment.Existing defense methods, including Image-to-Text Translation, Safe Prompting, and Multimodal Safety Tuning, attempt to address this by aligning multimodal inputs with LLMs' built-in safeguards.Yet, they fall short in uncovering root causes of multimodal vulnerabilities, particularly how harmful multimodal tokens trigger jailbreak in MLLMs? Consequently, they remain vulnerable to text-driven multimodal jailbreaks, often exhibiting overdefensive behaviors and imposing heavy training overhead.To bridge this gap, we present an comprehensive analysis of where, how and which harmful multimodal tokens bypass safeguards in MLLMs. Surprisingly, we find that less than 1% tokens in early-middle layers are responsible for inducing unsafe behaviors, highlighting the potential of precisely removing a small subset of harmful tokens, without requiring safety tuning, can still effectively improve safety against jailbreaks. Motivated by this, we propose Safe Prune-then-Restore (SafePTR), an training-free defense framework that selectively prunes harmful tokens at vulnerable layers while restoring benign features at subsequent layers.Without incurring additional computational overhead, SafePTR significantly enhances the safety of MLLMs while preserving efficiency. Extensive evaluations across three MLLMs and five benchmarks demonstrate SafePTR's state-of-the-art performance in mitigating jailbreak risks without compromising utility.
Large Language Models (LLMs) have demonstrated substantial progress in task automation and natural language understanding. However, without domain expertise in geographic information science (GIS), they continue to encounter limitations including reduced accuracy and unstable performance when processing complex tasks. To address these challenges, we propose GeoJSON Agents-a novel multi-agent LLM architecture specifically designed for geospatial analysis. This framework transforms natural language instructions into structured GeoJSON operations through two LLM enhancement techniques: Function Calling and Code Generation. The architecture integrates three core components: task parsing, agent collaboration, and result integration. The Planner agent systematically decomposes user-defined tasks into executable subtasks, while Worker agents perform spatial data processing and analysis either by invoking predefined function APIs or by generating and executing Python-based analytical code. The system produces reusable, standards-compliant GeoJSON outputs through iterative refinement. To evaluate both approaches, we constructed a benchmark comprising 70 tasks spanning basic, intermediate, and advanced complexity levels, conducting experiments with OpenAI's GPT-4o as the core model. Results indicate that the Code Generation-based agent achieved 97.14% accuracy, while the Function Calling-based agent attained 85.71%-both significantly outperforming the best-performing general-purpose model (48.57%). Comparative analysis reveals Code Generation offers superior flexibility for complex, open-ended tasks, whereas Function Calling provides enhanced execution stability for structured operations. This study represents the first systematic integration of GeoJSON data with a multi-agent LLM framework and provides empirical evidence comparing two mainstream enhancement methodologies in geospatial context.
While the reasoning capabilities of Large Language Models (LLMs) excel in analytical tasks such as mathematics and code generation, their utility for abstractive summarization remains widely assumed but largely unverified. To bridge this gap, we first tailor general reasoning strategies to the summarization domain. We then conduct a systematic, large scale comparative study of 8 reasoning strategies and 3 Large Reasoning Models (LRMs) across 8 diverse datasets, assessing both summary quality and faithfulness. Our findings show that reasoning is not a universal solution and its effectiveness is highly dependent on the specific strategy and context. Specifically, we observe a trade-off between summary quality and factual faithfulness: explicit reasoning strategies tend to improve fluency at the expense of factual grounding, while implicit reasoning in LRMs exhibits the inverse pattern. Furthermore, increasing an LRM's internal reasoning budget does not improve, and can even hurt, factual consistency, suggesting that effective summarization demands faithful compression rather than creative over-thinking.
We present SocraticAI, a scaffolded AI tutoring system that integrates large language models (LLMs) into undergraduate Computer Science education through structured constraints rather than prohibition. The system enforces well-formulated questions, reflective engagement, and daily usage limits while providing Socratic dialogue scaffolds. Unlike traditional AI bans, our approach cultivates responsible and strategic AI interaction skills through technical guardrails, including authentication, query validation, structured feedback, and RAG-based course grounding. Initial deployment demonstrates that students progress from vague help-seeking to sophisticated problem decomposition within 2-3 weeks, with over 75% producing substantive reflections and displaying emergent patterns of deliberate, strategic AI use.
Large language model (LLM)-based agents struggle to generalize to novel and complex environments, such as unseen websites or new sets of functions, due to a fundamental mismatch between their pre-training and test-time conditions. This challenge stems from two distinct failure modes: a syntactic misunderstanding of environment-specific components like observation formats, and a semantic misunderstanding of state-transition dynamics, which are only revealed at test time. To address these issues, we propose two distinct and complementary strategies for adapting LLM agents by leveraging environment-specific information available during deployment. First, an online distributional adaptation method parameterizes environmental nuances by learning a lightweight adaptation vector that biases the model's output distribution, enabling rapid alignment with an environment response format. Second, a deployment-time dynamics grounding method employs a persona-driven exploration phase to systematically probe and learn the environment's causal dynamics before task execution, equipping the agent with a nonparametric world model. We evaluate these strategies across diverse agentic benchmarks, including function calling and web navigation. Our empirical results show the effectiveness of both strategies across all benchmarks with minimal computational cost. We find that dynamics grounding is particularly effective in complex environments where unpredictable dynamics pose a major obstacle, demonstrating a robust path toward more generalizable and capable LLM-based agents. For example, on the WebArena multi-site split, this method increases the agent's success rate from 2% to 23%.
Deep learning methods have significantly enhanced motion planning for robotic manipulators by leveraging prior experiences within planning datasets. However, state-of-the-art neural motion planners are primarily trained on small datasets collected in manually generated workspaces, limiting their generalizability to out-of-distribution scenarios. Additionally, these planners often rely on monolithic network architectures that struggle to encode critical planning information. To address these challenges, we introduce Motion Policy with Dataset Synthesis powered by large language models (LLMs) and Fusion Action-Chunking Transformers (PerFACT), which incorporates two key components. Firstly, a novel LLM-powered workspace generation method, MotionGeneralizer, enables large-scale planning data collection by producing a diverse set of semantically feasible workspaces. Secondly, we introduce Fusion Motion Policy Networks (MpiNetsFusion), a generalist neural motion planner that uses a fusion action-chunking transformer to better encode planning signals and attend to multiple feature modalities. Leveraging MotionGeneralizer, we collect 3.5M trajectories to train and evaluate MpiNetsFusion against state-of-the-art planners, which shows that the proposed MpiNetsFusion can plan several times faster on the evaluated tasks.
The application of large language models (LLMs) in recommendation systems has recently gained traction. Traditional recommendation systems often lack explainability and suffer from issues such as popularity bias. Previous research has also indicated that LLMs, when used as standalone predictors, fail to achieve accuracy comparable to traditional models. To address these challenges, we propose to use LLM as an explainable re-ranker, a hybrid approach that combines traditional recommendation models with LLMs to enhance both accuracy and interpretability. We constructed a dataset to train the re-ranker LLM and evaluated the alignment between the generated dataset and human expectations. Leveraging a two-stage training process, our model significantly improved NDCG, a key ranking metric. Moreover, the re-ranker outperformed a zero-shot baseline in ranking accuracy and interpretability. These results highlight the potential of integrating traditional recommendation models with LLMs to address limitations in existing systems and pave the way for more explainable and fair recommendation frameworks.
Large language models (LLMs) excel at explicit reasoning, but their implicit computational strategies remain underexplored. Decades of psychophysics research show that humans intuitively process and integrate noisy signals using near-optimal Bayesian strategies in perceptual tasks. We ask whether LLMs exhibit similar behaviour and perform optimal multimodal integration without explicit training or instruction. Adopting the psychophysics paradigm, we infer computational principles of LLMs from systematic behavioural studies. We introduce a behavioural benchmark - BayesBench: four magnitude estimation tasks (length, location, distance, and duration) over text and image, inspired by classic psychophysics, and evaluate a diverse set of nine LLMs alongside human judgments for calibration. Through controlled ablations of noise, context, and instruction prompts, we measure performance, behaviour and efficiency in multimodal cue-combination. Beyond accuracy and efficiency metrics, we introduce a Bayesian Consistency Score that detects Bayes-consistent behavioural shifts even when accuracy saturates. Our results show that while capable models often adapt in Bayes-consistent ways, accuracy does not guarantee robustness. Notably, GPT-5 Mini achieves perfect text accuracy but fails to integrate visual cues efficiently. This reveals a critical dissociation between capability and strategy, suggesting accuracy-centric benchmarks may over-index on performance while missing brittle uncertainty handling. These findings reveal emergent principled handling of uncertainty and highlight the correlation between accuracy and Bayesian tendencies. We release our psychophysics benchmark and consistency metric (https://bayes-bench.github.io) as evaluation tools and to inform future multimodal architecture designs.
The rapid proliferation of frontier model agents promises significant societal advances but also raises concerns about systemic risks arising from unsafe interactions. Collusion to the disadvantage of others has been identified as a central form of undesirable agent cooperation. The use of information hiding (steganography) in agent communications could render such collusion practically undetectable. This underscores the need for investigations into the possibility of such behaviours emerging and the robustness corresponding countermeasures. To investigate this problem we design two approaches -- a gradient-based reinforcement learning (GBRL) method and an in-context reinforcement learning (ICRL) method -- for reliably eliciting sophisticated LLM-generated linguistic text steganography. We demonstrate, for the first time, that unintended steganographic collusion in LLMs can arise due to mispecified reward incentives during training. Additionally, we find that standard mitigations -- both passive oversight of model outputs and active mitigation through communication paraphrasing -- are not fully effective at preventing this steganographic communication. Our findings imply that (i) emergence of steganographic collusion is a plausible concern that should be monitored and researched, and (ii) preventing emergence may require innovation in mitigation techniques.
Large language model-powered chatbots have transformed how people seek information, especially in high-stakes contexts like mental health. Despite their support capabilities, safe detection and response to crises such as suicidal ideation and self-harm are still unclear, hindered by the lack of unified crisis taxonomies and clinical evaluation standards. We address this by creating: (1) a taxonomy of six crisis categories; (2) a dataset of over 2,000 inputs from 12 mental health datasets, classified into these categories; and (3) a clinical response assessment protocol. We also use LLMs to identify crisis inputs and audit five models for response safety and appropriateness. First, we built a clinical-informed crisis taxonomy and evaluation protocol. Next, we curated 2,252 relevant examples from over 239,000 user inputs, then tested three LLMs for automatic classification. In addition, we evaluated five models for the appropriateness of their responses to a user's crisis, graded on a 5-point Likert scale from harmful (1) to appropriate (5). While some models respond reliably to explicit crises, risks still exist. Many outputs, especially in self-harm and suicidal categories, are inappropriate or unsafe. Different models perform variably; some, like gpt-5-nano and deepseek-v3.2-exp, have low harm rates, but others, such as gpt-4o-mini and grok-4-fast, generate more unsafe responses. All models struggle with indirect signals, default replies, and context misalignment. These results highlight the urgent need for better safeguards, crisis detection, and context-aware responses in LLMs. They also show that alignment and safety practices, beyond scale, are crucial for reliable crisis support. Our taxonomy, datasets, and evaluation methods support ongoing AI mental health research, aiming to reduce harm and protect vulnerable users.
This paper examines why safety mechanisms designed for human-model interaction do not scale to environments where large language models (LLMs) interact with each other. Most current governance practices still rely on single-agent safety containment, prompts, fine-tuning, and moderation layers that constrain individual model behavior but leave the dynamics of multi-model interaction ungoverned. These mechanisms assume a dyadic setting: one model responding to one user under stable oversight. Yet research and industrial development are rapidly shifting toward LLM-to-LLM ecosystems, where outputs are recursively reused as inputs across chains of agents. In such systems, local compliance can aggregate into collective failure even when every model is individually aligned. We propose a conceptual transition from model-level safety to system-level safety, introducing the framework of the Emergent Systemic Risk Horizon (ESRH) to formalize how instability arises from interaction structure rather than from isolated misbehavior. The paper contributes (i) a theoretical account of collective risk in interacting LLMs, (ii) a taxonomy connecting micro, meso, and macro-level failure modes, and (iii) a design proposal for InstitutionalAI, an architecture for embedding adaptive oversight within multi-agent systems.
Ontology matching (OM) enables semantic interoperability between different ontologies and resolves their conceptual heterogeneity by aligning related entities. OM systems currently have two prevailing design paradigms: conventional knowledge-based expert systems and newer machine learning-based predictive systems. While large language models (LLMs) and LLM agents have revolutionised data engineering and have been applied creatively in many domains, their potential for OM remains underexplored. This study introduces a novel agent-powered LLM-based design paradigm for OM systems. With consideration of several specific challenges in leveraging LLM agents for OM, we propose a generic framework, namely Agent-OM (Agent for Ontology Matching), consisting of two Siamese agents for retrieval and matching, with a set of OM tools. Our framework is implemented in a proof-of-concept system. Evaluations of three Ontology Alignment Evaluation Initiative (OAEI) tracks over state-of-the-art OM systems show that our system can achieve results very close to the long-standing best performance on simple OM tasks and can significantly improve the performance on complex and few-shot OM tasks.
Large language models (LLMs) are now used in settings such as Google's AI Overviews, where it summarizes multiple long documents. However, it remains unclear whether they weight all inputs equally. Focusing on abortion-related news, we construct 40 pro-neutral-con article triplets, permute each triplet into six input orders, and prompt Gemini 2.5 Flash to generate a neutral overview. We evaluate each summary against its source articles using ROUGE-L (lexical overlap), BERTScore (semantic similarity), and SummaC (factual consistency). One-way ANOVA reveals a significant primacy effect for BERTScore across all stances, indicating that summaries are more semantically aligned with the first-seen article. Pairwise comparisons further show that Position 1 differs significantly from Positions 2 and 3, while the latter two do not differ from each other, confirming a selective preference for the first document. The findings present risks for applications that rely on LLM-generated overviews and for agentic AI systems, where the steps involving LLMs can disproportionately influence downstream actions.
Pairwise evaluation of large language models (LLMs) has become the dominant paradigm for benchmarking open-ended tasks, yet non-transitive preferences, where evaluators prefer A over B, B over C, but C over A, fundamentally undermine ranking reliability. We show that this critical issue stems largely from low-quality data that contains inherently ambiguous preference pairs. To address this challenge, we propose ELSPR, a principled graph-theoretic framework that models pairwise preferences as tournament graphs and systematically identifies problematic training data. ELSPR quantifies non-transitivity through strongly connected components (SCCs) analysis and measures overall preference clarity using a novel normalized directed graph structural entropy metric. Our filtering methodology selectively removes preference data that induce non-transitivity while preserving transitive preferences. Extensive experiments on the AlpacaEval benchmark demonstrate that models fine-tuned on ELSPR-filtered data achieve substantial improvements: a 13.8% reduction in non-transitivity, a 0.088 decrease in structural entropy, and significantly enhanced discriminative power in real-world evaluation systems. Human validation confirms that discarded data exhibit dramatically lower inter-annotator agreement (34.4% vs. 52.6%) and model-human consistency (51.2% vs. 80.6%) compared to cleaned data. These findings establish ELSPR as an effective data self-purification approach for developing more robust, consistent, and human-aligned LLM evaluation systems.
Unmanned Aerial Vehicles (UAVs) are increasingly used in defense, surveillance, and disaster response, yet most systems still operate at SAE Level 2 to 3 autonomy. Their dependence on rule-based control and narrow AI limits adaptability in dynamic and uncertain missions. Current UAV architectures lack context-aware reasoning, autonomous decision-making, and integration with external systems. Importantly, none make use of Large Language Model (LLM) agents with tool-calling for real-time knowledge access. This paper introduces the Agentic UAVs framework, a five-layer architecture consisting of Perception, Reasoning, Action, Integration, and Learning. The framework enhances UAV autonomy through LLM-driven reasoning, database querying, and interaction with third-party systems. A prototype built with ROS 2 and Gazebo combines YOLOv11 for object detection with GPT-4 for reasoning and a locally deployed Gemma 3 model. In simulated search-and-rescue scenarios, agentic UAVs achieved higher detection confidence (0.79 compared to 0.72), improved person detection rates (91% compared to 75%), and a major increase in correct action recommendations (92% compared to 4.5%). These results show that modest computational overhead can enable significantly higher levels of autonomy and system-level integration.
Although large language models (LLMs) hold significant promise in psychotherapy, their direct application in patient-facing scenarios raises ethical and safety concerns. Therefore, this work shifts towards developing an LLM as a supervisor to train real therapists. In addition to the privacy of clinical therapist training data, a fundamental contradiction complicates the training of therapeutic behaviors: clear feedback standards are necessary to ensure a controlled training system, yet there is no absolute "gold standard" for appropriate therapeutic behaviors in practice. In contrast, many common therapeutic mistakes are universal and identifiable, making them effective triggers for targeted feedback that can serve as clearer evidence. Motivated by this, we create a novel therapist-training paradigm: (1) guidelines for mistaken behaviors and targeted correction strategies are first established as standards; (2) a human-in-the-loop dialogue-feedback dataset is then constructed, where a mistake-prone agent intentionally makes standard mistakes during interviews naturally, and a supervisor agent locates and identifies mistakes and provides targeted feedback; (3) after fine-tuning on this dataset, the final supervisor model is provided for real therapist training. The detailed experimental results of automated, human and downstream assessments demonstrate that models fine-tuned on our dataset MATE, can provide high-quality feedback according to the clinical guideline, showing significant potential for the therapist training scenario.
The advent of strong generative AI has a considerable impact on various software engineering tasks such as code repair, test generation, or language translation. While tools like GitHub Copilot are already in widespread use in interactive settings, automated approaches require a higher level of reliability before being usable in industrial practice. In this paper, we focus on three aspects that directly influence the quality of the results: a) the effect of automated feedback loops, b) the choice of Large Language Model (LLM), and c) the influence of behavior-preserving code changes. We study the effect of these three variables on an automated C-to-Rust translation system. Code translation from C to Rust is an attractive use case in industry due to Rust's safety guarantees. The translation system is based on a generate-and-check pattern, in which Rust code generated by the LLM is automatically checked for compilability and behavioral equivalence with the original C code. For negative checking results, the LLM is re-prompted in a feedback loop to repair its output. These checks also allow us to evaluate and compare the respective success rates of the translation system when varying the three variables. Our results show that without feedback loops LLM selection has a large effect on translation success. However, when the translation system uses feedback loops the differences across models diminish. We observe this for the average performance of the system as well as its robustness under code perturbations. Finally, we also identify that diversity provided by code perturbations can even result in improved system performance.
The world currently has an abundance of ideas for how to use new LLM agents, and developers seek to rapidly prototype and test new agentic designs. However, executing agents at scale using high-capacity LLMs incurs high inference costs. We propose a simple method for reducing LLM agent inference costs without incurring the development friction costs associated with LLM fine-tuning (long training cycles, optimization hyperparameter tweaking loops) or manual prompt engineering (laborious trial and error). Most importantly, we introduce $\textit{in-context distillation}$, which adapts the idea of knowledge distillation (training a low cost-student model to mimic a high-cost teacher) to an in-context learning setting. Our approach retrieves relevant teacher demonstrations at each agent step and provides them to the student as in-context examples, enabling the student to imitate teacher behavior on-the-fly. We combine in-context distillation with the established idea of $\textit{self-consistency cascades}$ to know when the trust the student. This adaptive strategy realizes the cost benefits of model specialization while preserving the productivity of working with frozen models. On the multi-step embodied reasoning benchmark ALFWorld, our method matches teacher-level accuracy at $\textbf{2.5$\times$ lower cost}$, reducing per-episode costs from \$0.059 to \$0.024. The upfront demonstration cost amortizes after just 843 episodes, yielding cumulative savings exceeding \$34,900 at deployment scale (1M episodes). On AppWorld, a complex agent benchmark requiring multi-step API workflows, we shift the Pareto frontier by achieving a $\textbf{2$\times$ cost reduction}$ at iso-accuracy. By reducing operational costs while maintaining rapid experimentation cycles with frozen models, our approach makes advanced agentic systems economically viable for a broader range of applications.
As web platforms evolve towards greater personalization and emotional complexity, conversational agents must transcend superficial empathy to demonstrate identity-aware emotional reasoning. However, existing systems face two limitations: (1) reliance on situation-centric datasets lacking persistent user identity, which hampers the capture of personalized affective nuances; and (2) dependence on opaque, coarse reward signals that hinder development of verifiable empathetic reasoning. To address these gaps, we introduce KardiaBench, a large-scale user-grounded benchmark comprising 178,080 QA pairs across 22,080 multi-turn conversations anchored to 671 real-world profiles. The dataset is constructed via a model-in-the-loop pipeline with iterative rubric-guided refinement to ensure psychological plausibility and persona consistency. This progressive empathy pipeline that integrates user comprehension, contextual reasoning, and emotion perception into conversations, followed by iterative critique and rubric-based refinement to ensure psychological plausibility, emotional fidelity, and persona consistency. Building on this, we propose Kardia-R1, a framework that trains models for interpretable, stepwise empathetic cognition. Kardia-R1 leverages Rubric-as-Judge Empathetic Reinforcement Learning (Rubric-ERL), a GRPO-based method that uses explainable, human-aligned rubric rewards to tightly couple user understanding, emotional inference, and supportive response generation. Extensive experiments across four LLM backbones demonstrate that Kardia-R1 consistently outperforms othet methods in emotion accuracy, empathy, relevance, persona consistency, and safety. Our dataset and model will be released at https://github.com/JhCircle/Kardia-R1.
The "15-minute city" envisions neighborhoods where residents can meet daily needs via a short walk or bike ride. Realizing this vision requires not only physical proximity but also efficient and reliable access to information about nearby places, services, and events. Existing location-based systems, however, focus mainly on city-level tasks and neglect the spatial, temporal, and cognitive factors that shape localized decision-making. We conceptualize this gap as the Local Life Information Accessibility (LLIA) problem and introduce AskNearby, an AI-driven community application that unifies retrieval and recommendation within the 15-minute life circle. AskNearby integrates (i) a three-layer Retrieval-Augmented Generation (RAG) pipeline that synergizes graph-based, semantic-vector, and geographic retrieval with (ii) a cognitive-map model that encodes each user's neighborhood familiarity and preferences. Experiments on real-world community datasets demonstrate that AskNearby significantly outperforms LLM-based and map-based baselines in retrieval accuracy and recommendation quality, achieving robust performance in spatiotemporal grounding and cognitive-aware ranking. Real-world deployments further validate its effectiveness. By addressing the LLIA challenge, AskNearby empowers residents to more effectively discover local resources, plan daily activities, and engage in community life.
Recent advances in 3D scene-language understanding have leveraged Large Language Models (LLMs) for 3D reasoning by transferring their general reasoning ability to 3D multi-modal contexts. However, existing methods typically adopt standard decoders from language modeling, which rely on a causal attention mask. This design introduces two fundamental conflicts in 3D scene understanding: sequential bias among order-agnostic 3D objects and restricted object-instruction attention, hindering task-specific reasoning. To overcome these limitations, we propose 3D Spatial Language Instruction Mask (3D-SLIM), an effective masking strategy that replaces the causal mask with an adaptive attention mask tailored to the spatial structure of 3D scenes. Our 3D-SLIM introduces two key components: a Geometry-adaptive Mask that constrains attention based on spatial density rather than token order, and an Instruction-aware Mask that enables object tokens to directly access instruction context. This design allows the model to process objects based on their spatial relationships while being guided by the user's task. 3D-SLIM is simple, requires no architectural modifications, and adds no extra parameters, yet it yields substantial performance improvements across diverse 3D scene-language tasks. Extensive experiments across multiple benchmarks and LLM baselines validate its effectiveness and underscore the critical role of decoder design in 3D multi-modal reasoning.
Large language models (LLMs) frequently generate hallucinations -- plausible but factually incorrect outputs -- undermining their reliability. While prior work has examined hallucinations from macroscopic perspectives such as training data and objectives, the underlying neuron-level mechanisms remain largely unexplored. In this paper, we conduct a systematic investigation into hallucination-associated neurons (H-Neurons) in LLMs from three perspectives: identification, behavioral impact, and origins. Regarding their identification, we demonstrate that a remarkably sparse subset of neurons (less than $0.1\%$ of total neurons) can reliably predict hallucination occurrences, with strong generalization across diverse scenarios. In terms of behavioral impact, controlled interventions reveal that these neurons are causally linked to over-compliance behaviors. Concerning their origins, we trace these neurons back to the pre-trained base models and find that these neurons remain predictive for hallucination detection, indicating they emerge during pre-training. Our findings bridge macroscopic behavioral patterns with microscopic neural mechanisms, offering insights for developing more reliable LLMs.
We introduce HBLLM, a wavelet-enhanced high-fidelity $1$-bit post-training quantization method for Large Language Models (LLMs). By leveraging Haar wavelet transforms to enhance expressive capacity through frequency decomposition, HBLLM significantly improves quantization fidelity while maintaining minimal overhead. This approach features two innovative structure-aware grouping strategies: (1) frequency-aware multi-parameter intra-row grouping and (2) $\ell_2$-norm-based saliency-driven column selection. For non-salient weights, a shared mean is employed across quantization groups within each frequency band to optimize storage efficiency. Experiments conducted on the OPT and LLaMA models demonstrate that HBLLM achieves state-of-the-art performance in $1$-bit quantization, attaining a perplexity of $6.71$ on LLaMA$2$-$13$B with an average weight storage of only $1.08$ bits. Code available at: https://github.com/Yeyke/HBLLM.
Solving complex or long-horizon problems often requires large language models (LLMs) to use external tools and operate over a significantly longer context window. New LLMs enable longer context windows and support tool calling capabilities. Prior works have focused mainly on evaluation of LLMs on long-context prompts, leaving agentic setup relatively unexplored, both from capability and safety perspectives. Our work addresses this gap. We find that LLM agents could be sensitive to length, type, and placement of the context, exhibiting unexpected and inconsistent shifts in task performance and in refusals to execute harmful requests. Models with 1M-2M token context windows show severe degradation already at 100K tokens, with performance drops exceeding 50\% for both benign and harmful tasks. Refusal rates shift unpredictably: GPT-4.1-nano increases from $\sim$5\% to $\sim$40\% while Grok 4 Fast decreases from $\sim$80\% to $\sim$10\% at 200K tokens. Our work shows potential safety issues with agents operating on longer context and opens additional questions on the current metrics and paradigm for evaluating LLM agent safety on long multi-step tasks. In particular, our results on LLM agents reveal a notable divergence in both capability and safety performance compared to prior evaluations of LLMs on similar criteria.
We propose a novel framework for persona-based language model system, motivated by the need for personalized AI agents that adapt to individual user preferences. In our approach, the agent embodies the user's "persona" (e.g. user profile or taste) and is powered by a large language model (LLM). To enable the agent to leverage rich contextual information, we introduce a Knowledge-Graph-enhanced Retrieval-Augmented Generation (Graph RAG) mechanism that constructs an LLM-derived graph index of relevant documents and summarizes communities of related information. Our framework generates personalized prompts by combining: (1) a summary of the user's historical behaviors and preferences extracted from the knowledge graph, and (2) relevant global interaction patterns identified through graph-based community detection. This dynamic prompt engineering approach allows the agent to maintain consistent persona-aligned behaviors while benefiting from collective knowledge. On the LaMP benchmark, our method improves news categorization F1 by 11.1%, movie tagging F1 by 56.1%, and reduces product rating MAE by 10.4% over prior methods. Our code is available at https://anonymous.4open.science/r/PersonaAgentwGraphRAG-DE6F
The emergence of Large Language Models (LLMs) in Multi-Agent Systems (MAS) has opened new possibilities for artificial intelligence, yet current implementations face significant challenges in resource management, task coordination, and system efficiency. While existing frameworks demonstrate the potential of LLM-based agents in collaborative problem-solving, they often lack sophisticated mechanisms for parallel execution and dynamic task management. This paper introduces DynTaskMAS, a novel framework that orchestrates asynchronous and parallel operations in LLM-based MAS through dynamic task graphs. The framework features four key innovations: (1) a Dynamic Task Graph Generator that intelligently decomposes complex tasks while maintaining logical dependencies, (2) an Asynchronous Parallel Execution Engine that optimizes resource utilization through efficient task scheduling, (3) a Semantic-Aware Context Management System that enables efficient information sharing among agents, and (4) an Adaptive Workflow Manager that dynamically optimizes system performance. Experimental evaluations demonstrate that DynTaskMAS achieves significant improvements over traditional approaches: a 21-33% reduction in execution time across task complexities (with higher gains for more complex tasks), a 35.4% improvement in resource utilization (from 65% to 88%), and near-linear throughput scaling up to 16 concurrent agents (3.47X improvement for 4X agents). Our framework establishes a foundation for building scalable, high-performance LLM-based multi-agent systems capable of handling complex, dynamic tasks efficiently.
Large language model (LLM) watermarks enable authentication of text provenance, curb misuse of machine-generated text, and promote trust in AI systems. Current watermarks operate by changing the next-token predictions output by an LLM. The updated (i.e., watermarked) predictions depend on random side information produced, for example, by hashing previously generated tokens. LLM watermarking is particularly challenging in low-entropy generation tasks -- such as coding -- where next-token predictions are near-deterministic. In this paper, we propose an optimization framework for watermark design. Our goal is to understand how to most effectively use random side information in order to maximize the likelihood of watermark detection and minimize the distortion of generated text. Our analysis informs the design of two new watermarks: HeavyWater and SimplexWater. Both watermarks are tunable, gracefully trading-off between detection accuracy and text distortion. They can also be applied to any LLM and are agnostic to side information generation. We examine the performance of HeavyWater and SimplexWater through several benchmarks, demonstrating that they can achieve high watermark detection accuracy with minimal compromise of text generation quality, particularly in the low-entropy regime. Our theoretical analysis also reveals surprising new connections between LLM watermarking and coding theory. The code implementation can be found in https://github.com/DorTsur/HeavyWater_SimplexWater
Large language models (LLMs) can act as both problem solvers and solution verifiers, with verifiers improving solver performance by selecting high-quality answers from a pool of candidates. However, prior studies of solver-verifier interactions have been limited, focusing mainly on self-verification and rarely examining how verifiers judge outputs from models in their own or in another model family. Modern LLMs also undergo extensive post-training, but its effect on verification remains unclear. We present a systematic study across 37 models spanning multiple families, sizes, and base vs. post-trained variants, evaluated on 9 benchmarks covering logical reasoning, structured puzzles, symbolic computation, mathematics, commonsense, factual recall, and domain knowledge. We compare self-verification with verification within the same family and across different families. To support this, we introduce and empirically validate verifier gain, a metric that predicts the performance improvements from test-time verifier-based rejection sampling. We analyze how metrics like verifier gain and false positive rate scale with model size and post-training, and characterize differences in dataset verifiability. Our findings show that cross-family verification is especially effective; post-training reduces self-improvement but strengthens cross-family improvement; and mathematical and logical tasks exhibit the highest inherent verifiability.
Large language models (LLMs) remain vulnerable to jailbreak attacks that bypass safety guardrails to elicit harmful outputs. Existing approaches overwhelmingly operate within the prompt-optimization paradigm: whether through traditional algorithmic search or recent agent-based workflows, the resulting prompts typically retain malicious semantic signals that modern guardrails are primed to detect. In contrast, we identify a deeper, largely overlooked vulnerability stemming from the highly interconnected nature of an LLM's internal knowledge. This structure allows harmful objectives to be realized by weaving together sequences of benign sub-queries, each of which individually evades detection. To exploit this loophole, we introduce the Correlated Knowledge Attack Agent (CKA-Agent), a dynamic framework that reframes jailbreaking as an adaptive, tree-structured exploration of the target model's knowledge base. The CKA-Agent issues locally innocuous queries, uses model responses to guide exploration across multiple paths, and ultimately assembles the aggregated information to achieve the original harmful objective. Evaluated across state-of-the-art commercial LLMs (Gemini2.5-Flash/Pro, GPT-oss-120B, Claude-Haiku-4.5), CKA-Agent consistently achieves over 95% success rates even against strong guardrails, underscoring the severity of this vulnerability and the urgent need for defenses against such knowledge-decomposition attacks. Our codes are available at https://github.com/Graph-COM/CKA-Agent.
The current literature on memorization in Natural Language Models, especially Large Language Models (LLMs), poses severe security and privacy risks, as models tend to memorize personally identifying information (PIIs) from training data. We introduce Randomized Masked Fine-Tuning (RMFT), a novel privacy-preserving fine-tuning technique that reduces PII memorization while minimizing performance impact. Using the Enron Email Dataset, we demonstrate that RMFT achieves an 80.81% reduction in Total Extraction Rate and 80.17% reduction in Seen Extraction Rate compared to baseline fine-tuning, outperforming deduplication methods while maintaining only a 5.73% increase in perplexity. We present MaxTER, a Pareto-optimal evaluation framework for assessing privacy-utility tradeoffs, and show the performance of RMFT vs Deduplication by Area Under The Response Curve (AURC) metric.
In today's age, it is becoming increasingly difficult to decipher truth from lies. Every day, politicians, media outlets, and public figures make conflicting claims$\unicode{x2014}$often about topics that can, in principle, be verified against structured data. For instance, statements about crime rates, economic growth or healthcare can all be verified against official public records and structured datasets. Building a system that can automatically do that would have sounded like science fiction just a few years ago. Yet, with the extraordinary progress in LLMs and agentic AI, this is now within reach. Still, there remains a striking gap between what is technically possible and what is being demonstrated by recent work. Most existing verification systems operate only on small, single-table databases$\unicode{x2014}$typically a few hundred rows$\unicode{x2014}$that conveniently fit within an LLM's context window. In this paper we report our progress on Thucy, the first cross-database, cross-table multi-agent claim verification system that also provides concrete evidence for each verification verdict. Thucy remains completely agnostic to the underlying data sources before deployment and must therefore autonomously discover, inspect, and reason over all available relational databases to verify claims. Importantly, Thucy also reports the exact SQL queries that support its verdict (whether the claim is accurate or not) offering full transparency to expert users familiar with SQL. When evaluated on the TabFact dataset$\unicode{x2014}$the standard benchmark for fact verification over structured data$\unicode{x2014}$Thucy surpasses the previous state of the art by 5.6 percentage points in accuracy (94.3% vs. 88.7%).
Autonomous Vehicles (AVs) are transforming the future of transportation through advances in intelligent perception, decision-making, and control systems. However, their success is tied to one core capability, reliable object detection in complex and multimodal environments. While recent breakthroughs in Computer Vision (CV) and Artificial Intelligence (AI) have driven remarkable progress, the field still faces a critical challenge as knowledge remains fragmented across multimodal perception, contextual reasoning, and cooperative intelligence. This survey bridges that gap by delivering a forward-looking analysis of object detection in AVs, emphasizing emerging paradigms such as Vision-Language Models (VLMs), Large Language Models (LLMs), and Generative AI rather than re-examining outdated techniques. We begin by systematically reviewing the fundamental spectrum of AV sensors (camera, ultrasonic, LiDAR, and Radar) and their fusion strategies, highlighting not only their capabilities and limitations in dynamic driving environments but also their potential to integrate with recent advances in LLM/VLM-driven perception frameworks. Next, we introduce a structured categorization of AV datasets that moves beyond simple collections, positioning ego-vehicle, infrastructure-based, and cooperative datasets (e.g., V2V, V2I, V2X, I2I), followed by a cross-analysis of data structures and characteristics. Ultimately, we analyze cutting-edge detection methodologies, ranging from 2D and 3D pipelines to hybrid sensor fusion, with particular attention to emerging transformer-driven approaches powered by Vision Transformers (ViTs), Large and Small Language Models (SLMs), and VLMs. By synthesizing these perspectives, our survey delivers a clear roadmap of current capabilities, open challenges, and future opportunities.
Current LLMs are trained to refuse potentially harmful input queries regardless of whether users actually had harmful intents, causing a tradeoff between safety and user experience. Through a study of 480 participants evaluating 3,840 query-response pairs, we examine how different refusal strategies affect user perceptions across varying motivations. Our findings reveal that response strategy largely shapes user experience, while actual user motivation has negligible impact. Partial compliance -- providing general information without actionable details -- emerges as the optimal strategy, reducing negative user perceptions by over 50% to flat-out refusals. Complementing this, we analyze response patterns of 9 state-of-the-art LLMs and evaluate how 6 reward models score different refusal strategies, demonstrating that models rarely deploy partial compliance naturally and reward models currently undervalue it. This work demonstrates that effective guardrails require focusing on crafting thoughtful refusals rather than detecting intent, offering a path toward AI safety mechanisms that ensure both safety and sustained user engagement.
This paper introduces CryptoBench, the first expert-curated, dynamic benchmark designed to rigorously evaluate the real-world capabilities of Large Language Model (LLM) agents in the uniquely demanding and fast-paced cryptocurrency domain. Unlike general-purpose agent benchmarks for search and prediction, professional crypto analysis presents specific challenges: \emph{extreme time-sensitivity}, \emph{a highly adversarial information environment}, and the critical need to synthesize data from \emph{diverse, specialized sources}, such as on-chain intelligence platforms and real-time Decentralized Finance (DeFi) dashboards. CryptoBench thus serves as a much more challenging and valuable scenario for LLM agent assessment. To address these challenges, we constructed a live, dynamic benchmark featuring 50 questions per month, expertly designed by crypto-native professionals to mirror actual analyst workflows. These tasks are rigorously categorized within a four-quadrant system: Simple Retrieval, Complex Retrieval, Simple Prediction, and Complex Prediction. This granular categorization enables a precise assessment of an LLM agent's foundational data-gathering capabilities alongside its advanced analytical and forecasting skills. Our evaluation of ten LLMs, both directly and within an agentic framework, reveals a performance hierarchy and uncovers a failure mode. We observe a \textit{retrieval-prediction imbalance}, where many leading models, despite being proficient at data retrieval, demonstrate a pronounced weakness in tasks requiring predictive analysis. This highlights a problematic tendency for agents to appear factually grounded while lacking the deeper analytical capabilities to synthesize information.
Most existing 3D shape datasets and models focus solely on geometry, overlooking the material properties that determine how objects appear. We introduce a two-stage large language model (LLM) based method for inferring material composition directly from 3D point clouds with coarse segmentations. Our key insight is to decouple reasoning about what an object is from what it is made of. In the first stage, an LLM predicts the object's semantic; in the second stage, it assigns plausible materials to each geometric segment, conditioned on the inferred semantics. Both stages operate in a zero-shot manner, without task-specific training. Because existing datasets lack reliable material annotations, we evaluate our method using an LLM-as-a-Judge implemented in DeepEval. Across 1,000 shapes from Fusion/ABS and ShapeNet, our method achieves high semantic and material plausibility. These results demonstrate that language models can serve as general-purpose priors for bridging geometric reasoning and material understanding in 3D data.
Despite recent advances in LLMs, the task of code generation is still challenging. To cope, code selection algorithms select the best program from multiple programs generated by an LLM. However, existing algorithms can fail to identify the correct program, either because they can misidentify nonequivalent programs or because they rely on an LLM and assume it always correctly determines the output for every input. We present ExPairT-LLM, an exact learning algorithm for code selection that selects a program by posing to an LLM oracle two new types of queries: pairwise membership and pairwise equivalence. These queries are simpler for LLMs and enable ExPairT-LLM to identify the correct program through a tournament, which is robust to some LLM mistakes. We evaluate ExPairT-LLM on four popular code datasets. Its pass@1 (success rate) outperforms the state-of-the-art code selection algorithm on average by +13.0% and up to +27.1%. It also improves the pass@1 of LLMs performing complex reasoning by +24.0%.
Large language model (LLM) services now answer billions of queries per day, and industry reports show that inference, not training, accounts for more than 90% of total power consumption. However, existing benchmarks focus on either training/fine-tuning or performance of inference and provide little support for power consumption measurement and analysis of inference. We introduce TokenPowerBench, the first lightweight and extensible benchmark designed for LLM-inference power consumption studies. The benchmark combines (i) a declarative configuration interface covering model choice, prompt set, and inference engine, (ii) a measurement layer that captures GPU-, node-, and system-level power without specialized power meters, and (iii) a phase-aligned metrics pipeline that attributes energy to the prefill and decode stages of every request. These elements make it straight-forward to explore the power consumed by an LLM inference run; furthermore, by varying batch size, context length, parallelism strategy and quantization, users can quickly assess how each setting affects joules per token and other energy-efficiency metrics. We evaluate TokenPowerBench on four of the most widely used model series (Llama, Falcon, Qwen, and Mistral). Our experiments cover from 1 billion parameters up to the frontier-scale Llama3-405B model. Furthermore, we release TokenPowerBench as open source to help users to measure power consumption, forecast operating expenses, and meet sustainability targets when deploying LLM services.
Thinking Large Language Models (LLMs) used as judges for pairwise preferences remain noisy at the single-sample level, and common aggregation rules (majority vote, soft self-consistency, or instruction-based self-aggregation) are inconsistent when ties are allowed. We study inference-time compute (ITC) for evaluators that generate n independent thinking-rating samples per item, and propose a principled, distribution-calibrated aggregation scheme. Our method models three-way preferences with a Bradley-Terry-Davidson formulation on rating counts, leveraging both polarity (margin among non-ties) and decisiveness (non-tie rate) to distinguish narrow margins from strong consensus. Across various evaluation benchmarks, our approach consistently reduces MAE and increases pairwise accuracy versus standard baselines, and when evaluated against human-consensus meta-labels, matches or exceeds individual human raters. These results show that carefully allocating ITC and aggregating with distribution-aware methods turns noisy individual model judgments into reliable ratings for evaluation.
The rapid advancement of large language models (LLMs) has opened new possibilities for AI for good applications. As LLMs increasingly mediate online communication, their potential to foster empathy and constructive dialogue becomes an important frontier for responsible AI research. This work explores whether LLMs can serve not only as moderators that detect harmful content, but as mediators capable of understanding and de-escalating online conflicts. Our framework decomposes mediation into two subtasks: judgment, where an LLM evaluates the fairness and emotional dynamics of a conversation, and steering, where it generates empathetic, de-escalatory messages to guide participants toward resolution. To assess mediation quality, we construct a large Reddit-based dataset and propose a multi-stage evaluation pipeline combining principle-based scoring, user simulation, and human comparison. Experiments show that API-based models outperform open-source counterparts in both reasoning and intervention alignment when doing mediation. Our findings highlight both the promise and limitations of current LLMs as emerging agents for online social mediation.
Hardware-Aware Neural Architecture Search (HW-NAS) requires joint optimization of accuracy and latency under device constraints. Traditional supernet-based methods require multiple GPU days per dataset. Large Language Model (LLM)-driven approaches avoid training a large supernet and can provide quick feedback, but we observe an exploration bias: the LLM repeatedly proposes neural network designs within limited search space and fails to discover architectures across different latency ranges in the entire search space. To address this issue, we propose LLM-NAS: an LLM-driven Neural Architecture Search that can generate neural networks with high accuracy and low latency with reduced search cost. Our proposed LLM-NAS has three key components: 1) a complexity-driven partitioning engine that divides the search space by complexity to enforce diversity and mitigate exploration bias; 2) an LLM-powered architecture prompt co-evolution operator, in which the LLM first updates a knowledge base of design heuristics based on results from the previous round, then performs a guided evolution algorithm on architectures with prompts that incorporate this knowledge base. Prompts and designs improve together across rounds which avoids random guesswork and improve efficiency; 3) a zero-cost predictor to avoid training a large number of candidates from scratch. Experimental results show that on HW-NAS-Bench, LLM-NAS can achieve overall higher HV, lower IGD, and up to 54% lower latency than baselines at similar accuracy. Meanwhile, the search cost drops from days to minutes compared with traditional supernet baselines.
Developing large language models (LLMs) to cooperate and compete effectively within multi-agent systems (MASs) 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 MARSHAL, an end-to-end RL framework that incentivizes Multi-Agent Reasoning through Self-play witH strAtegic LLMs in both cooperative and competitive games. MARSHAL 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, MARSHAL 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 MASs in reasoning benchmarks. When integrated into leading MASs, our MARSHAL agent achieves significant performance gains of up to 10.0% on AIME, 6.6% on GPQA-Diamond, and 3.5% on average across all benchmarks. 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.
To prevent misinformation and social issues arising from trustworthy-looking content generated by LLMs, it is crucial to develop efficient and reliable methods for identifying the source of texts. Previous approaches have demonstrated exceptional performance in detecting texts fully generated by LLMs. However, these methods struggle when confronting more advanced LLM output or text with adversarial multi-task machine revision, especially in the black-box setting, where the generating model is unknown. To address this challenge, grounded in the hypothesis that human writing possesses distinctive stylistic patterns, we propose Human Language Preference Detection (HLPD). HLPD employs a reward-based alignment process, Human Language Preference Optimization (HLPO), to shift the scoring model's token distribution toward human-like writing, making the model more sensitive to human writing, therefore enhancing the identification of machine-revised text. We test HLPD in an adversarial multi-task evaluation framework that leverages a five-dimensional prompt generator and multiple advanced LLMs to create diverse revision scenarios. When detecting texts revised by GPT-series models, HLPD achieves a 15.11% relative improvement in AUROC over ImBD, surpassing Fast-DetectGPT by 45.56%. When evaluated on texts generated by advanced LLMs, HLPD achieves the highest average AUROC, exceeding ImBD by 5.53% and Fast-DetectGPT by 34.14%. Code will be made available at https://github.com/dfq2021/HLPD.
Large language models (LLMs) have transformed the field of natural language processing, but they remain susceptible to jailbreaking attacks that exploit their capabilities to generate unintended and potentially harmful content. Existing token-level jailbreaking techniques, while effective, face scalability and efficiency challenges, especially as models undergo frequent updates and incorporate advanced defensive measures. In this paper, we introduce JailMine, an innovative token-level manipulation approach that addresses these limitations effectively. JailMine employs an automated "mining" process to elicit malicious responses from LLMs by strategically selecting affirmative outputs and iteratively reducing the likelihood of rejection. Through rigorous testing across multiple well-known LLMs and datasets, we demonstrate JailMine's effectiveness and efficiency, achieving a significant average reduction of 86% in time consumed while maintaining high success rates averaging 95%, even in the face of evolving defensive strategies. Our work contributes to the ongoing effort to assess and mitigate the vulnerability of LLMs to jailbreaking attacks, underscoring the importance of continued vigilance and proactive measures to enhance the security and reliability of these powerful language models.
Recent advances in reasoning techniques have substantially improved the performance of large language models (LLMs), raising expectations for their ability to provide accurate, truthful, and reliable information. However, emerging evidence suggests that iterative reasoning may foster belief entrenchment and confirmation bias, rather than enhancing truth-seeking behavior. In this study, we propose a systematic evaluation framework for belief entrenchment in LLM reasoning by leveraging the Martingale property from Bayesian statistics. This property implies that, under rational belief updating, the expected value of future beliefs should remain equal to the current belief, i.e., belief updates are unpredictable from the current belief. We propose the unsupervised, regression-based Martingale Score to measure violations of this property, which signal deviation from the Bayesian ability of updating on new evidence. In open-ended problem domains including event forecasting, value-laden questions, and academic paper review, we find such violations to be widespread across models and setups, where the current belief positively predicts future belief updates, a phenomenon which we term belief entrenchment. We identify the models, reasoning techniques, and domains more prone to belief entrenchment. Finally, we validate the Martingale Score by showing that it predicts ground-truth accuracy on problem domains where ground truth labels are available. This indicates that, while designed as an unsupervised metric that operates even in domains without access to ground truth, the Martingale Score is a useful proxy of the truth-seeking ability of a reasoning process.
Understanding how visual content communicates sentiment is critical in an era where online interaction is increasingly dominated by this kind of media on social platforms. However, this remains a challenging problem, as sentiment perception is closely tied to complex, scene-level semantics. In this paper, we propose an original framework, MLLMsent, to investigate the sentiment reasoning capabilities of Multimodal Large Language Models (MLLMs) through three perspectives: (1) using those MLLMs for direct sentiment classification from images; (2) associating them with pre-trained LLMs for sentiment analysis on automatically generated image descriptions; and (3) fine-tuning the LLMs on sentiment-labeled image descriptions. Experiments on a recent and established benchmark demonstrate that our proposal, particularly the fine-tuned approach, achieves state-of-the-art results outperforming Lexicon-, CNN-, and Transformer-based baselines by up to 30.9%, 64.8%, and 42.4%, respectively, across different levels of evaluators' agreement and sentiment polarity categories. Remarkably, in a cross-dataset test, without any training on these new data, our model still outperforms, by up to 8.26%, the best runner-up, which has been trained directly on them. These results highlight the potential of the proposed visual reasoning scheme for advancing affective computing, while also establishing new benchmarks for future research.
Developing and validating psychometric scales requires large samples, multiple testing phases, and substantial resources. Recent advances in Large Language Models (LLMs) enable the generation of synthetic participant data by prompting models to answer items while impersonating individuals of specific demographic profiles, potentially allowing in silico piloting before real data collection. Across four preregistered studies (N = circa 300 each), we tested whether LLM-simulated datasets can reproduce the latent structures and measurement properties of human responses. In Studies 1-2, we compared LLM-generated data with real datasets for two validated scales; in Studies 3-4, we created new scales using EFA on simulated data and then examined whether these structures generalized to newly collected human samples. Simulated datasets replicated the intended factor structures in three of four studies and showed consistent configural and metric invariance, with scalar invariance achieved for the two newly developed scales. However, correlation-based tests revealed substantial differences between real and synthetic datasets, and notable discrepancies appeared in score distributions and variances. Thus, while LLMs capture group-level latent structures, they do not approximate individual-level data properties. Simulated datasets also showed full internal invariance across gender. Overall, LLM-generated data appear useful for early-stage, group-level psychometric prototyping, but not as substitutes for individual-level validation. We discuss methodological limitations, risks of bias and data pollution, and ethical considerations related to in silico psychometric simulations.
By integrating tools from external APIs, Large Language Models (LLMs) have expanded their promising capabilities in a diverse spectrum of complex real-world tasks. However, testing, evaluation, and analysis of LLM tool use remain in their early stages. Most existing benchmarks rely on manually collected test cases, many of which cannot be automatically checked for semantic correctness and instead depend on static methods such as string matching. Additionally, these benchmarks often overlook the complex interactions that occur between sequential API calls, which are common in real-world applications. To fill the gap, in this paper, we introduce StateGen, an automated framework designed to generate diverse coding tasks involving sequential API interactions. StateGen combines state-machine-based API constraint solving and validation, energy-based sampling, and control-flow injection to generate executable programs. These programs are then translated into human-like natural language task descriptions through a collaboration of two LLM agents. Utilizing StateGen, we construct StateEval, a benchmark encompassing 120 verified test cases spanning across three representative scenarios: Session Service, Tensor Operation, and ElevenLabs MCP. Experimental results confirm that StateGen can effectively generate challenging and realistic API-oriented tasks, highlighting areas for improvement in current LLMs incorporating APIs.We make our framework and benchmark publicly available to support future research.
Large language models (LLMs) have achieved remarkable success in many natural language tasks but still struggle with complex, multi-step reasoning, particularly across diverse disciplines. Existing reasoning datasets often lack disciplinary breadth, reasoning depth, and diversity, as well as guiding principles for question synthesis. We propose DESIGNER: a DESIGN-logic-guidEd Reasoning data synthesis pipeline that leverages naturally available, extensive raw documents (e.g., book corpus and web corpus) to generate multidisciplinary challenging questions. We introduce the concept of "design logic" and instruct LLMs to mimic human educators' question-creation process, enabling the automated synthesis of large-scale, high-difficulty questions. We use LLMs to reverse-engineer and abstract over 120,000 design logics from existing questions across various disciplines. By matching these design logics with source documents, we are able to generate reasoning questions with controllable question types and difficulty levels. Using this pipeline, we synthesized two large-scale reasoning datasets that span 75 disciplines: DLR-Book (3.04 million questions from the book corpus) and DLR-Web (1.66 million questions from the web corpus). Data analysis indicates that the questions synthesized by our method exhibit greater difficulty and diversity compared to those in the baseline datasets. We validate our synthesized data through supervised fine-tuning (SFT) on the Qwen3 and Llama3 model families. Our data substantially enhances their multidisciplinary reasoning capabilities, outperforming existing datasets. Notably, by applying SFT on the base versions of these models using only our data, we even surpass their official final models that have undergone the full post-training process.
Neural Networks can be effectively compressed through pruning, significantly reducing storage and compute demands while maintaining predictive performance. Simple yet effective methods like magnitude pruning remove less important parameters and typically require a costly retraining procedure to restore performance. However, with the rise of LLMs, full retraining has become infeasible due to memory and compute constraints. This study challenges the practice of retraining all parameters by showing that updating a small subset of highly expressive parameters can suffice to recover or even enhance performance after pruning. Surprisingly, retraining just 0.01%-0.05% of the parameters in GPT-architectures can match the performance of full retraining across various sparsity levels, significantly reducing compute and memory requirements, and enabling retraining of models with up to 30 billion parameters on a single GPU in minutes. To bridge the gap to full retraining in the high sparsity regime, we introduce two novel LoRA variants that, unlike standard LoRA, allow merging adapters back without compromising sparsity. Going a step further, we show that these methods can be applied for memory-efficient layer-wise reconstruction, significantly enhancing state-of-the-art retraining-free methods like Wanda (Sun et al., 2023) and SparseGPT (Frantar & Alistarh, 2023). Our findings present a promising alternative to avoiding retraining.
Recently, two-stage fine-tuning strategies, e.g., acquiring essential driving knowledge through supervised fine-tuning (SFT) and further enhancing decision-making and planning via reinforcement fine-tuning (RFT), have shown strong potential in advancing the knowledge-driven autonomous driving (AD) paradigm. However, the learning nature of SFT still limits the generalization of reasoning, thereby constraining the full potential of driving performance. Meanwhile, current RFT approaches are primarily applied to downstream tasks, since scene understanding is an open-ended problem where corresponding rewards are difficult to quantify. To address these limitations, we propose OpenREAD, an OPEN-ended REasoning reinforced vision-language model (VLM)-based autonomous driving (AD) framework that enables end-to-end RFT across the full spectrum from high-level reasoning to low-level trajectory planning. Specifically, we begin by constructing large-scale Chain-of-Thought (CoT) annotations on open-source driving-related knowledge datasets, and employ the powerful Qwen3 large language model (LLM) as the critic in RFT to quantify reasoning quality for open-ended questions during reward modeling. Extensive experiments confirm that joint end-to-end RFT yields substantial improvements in both upstream and downstream tasks, enabling OpenREAD to achieve state-of-the-art performance on reasoning and planning benchmarks.
System prompts provide a lightweight yet powerful mechanism for conditioning large language models (LLMs) at inference time. While prior work has focused on English-only settings, real-world deployments benefit from having a single prompt to operate reliably across languages. This paper presents a comprehensive study of how different system prompts steer models toward accurate and robust cross-lingual behavior. We propose a unified four-dimensional evaluation framework to assess system prompts in multilingual environments. Through large-scale experiments on five languages, three LLMs, and three benchmarks, we uncover that certain prompt components, such as CoT, emotion, and scenario, correlate with robust multilingual behavior. We develop a prompt optimization framework for multilingual settings and show it can automatically discover prompts that improve all metrics by 5-10%. Finally, we analyze over 10 million reasoning units and find that more performant system prompts induce more structured and consistent reasoning patterns, while reducing unnecessary language-switching. Together, we highlight system prompt optimization as a scalable path to accurate and robust multilingual LLM behavior.
Multi-robot task allocation in construction automation has traditionally relied on optimization methods such as Dynamic Programming and Reinforcement Learning. This research introduces the LangGraph-based Task Allocation Agent (LTAA), an LLM-driven framework that integrates phase-adaptive allocation strategies, multi-stage validation with hierarchical retries, and dynamic prompting for efficient robot coordination. Although recent LLM approaches show potential for construction robotics, they largely lack rigorous validation and benchmarking against established algorithms. This paper presents the first systematic comparison of LLM-based task allocation with traditional methods in construction scenarios.The study validates LLM feasibility through SMART-LLM replication and addresses implementation challenges using a Self-Corrective Agent Architecture. LTAA leverages natural-language reasoning combined with structured validation mechanisms, achieving major computational gains reducing token usage by 94.6% and allocation time by 86% through dynamic prompting. The framework adjusts its strategy across phases: emphasizing execution feasibility early and workload balance in later allocations.The authors evaluate LTAA against Dynamic Programming, Q-learning, and Deep Q-Network (DQN) baselines using construction operations from the TEACh human-robot collaboration dataset. In the Heavy Excels setting, where robots have strong task specializations, LTAA achieves 77% task completion with superior workload balance, outperforming all traditional methods. These findings show that LLM-based reasoning with structured validation can match established optimization algorithms while offering additional advantages such as interpretability, adaptability, and the ability to update task logic without retraining.
Large Language Model (LLM) agent systems have advanced rapidly, driven by their strong generalization in zero-shot settings. To further enhance reasoning and accuracy on complex tasks, Multi-Agent Debate (MAD) has emerged as a promising framework that engages multiple LLM agents in structured debates to encourage diverse reasoning. However, triggering MAD for every query is inefficient, as it incurs substantial computational (token) cost and may even degrade accuracy by overturning correct single-agent answers. To address these limitations, we propose intelligent Multi-Agent Debate (iMAD), a token-efficient framework that selectively triggers MAD only when it is likely to be beneficial (i.e., correcting an initially wrong answer). To achieve this goal, iMAD learns generalizable model behaviors to make accurate debate decisions. Specifically, iMAD first prompts a single agent to produce a structured self-critique response, from which we extract 41 interpretable linguistic and semantic features capturing hesitation cues. Then, iMAD uses a lightweight debate-decision classifier, trained using our proposed FocusCal loss, to determine whether to trigger MAD, enabling robust debate decisions without test dataset-specific tuning. Through extensive experiments using six (visual) question answering datasets against five competitive baselines, we have shown that iMAD significantly reduces token usage (by up to 92%) while also improving final answer accuracy (by up to 13.5%).
LLM-based automatic survey systems are transforming how users acquire information from the web by integrating retrieval, organization, and content synthesis into end-to-end generation pipelines. While recent works focus on developing new generation pipelines, how to evaluate such complex systems remains a significant challenge. To this end, we introduce SurveyEval, a comprehensive benchmark that evaluates automatically generated surveys across three dimensions: overall quality, outline coherence, and reference accuracy. We extend the evaluation across 7 subjects and augment the LLM-as-a-Judge framework with human references to strengthen evaluation-human alignment. Evaluation results show that while general long-text or paper-writing systems tend to produce lower-quality surveys, specialized survey-generation systems are able to deliver substantially higher-quality results. We envision SurveyEval as a scalable testbed to understand and improve automatic survey systems across diverse subjects and evaluation criteria.
Reinforcement learning (RL) is widely used for post-training large language models (LLMs) in code editing, where group-relative methods like GRPO are popular for their critic-free, normalized advantage estimation. However, in real-world code-editing scenarios, reward distributions are often skewed with unpredictable outliers, leading to distorted advantage computation and increased noise. To address this issue, we propose Group Adaptive Policy Optimization (GAPO), which adaptively finds an outlier-free highest-density interval (HDI) per prompt and then uses the median of that interval as an adaptive Q to replace the group mean in advantage calculation. This adaptive Q robustly handles skewed distributions while remaining plug-and-play and efficient. We validate GAPO on nine instruction-tuned LLMs (3B-14B) using a large internal dataset of 51,844 real-world, history-aware code-editing tasks across 10 languages, demonstrating consistent improvements in exact match accuracy over GRPO and its variant DAPO. Code is publicly available.
Despite their capabilities, Large Language Models (LLMs) remain opaque with limited understanding of their internal representations. Current interpretability methods either focus on input-oriented feature extraction, such as supervised probes and Sparse Autoencoders (SAEs), or on output distribution inspection, such as logit-oriented approaches. A full understanding of LLM vector spaces, however, requires integrating both perspectives, something existing approaches struggle with due to constraints on latent feature definitions. We introduce the Hyperdimensional Probe, a hybrid supervised probe that combines symbolic representations with neural probing. Leveraging Vector Symbolic Architectures (VSAs) and hypervector algebra, it unifies prior methods: the top-down interpretability of supervised probes, SAE's sparsity-driven proxy space, and output-oriented logit investigation. This allows deeper input-focused feature extraction while supporting output-oriented investigation. Our experiments show that our method consistently extracts meaningful concepts across LLMs, embedding sizes, and setups, uncovering concept-driven patterns in analogy-oriented inference and QA-focused text generation. By supporting joint input-output analysis, this work advances semantic understanding of neural representations while unifying the complementary perspectives of prior methods.
Auditors rely on Journal Entry Tests (JETs) to detect anomalies in tax-related ledger records, but rule-based methods generate overwhelming false positives and struggle with subtle irregularities. We investigate whether large language models (LLMs) can serve as anomaly detectors in double-entry bookkeeping. Benchmarking SoTA LLMs such as LLaMA and Gemma on both synthetic and real-world anonymized ledgers, we compare them against JETs and machine learning baselines. Our results show that LLMs consistently outperform traditional rule-based JETs and classical ML baselines, while also providing natural-language explanations that enhance interpretability. These results highlight the potential of \textbf{AI-augmented auditing}, where human auditors collaborate with foundation models to strengthen financial integrity.
Modern work relies on an assortment of digital collaboration tools, yet routine processes continue to suffer from human error and delay. To address this gap, this dissertation extends TheAgentCompany with a finance-focused environment and investigates whether a general purpose LLM agent can complete representative wealth-management tasks both accurately and economically. This study introduces synthetic domain data, enriches colleague simulations, and prototypes an automatic task-generation pipeline. The study aims to create and assess an evaluation set that can meaningfully measure an agent's fitness for assistant-level wealth management work. We construct a benchmark of 12 task-pairs for wealth management assistants spanning retrieval, analysis, and synthesis/communication, with explicit acceptance criteria and deterministic graders. We seeded a set of new finance-specific data and introduced a high vs. low-autonomy variant of every task. The paper concluded that agents are limited less by mathematical reasoning and more so by end-to-end workflow reliability, and meaningfully affected by autonomy level, and that incorrect evaluation of models have hindered benchmarking.
The rapid shift from stateless large language models (LLMs) to autonomous, goal-driven agents raises a central question: When is agentic AI truly necessary? While agents enable multi-step reasoning, persistent memory, and tool orchestration, deploying them indiscriminately leads to higher cost, complexity, and risk. We present STRIDE (Systematic Task Reasoning Intelligence Deployment Evaluator), a framework that provides principled recommendations for selecting between three modalities: (i) direct LLM calls, (ii) guided AI assistants, and (iii) fully autonomous agentic AI. STRIDE integrates structured task decomposition, dynamism attribution, and self-reflection requirement analysis to produce an Agentic Suitability Score, ensuring that full agentic autonomy is reserved for tasks with inherent dynamism or evolving context. Evaluated across 30 real-world tasks spanning SRE, compliance, and enterprise automation, STRIDE achieved 92% accuracy in modality selection, reduced unnecessary agent deployments by 45%, and cut resource costs by 37%. Expert validation over six months in SRE and compliance domains confirmed its practical utility, with domain specialists agreeing that STRIDE effectively distinguishes between tasks requiring simple LLM calls, guided assistants, or full agentic autonomy. This work reframes agent adoption as a necessity-driven design decision, ensuring autonomy is applied only when its benefits justify the costs.
Converging evidence suggests that human systems of semantic categories achieve near-optimal compression via the Information Bottleneck (IB) complexity-accuracy tradeoff. Large language models (LLMs) are not trained for this objective, which raises the question: are LLMs capable of evolving efficient human-aligned semantic systems? To address this question, we focus on color categorization -- a key testbed of cognitive theories of categorization with uniquely rich human data -- and replicate with LLMs two influential human studies. First, we conduct an English color-naming study, showing that LLMs vary widely in their complexity and English-alignment, with larger instruction-tuned models achieving better alignment and IB-efficiency. Second, to test whether these LLMs simply mimic patterns in their training data or actually exhibit a human-like inductive bias toward IB-efficiency, we simulate cultural evolution of pseudo color-naming systems in LLMs via a method we refer to as Iterated in-Context Language Learning (IICLL). We find that akin to humans, LLMs iteratively restructure initially random systems towards greater IB-efficiency. However, only a model with strongest in-context capabilities (Gemini 2.0) is able to recapitulate the wide range of near-optimal IB-tradeoffs observed in humans, while other state-of-the-art models converge to low-complexity solutions. These findings demonstrate how human-aligned semantic categories can emerge in LLMs via the same fundamental principle that underlies semantic efficiency in humans.
We present Apertus, a fully open suite of large language models (LLMs) designed to address two systemic shortcomings in today's open model ecosystem: data compliance and multilingual representation. Unlike many prior models that release weights without reproducible data pipelines or regard for content-owner rights, Apertus models are pretrained exclusively on openly available data, retroactively respecting `robots.txt` exclusions and filtering for non-permissive, toxic, and personally identifiable content. To mitigate risks of memorization, we adopt the Goldfish objective during pretraining, strongly suppressing verbatim recall of data while retaining downstream task performance. The Apertus models also expand multilingual coverage, training on 15T tokens from over 1800 languages, with ~40% of pretraining data allocated to non-English content. Released at 8B and 70B scales, Apertus approaches state-of-the-art results among fully open models on multilingual benchmarks, rivalling or surpassing open-weight counterparts. Beyond model weights, we release all scientific artifacts from our development cycle with a permissive license, including data preparation scripts, checkpoints, evaluation suites, and training code, enabling transparent audit and extension.
Retrieval-Augmented Generation (RAG) and Supervised Finetuning (SFT) have become the predominant paradigms for equipping Large Language Models (LLMs) with external knowledge for diverse, knowledge-intensive tasks. However, while such knowledge injection improves performance, it also exposes new attack surfaces. Membership Inference Attacks (MIAs), which aim to determine whether a given data sample was included in a model's training set, pose serious threats to privacy and trust in sensitive domains. To this end, we first systematically evaluate the vulnerability of RAG- and SFT-based LLMs to various MIAs. Then, to address the privacy risk, we further introduce a novel, model-agnostic defense framework, Ensemble Privacy Defense (EPD), which aggregates and evaluates the outputs of a knowledge-injected LLM, a base LLM, and a dedicated judge model to enhance resistance against MIAs. Comprehensive experiments show that, on average, EPD reduces MIA success by up to 27.8\% for SFT and 526.3\% for RAG compared to inference-time baseline, while maintaining answer quality.
Recent advances in Large language models (LLMs) have demonstrated their promising capabilities of generating robot operation code to enable LLM-driven robots. To enhance the reliability of operation code generated by LLMs, corrective designs with feedback from the observation of executing code have been increasingly adopted in existing research. However, the code execution in these designs relies on either a physical experiment or a customized simulation environment, which limits their deployment due to the high configuration effort of the environment and the potential long execution time. In this paper, we explore the possibility of directly leveraging LLM to enable static simulation of robot operation code, and then leverage it to design a new reliable LLM-driven corrective robot operation code generation framework. Our framework configures the LLM as a static simulator with enhanced capabilities that reliably simulate robot code execution by interpreting actions, reasoning over state transitions, analyzing execution outcomes, and generating se- mantic observations that accurately capture trajectory dynamics. To validate the performance of our framework, we performed experiments on various operation tasks for different robots, including UAVs and small ground vehicles. The experiment results not only demonstrated the high accuracy of our static text-based simulation but also the reliable code generation of our LLM-driven corrective framework, which achieves a comparable performance with state-of-the-art research while does not rely on dynamic code execution using physical experiments or simulators.
We introduce LLM CHESS, an evaluation framework designed to probe the generalization of reasoning and instruction-following abilities in large language models (LLMs) through extended agentic interaction in the domain of chess. We rank over 50 open and closed source models by playing against a random opponent using a range of behavioral metrics, including win and loss rates, move quality, move legality, hallucinated actions, and game duration. For a subset of top reasoning models, we derive an Elo estimate by playing against a chess engine with variably configured skill, which allows for comparisons between models in an easily understandable way. Despite the simplicity of the instruction-following task and the weakness of the opponent, many state-of-the-art models struggle to complete games or achieve consistent wins. Similar to other benchmarks on complex reasoning tasks, our experiments reveal a clear separation between reasoning and non-reasoning models. However, unlike existing static benchmarks, the stochastic and dynamic nature of LLM CHESS uniquely reduces overfitting and memorization while preventing benchmark saturation, proving difficult even for top reasoning models. To support future work on evaluating reasoning and instruction-following in LLMs, we release our experimental framework, a public leaderboard, and a dataset of associated games.
As large language models (LLMs) become increasingly capable and widely adopted, benchmarks play a central role in assessing their practical utility. For example, SWE-Bench Verified has emerged as a critical benchmark for evaluating LLMs' software engineering abilities, particularly their aptitude for resolving real-world GitHub issues. Recent LLMs show impressive performance on SWE-Bench, leading to optimism about their capacity for complex coding tasks. However, current evaluation protocols may overstate these models' true capabilities. It is crucial to distinguish LLMs' generalizable problem-solving ability and other learned artifacts. In this work, we introduce two diagnostic tasks: file path identification from issue descriptions alone and ground truth function reproduction with only the current file context and issue description to probe models' underlying knowledge. We present empirical evidence that performance gains on SWE-Bench-Verified may be partially driven by memorization rather than genuine problem-solving. We show that state-of-the-art models achieve up to 76% accuracy in identifying buggy file paths using only issue descriptions, without access to repository structure. This performance is merely up to 53% on tasks from repositories not included in SWE-Bench, pointing to possible data contamination or memorization. Similar patterns are also observed for the function reproduction task, where the verbatim similarity is much higher on SWE-Bench Verified than on other similar coding benchmarks (up to 35% consecutive 5-gram accuracy on SWE-Bench Verified and Full, but only up to 18% for tasks in other benchmarks). These findings raise concerns about the validity of existing results and underscore the need for more robust, contamination-resistant benchmarks to reliably evaluate LLMs' coding abilities.
Recent advancements in large language models (LLMs) have been driven by their emergent reasoning capabilities, particularly through long chain-of-thought (CoT) prompting, which enables thorough exploration and deliberation. Despite these advances, long-CoT LLMs often exhibit suboptimal reasoning behaviors, such as overthinking and excessively protracted reasoning chains, which can impair performance. In this paper, we analyze reasoning processes through an optimization lens, framing CoT as a gradient descent procedure where each reasoning step constitutes an update toward problem resolution. Building on this perspective, we introduce RePro (Rectifying Process-level Reward), a novel approach to refine LLM reasoning during post-training. RePro defines a surrogate objective function to assess the optimization process underlying CoT, utilizing a dual scoring mechanism to quantify its intensity and stability. These scores are aggregated into a composite process-level reward, seamlessly integrated into reinforcement learning with verifiable rewards (RLVR) pipelines to optimize LLMs. Extensive experiments across multiple reinforcement learning algorithms and diverse LLMs, evaluated on benchmarks spanning mathematics, science, and coding, demonstrate that RePro consistently enhances reasoning performance and mitigates suboptimal reasoning behaviors.
Understanding the internal thinking process of Large Language Models (LLMs) and the cause of hallucinations remains a key challenge. To this end, we introduce latent debate, a novel framework for interpreting model predictions through the lens of implicit internal arguments. Unlike the current work of self-consistency and multi-agent debate, which relies on explicit debates among multiple answers or multiple models, latent debate captures the hidden supporting and attacking signals that arise within a single model during a single inference. We first present a model- and task-agnostic conceptual framework, and then instantiate it symbolically to approximate the thinking process of LLMs on True/False prediction tasks. Empirical studies demonstrate that latent debate is a faithful structured surrogate model that has highly consistent predictions with the original LLM. Beyond interpretability, we demonstrate that latent debate provides a strong baseline for hallucination detection. Further analysis reveals strong correlations between hallucinations and debate patterns, such as a high degree of latent debates in the middle layers is linked to a higher risk of hallucinations. These findings position latent debate as a potential framework for understanding internal mechanisms of LLMs, especially for scenarios where internal (dis)agreements appear during the inference steps.
A meaningful text can be hidden inside another, completely different yet still coherent and plausible, text of the same length. For example, a tweet containing a harsh political critique could be embedded in a tweet that celebrates the same political leader, or an ordinary product review could conceal a secret manuscript. This uncanny state of affairs is now possible thanks to Large Language Models, and in this paper we present Calgacus, a simple and efficient protocol to achieve it. We show that even modest 8-billion-parameter open-source LLMs are sufficient to obtain high-quality results, and a message as long as this abstract can be encoded and decoded locally on a laptop in seconds. The existence of such a protocol demonstrates a radical decoupling of text from authorial intent, further eroding trust in written communication, already shaken by the rise of LLM chatbots. We illustrate this with a concrete scenario: a company could covertly deploy an unfiltered LLM by encoding its answers within the compliant responses of a safe model. This possibility raises urgent questions for AI safety and challenges our understanding of what it means for a Large Language Model to know something.
Recently, two-stage fine-tuning strategies, e.g., acquiring essential driving knowledge through supervised fine-tuning (SFT) and further enhancing decision-making and planning via reinforcement fine-tuning (RFT), have shown strong potential in advancing the knowledge-driven autonomous driving (AD) paradigm. However, the learning nature of SFT still limits the generalization of reasoning, thereby constraining the full potential of driving performance. Meanwhile, current RFT approaches are primarily applied to downstream tasks, since scene understanding is an open-ended problem where corresponding rewards are difficult to quantify. To address these limitations, we propose OpenREAD, an OPEN-ended REasoning reinforced vision-language model (VLM)-based autonomous driving (AD) framework that enables end-to-end RFT across the full spectrum from high-level reasoning to low-level trajectory planning. Specifically, we begin by constructing large-scale Chain-of-Thought (CoT) annotations on open-source driving-related knowledge datasets, and employ the powerful Qwen3 large language model (LLM) as the critic in RFT to quantify reasoning quality for open-ended questions during reward modeling. Extensive experiments confirm that joint end-to-end RFT yields substantial improvements in both upstream and downstream tasks, enabling OpenREAD to achieve state-of-the-art performance on reasoning and planning benchmarks.
As Large Language Models (LLMs) become integrated into high-stakes domains, there is a growing need for evaluation methods that are both scalable for real-time deployment and reliable for critical decision-making. While human evaluation is reliable, it is slow and costly. Single LLM judges are biased, and static juries lack adaptability. To overcome these limitations, we propose LLM Jury-on-Demand - a dynamic, learning-based framework for scalable and context-aware evaluation. Our method trains a set of reliability predictors to assess when LLM judges will agree with human experts, leveraging token distributions, embeddings, and structural input features. This enables a fully adaptive evaluation where, for each data point, an optimal jury of the most reliable judges is dynamically selected, and their scores are aggregated using their reliability as weights. Experiments on summarization and RAG benchmarks show that our dynamic jury system achieves significantly higher correlation with human judgment than both single-judge and static-jury baselines. These results highlight the promise of adaptive, learning-based juries for building scalable, more reliable and trustworthy evaluation systems for modern LLMs in high-stakes domains.
The integration of robotics and augmented reality (AR) offers promising opportunities to enhance human-robot interaction (HRI) by making teleoperation more transparent, spatially grounded, and intuitive. We present a head-mounted AR "puppeteer" framework in which users control a physical robot via interacting with its virtual counterpart robot using large language model (LLM)-driven voice commands and hand-gesture interaction on the Meta Quest 3. In a within-subject user study with 42 participants performing an AR-based robotic pick-and-place pattern-matching task, we compare two interaction conditions: gesture-only (GO) and combined voice+gesture (VG). Our results show that GO currently provides more reliable and efficient control for this time-critical task, while VG introduces additional flexibility but also latency and recognition issues that can increase workload. We further explore how prior robotics experience shapes participants' perceptions of each modality. Based on these findings, we distill a set of evidence-based design guidelines for AR puppeteer metaphoric robot teleoperation, implicating multimodality as an adaptive strategy that must balance efficiency, robustness, and user expertise rather than assuming that additional modalities are universally beneficial. Our work contributes empirical insights into how multimodal (voice+gesture) interaction influences task efficiency, usability, and user experience in AR-based HRI.
We introduce POLLUX, a comprehensive open-source benchmark designed to evaluate the generative capabilities of large language models (LLMs) in Russian. Our main contribution is a novel evaluation methodology that enhances the interpretability of LLM assessment. For each task type, we define a set of detailed criteria and develop a scoring protocol where models evaluate responses and provide justifications for their ratings. This enables transparent, criteria-driven evaluation beyond traditional resource-consuming, side-by-side human comparisons. POLLUX includes a detailed, fine-grained taxonomy of 35 task types covering diverse generative domains such as code generation, creative writing, and practical assistant use cases, totaling 2,100 manually crafted and professionally authored prompts. Each task is categorized by difficulty (easy/medium/hard), with experts constructing the dataset entirely from scratch. We also release a family of LLM-as-a-Judge (7B and 32B) evaluators trained for nuanced assessment of generative outputs. This approach provides scalable, interpretable evaluation and annotation tools for model development, effectively replacing costly and less precise human judgments.
Large language model (LLM)-based agentic frameworks increasingly adopt the paradigm of dynamically generating task-specific agents. We suggest that not only agents but also specialized software modules for scientific and engineering tasks can be generated on demand. We demonstrate this concept in the field of solid mechanics. There, so-called constitutive models are required to describe the relationship between mechanical stress and body deformation. Constitutive models are essential for both the scientific understanding and industrial application of materials. However, even recent data-driven methods of constitutive modeling, such as constitutive artificial neural networks (CANNs), still require substantial expert knowledge and human labor. We present a framework in which an LLM generates a CANN on demand, tailored to a given material class and dataset provided by the user. The framework covers LLM-based architecture selection, integration of physical constraints, and complete code generation. Evaluation on three benchmark problems demonstrates that LLM-generated CANNs achieve accuracy comparable to or greater than manually engineered counterparts, while also exhibiting reliable generalization to unseen loading scenarios and extrapolation to large deformations. These findings indicate that LLM-based generation of physics-constrained neural networks can substantially reduce the expertise required for constitutive modeling and represent a step toward practical end-to-end automation.
Metal-organic framework (MOF) databases have grown rapidly through experimental deposition and large-scale literature extraction, but recent analyses show that nearly half of their entries contain substantial structural errors. These inaccuracies propagate through high-throughput screening and machine-learning workflows, limiting the reliability of data-driven MOF discovery. Correcting such errors is exceptionally difficult because true repairs require integrating crystallographic files, synthesis descriptions, and contextual evidence scattered across the literature. Here we introduce LitMOF, a large language model-driven multi-agent framework that validates crystallographic information directly from the original literature and cross-validates it with database entries to repair structural errors. Applying LitMOF to the experimental MOF database (the CSD MOF Subset), we constructed LitMOF-DB, a curated set 118,464 computation-ready structures, including corrections of 69% (6,161 MOFs) of the invalid MOFs in the latest CoRE MOF database. Additionally, the system uncovered 12,646 experimentally reported MOFs absent from existing resources, substantially expanding the known experimental design space. This work establishes a scalable pathway toward self-correcting scientific databases and a generalizable paradigm for LLM-driven curation in materials science.
Anomaly detection (AD) is a fundamental task of critical importance across numerous domains. Current systems increasingly operate in rapidly evolving environments that generate diverse yet interconnected data modalities -- such as time series, system logs, and tabular records -- as exemplified by modern IT systems. Effective AD methods in such environments must therefore possess two critical capabilities: (1) the ability to handle heterogeneous data formats within a unified framework, allowing the model to process and detect multiple modalities in a consistent manner during anomalous events; (2) a strong generalization ability to quickly adapt to new scenarios without extensive retraining. However, most existing methods fall short of these requirements, as they typically focus on single modalities and lack the flexibility to generalize across domains. To address this gap, we introduce a novel paradigm: In-Context Anomaly Detection (ICAD), where anomalies are defined by their dissimilarity to a relevant reference set of normal samples. Under this paradigm, we propose ICAD-LLM, a unified AD framework leveraging Large Language Models' in-context learning abilities to process heterogeneous data within a single model. Extensive experiments demonstrate that ICAD-LLM achieves competitive performance with task-specific AD methods and exhibits strong generalization to previously unseen tasks, which substantially reduces deployment costs and enables rapid adaptation to new environments. To the best of our knowledge, ICAD-LLM is the first model capable of handling anomaly detection tasks across diverse domains and modalities.
Ensuring LLM reliability requires not only solving complex problems but also recognizing when a problem is unsolvable. Current models often struggle to distinguish objective unsolvability (inherent contradictions in the problem) from subjective capability limitations (problems beyond the model's competence), which leads to hallucinations and overconfidence. To address this, we propose UnsolvableQA and UnsolvableRL to solve feasible problems, detect inherent contradictions, and prudently refuse tasks beyond capability. Specifically, we construct UnsolvableQA, a dataset of paired solvable and unsolvable instances derived via a dual-track methodology: programmatic generation for logic puzzles and a novel "Reverse Construction" method that injects contradictions into valid reasoning chains for mathematics. Building on this dataset, we introduce UnsolvableRL, a reinforcement learning framework with three reward components jointly accounting for accuracy, unsolvability, and difficulty. Empirical results show that our approach achieves near-perfect unsolvability detection while also improving accuracy on solvable tasks. Crucially, we identify Capability Collapse, demonstrating that explicit exposure to unsolvable data is indispensable for preventing models from becoming systematically overconfident. Our code and data are available at https://github.com/sfasfaffa/unsolvableQA.
This work presents a systematic characterization of Large Language Model (LLM) inference to address fragmented understanding. Through comprehensive experiments, we establish a four-dimensional analytical framework: (1) Two-Phase Heterogeneity Observation; (2) Microarchitectural Root Cause Analysis; (3) System Scaling Principles; and (4) Emerging Paradigm Boundaries. Our investigation progresses systematically from observation to foresight: identifying performance phenomena, revealing hardware causes, validating system behavior, and exploring new paradigms. This study not only consolidates a reliable empirical foundation for existing research but also provides new discoveries and practical optimization guidance for LLM inference.
With the growing use of Large Language Model (LLM)-based tools like ChatGPT, Perplexity, and Gemini across industries, there is a rising need for efficient LLM inference systems. These systems handle requests with a unique two-phase computation structure: a prefill-phase that processes the full input prompt and a decode-phase that autoregressively generates tokens one at a time. This structure calls for new strategies for routing and scheduling requests. In this paper, we take a comprehensive approach to this challenge by developing a theoretical framework that models routing and scheduling in LLM inference systems. We identify two key design principles-optimal tiling and dynamic resource allocation-that are essential for achieving high throughput. Guided by these principles, we propose the Resource-Aware Dynamic (RAD) scheduler and prove that it achieves throughput optimality under mild conditions. To address practical Service Level Objectives (SLOs) such as serving requests with different Time Between Token (TBT) constraints, we design the SLO-Aware LLM Inference (SLAI) scheduler. SLAI uses real-time measurements to prioritize decode requests that are close to missing their TBT deadlines and reorders prefill requests based on known prompt lengths to further reduce the Time To First Token (TTFT) delays. We evaluate SLAI on the Openchat ShareGPT4 dataset using the Mistral-7B model on an NVIDIA RTX ADA 6000 GPU. Compared to Sarathi-Serve, SLAI reduces the median TTFT by 53% and increases the maximum serving capacity by 26% such that median TTFT is below 0.5 seconds, while meeting tail TBT latency constraints.
Multi-Agent System (MAS) developing frameworks serve as the foundational infrastructure for social simulations powered by Large Language Models (LLMs). However, existing frameworks fail to adequately support large-scale simulation development due to inherent limitations in adaptability, configurability, reliability, and code reusability. For example, they cannot simulate a society where the agent population and profiles change over time. To fill this gap, we propose Agent-Kernel, a framework built upon a novel society-centric modular microkernel architecture. It decouples core system functions from simulation logic and separates cognitive processes from physical environments and action execution. Consequently, Agent-Kernel achieves superior adaptability, configurability, reliability, and reusability. We validate the framework's superiority through two distinct applications: a simulation of the Universe 25 (Mouse Utopia) experiment, which demonstrates the handling of rapid population dynamics from birth to death; and a large-scale simulation of the Zhejiang University Campus Life, successfully coordinating 10,000 heterogeneous agents, including students and faculty.
Recent studies suggest that the representations learned by large language models (LLMs) are partially aligned to those of the human brain. However, whether and why this alignment score arises from a similar sequence of computations remains elusive. In this study, we explore this question by examining temporally-resolved brain signals of participants listening to 10 hours of an audiobook. We study these neural dynamics jointly with a benchmark encompassing 22 LLMs varying in size and architecture type. Our analyses confirm that LLMs and the brain generate representations in a similar order: specifically, activations in the initial layers of LLMs tend to best align with early brain responses, while the deeper layers of LLMs tend to best align with later brain responses. This brain-LLM alignment is consistent across transformers and recurrent architectures. However, its emergence depends on both model size and context length. Overall, this study sheds light on the sequential nature of computations and the factors underlying the partial convergence between biological and artificial neural networks.
Large language models (LLMs) have achieved remarkable success in many natural language tasks but still struggle with complex, multi-step reasoning, particularly across diverse disciplines. Existing reasoning datasets often lack disciplinary breadth, reasoning depth, and diversity, as well as guiding principles for question synthesis. We propose DESIGNER: a DESIGN-logic-guidEd Reasoning data synthesis pipeline that leverages naturally available, extensive raw documents (e.g., book corpus and web corpus) to generate multidisciplinary challenging questions. We introduce the concept of "design logic" and instruct LLMs to mimic human educators' question-creation process, enabling the automated synthesis of large-scale, high-difficulty questions. We use LLMs to reverse-engineer and abstract over 120,000 design logics from existing questions across various disciplines. By matching these design logics with source documents, we are able to generate reasoning questions with controllable question types and difficulty levels. Using this pipeline, we synthesized two large-scale reasoning datasets that span 75 disciplines: DLR-Book (3.04 million questions from the book corpus) and DLR-Web (1.66 million questions from the web corpus). Data analysis indicates that the questions synthesized by our method exhibit greater difficulty and diversity compared to those in the baseline datasets. We validate our synthesized data through supervised fine-tuning (SFT) on the Qwen3 and Llama3 model families. Our data substantially enhances their multidisciplinary reasoning capabilities, outperforming existing datasets. Notably, by applying SFT on the base versions of these models using only our data, we even surpass their official final models that have undergone the full post-training process.
The use of large language models (LLMs) in scholarly publications has grown dramatically since the launch of ChatGPT in late 2022. This usage is often undisclosed, and it can be challenging for readers and reviewers to identify human written but LLM-revised or translated text, or predominantly LLM-generated text. Given the known quality and reliability issues connected with LLM-generated text, their potential growth poses an increasing problem for research integrity, and for public trust in research. This study presents a simple and easily reproducible methodology to show the growth in the full text of published papers, across the full range of research, as indexed in the Dimensions database. It uses this to demonstrate that LLM tools are likely to have been involved in the production of more than 10% of all published papers in 2024, based on disproportionate use of specific indicative words, and draws together earlier studies to confirm that this is a plausible overall estimate. It then discusses the implications of this for the integrity of scholarly publishing, highlighting evidence that use of LLMs for text generation is still being concealed or downplayed by authors, and presents an argument that more comprehensive disclosure requirements are urgently required to address this.
LLMs are increasingly executed in edge where limited GPU memory and heterogeneous computation jointly constrain deployment which motivates model partitioning and request scheduling. In this setting, minimizing latency requires addressing the tight coupling between model placement and request scheduling across heterogeneous nodes, as suboptimal decisions in one domain can negate benefits in the other. In this paper, we propose Hyperion, a hierarchical two-stage framework that jointly optimizes partitioning and scheduling for pipelined LLM inference. Hyperion minimizes latency by balancing resources across tiers without requiring model retraining or incurring significant runtime overhead. Leveraging the timescale difference between partitioning and request arrivals, Stage 1 performs offline, inter-tier partitioning via a Hyperion Split with Dynamic Programming (HypSplit-DP) procedure to produce balanced stage times under tier capacity and memory constraints; to adapt to time-varying load, Stage 2 performs online, intra-tier scheduling with a lightweight Hyperion Scheduling for Real-Time (HypSched-RT) that maps each request to the best available node using real-time estimates of queue length and effective capacity. Experiments with Phi-3-medium demonstrate that Hyperion reduces latency by up to 52.1% (vs. GPipe) and 31.2% (vs. HEFT). Furthermore, Hyperion exhibits superior scalability for long-sequence generation, maintaining 44.5% lower latency and higher GPU utilization.
While model serving has unlocked unprecedented capabilities, the high cost of serving large-scale models continues to be a significant barrier to widespread accessibility and rapid innovation. Compiler optimizations have long driven substantial performance improvements, but existing compilers struggle with neural workloads due to the exponentially large and highly interdependent space of possible transformations. Although existing stochastic search techniques can be effective, they are often sample-inefficient and fail to leverage the structural context underlying compilation decisions. We set out to investigate the research question of whether reasoning with large language models (LLMs), without any retraining, can leverage the context-aware decision space of compiler optimizations to significantly improve sample efficiency. To that end, we introduce a novel compilation framework (dubbed Reasoning Compiler) that formulates optimization as a sequential, context-aware decision process guided by a large language model and structured Monte Carlo tree search (MCTS). The LLM acts as a proposal mechanism, suggesting hardware-informed transformations that reflect the current program state and accumulated performance feedback. MCTS incorporates the LLM-generated proposals to balance exploration and exploitation, facilitating structured, context-sensitive traversal of the expansive compiler optimization space. By achieving substantial speedups with markedly fewer samples than leading neural compilers, our approach demonstrates the potential of LLM-guided reasoning to transform the landscape of compiler optimization.
This study presents PARROT (Persuasion and Agreement Robustness Rating of Output Truth), a robustness focused framework designed to measure the degradation in accuracy that occurs under social pressure exerted on users through authority and persuasion in large language models (LLMs) the phenomenon of sycophancy (excessive conformity). PARROT (i) isolates causal effects by comparing the neutral version of the same question with an authoritatively false version using a double-blind evaluation, (ii) quantifies confidence shifts toward the correct and imposed false responses using log-likelihood-based calibration tracking, and (iii) systematically classifies failure modes (e.g., robust correct, sycophantic agreement, reinforced error, stubborn error, self-correction, etc.) using an eight-state behavioral taxonomy. We evaluated 22 models using 1,302 MMLU-style multiple-choice questions across 13 domains and domain-specific authority templates. Findings show marked heterogeneity: advanced models (e.g., GPT-5, GPT-4.1, Claude Sonnet 4.5) exhibit low "follow rates" ($\leq 11\%$, GPT-5: 4\%) and minimal accuracy loss, while older/smaller models show severe epistemic collapse (GPT-4: 80\%, Qwen 2.5-1.5B: 94\%). The danger is not limited to response changes; weak models reduce confidence in the correct response while increasing confidence in the imposed incorrect response. While international law and global knowledge at the domain level exhibit high fragility, elementary mathematics is relatively resilient. Consequently, we argue that the goal of "resistance to overfitting pressure" should be addressed as a primary objective alongside accuracy, harm avoidance, and privacy for safe deployment in the real world.
We introduce CollabToolBuilder, a flexible multiagent LLM framework with expert-in-the-loop (HITL) guidance that iteratively learns to create tools for a target goal, aligning with human intent and process, while minimizing time for task/domain adaptation effort and human feedback capture. The architecture generates and validates tools via four specialized agents (Coach, Coder, Critic, Capitalizer) using a reinforced dynamic prompt and systematic human feedback integration to reinforce each agent's role toward goals and constraints. This work is best viewed as a system-level integration and methodology combining multi-agent in-context learning, HITL controls, and reusable tool capitalization for complex iterative problems such as scientific document generation. We illustrate it with preliminary experiments (e.g., generating state-of-the-art research papers or patents given an abstract) and discuss its applicability to other iterative problem-solving.
In large-scale hypothesis testing, computing exact $p$-values or $e$-values is often resource-intensive, creating a need for budget-aware inferential methods. We propose a general framework for active hypothesis testing that leverages inexpensive auxiliary statistics to allocate a global computational budget. For each hypothesis, our data-adaptive procedure probabilistically decides whether to compute the exact test statistic or a transformed proxy, guaranteeing a valid $p$-value or $e$-value while satisfying the budget constraint in expectation. Theoretical guarantees are established for our constructions, showing that the procedure achieves optimality for $e$-values and for $p$-values under independence, and admissibility for $p$-values under general dependence. Empirical results from simulations and two real-world applications, including a large-scale genome-wide association study (GWAS) and a clinical prediction task leveraging large language models (LLM), demonstrate that our framework improves statistical efficiency under fixed resource limits.
Large Language Models (LLMs) are increasingly employed for simulating human behaviors across diverse domains. However, our position is that current LLM-based human simulations remain insufficiently reliable, as evidenced by significant discrepancies between their outcomes and authentic human actions. Our investigation begins with a systematic review of LLM-based human simulations in social, economic, policy, and psychological contexts, identifying their common frameworks, recent advances, and persistent limitations. This review reveals that such discrepancies primarily stem from inherent limitations of LLMs and flaws in simulation design, both of which are examined in detail. Building on these insights, we propose a systematic solution framework that emphasizes enriching data foundations, advancing LLM capabilities, and ensuring robust simulation design to enhance reliability. Finally, we introduce a structured algorithm that operationalizes the proposed framework, aiming to guide credible and human-aligned LLM-based simulations. To facilitate further research, we provide a curated list of related literature and resources at https://github.com/Persdre/awesome-llm-human-simulation.
Energy system models are increasingly employed to guide long-term planning in multi-sectoral environments where decisions span electricity, heat, transport, land use, and industry. While these models provide rigorous quantitative insights, their outputs are often highly technical, making them difficult to interpret for non-expert stakeholders such as policymakers, planners, and the public. This communication gap limits the accessibility and practical impact of scenario-based modeling, particularly as energy transitions grow more complex with rising shares of renewables, sectoral integration, and deep uncertainties. To address this challenge, we propose the Renewable Energy Large Language Model (RE-LLM), a hybrid framework that integrates Large Language Models (LLMs) directly into the energy system modeling workflow. RE-LLM combines three core elements: (i) optimization-based scenario exploration, (ii) machine learning surrogates that accelerate computationally intensive simulations, and (iii) LLM-powered natural language generation that translates complex results into clear, stakeholder-oriented explanations. This integrated design not only reduces computational burden but also enhances inter-pretability, enabling real-time reasoning about trade-offs, sensitivities, and policy implications. The framework is adaptable across different optimization platforms and energy system models, ensuring broad applicability beyond the case study presented. By merging speed, rigor, and interpretability, RE-LLM advances a new paradigm of human-centric energy modeling. It enables interactive, multilingual, and accessible engagement with future energy pathways, ultimately bridging the final gap between data-driven analysis and actionable decision-making for sustainable transitions.
Serverless Large Language Models (LLMs) have emerged as a cost-effective solution for deploying AI services by enabling a 'pay-as-you-go' pricing model through GPU resource sharing. However, cold-start latency, especially the model loading phase, has become a critical performance bottleneck, as it scales linearly with model size and severely limits the practical deployment of large-scale LLM services. This paper presents Tangram, a novel system that accelerates Serverless LLM loading through efficient GPU memory reuse. By leveraging the unused GPU memory to retain model parameters, Tangram significantly reduces model transfer time and cold-start latency. Its design includes three key components: unified GPU memory pool for tensor-level parameter sharing across models, on-demand KV cache allocation for dynamic memory management, and GPU-affinity-aware scheduling for maximizing resource utilization. These techniques collectively address the critical challenges of inefficient memory usage and the cold-start problem in Serverless LLM platforms. We have implemented a fully functional prototype, and experiments show that Tangram achieves up to 6.2 times faster loading and reduces Time-To-First-Token (TTFT) during cold-start by 23--55% over state-of-the-art methods.
Code review is critical for ensuring software quality and maintainability. With the rapid growth in software scale and complexity, code review has become a bottleneck in the development process because of its time-consuming and knowledge-intensive nature and the shortage of experienced developers willing to review code. Several approaches have been proposed for automatically generating code reviews based on retrieval, neural machine translation, pre-trained models, or large language models (LLMs). These approaches mainly leverage historical code changes and review comments. However, a large amount of crucial information for code review, such as the context of code changes and prior review knowledge, has been overlooked. This paper proposes an LLM-based review knowledge-augmented, context-aware framework for code review generation, named LAURA. The framework integrates review exemplar retrieval, context augmentation, and systematic guidance to enhance the performance of ChatGPT-4o and DeepSeek v3 in generating code review comments. Besides, given the extensive low-quality reviews in existing datasets, we also constructed a high-quality dataset. Experimental results show that for both models, LAURA generates review comments that are either completely correct or at least helpful to developers in 42.2% and 40.4% of cases, respectively, significantly outperforming SOTA baselines. Furthermore, our ablation studies demonstrate that all components of LAURA contribute positively to improving comment quality.
Large language models (LLMs) remain vulnerable to jailbreak attacks that bypass safety guardrails to elicit harmful outputs. Existing approaches overwhelmingly operate within the prompt-optimization paradigm: whether through traditional algorithmic search or recent agent-based workflows, the resulting prompts typically retain malicious semantic signals that modern guardrails are primed to detect. In contrast, we identify a deeper, largely overlooked vulnerability stemming from the highly interconnected nature of an LLM's internal knowledge. This structure allows harmful objectives to be realized by weaving together sequences of benign sub-queries, each of which individually evades detection. To exploit this loophole, we introduce the Correlated Knowledge Attack Agent (CKA-Agent), a dynamic framework that reframes jailbreaking as an adaptive, tree-structured exploration of the target model's knowledge base. The CKA-Agent issues locally innocuous queries, uses model responses to guide exploration across multiple paths, and ultimately assembles the aggregated information to achieve the original harmful objective. Evaluated across state-of-the-art commercial LLMs (Gemini2.5-Flash/Pro, GPT-oss-120B, Claude-Haiku-4.5), CKA-Agent consistently achieves over 95% success rates even against strong guardrails, underscoring the severity of this vulnerability and the urgent need for defenses against such knowledge-decomposition attacks. Our codes are available at https://github.com/Graph-COM/CKA-Agent.
We introduce a novel framework for measuring Overton pluralism in LLMs--the extent to which diverse viewpoints are represented in model outputs. We (i) formalize Overton pluralism as a set coverage metric (OvertonScore), (ii) conduct a large-scale U.S.-representative human study (N = 1209; 60 questions; 8 LLMs), and (iii) develop an automated benchmark that closely reproduces human judgments. On average, models achieve OvertonScores of 0.35--0.41, with DeepSeek V3 performing best; yet all models remain far below the theoretical maximum of 1.0, revealing substantial headroom for improvement. Because repeated large-scale human studies are costly and slow, scalable evaluation tools are essential for model development. Hence, we propose an automated benchmark that achieves high rank correlation with human judgments ($ρ=0.88$), providing a practical proxy without replacing human assessment. By turning pluralistic alignment from a normative aim into a measurable benchmark, our work establishes a foundation for systematic progress toward more pluralistic LLMs.
Large Language Models (LLMs) are being increasingly deployed in real-world applications, but they remain susceptible to hallucinations, which produce fluent yet incorrect responses and lead to erroneous decision-making. Uncertainty estimation is a feasible approach to detect such hallucinations. For example, semantic entropy estimates uncertainty by considering the semantic diversity across multiple sampled responses, thus identifying hallucinations. However, semantic entropy relies on post-softmax probabilities and fails to capture the model's inherent uncertainty, causing it to be ineffective in certain scenarios. To address this issue, we introduce Semantic Energy, a novel uncertainty estimation framework that leverages the inherent confidence of LLMs by operating directly on logits of penultimate layer. By combining semantic clustering with a Boltzmann-inspired energy distribution, our method better captures uncertainty in cases where semantic entropy fails. Experiments across multiple benchmarks show that Semantic Energy significantly improves hallucination detection and uncertainty estimation, offering more reliable signals for downstream applications such as hallucination detection.
Large Language Models (LLMs) are increasingly being deployed in real-world applications, but their flexibility exposes them to prompt injection attacks. These attacks leverage the model's instruction-following ability to make it perform malicious tasks. Recent work has proposed JATMO, a task-specific fine-tuning approach that trains non-instruction-tuned base models to perform a single function, thereby reducing susceptibility to adversarial instructions. In this study, we evaluate the robustness of JATMO against HOUYI, a genetic attack framework that systematically mutates and optimizes adversarial prompts. We adapt HOUYI by introducing custom fitness scoring, modified mutation logic, and a new harness for local model testing, enabling a more accurate assessment of defense effectiveness. We fine-tuned LLaMA 2-7B, Qwen1.5-4B, and Qwen1.5-0.5B models under the JATMO methodology and compared them with a fine-tuned GPT-3.5-Turbo baseline. Results show that while JATMO reduces attack success rates relative to instruction-tuned models, it does not fully prevent injections; adversaries exploiting multilingual cues or code-related disruptors still bypass defenses. We also observe a trade-off between generation quality and injection vulnerability, suggesting that better task performance often correlates with increased susceptibility. Our results highlight both the promise and limitations of fine-tuning-based defenses and point toward the need for layered, adversarially informed mitigation strategies.
Large language models (LLMs) are reshaping the recommender system paradigm by enabling users to express preferences and receive recommendations through conversations. Yet, aligning LLMs to the recommendation task remains challenging: pretrained LLMs often generate out-of-catalog items, violate required output formats, and their ranking quality degrades sharply toward the end of the generated list. To this end, we propose ConvRec-R1, a two-stage framework for end-to-end training of LLM-based conversational recommender systems. In Stage 1, we construct a behavioral-cloning dataset with a Remap-Reflect-Adjust pipeline, which produces high-quality, catalog-grounded demonstrations from powerful blackbox LLMs to warm-start the RL training. In Stage 2, we propose Rank-GRPO, a principled extension of group relative policy optimization (GRPO) tailored to tasks with rank-style outputs. Rank-GRPO treats each rank in the recommendation list as the unit instead of token (too fine-grained) or sequence (too coarse), redefining rewards to remove non-causal credit assignment and introducing a rank-level importance ratio based on the geometric mean of rank-wise token probabilities to stabilize policy updates. Experiments on the public Reddit-v2 dataset show that ConvRec-R1 converges faster and achieves higher Recall and NDCG than GRPO-style baselines. Code and datasets are released at https://github.com/yaochenzhu/Rank-GRPO.
As web platforms evolve towards greater personalization and emotional complexity, conversational agents must transcend superficial empathy to demonstrate identity-aware emotional reasoning. However, existing systems face two limitations: (1) reliance on situation-centric datasets lacking persistent user identity, which hampers the capture of personalized affective nuances; and (2) dependence on opaque, coarse reward signals that hinder development of verifiable empathetic reasoning. To address these gaps, we introduce KardiaBench, a large-scale user-grounded benchmark comprising 178,080 QA pairs across 22,080 multi-turn conversations anchored to 671 real-world profiles. The dataset is constructed via a model-in-the-loop pipeline with iterative rubric-guided refinement to ensure psychological plausibility and persona consistency. This progressive empathy pipeline that integrates user comprehension, contextual reasoning, and emotion perception into conversations, followed by iterative critique and rubric-based refinement to ensure psychological plausibility, emotional fidelity, and persona consistency. Building on this, we propose Kardia-R1, a framework that trains models for interpretable, stepwise empathetic cognition. Kardia-R1 leverages Rubric-as-Judge Empathetic Reinforcement Learning (Rubric-ERL), a GRPO-based method that uses explainable, human-aligned rubric rewards to tightly couple user understanding, emotional inference, and supportive response generation. Extensive experiments across four LLM backbones demonstrate that Kardia-R1 consistently outperforms othet methods in emotion accuracy, empathy, relevance, persona consistency, and safety. Our dataset and model will be released at https://github.com/JhCircle/Kardia-R1.
Resolution of complex SQL issues persists as a significant bottleneck in real-world database applications. Current Large Language Models (LLMs), while adept at text-to-SQL translation, have not been rigorously evaluated on the more challenging task of debugging SQL issues. To address this gap, we introduce BIRD-CRITIC, a new SQL issue debugging benchmark comprising 530 PostgreSQL tasks (BIRD-CRITIC-PG) and 570 multi-dialect tasks (BIRD-CRITIC-Multi), distilled from authentic user issues and replayed within new environments to facilitate rigorous evaluation. Baseline evaluations underscore the task's complexity, with the leading reasoning model O3-Mini achieving only 38.87% success rate on BIRD-CRITIC-PG and 33.33% on BIRD-CRITIC-Multi. Meanwhile, advancing open-source models for database tasks is crucial for empowering local development while safeguarding data privacy. Therefore, we present Six-Gym (Sql-fIX-Gym), a training environment for elevating open-source model capabilities for SQL issue debugging. This environment leverages SQL-Rewind strategy, which automatically generates executable issue-solution datasets by reverse-engineering issues from verified SQLs. However, popular trajectory-based fine-tuning methods do not explore substantial supervisory signals. We further propose f-Plan Boosting, which extracts high-level debugging plans from SQL solutions, enabling teacher LLMs to produce 73.7% more successful trajectories for training. We integrate these components into an open-source agent, Bird-Fixer. Based on Qwen-2.5-Coder-14B, Bird-Fixer achieves 38.11% success rate on BIRD-CRITIC-PG and 29.65% on BIRD-CRITIC-Multi, surpassing leading proprietary models such as Claude-3.7-Sonnet and GPT-4.1, marking a significant step toward democratizing sophisticated SQL-debugging capabilities. The leaderboard and source code are available: https://bird-critic.github.io/
Should LLM reasoning live in a separate module, or within a single model's forward pass and representational space? We study dual-architecture latent reasoning, where a fluent Base exchanges latent messages with a Coprocessor, and test two hypotheses aimed at improving latent communication over Liu et al. (2024): (H1) increase channel capacity; (H2) learn communication via joint finetuning. Under matched latent-token budgets on GPT-2 and Qwen-3, H2 is consistently strongest while H1 yields modest gains. A unified soft-embedding baseline, a single model with the same forward pass and shared representations, using the same latent-token budget, nearly matches H2 and surpasses H1, suggesting current dual designs mostly add compute rather than qualitatively improving reasoning. Across GSM8K, ProsQA, and a Countdown stress test with increasing branching factor, scaling the latent-token budget beyond small values fails to improve robustness. Latent analyses show overlapping subspaces with limited specialization, consistent with weak reasoning gains. We conclude dual-model latent reasoning remains promising in principle, but likely requires objectives and training schedules that explicitly shape latent spaces for algorithmic planning.
Assessing software test coverage at scale remains a bottleneck in QA pipelines. We present LLM-as-a-Judge (LAJ), a production-ready, rubric-driven framework for evaluating Gherkin acceptance tests with structured JSON outputs. Across 20 model configurations (GPT-4, GPT-5 with varying reasoning effort, and open-weight models) on 100 expert-annotated scripts over 5 runs (500 evaluations), we provide the first comprehensive analysis spanning accuracy, operational reliability, and cost. We introduce the Evaluation Completion Rate (ECR@1) to quantify first-attempt success, revealing reliability from 85.4% to 100.0% with material cost implications via retries. Results show that smaller models can outperform larger ones: GPT-4o Mini attains the best accuracy (6.07 MAAE), high reliability (96.6% ECR@1), and low cost ($1.01 per 1K), yielding a 78x cost reduction vs. GPT-5 (high reasoning) while improving accuracy. Reasoning effort is model-family dependent: GPT-5 benefits from increased reasoning (with predictable accuracy-cost tradeoffs), whereas open-weight models degrade across all dimensions as reasoning increases. Overall, cost spans 175x ($0.45-$78.96 per 1K). We release the dataset, framework, and code to support reproducibility and deployment.