ICLR 2026 论文列表

Harnessing the power of LLMs requires a delicate dance between being helpful and harmless, leading to two critical challenges: vulnerability to adversarial attacks that elicit unsafe content, and a tendency for overrefusal on benign but sensitive prompts. Current approaches often navigate this dance with safeguard models that completely reject any content that contains unsafe portions. This approach cuts the music entirely—it may exacerbate overrefusals and fails to provide nuanced guidance for queries it refuses. To teach models a more coordinated choreography, we propose WaltzRL, a novel multi-agent reinforcement learning framework that formulates safety alignment as a collaborative, positive-sum game. WaltzRL jointly trains a conversation agent and a feedback agent, where the latter is incentivized to provide useful suggestions that improve the safety and helpfulness of the conversation agent's responses. At the core of WaltzRL is a Dynamic Improvement Reward (DIR) that evolves over time based on how well the conversation agent incorporates the feedback. At inference time, unsafe or overrefusing responses from the conversation agent are improved rather than discarded. The feedback agent is deployed together with the conversation agent and only engages adaptively when needed, preserving helpfulness and low latency on safe queries. Our experiments, conducted across five diverse datasets, demonstrate that WaltzRL significantly reduces both unsafe responses (e.g., from 39.0% to 4.6% on WildJailbreak) and overrefusals (from 45.3% to 9.9% on OR-Bench) compared to various baselines. By enabling the conversation and feedback agents to co-evolve and adaptively apply feedback, WaltzRL enhances LLM safety without degrading general capabilities, thereby advancing the Pareto front between helpfulness and harmlessness.

This work presents MAC-Flow, a simple yet expressive framework for multi-agent coordination. We argue that requirements of effective coordination are twofold: *(i)* a rich representation of the diverse joint behaviors present in offline data and *(ii)* the ability to act efficiently in real time. However, prior approaches often sacrifice one for the other, *i.e.*, denoising diffusion-based solutions capture complex coordination but are computationally slow, while Gaussian policy-based solutions are fast but brittle in handling multi-agent interaction. MAC-Flow addresses this trade-off by first learning a flow-based representation of joint behaviors, and then distilling it into decentralized one-step policies that preserve coordination while enabling fast execution. Across four different benchmarks, including $12$ environments and $34$ datasets, MAC-Flow alleviates the trade-off between performance and computational cost, specifically achieving about $\boldsymbol{\times14.5}$ faster inference compared to diffusion-based MARL methods, while maintaining good performance. At the same time, its inference speed is similar to that of prior Gaussian policy-based offline MARL methods.

Recent advances in large language models (LLMs) and agent system designs have empowered agents with unprecedented levels of capability. However, existing agent benchmarks are showing a trend of rapid ceiling-hitting by newly developed agents, making it difficult to meet the demands for evaluating agent abilities. To address this problem, we propose the Trajectory-based Reproducible Agent- benchmark Complexity Evolution (TRACE) framework. This framework takes an original task from an existing benchmark and encourages agents to freely explore and evolve it into a new task with higher difficulty while recording traceable agent trajectories. The framework proceeds in three stages: (1) evolutionary proposal mining, which provides task evolution proposals through preliminary exploration and divergent thinking; (2) problem formation and free exploration, where proposals are conceptualized into feasible problem candidates and the agents then explore them freely while recording their execution trajectories; and (3) multi-level validation, which ensures that the evolved tasks are accompanied by validatable and reproducible trajectories. Experiments on the GAIA benchmark demonstrate that the TRACE framework consistently enhances task complexity while improving the reliability of correctness through validatable execution trajectories. This work marks a paradigm shift from static, manually curated benchmarks to dynamic, self-evolving evaluation systems, providing a sustainable and challenging runway for agent development

Large language model (LLM) agents exhibit strong mathematical problem-solving abilities and can even solve International Mathematical Olympiad (IMO) level problems with the assistance of formal proof systems. However, due to weak heuristics for auxiliary constructions, AI for geometry problem solving remains dominated by expert models such as AlphaGeometry 2, which rely heavily on large-scale data synthesis and search for both training and evaluation. In this work, we make the first attempt to build a medalist-level LLM agent for geometry and present InternGeometry. InternGeometry overcomes the heuristic limitations in geometry by iteratively proposing propositions and auxiliary constructions, verifying them with a symbolic engine, and reflecting on the engine’s feedback to guide subsequent proposals. A dynamic memory mechanism enables InternGeometry to conduct more than two hundred interactions with the symbolic engine per problem. To further accelerate learning, we introduce Complexity-Boosting Reinforcement Learning (CBRL), which gradually increases the complexity of synthesized problems across training stages. Built on InternThinker-32B, InternGeometry solves 44 of 50 IMO geometry problems (2000–2024), exceeding the average gold medalist score (40.9), using only 13K training examples, just $0.004\%$ of the data used by AlphaGeometry 2, demonstrating the potential of LLM agents on expert-level geometry tasks. InternGeometry can also propose novel auxiliary constructions for IMO problems that do not appear in human solutions. We will release the model, data, and symbolic engine to support future research.

Large Language Models (LLMs) have achieved remarkable success in software engineering tasks when trained with executable runtime environments, particularly in resolving GitHub issues. However, such runtime environments are often unavailable in other domains, especially cybersecurity, where challenge configurations and execution contexts are ephemeral or restricted. We present Cyber-Zero, the first runtime-free framework for synthesizing high-quality agent trajectories to train cybersecurity LLMs. Cyber-Zero leverages publicly available CTF writeups and employs persona-driven LLM simulation to reverse-engineer runtime behaviors and generate realistic, long-horizon interaction sequences without actual environments. Using trajectories synthesized by Cyber-Zero, we train LLM-based agents that achieve up to 13.1% absolute performance gains over baseline models on three prominent CTF benchmarks: InterCode-CTF, NYU CTF Bench, and Cybench. Our best model, Cyber-Zero-32B, establishes new state-of-the-art performance among open-weight models, matching the capabilities of proprietary systems like DeepSeek-V3-0324 and Claude-3.5-Sonnet while offering superior cost-effectiveness, and demonstrating that runtime-free trajectory synthesis can effectively democratize the development of state-of-the-art cybersecurity agents.

Large language models face challenges in long-context question answering, where key evidence of a query may be dispersed across millions of tokens. Existing works equip large language models with a memory corpus that is dynamically updated during a single-pass document scan, also known as the "memorize while reading" methods. While this approach scales efficiently, it suffers from irreversible forward-only processing, information loss through overwriting, and sparse reinforcement learning signals. To tackle these challenges, we present ReMemR1, a memory-augmented agent with callback-enhanced memory that allows selective retrieval from the entire memory history and allows non-linear reasoning and revisiting of early evidence. To further strengthen training, we propose Reinforcement Learning with Multi-Level Rewards (RLMLR), which combines final-answer rewards with dense, step-level signals that guide effective memory use. Together, these contributions mitigate information degradation, improve supervision, and support multi-hop memory utilizing. Experiments on long-document QA show significant gains over existing memory-based approaches, which validates ReMemR1 as an effective solution for long-context reasoning agents.

Large Language Models (LLMs) achieve remarkable success, but their massive parameter counts present significant deployment challenges. Post-training tensor decomposition offers a promising, data-free compression strategy by exploiting structural redundancies within the model weights. However, existing tensor methods face a critical limitation: the dense core tensor bottleneck. While these methods find a shared low-rank basis, the resulting dense core tensor grows polynomially with the chosen ranks, becoming a new storage bottleneck and capping the maximum achievable compression. To overcome this fundamental barrier, we introduce LeSTD (\textbf{Le}arning-based \textbf{S}parse \textbf{T}ensor \textbf{D}ecomposition), a novel two-stage framework for the high-ratio compression of Multi-Head Attention (MHA) blocks. LeSTD first employs an iterative algorithm to identify a high-quality, and shared orthogonal basis that jointly represents all attention heads. Subsequently, it introduces a principled, importance-based pruning algorithm that learns an ultra-sparse core tensor by systematically removing the least salient elements and refitting the remaining ones to preserve model fidelity. By decoupling basis optimization from core sparsification, LeSTD breaks the compression ceiling imposed by the dense core, enabling significantly higher compression ratios than prior methods.

The rapid advancement of AI-generated content (AIGC) has enabled the synthesis of visually convincing images; however, many such outputs exhibit subtle \textbf{semantic anomalies}, including unrealistic object configurations, violations of physical laws, or commonsense inconsistencies, which compromise the overall plausibility of the generated scenes. Detecting these semantic-level anomalies is essential for assessing the trustworthiness of AIGC media, especially in AIGC image analysis, explainable deepfake detection and semantic authenticity assessment.In this paper, we formalize \textbf{semantic anomaly detection and reasoning} for AIGC images and introduce \textbf{AnomReason}, a large-scale benchmark with structured annotations as quadruples \emph{(Name, Phenomenon, Reasoning, Severity)}. Annotations are produced by a modular multi-agent pipeline (\textbf{AnomAgent}) with lightweight human-in-the-loop verification, enabling scale while preserving quality. At construction time, AnomAgent processed approximately 4.17\,B GPT-4o tokens, providing scale evidence for the resulting structured annotations. We further show that models fine-tuned on AnomReason achieve consistent gains over strong vision-language baselines under our proposed semantic matching metric (\textit{SemAP} and \textit{SemF1}). Applications to {explainable deepfake detection} and {semantic reasonableness assessment of image generators} demonstrate practical utility. In summary, AnomReason and AnomAgent serve as a foundation for measuring and improving the semantic plausibility of AI-generated images. We will release code, metrics, data, and task-aligned models to support reproducible research on semantic authenticity and interpretable AIGC forensics.

As AI becomes more personal, e.g., Agentic AI, there is an increasing need for personalizing models for various use cases. Personalized federated learning (PFL) enables each client to collaboratively leverage other clients' knowledge for better adaptation to the task of interest, without privacy risks. Despite its potential, existing PFL methods remain confined to rather simplified scenarios where data and models are the same across clients. To move towards realistic scenarios, we propose FedMosaic, a method that jointly addresses data and model heterogeneity with a task-relevance-aware model aggregation strategy to reduce parameter interference, and a dimension-invariant module that enables knowledge sharing across heterogeneous architectures without huge computational cost. To mimic the real-world task diversity, we propose a multi-modal PFL benchmark spanning 40 distinct tasks with distribution shifts over time. The empirical study shows that FedMosaic outperforms the state-of-the-art PFL methods, excelling in both personalization and generalization capabilities under challenging, realistic scenarios.

Large-language-models (LLMs) demonstrate enormous utility in long-context tasks which require processing prompts that consist of tens to hundreds of thousands of tokens. However, existing LLM training libraries do not provide easy to use abstractions to optimize for long-context training, instead focusing on optimizations for models with large parameter counts through ZeRO-3/FSDP, Tensor and Pipeline parallelism. This forces users to rewrite LLM training libraries to incorporate compositions of various complex long-context optimizations, such as sequence-parallelism, to training pipelines; a process that requires in-depth expertise, reducing developer productivity. To tackle these challenges, we introduce AutSP: the first automated solution to automatically optimize LLM training for longer-contexts. AutoSP compiles models and applies a targeted set of optimizations: automated sequence parallelism, and long-context aware activation-checkpointing, to drastically enhance LLM trainability at negligible cost to throughput. Our evaluation demonstrates AutoSP's capability on both NVIDIA and AMD hardware, increasing training contexts by upto 2.7$\times$ and 2.5$\times$ respectively at negligible cost to runtime performance over competitive hand-written baselines.

Chain-of-thought prompting has popularized step-by-step reasoning in large language models, yet model performance still degrades as problem complexity and context length grow. By decomposing difficult tasks with long contexts into shorter, manageable ones, recent multi-agent paradigms offer a promising near-term solution to this problem. However, the fundamental capacities of such systems are poorly understood. In this work, we propose a theoretical framework to analyze the expressivity of multi-agent systems. We apply our framework to three algorithmic families: state tracking, recall, and $k$-hop reasoning. We derive bounds on (i) the number of agents required to solve the task exactly, (ii) the quantity and structure of inter-agent communication, and (iii) the achievable speedups as problem size and context scale. Our results identify regimes where communication is provably beneficial, delineate tradeoffs between agent count and bandwidth, and expose intrinsic limitations when either resource is constrained. We complement our theoretical analysis with a set of experiments on pretrained LLMs using controlled synthetic benchmarks. Empirical outcomes confirm the tradeoffs between key quantities predicted by our theory. Collectively, our analysis offers principled guidance for designing scalable multi-agent reasoning systems.

Travel planning stands out among real-world applications of \emph{Language Agents} because it couples significant practical demand with a rigorous constraint-satisfaction challenge. However, existing benchmarks primarily operate on a slot-filling paradigm, restricting agents to synthetic queries with pre-defined constraint menus, which fails to capture the open-ended nature of natural language interaction, where user requirements are compositional, diverse, and often implicitly expressed. To address this gap, we introduce \emph{ChinaTravel}, with four key contributions: 1) a practical sandbox aligned with the multi-day, multi-POI travel planning, 2) a compositionally generalizable domain-specific language (DSL) for scalable evaluation, covering feasibility, constraint satisfaction, and preference comparison 3) an open-ended dataset that integrates diverse travel requirements and implicit intent from 1154 human participants, and 4) fine-grained analysis reveal the potential of neuro-symbolic agents in travel planning, achieving a 37.0\% constraint satisfaction rate on human queries, a 10$\times$ improvement over purely neural models, \blue{yet highlighting significant challenges in compositional generalization}. Overall, ChinaTravel provides a foundation for advancing language agents through compositional constraint validation in complex, real-world planning scenarios.

With the growing popularity of LLM agents and RAG, it has become increasingly important to retrieve documents that are essential for solving a task, even when their connection to the task is indirect or implicit. Addressing this problem requires fine-grained reasoning to accurately assess the relevance between the task and each candidate document. This capability, however, poses a significant challenge for existing IR techniques. Despite recent progress in reasoning-enhanced IR, existing approaches still face significant challenges in applicability, scalability, and efficiency. In this work, we propose **Retro\***, a novel approach for reasoning-intensive document retrieval. Our method introduces a rubric-based relevance **scoring mechanism**, enabling the model to reason about the relationship between a task and a document based on explicitly defined criteria, whereby producing a fine-grained, interpretable relevance score. Retro\* also supports **test-time scaling** by combining multiple reasoning trajectories via score integration, which produces more reliable relevance estimates. To optimize Retro\*'s reasoning capabilities, we introduce a novel **reinforcement learning** algorithm tailored for its relevance scoring mechanism, which employs two composite rewards to fully exploit the trajectories of each training sample. Our experiments show that Retro\* outperforms existing document retrieval methods with notable advantages, leading to **state-of-the-art** performance on the BRIGHT benchmark.

Scientific discovery is often slowed by the manual development of computational tools needed to analyze complex experimental data. Building such tools is costly and time-consuming because scientists must iteratively review literature, test mod- eling and scientific assumptions against empirical data, and implement these in- sights into efficient software. Large language models (LLMs) have demonstrated strong capabilities in synthesizing literature, reasoning with empirical data, and generating domain-specific code, offering new opportunities to accelerate com- putational method development. Existing LLM-based systems either focus on performing scientific analyses using existing computational methods or on de- veloping computational methods or models for general machine learning without effectively integrating the often unstructured knowledge specific to scientific do- mains. Here, we introduce TusoAI, an agentic AI system that takes a scientific task description with an evaluation function and autonomously develops and optimizes computational methods for the application. TusoAI integrates domain knowledge into a knowledge tree representation and performs iterative, domain-specific op- timization and model diagnosis, improving performance over a pool of candidate solutions. We conducted comprehensive benchmark evaluations demonstrating that TusoAI outperforms state-of-the-art expert methods, MLE agents, and scien- tific AI agents across diverse tasks. Applying TusoAI to two key open problems in genetics improved existing computational methods and uncovered new biology missed by previous methods.

Autoformalization—the process of converting natural language mathematical statements into machine-verifiable formal code—plays a critical role in ensuring the reliability of mathematical reasoning generated by large language models (LLMs). Recent studies show that LLMs exhibit strong potential in automating this process, producing formal code for systems such as Lean 4, Coq, and Isabelle. Despite prominent advances, existing LLM-based autoformalization methods remain limited: they lack the ability to provide reliable semantic consistency checks to ensure that the formal code accurately preserves the meaning of the original statement. Furthermore, such methods are unable to support iterative improvement through corrective feedback. To address these limitations, we propose Loc-Decomp, a novel framework that integrates an automatic semantic consistency checker and the Lean 4 compiler to iteratively refine LLM-generated formalizations, ensuring both semantic consistency and syntactic correctness. Our approach introduces three key innovations: __(1)__ A structured and COT-like formalization template that decomposes complex formalization tasks into modular, foundational components, and systematically assembles them—like building blocks—into a complete formal expression. __(2)__ A semantic self-checking mechanism based on a divide-conquer-merge strategy to detect subtle inconsistencies between the formalization and the original statement. __(3)__ An iterative feedback-driven refinement loop that leverages both semantic and syntactic error signals to guide the LLM in progressively improving the formal output. By integrating these innovations, Loc-Decomp significantly enhances the accuracy of LLM-driven formalization, reduces reliance on human intervention, and moves closer to truly reliable automated reasoning. Extensive experiments on high-school-level and undergraduate-level datasets demonstrate that our approach achieves a significantly higher formalization success rate compared to baseline methods and state-of-the-art (SOTA) models. On the PutnamBench dataset, for instance, our method attains a success rate of 93.09\%, representing an improvement of 18 percentage points over the previous SOTA SFT-based model.

Large language model (LLM)-based agents are increasingly used to solve complex tasks involving tool use, such as web browsing, code execution, and data analysis. However, current evaluation benchmarks do not adequately assess their ability to solve real-world tasks that require synthesizing information from multiple sources and inferring insights beyond simple fact retrieval. To address this, we introduce DEEPSYNTH , a novel benchmark designed to evaluate agents on realistic, time-consuming problems that combine information gathering, synthesis, and structured reasoning to produce insights. DEEPSYNTH contains 120 tasks collected across 7 domains and data sources covering 42 countries. DEEPSYNTH is constructed using a multi-stage data collection pipeline that requires annotators to collect official data sources, create hypotheses, perform manual analysis and design tasks with verifiable answers. When evaluated on DEEPSYNTH, 9 state-of-the-art LLMs and deep research agents achieve a maximum F1 score of 8.97. Our analysis reveals that current agents struggle with hallucinations and reasoning over large information spaces, highlighting \ourdata as a crucial benchmark for guiding future research.

Large Language Models (LLMs) have made remarkable progress in mathematical reasoning, but still continue to struggle with high-precision tasks like numerical computation and formal symbolic manipulation. Integrating external tools has emerged as a promising approach to bridge this gap. Despite recent advances, existing methods struggle with three key challenges: constructing tool-integrated reasoning data, performing fine-grained optimization, and enhancing inference. To overcome these limitations, we propose THOR (Tool-Integrated Hierarchical Optimization via RL). First, we introduce TIRGen, a multi-agent based pipeline for constructing high-quality datasets of tool-integrated reasoning paths, aligning with the policy and generalizing well across diverse models. Second, to perform fine-grained hierarchical optimization, we introduce an RL strategy that jointly optimizes for both episode-level problem solving and step-level code generation. This is motivated by our key insight that the success of an intermediate tool call is a strong predictor of the final answer's correctness. Finally, THOR incorporates a self-correction mechanism that leverages immediate tool feedback to dynamically revise erroneous reasoning paths during inference. Our approach demonstrates strong generalization across diverse models, performing effectively in both reasoning and non-reasoning models. It further achieves state-of-the-art performance for models of a similar scale on multiple mathematical benchmarks, while also delivering consistent improvements on code benchmarks. Our code will be publicly available at https://github.com/JingMog/THOR.

Large Language Models (LLMs) are improving at an exceptional rate. With the advent of agentic workflows, multi-turn dialogue has become the de facto mode of interaction with LLMs for completing long and complex tasks. While LLM capabilities continue to improve, they remain increasingly susceptible to jailbreaking, especially in multi-turn scenarios where harmful intent can be subtly injected across the conversation to produce nefarious outcomes. While single-turn attacks have been extensively explored, adaptability, efficiency and effectiveness continue to remain key challenges for their multi-turn counterparts. To address these gaps, we present PLAGUE, a novel plug-and-play framework for designing multi-turn attacks inspired by lifelong-learning agents. PLAGUE dissects the lifetime of a multi-turn attack into three carefully designed phases (Primer, Planner and Finisher) that enable a systematic and information-rich exploration of the multi-turn attack family. Evaluations show that red-teaming agents designed using PLAGUE achieve state-of-the-art jailbreaking results, improving attack success rates (ASR) by more than 30% across leading models in a lesser or comparable query budget. Particularly, PLAGUE enables an ASR (based on StrongReject) of 81.4% on OpenAI's o3 and 67.3% on Claude's Opus 4.1, two models that are considered highly resistant to jailbreaks in safety literature. Our work offers tools and insights to understand the importance of plan initialization, context optimization, and lifelong learning in crafting multi-turn attacks for a comprehensive model vulnerability evaluation.

Agentic Retrieval-Augmented Generation （ RAG ）是一种强大的技术，可以整合大型语言模型（ LLM ）缺乏的外部信息，从而更好地解决问题和回答问题。然而，次优搜索行为广泛存在，例如过度搜索（检索已知信息）和搜索不足（必要时无法搜索） ，从而导致不必要的开销和不可靠的输出。 目前的训练方法通常依赖于强化学习（ RL ）框架中基于结果的奖励，缺乏解决这些低效问题所需的细粒度控制。为了克服这个问题，我们引入了$\ textbf {Hi} $ erarchical $\ textbf {P} $ rocess Rewards for Efficient agentic $\ textbf {RAG} $ (HiPRAG) ，这是一种新颖的培训方法，将细粒度、基于知识的过程奖励融入到RL培训中。 我们的方法通过将客服代表的推理轨迹分解为离散的、可解析的步骤，即时评估每个搜索决策的必要性。然后，我们应用分层奖励函数，该函数根据最优搜索和非搜索步骤的比例，在常用的结果和格式奖励之上提供额外的奖励。 在七个不同QA基准的Qwen2.5和Llama-3.2模型上进行的实验表明，我们的方法达到了65.4\ % （ 3B ）和67.2\ % （ 7B ）的平均准确度，优于强大的代理RAG基线。这在显著提高搜索效率的同时实现，将过度搜索率从先前工作的27\ %以上降至仅2.3\ % ，同时降低了搜索不足率。 这些结果证明了优化推理过程本身的有效性，而不仅仅是最终结果。进一步的实验和分析表明， HiPRAG在广泛的RL算法、模型族、大小和类型中显示出良好的泛化性。这项工作展示了通过RL进行细粒度控制的重要性和潜力，以提高搜索代理推理的效率和最优性。