Current benchmarks for AI clinician systems, often based on multiple-choice exams or manual rubrics, fail to capture the depth, robustness, and safety required for real-world clinical practice. To address this, we introduce the GAPS framework, a multidimensional paradigm for evaluating \textbf{G}rounding (cognitive depth), \textbf{A}dequacy (answer completeness), \textbf{P}erturbation (robustness), and \textbf{S}afety. Critically, we developed a fully automated, guideline-anchored pipeline to construct a GAPS-aligned benchmark end-to-end, overcoming the scalability and subjectivity limitations of prior work. Our pipeline assembles an evidence neighborhood, creates dual graph and tree representations, and automatically generates questions across G-levels. Rubrics are synthesized by a DeepResearch agent that mimics GRADE-consistent, PICO-driven evidence review in a ReAct loop. Scoring is performed by an ensemble of large language model (LLM) judges. Validation confirmed our automated questions are high-quality and align with clinician judgment. Evaluating state-of-the-art models on the benchmark revealed key failure modes: performance degrades sharply with increased reasoning depth (G-axis), models struggle with answer completeness (A-axis), and they are highly vulnerable to adversarial perturbations (P-axis) as well as certain safety issues (S-axis). This automated, clinically-grounded approach provides a reproducible and scalable method for rigorously evaluating AI clinician systems and guiding their development toward safer, more reliable clinical practice.
Generative AI systems are increasingly assisting and acting on behalf of end users in practical settings, from digital shopping assistants to next-generation autonomous cars. In this context, safety is no longer about blocking harmful content, but about preempting downstream hazards like financial or physical harm. Yet, most AI guardrails continue to rely on output classification based on labeled datasets and human-specified criteria,making them brittle to new hazardous situations. Even when unsafe conditions are flagged, this detection offers no path to recovery: typically, the AI system simply refuses to act--which is not always a safe choice. In this work, we argue that agentic AI safety is fundamentally a sequential decision problem: harmful outcomes arise from the AI system's continually evolving interactions and their downstream consequences on the world. We formalize this through the lens of safety-critical control theory, but within the AI model's latent representation of the world. This enables us to build predictive guardrails that (i) monitor an AI system's outputs (actions) in real time and (ii) proactively correct risky outputs to safe ones, all in a model-agnostic manner so the same guardrail can be wrapped around any AI model. We also offer a practical training recipe for computing such guardrails at scale via safety-critical reinforcement learning. Our experiments in simulated driving and e-commerce settings demonstrate that control-theoretic guardrails can reliably steer LLM agents clear of catastrophic outcomes (from collisions to bankruptcy) while preserving task performance, offering a principled dynamic alternative to today's flag-and-block guardrails.
Assistive agents should not only take actions on behalf of a human, but also step out of the way and cede control when there are important decisions to be made. However, current methods for building assistive agents, whether via mimicking expert humans or via RL finetuning on an inferred reward, often encourage agents to complete tasks on their own rather than truly assisting the human attain their objectives. Additionally, these methods often require costly explicit human feedback to provide a training signal. We propose a new approach to tuning assistive language models based on maximizing the human's empowerment, their ability to effect desired changes in the environment. Our empowerment-maximizing method, Empower, only requires offline text data, providing a self-supervised method for fine-tuning language models to better assist humans. To study the efficacy of our approach, we conducted an 18-person user study comparing our empowerment assistant with a strong baseline. Participants preferred our assistant 78% of the time (p=0.015), with a 31% higher acceptance rate and 38% fewer suggestions. Additionally, we introduce a new environment for evaluating multi-turn code assistance using simulated humans. Using this environment, we show that agents trained with Empower increase the success rate of a simulated human programmer on challenging coding questions by an average of 192% over an SFT baseline. With this empowerment objective, we provide a framework for useful aligned AI agents at scale using only offline data without the need for any additional human feedback or verifiable rewards.
The emergence of large language models (LLMs) has opened new opportunities for cre- ating dynamic non-player characters (NPCs) in gaming environments, enabling both func- tional task execution and persona-consistent dialogue generation. In this paper, we (Tu_Character_lab) report our participation in the Commonsense Persona-Grounded Dialogue Challenge (CPDC) 2025 Round 2, which eval- uates agents across three tracks: task-oriented dialogue, context-aware dialogue, and their integration. Our approach combines two complementary strategies: (i) lightweight prompting techniques in the API track, including a Deflanderization prompting method to suppress excessive role-play and improve task fidelity, and (ii) fine-tuned large models in the GPU track, leveraging Qwen3-14B with supervisedfinetuning (SFT) and Low-Rank Adaptation(LoRA). Our best submissions ranked 2nd on Task 1, 2nd on Task 3 (API track), and 4th on Task 3 (GPU track).
Aligning pretrained audio encoders and Large Language Models (LLMs) offers a promising, parameter-efficient path to building powerful multimodal agents. However, existing methods often require costly full-model finetuning or rely on static adapters that may lack expressive power. Drawing inspiration from the Platonic Representation Hypothesis, we introduce SteerMoE, a novel and modular framework for audio-language alignment. SteerMoE freezes both the audio encoder and the LLM decoder, training only a lightweight steering module integrated within the encoder's layers. This module uses a Mixture-of-Experts (MoE) router to dynamically select and apply learned steering vectors, progressively transforming continuous audio representations into a space comprehensible to the LLM. By operating entirely in the continuous embedding space, our approach requires no modifications to the LLM's vocabulary and preserves its advanced reasoning and agentic capabilities. We demonstrate through experiments on ASR, audio understanding, and a qualitative function-calling task that SteerMoE achieves strong performance while remaining highly modular and computationally efficient, offering a robust new paradigm for developing sophisticated audio-language systems.
Large Language Model (LLM) based agents integrated into web browsers (often called agentic AI browsers) offer powerful automation of web tasks. However, they are vulnerable to indirect prompt injection attacks, where malicious instructions hidden in a webpage deceive the agent into unwanted actions. These attacks can bypass traditional web security boundaries, as the AI agent operates with the user privileges across sites. In this paper, we present a novel fuzzing framework that runs entirely in the browser and is guided by an LLM to automatically discover such prompt injection vulnerabilities in real time.
Language is fundamental to human cooperation, facilitating not only the exchange of information but also the coordination of actions through shared interpretations of situational contexts. This study explores whether the Generative Agent-Based Model (GABM) Concordia can effectively model Theory of Mind (ToM) within simulated real-world environments. Specifically, we assess whether this framework successfully simulates ToM abilities and whether GPT-4 can perform tasks by making genuine inferences from social context, rather than relying on linguistic memorization. Our findings reveal a critical limitation: GPT-4 frequently fails to select actions based on belief attribution, suggesting that apparent ToM-like abilities observed in previous studies may stem from shallow statistical associations rather than true reasoning. Additionally, the model struggles to generate coherent causal effects from agent actions, exposing difficulties in processing complex social interactions. These results challenge current statements about emergent ToM-like capabilities in LLMs and highlight the need for more rigorous, action-based evaluation frameworks.
Persuasion, a fundamental social capability for humans, remains a challenge for AI systems such as large language models (LLMs). Current studies often overlook the strategic use of information asymmetry in message design or rely on strong assumptions regarding pre-commitment. In this work, we explore the application of Bayesian Persuasion (BP) in natural language within single-turn dialogue settings, to enhance the strategic persuasion capabilities of LLMs. Our framework incorporates a commitment-communication mechanism, where the persuader explicitly outlines an information schema by narrating their potential types (e.g., honest or dishonest), thereby guiding the persuadee in performing the intended Bayesian belief update. We evaluate two variants of our approach: Semi-Formal-Natural-Language (SFNL) BP and Fully-Natural-Language (FNL) BP, benchmarking them against both naive and strong non-BP (NBP) baselines within a comprehensive evaluation framework. This framework covers a diverse set of persuadees -- including LLM instances with varying prompts and fine-tuning and human participants -- across tasks ranging from specially designed persuasion scenarios to general everyday situations. Experimental results on LLM-based agents reveal three main findings: (1) LLMs guided by BP strategies consistently achieve higher persuasion success rates than NBP baselines; (2) SFNL exhibits greater credibility and logical coherence, while FNL shows stronger emotional resonance and robustness in naturalistic conversations; (3) with supervised fine-tuning, smaller models can attain BP performance comparable to that of larger models.
Recent attempts to integrate large language models (LLMs) into recommender systems have gained momentum, but most remain limited to simple text generation or static prompt-based inference, failing to capture the complexity of user preferences and real-world interactions. This study proposes the Multi-Aspect Driven LLM Agent MADRec, an autonomous LLM-based recommender that constructs user and item profiles by unsupervised extraction of multi-aspect information from reviews and performs direct recommendation, sequential recommendation, and explanation generation. MADRec generates structured profiles via aspect-category-based summarization and applies Re-Ranking to construct high-density inputs. When the ground-truth item is missing from the output, the Self-Feedback mechanism dynamically adjusts the inference criteria. Experiments across multiple domains show that MADRec outperforms traditional and LLM-based baselines in both precision and explainability, with human evaluation further confirming the persuasiveness of the generated explanations.
Document AI (DAI) has emerged as a vital application area, and is significantly transformed by the advent of large language models (LLMs). While earlier approaches relied on encoder-decoder architectures, decoder-only LLMs have revolutionized DAI, bringing remarkable advancements in understanding and generation. This survey provides a comprehensive overview of DAI's evolution, highlighting current research attempts and future prospects of LLMs in this field. We explore key advancements and challenges in multimodal, multilingual, and retrieval-augmented DAI, while also suggesting future research directions, including agent-based approaches and document-specific foundation models. This paper aims to provide a structured analysis of the state-of-the-art in DAI and its implications for both academic and practical applications.
Large Language Models (LLMs) often exhibit factual inconsistencies and logical decay in extended, multi-turn dialogues, a challenge stemming from their reliance on static, pre-trained knowledge and an inability to reason adaptively over the dialogue history. Prevailing mitigation strategies, such as Retrieval-Augmented Generation (RAG) and agentic working memories, improve information recall but still engage with fundamentally static knowledge sources and follow pre-defined single reasoning path. This hinders their ability to preserve factual and logical consistency of their responses in multi-turn dialogues while the context evolves over time. To address this issue, we propose D-SMART, a model-agnostic framework designed to maintain multi-turn dialogue consistency by enabling LLMs to build and reason over a dynamic, structured representation of the conversational context. This is achieved via two synergistic components: (1) a Dynamic Structured Memory (DSM), which incrementally constructs and maintains an authoritative, OWL-compliant knowledge graph of the conversation; and (2) a Reasoning Tree (RT), which executes inferences as an explicit and traceable multi-step search over the graph. As the popular-used quality score (judged by GPT-4) can overlook logical flaws, we introduce new NLI-based metrics to better measure multi-turn dialogue consistency. Comprehensive experiments on the MT-Bench-101 benchmark show that D-SMART significantly outperforms state-of-the-art baselines, elevating the dialogue consistency score by over 48\% for both proprietary and open-source models, and notably improves the quality score of the latter by up to 10.1\%.
Enhancing customer experience is essential for business success, particularly as service demands grow in scale and complexity. Generative artificial intelligence and Large Language Models (LLMs) have empowered intelligent interaction systems to deliver efficient, personalized, and 24/7 support. In practice, intelligent interaction systems encounter several challenges: (1) Constructing high-quality data for cold-start training is difficult, hindering self-evolution and raising labor costs. (2) Multi-turn dialogue performance remains suboptimal due to inadequate intent understanding, rule compliance, and solution extraction. (3) Frequent evolution of business rules affects system operability and transferability, constraining low-cost expansion and adaptability. (4) Reliance on a single LLM is insufficient in complex scenarios, where the absence of multi-agent frameworks and effective collaboration undermines process completeness and service quality. (5) The open-domain nature of multi-turn dialogues, lacking unified golden answers, hampers quantitative evaluation and continuous optimization. To address these challenges, we introduce WOWService, an intelligent interaction system tailored for industrial applications. With the integration of LLMs and multi-agent architectures, WOWService enables autonomous task management and collaborative problem-solving. Specifically, WOWService focuses on core modules including data construction, general capability enhancement, business scenario adaptation, multi-agent coordination, and automated evaluation. Currently, WOWService is deployed on the Meituan App, achieving significant gains in key metrics, e.g., User Satisfaction Metric 1 (USM 1) -27.53% and User Satisfaction Metric 2 (USM 2) +25.51%, demonstrating its effectiveness in capturing user needs and advancing personalized service.
Inspired by the success of reinforcement learning (RL) in Large Language Model (LLM) training for domains like math and code, recent works have begun exploring how to train LLMs to use search engines more effectively as tools for retrieval-augmented generation. Although these methods achieve performance improvement across QA benchmarks, many prioritize final answer correctness while overlooking the quality of intermediate reasoning steps, which may lead to chain-of-thought unfaithfulness. In this paper, we first introduce a comprehensive evaluation framework for evaluating RL-based search agents, covering three distinct faithfulness metrics: information-think faithfulness, think-answer faithfulness, and think-search faithfulness. Our evaluations reveal that a prototypical RL-based search agent, Search-R1, has significant room for improvement in this regard. To foster faithful reasoning, we introduce VERITAS (Verifying Entailed Reasoning through Intermediate Traceability in Agentic Search), a novel framework that integrates fine-grained faithfulness rewards into the reinforcement learning process. Our experiments show that models trained with VERITAS not only significantly improve reasoning faithfulness, but also achieve comparable task performance across seven QA benchmarks.
Rule-based network intrusion detection systems play a crucial role in the real-time detection of Web attacks. However, most existing works primarily focus on automatically generating detection rules for new attacks, often overlooking the relationships between new attacks and existing rules, which leads to significant redundancy within the ever-expanding ruleset. To address this issue, we propose GRIDAI, a novel end-to-end framework for the automated Generation and Repair of Intrusion Detection rules through collaboration among multiple LLM-based agents. Unlike traditional methods, GRIDAI first assesses the nature of incoming attack samples. If the sample represents a new attack type, it is used to generate a new rule. Otherwise, the sample is identified as a variant of an attack already covered by an existing rule and used to repair the rule by updating the corresponding signature, thereby enhancing its generalization capability. Additionally, to mitigate syntactic and semantic errors in rules caused by LLM hallucinations, we incorporate a tool-based real-time validation mechanism and a representative attack sample maintained for each rule, enabling fully automated rule generation and repair. Comprehensive experiments were conducted on a public dataset containing seven types of attacks and a private dataset with 43 attack types. The results demonstrate that GRIDAI accurately identifies the relationships between new attack samples and existing rules, efficiently generates and repairs rules to handle new attacks and variants, and effectively mitigates the impact of LLM hallucinations.
Network protocol testing is fundamental for modern network infrastructure. However, traditional network protocol testing methods are labor-intensive and error-prone, requiring manual interpretation of specifications, test case design, and translation into executable artifacts, typically demanding one person-day of effort per test case. Existing model-based approaches provide partial automation but still involve substantial manual modeling and expert intervention, leading to high costs and limited adaptability to diverse and evolving protocols. In this paper, we propose a first-of-its-kind system called NeTestLLM that takes advantage of multi-agent Large Language Models (LLMs) for end-to-end automated network protocol testing. NeTestLLM employs hierarchical protocol understanding to capture complex specifications, iterative test case generation to improve coverage, a task-specific workflow for executable artifact generation, and runtime feedback analysis for debugging and refinement. NeTestLLM has been deployed in a production environment for several months, receiving positive feedback from domain experts. In experiments, NeTestLLM generated 4,632 test cases for OSPF, RIP, and BGP, covering 41 historical FRRouting bugs compared to 11 by current national standards. The process of generating executable artifacts also improves testing efficiency by a factor of 8.65x compared to manual methods. NeTestLLM provides the first practical LLM-powered solution for automated end-to-end testing of heterogeneous network protocols.
Recent advances in Large Language Models (LLMs) demonstrate that chain-of-thought prompting and deep reasoning substantially enhance performance on complex tasks, and multi-agent systems can further improve accuracy by enabling model debates. However, applying deep reasoning to all problems is computationally expensive. To mitigate these costs, we propose a complementary agent system integrating small and large LLMs. The small LLM first generates an initial answer, which is then verified by the large LLM. If correct, the answer is adopted directly; otherwise, the large LLM performs in-depth reasoning. Experimental results show that, for simple problems, our approach reduces the computational cost of the large LLM by more than 50% with negligible accuracy loss, while consistently maintaining robust performance on complex tasks.
The advent of large language models (LLMs) has enabled agents to represent virtual humans in societal simulations, facilitating diverse interactions within complex social systems. However, existing LLM-based agents exhibit severe limitations in affective cognition: They fail to simulate the bounded rationality essential for bridging virtual and real-world services; They lack empirically validated integration mechanisms embedding emotions within agent decision architectures. This paper constructs an emotional cognition framework incorporating desire generation and objective management, designed to achieve emotion alignment between LLM-based agents and humans, modeling the complete decision-making process of LLM-based agents, encompassing state evolution, desire generation, objective optimization, decision generation, and action execution. This study implements the proposed framework within our proprietary multi-agent interaction environment. Experimental results demonstrate that agents governed by our framework not only exhibit behaviors congruent with their emotional states but also, in comparative assessments against other agent types, demonstrate superior ecological validity and generate decision outcomes that significantly more closely approximate human behavioral patterns.
A key challenge in transportation planning is that the collective preferences of heterogeneous travelers often diverge from the policies produced by model-driven decision tools. This misalignment frequently results in implementation delays or failures. Here, we investigate whether large language models (LLMs), noted for their capabilities in reasoning and simulating human decision-making, can help inform and address this alignment problem. We develop a multi-agent simulation in which LLMs, acting as agents representing residents from different communities in a city, participate in a referendum on a set of transit policy proposals. Using chain-of-thought reasoning, LLM agents provide ranked-choice or approval-based preferences, which are aggregated using instant-runoff voting (IRV) to model democratic consensus. We implement this simulation framework with both GPT-4o and Claude-3.5, and apply it for Chicago and Houston. Our findings suggest that LLM agents are capable of approximating plausible collective preferences and responding to local context, while also displaying model-specific behavioral biases and modest divergences from optimization-based benchmarks. These capabilities underscore both the promise and limitations of LLMs as tools for solving the alignment problem in transportation decision-making.
Social and behavioral scientists increasingly aim to study how humans interact, collaborate, and make decisions alongside artificial intelligence. However, the experimental infrastructure for such work remains underdeveloped: (1) few platforms support real-time, multi-party studies at scale; (2) most deployments require bespoke engineering, limiting replicability and accessibility, and (3) existing tools do not treat AI agents as first-class participants. We present Deliberate Lab, an open-source platform for large-scale, real-time behavioral experiments that supports both human participants and large language model (LLM)-based agents. We report on a 12-month public deployment of the platform (N=88 experimenters, N=9195 experiment participants), analyzing usage patterns and workflows. Case studies and usage scenarios are aggregated from platform users, complemented by in-depth interviews with select experimenters. By lowering technical barriers and standardizing support for hybrid human-AI experimentation, Deliberate Lab expands the methodological repertoire for studying collective decision-making and human-centered AI.
We present Sentinel, the first framework for formally evaluating the physical safety of Large Language Model(LLM-based) embodied agents across the semantic, plan, and trajectory levels. Unlike prior methods that rely on heuristic rules or subjective LLM judgments, Sentinel grounds practical safety requirements in formal temporal logic (TL) semantics that can precisely specify state invariants, temporal dependencies, and timing constraints. It then employs a multi-level verification pipeline where (i) at the semantic level, intuitive natural language safety requirements are formalized into TL formulas and the LLM agent's understanding of these requirements is probed for alignment with the TL formulas; (ii) at the plan level, high-level action plans and subgoals generated by the LLM agent are verified against the TL formulas to detect unsafe plans before execution; and (iii) at the trajectory level, multiple execution trajectories are merged into a computation tree and efficiently verified against physically-detailed TL specifications for a final safety check. We apply Sentinel in VirtualHome and ALFRED, and formally evaluate multiple LLM-based embodied agents against diverse safety requirements. Our experiments show that by grounding physical safety in temporal logic and applying verification methods across multiple levels, Sentinel provides a rigorous foundation for systematically evaluating LLM-based embodied agents in physical environments, exposing safety violations overlooked by previous methods and offering insights into their failure modes.
Large language models (LLMs) augmented with multi-step reasoning and action generation abilities have shown promise in leveraging external tools to tackle complex tasks that require long-horizon planning. However, existing approaches either rely on implicit planning in the reasoning stage or introduce explicit planners without systematically addressing how to optimize the planning stage. As evidence, we observe that under vanilla reinforcement learning (RL), planning tokens exhibit significantly higher entropy than other action tokens, revealing uncertain decision points that remain under-optimized. To address this, we propose DeepPlanner, an end-to-end RL framework that effectively enhances the planning capabilities of deep research agents. Our approach shapes token-level advantage with an entropy-based term to allocate larger updates to high entropy tokens, and selectively upweights sample-level advantages for planning-intensive rollouts. Extensive experiments across seven deep research benchmarks demonstrate that DeepPlanner improves planning quality and achieves state-of-the-art results under a substantially lower training budget.
Recently, several multi-turn dialogue benchmarks have been proposed to evaluate the conversational abilities of large language models (LLMs). As LLMs are increasingly recognized as a key technology for advancing intelligent education, owing to their ability to deeply understand instructional contexts and provide personalized guidance, the construction of dedicated teacher-student dialogue benchmarks has become particularly important. To this end, we present EduDial, a comprehensive multi-turn teacher-student dialogue dataset. EduDial covers 345 core knowledge points and consists of 34,250 dialogue sessions generated through interactions between teacher and student agents. Its design is guided by Bloom's taxonomy of educational objectives and incorporates ten questioning strategies, including situational questioning, zone of proximal development (ZPD) questioning, and metacognitive questioning-thus better capturing authentic classroom interactions. Furthermore, we design differentiated teaching strategies for students at different cognitive levels, thereby providing more targeted teaching guidance. Building on EduDial, we further develop EduDial-LLM 32B via training and propose an 11-dimensional evaluation framework that systematically measures the teaching abilities of LLMs, encompassing both overall teaching quality and content quality. Experiments on 17 mainstream LLMs reveal that most models struggle in student-centered teaching scenarios, whereas our EduDial-LLM achieves significant gains, consistently outperforming all baselines across all metrics. The code is available at https://github.com/Mind-Lab-ECNU/EduDial/tree/main.
Multi-agent large language model (LLM) systems are increasingly adopted for complex language processing tasks that require communication and coordination among agents. However, these systems often suffer substantial overhead from repeated reprocessing of overlapping contexts across agents. In typical pipelines, once an agent receives a message from its predecessor, the full context-including prior turns-must be reprocessed from scratch, leading to inefficient processing. While key-value (KV) caching is an effective solution for avoiding redundant computation in single-agent settings where prefixes remain unchanged, it cannot be directly reused in multi-agent scenarios due to diverging prefixes introduced by agent-specific context extensions. We identify that the core challenge lies in the offset variance of KV-caches across agents. To address this, we propose KVCOMM, a training-free framework that enables efficient prefilling in multi-agent inference by reusing KV-caches and aligning cache offsets of overlapping contexts under diverse prefix contexts. KVCOMM estimates and adjusts KV-caches for shared content by referencing a pool of cached examples-termed anchors-that store observed cache deviations under varying prefixes. The anchor pool is maintained and updated online, allowing dynamic adaptation to distinct user requests and context structures. KVCOMM achieves over 70% reuse rate across diverse multi-agent workloads, including retrieval-augmented generation, math reasoning, and collaborative coding tasks, all without quality degradation. Particularly, when each fully-connected agent receives 1K input tokens with 512 prefix tokens and 512 output tokens under a five-agent setting, KVCOMM achieves up to 7.8x speedup compared to the standard prefill pipeline, reducing TTFT from ~430 ms to ~55 ms.
Multimodal Large Language Models (MLLMs) in real-world applications require access to external knowledge sources and must remain responsive to the dynamic and ever-changing real-world information in order to address information-seeking and knowledge-intensive user queries. Existing approaches, such as retrieval augmented generation (RAG) methods, search agents, and search equipped MLLMs, often suffer from rigid pipelines, excessive search calls, and poorly constructed search queries, which result in inefficiencies and suboptimal outcomes. To address these limitations, we present DeepMMSearch-R1, the first multimodal LLM capable of performing on-demand, multi-turn web searches and dynamically crafting queries for both image and text search tools. Specifically, DeepMMSearch-R1 can initiate web searches based on relevant crops of the input image making the image search more effective, and can iteratively adapt text search queries based on retrieved information, thereby enabling self-reflection and self-correction. Our approach relies on a two-stage training pipeline: a cold start supervised finetuning phase followed by an online reinforcement learning optimization. For training, we introduce DeepMMSearchVQA, a novel multimodal VQA dataset created through an automated pipeline intermixed with real-world information from web search tools. This dataset contains diverse, multi-hop queries that integrate textual and visual information, teaching the model when to search, what to search for, which search tool to use and how to reason over the retrieved information. We conduct extensive experiments across a range of knowledge-intensive benchmarks to demonstrate the superiority of our approach. Finally, we analyze the results and provide insights that are valuable for advancing multimodal web-search.
We present Ax-Prover, a multi-agent system for automated theorem proving in Lean that can solve problems across diverse scientific domains and operate either autonomously or collaboratively with human experts. To achieve this, Ax-Prover approaches scientific problem solving through formal proof generation, a process that demands both creative reasoning and strict syntactic rigor. Ax-Prover meets this challenge by equipping Large Language Models (LLMs), which provide knowledge and reasoning, with Lean tools via the Model Context Protocol (MCP), which ensure formal correctness. To evaluate its performance as an autonomous prover, we benchmark our approach against frontier LLMs and specialized prover models on two public math benchmarks and on two Lean benchmarks we introduce in the fields of abstract algebra and quantum theory. On public datasets, Ax-Prover is competitive with state-of-the-art provers, while it largely outperform them on the new benchmarks. This shows that, unlike specialized systems that struggle to generalize, our tool-based agentic theorem prover approach offers a generalizable methodology for formal verification across diverse scientific domains. Furthermore, we demonstrate Ax-Prover's assistant capabilities in a practical use case, showing how it enabled an expert mathematician to formalize the proof of a complex cryptography theorem.
Large Language Models (LLMs) are increasingly being deployed as the reasoning engines for agentic AI systems, yet they exhibit a critical flaw: a rigid adherence to explicit rules that leads to decisions misaligned with human common sense and intent. This "rule-rigidity" is a significant barrier to building trustworthy autonomous agents. While prior work has shown that supervised fine-tuning (SFT) with human explanations can mitigate this issue, SFT is computationally expensive and inaccessible to many practitioners. To address this gap, we introduce the Rule-Intent Distinction (RID) Framework, a novel, low-compute meta-prompting technique designed to elicit human-aligned exception handling in LLMs in a zero-shot manner. The RID framework provides the model with a structured cognitive schema for deconstructing tasks, classifying rules, weighing conflicting outcomes, and justifying its final decision. We evaluated the RID framework against baseline and Chain-of-Thought (CoT) prompting on a custom benchmark of 20 scenarios requiring nuanced judgment across diverse domains. Our human-verified results demonstrate that the RID framework significantly improves performance, achieving a 95% Human Alignment Score (HAS), compared to 80% for the baseline and 75% for CoT. Furthermore, it consistently produces higher-quality, intent-driven reasoning. This work presents a practical, accessible, and effective method for steering LLMs from literal instruction-following to liberal, goal-oriented reasoning, paving the way for more reliable and pragmatic AI agents.
With advancements in reasoning capabilities, Large Language Models (LLMs) are increasingly employed for automated judgment tasks. While LLMs-as-Judges offer promise in automating evaluations, current approaches often rely on simplistic aggregation methods (e.g., majority voting), which can fail even when individual agents provide correct answers. To address this, we propose a multi-agent debate judge framework where agents collaboratively reason and iteratively refine their responses. We formalize the debate process mathematically, analyzing agent interactions and proving that debate amplifies correctness compared to static ensembles. To enhance efficiency, we introduce a stability detection mechanism that models judge consensus dynamics via a time-varying Beta-Binomial mixture, with adaptive stopping based on distributional similarity (Kolmogorov-Smirnov test). This mechanism models the judges' collective correct rate dynamics using a time-varying mixture of Beta-Binomial distributions and employs an adaptive stopping criterion based on distributional similarity (Kolmogorov-Smirnov statistic). Experiments across multiple benchmarks and models demonstrate that our framework improves judgment accuracy over majority voting while maintaining computational efficiency.
Large Language Models face challenges in long-horizon agentic tasks as their constrained memory is easily overwhelmed by distracting or irrelevant context. Existing working memory methods typically rely on external, heuristic mechanisms that are decoupled from the agent's core policy. In this work, we reframe working memory management as a learnable, intrinsic capability. We propose a novel framework, Memory-as-Action, where an agent actively manages its working memory by executing explicit editing operations as part of a unified policy. This formulation allows an agent, trained via reinforcement learning, to balance memory curation against long-term task objectives under given resource constraints. However, such memory editing actions break the standard assumption of a continuously growing prefix in LLM interactions, leading to what we call trajectory fractures. These non-prefix changes disrupt the causal continuity required by standard policy gradient methods, making those methods inapplicable. To address this, we propose a new algorithm, Dynamic Context Policy Optimization, which enables stable end-to-end reinforcement learning by segmenting trajectories at memory action points and applying trajectory-level advantages to the resulting action segments. Our results demonstrate that jointly optimizing for task reasoning and memory management in an end-to-end fashion not only reduces overall computational consumption but also improves task performance, driven by adaptive context curation strategies tailored to the model's intrinsic capabilities.
Reliable handling of code diffs is central to agents that edit and refactor repositories at scale. We introduce Diff-XYZ, a compact benchmark for code-diff understanding with three supervised tasks: apply (old code $+$ diff $\rightarrow$ new code), anti-apply (new code $-$ diff $\rightarrow$ old code), and diff generation (new code $-$ old code $\rightarrow$ diff). Instances in the benchmark are triples $\langle \textit{old code}, \textit{new code}, \textit{diff} \rangle$ drawn from real commits in CommitPackFT, paired with automatic metrics and a clear evaluation protocol. We use the benchmark to do a focused empirical study of the unified diff format and run a cross-format comparison of different diff representations. Our findings reveal that different formats should be used depending on the use case and model size. For example, representing diffs in search-replace format is good for larger models in the diff generation scenario, yet not suited well for diff analysis and smaller models. The Diff-XYZ benchmark is a reusable foundation for assessing and improving diff handling in LLMs that can aid future development of diff formats and models editing code. The dataset is published on HuggingFace Hub: https://huggingface.co/datasets/JetBrains-Research/diff-xyz.
The polarization of opinions, information segregation, and cognitive biases on social media have attracted significant academic attention. In real-world networks, information often spans multiple interrelated topics, posing challenges for opinion evolution and highlighting the need for frameworks that simulate interactions among topics. Existing studies based on large language models (LLMs) focus largely on single topics, limiting the capture of cognitive transfer in multi-topic, cross-domain contexts. Traditional numerical models, meanwhile, simplify complex linguistic attitudes into discrete values, lacking interpretability, behavioral consistency, and the ability to integrate multiple topics. To address these issues, we propose Multi-topic Opinion Simulation (MTOS), a social simulation framework integrating multi-topic contexts with LLMs. MTOS leverages LLMs alongside short-term and long-term memory, incorporates multiple user-selection interaction mechanisms and dynamic topic-selection strategies, and employs a belief decay mechanism to enable perspective updates across topics. We conduct extensive experiments on MTOS, varying topic numbers, correlation types, and performing ablation studies to assess features such as group polarization and local consistency. Results show that multi-topic settings significantly alter polarization trends: positively correlated topics amplify echo chambers, negatively correlated topics inhibit them, and irrelevant topics also mitigate echo chamber effects through resource competition. Compared with numerical models, LLM-based agents realistically simulate dynamic opinion changes, reproduce linguistic features of news texts, and capture complex human reasoning, improving simulation interpretability and system stability.