ICML 2026 Papers

Agentic reinforcement learning (RL) for web search is prohibitively expensive due to long context lengths and costly environment interactions, and this inefficiency is further exacerbated by group-based optimization, which discards learning signals from entire rollout groups with zero reward variance. In this work, we propose EVOLVING ROLLOUTS, an RL framework for web-search agents that moves beyond episodic training and distills collected rollouts into in-context guidance for future policy behavior. By extracting the reward-labeled trajectories into strategic experiences, our method augments standard parameter-space optimization with implicit context-space optimization guided by prior experience. This enables the agent to recover learning signals from zero-variance rollouts, thereby fostering co-evolution between the policy and the experience repository. EVOLVING ROLLOUTS improves sample efficiency and task performance across representative web search benchmarks, with Qwen3-8B surpassing the much larger Qwen3-30B-A3B in average performance across GAIA, xBench, and HLE, and Qwen3-4B attaining comparable results on GAIA and HLE.

Static program analysis is a foundational technique in software engineering for reasoning about program behavior. Traditional static analysis algorithms model programs as logical systems with well-defined semantics, but rely on uniform, hard-coded heap abstractions. This limits their precision and flexibility, especially in dynamic languages like JavaScript, where heap structures are heterogeneous and difficult to analyze statically. In this work, we introduce ABSINT-AI, a language-model-guided static analysis framework that augments abstract interpretation with adaptive, per-object heap abstractions for Javascript. This enables the analysis to leverage high-level cues, such as naming conventions and access patterns, without requiring brittle, hand-engineered heuristics. Importantly, the LM agent operates within a bounded interface and never directly manipulates program state, preserving the soundness guarantees of abstract interpretation. To evaluate our approach, we focus on a soundness-critical task: determining whether object property accesses may result in undefined or null dereferences. This task directly models a common requirement in compiler optimizations, where proving that an access is safe enables the removal of dynamic checks or simplifies code motion. On this task, ABSINT-AI reduces false positives by up to 34% compared to traditional static analyses with fixed heap abstractions, while preserving formal guarantees. Our ablations show that the LM’s ability to interact agentically with the analysis environment is crucial, outperforming non-agentic LM predictions by 25%.

Achieving cross-task generalization remains a critical challenge in Multi-Agent Reinforcement Learning (MARL), fundamentally relying on effective inductive biases. However, existing entity-level biases often overlook collaborative patterns, whereas task-level biases lack sufficient coverage for novel scenarios. To address this, we introduce a role-level inductive bias as an intermediate abstraction that integrates entity-level flexibility with task-level inter-agent collaboration. To instantiate this, we propose Gaussian-mixture-model-based Transferable Role discovery (GTR). Specifically, GTR constructs a structured role space to ensure diverse role assignment, further achieves role decoupling via regularization, and ultimately utilizes these roles for efficient generalization. Empirical results demonstrate that GTR achieves superior zero-shot and few-shot transfer performance on unseen tasks compared to state-of-the-art methods.

LLM–driven evolutionary systems have shown promise for automated science discovery, yet existing approaches such as AlphaEvolve rely on full-code histories that are context-inefficient and potentially provide weak evolutionary guidance. In this work, we first formalize the evolutionary agents as a general Expectation–Maximization framework, where the language model samples candidate programs (E-step) and the system updates the control context based on evaluation feedback (M-step). Under this view, constructing context via full-code snapshots constitutes a suboptimal M-step, as redundant implement details dilutes core algorithmic ideas, making it difficult to provide clear inspirations for evolution. To address this, we propose DeltaEvolve, a momentum-driven evolutionary framework that replaces full-code history with structured semantic delta capturing how and why modifications between successive nodes affect performance. As programs are often decomposable, semantic delta usually contains many effective components which are transferable and more informative to drive improvement. By organizing semantic delta through multi-level database and progressive disclosure mechanism, input tokens are further reduced. Empirical evaluations on tasks across diverse scientific domains show that our framework can discover better solution with less token consumption over full-code-based evolutionary agents.

Chain-of-Thought (CoT) improves large language models (LLMs) on difficult reasoning tasks, but it often incurs long natural-language rationales that are poorly aligned with efficient machine reasoning. We propose Communicative Language Symbolism Routing (CLSR), a test-time framework in which multiple LLM agents autonomously invent, evolve, and share compact Language Symbolism Frameworks (LSFs), while a latent-free router adaptively selects and composes these languages per query to optimize the accuracy--token trade-off. Unlike prompt optimization that refines surface instructions, CLSR treats each LSF as a reusable symbolic protocol with compact symbols, usage rules, and a message-passing contract, and improves it through an evolutionary loop driven by correctness and token cost. At inference time, the router may invoke a single low-cost LSF call, ensemble multiple LSFs, or execute a multi-round LSF composition protocol on harder queries. Across challenging benchmarks, CLSR reduces latency-oriented generated token completion by $3\sim 6\times$ compared to standard CoT while maintaining accuracy. We further derive an information-theoretic lower bound on token cost under arbitrary symbolism and show that, under an interpreter-realizability premise, multi-round LSF protocols conditionally subsume program-execution pipelines.

Role specialization in multi-LLM agent systems is often realized via multi-LoRA, where agents share a pretrained backbone and differ only through lightweight adapters. Despite sharing base model weights, each agent independently builds and stores its own KV cache for the same long, tool-augmented trajectories, incurring substantial memory and compute overhead. Existing KV cache sharing methods largely overlook this multi-LoRA setting. We observe that, across agents, cache differences are dominated by adapter outputs, while activations from the shared pretrained backbone remain highly similar. Based on this observation, we propose LRAgent, a KV cache sharing framework for multi-LoRA agents that decomposes the cache into a shared base component from the pretrained weights and an adapter-dependent component from LoRA weights. LRAgent reduces memory overhead by sharing the base component and storing the adapter component in its inherent low-rank form, and further reduces compute overhead, enabled by shared-$A$ multi-LoRA architectures, by also sharing the low-rank cache and avoiding redundant computations for contexts already processed by other agents. To efficiently reconstruct adapter contributions at runtime, we introduce Flash-LoRA-Attention, a kernel that reorders attention computation to avoid materializing the low-rank cache to full dimension. LRAgent achieves throughput and time-to-first-token latency close to fully shared caching, while preserving accuracy near the non-shared caching baseline across agentic question-answering benchmarks.

Large Language Models (LLMs) increasingly act as function call agents that invoke external tools to tackle tasks beyond their static knowledge. However, they typically invoke tools one at a time without a global view of task structure. As tools often depend on one another, this leads to error accumulation and poor scalability, particularly when scaling to hundreds or thousands of tools. To address these limitations, we propose NaviAgent, an explicit bilevel architecture that decouples task planning from tool execution through graph‑based modeling of tool relations. At the planning level, the LLM‑based agent decides whether to respond directly, clarify intent, or retrieve and execute a toolchain independent of inter‑tool complexity. At the execution level, a Tool World Navigation Model (TWNM) encodes structural and behavioral relations among tools, steering the agent to compose scalable and robust invocation sequences. Incorporating feedback from real tool interactions, NaviAgent achieves closed‑loop alignment between planning and execution, enabling adaptive navigation in large‑scale tool ecosystems. Evaluations on API-Bank and ToolBench show consistent improvements in task success rate (TSR), with TWNM yielding an average gain of 13.1 points on complex tasks. Further tests on 50 real APIs across 7 domains show consistent gains of 4.3-12.0 points, with fewer steps and latency, demonstrating robust generalization under real-world dynamics.

Estimating uncertainty for AI agents in real-world multi-turn tool-using interaction with humans is difficult because failures are often triggered by sparse critical episodes (e.g., looping, incoherent tool use, or user-agent miscoordination) even when local generation appears confident. Existing uncertainty proxies focus on single-shot text generation and therefore miss these trajectory-level breakdown signals. We introduce TRACER, a trajectory-level uncertainty metric for dual-control Tool-Agent-User interaction. TRACER combines content-aware surprisal with situational-awareness signals, semantic and lexical repetition, and tool-grounded coherence gaps, and aggregates them using a tail-focused risk functional with a MAX-composite step risk to surface decisive anomalies. We evaluate TRACER on $\tau^2$-bench (Barres et al., 2025) by predicting task failure and selective task execution. To this end, TRACER improves AUROC by up to 37.1\% and AUARC by up to 55\% over baselines, enabling earlier and more accurate detection of uncertainty in complex conversational tool-use settings. Our codebase is available at: https://github.com/sinatayebati/agent-tracer.

Knowledge distillation (KD) transfers knowledge from a large teacher model to a smaller student. In language modeling, the student is trained either on tokens sampled from the teacher (\textbf{hard labels}) or the teacher’s full next-token distribution (\textbf{soft labels}). Despite soft labels appear strictly richer, we find that mixing hard and soft labels consistently yields better results. Crucially, we show that this gain cannot be explained by closer teacher matching during training. Instead, it comes from reduced exposure bias---the mismatch between training and inference distributions. To explain this phenomenon, we introduce the Bridge–Garden Decomposition theory, which categorizes generation steps into two types: \textit{Bridges}, where the next token must be \textit{exact}, and \textit{Gardens}, where it can be \textit{flexible}. We show that hard-only KD excels in Bridges by avoiding risky deviations, while soft-only KD preserves diversity in Gardens. A hybrid strategy handles both cases and, as a result, reduces exposure bias across the sequence. Guided by this theory, we develop a family of Bridge--Garden hybrid supervision methods that adaptively balance hard and soft labels. Across seven teacher--student pairs (including Qwen, Llama, Gemma, and DeepSeek) and benchmarks in reasoning and coding, our approach outperforms divergence-based and on-policy KD baselines while reducing training cost by \textbf{9.7$\times$}, enabling efficient model compression.

Reinforcement learning (RL) has become a de facto paradigm for building LLM-based agents that act, interact, and reason over extended task horizons. However, in active reasoning where agents must elicit new observations through interaction with the environment to solve the task, we find that outcome-based RL can induce a systematic failure mode which we call information self-locking (SeL): agents fail both to elicit informative feedback and to internalize obtained evidence. To understand the issue, we trace agentic behaviors into two coupled capabilities: Action Selection (AS), which determines observation streams, and Belief Tracking (BT), which updates the agent’s internal task understanding. Theoretical and empirical analyses reveal a bidirectional bottleneck that leads to SeL: weak BT obscures the credit of informative actions, while weak AS deprives BT of useful evidence. This coupling weakens the learning signal for both capabilities and leads to SeL. To mitigate this issue, we propose AREW, a simple yet effective Advantage Reweighting method that uses easy-to-obtain directional critiques to reallocate credit within trajectories. Extensive experiments across 9 agentic tasks of varying complexity show that AREW significantly mitigates SeL, yielding up to 60-point gains in final performance. Code is available at https://github.com/unimpor/T3.

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 achieve her objectives. 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 examples of human-written text, 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. With this empowerment objective, we provide a framework for useful aligned AI agents without the need for human feedback or verifiable rewards.

Securing Contextual Integrity (CI) is critical for privacy-preserving Large Language Model (LLM) agent execution. However, existing agents struggle to balance the agility of direct generation against the prohibitive latency of CI-constrained thinking. To address this, we propose PrivGate, a framework that selectively invokes explicit reasoning based on internal privacy signals. Our approach is grounded in the discovery of a privacy manifold, where models linearly encode privacy sensitivity within their residual streams, even during non-compliant generation. Leveraging this structure, PrivGate employs Latent Gating, a training-free mechanism that requires no fine-tuning of the base LLM and triggers explicit reasoning only when high latent risk is detected, thereby optimizing the efficiency-privacy trade-off by minimizing unnecessary compute. On the contextual PrivacyLens benchmark, PrivGate maintains consistently high performance in out-of-distribution risk identification, validating the generalizability of the discovered manifold. End-to-end evaluations show that PrivGate achieves a 62.6% average relative reduction in privacy leakage with 15.9% token overhead, offering a practical pathway to reconcile rigorous CI requirements with the performance demands of LLM agents.

To support long-term interaction in complex environments, LLM agents require memory systems that manage historical experiences. Existing approaches either retain full interaction histories via passive context extension, leading to substantial redundancy, or rely on iterative reasoning to filter noise, incurring high token costs. To address this challenge, we introduce SimpleMem, an efficient memory framework based on semantic lossless compression. We propose a three-stage pipeline designed to maximize information density and token utilization: (1) Semantic Structured Compression, which distills unstructured interactions into compact, multi-view indexed memory units; (2) Online Semantic Synthesis, an intra-session process that instantly integrates related context into unified abstract representations to eliminate redundancy; and (3) Intent-Aware Retrieval Planning, which infers search intent to dynamically determine retrieval scope and construct precise context efficiently. Experiments on benchmark datasets show that our method consistently outperforms baseline approaches in accuracy, retrieval efficiency, and inference cost, achieving an average F1 improvement of 26.4% in LoCoMo while reducing inference-time token consumption by up to 30×, demonstrating a superior balance between performance and efficiency. Code is available at https://github.com/aiming-lab/SimpleMem.

Multi-agent debate (MAD) systems are often evaluated using binary correctness or peer agreement, yet such evaluations obscure how individual agents revise their answers during social interaction. We study the microscopic dynamics of answer revision in large language models (LLMs) using ConceptARC, a 2D grid-reasoning benchmark that admits quantitative distance measures between candidate solutions. By exposing a target model to controlled configurations of peer answers, we analyze how the likelihood and direction of revision depend on both social context and the distance between answers and the ground truth. We find that agents are more likely to revise when their answers are farther from the correct solution, and that revisions of incorrect answers exhibit a systematic contraction toward the ground truth, even when the final answer remains incorrect. Conversely, correct answers can be overturned by social pressure, particularly when wrong peers are near-correct. Together, these results show that multi-agent interaction induces structured, distance-aware movements in solution space that are invisible under binary correctness, clarifying when social reasoning leads to improvement, stability, or gradual regression in solution quality.

Existing benchmarks for computational materials discovery primarily evaluate static predictive tasks or isolated computational sub-tasks. While valuable, these evaluations neglect the inherently iterative and adaptive nature of scientific discovery. We introduce **MA**terials **D**iscovery **E**nvironments (**MADE**), a novel framework for benchmarking end-to-end autonomous materials discovery pipelines. MADE simulates closed loop discovery campaigns in which an agent or algorithm proposes, evaluates, and refines candidate materials under a constrained oracle budget, capturing the sequential and resource-limited nature of real discovery workflows. We formalize discovery as a search for thermodynamically stable compounds relative to a given convex hull, and evaluate efficacy and efficiency via comparison to baseline algorithms. The framework is flexible; users can compose discovery agents from interchangeable components such as generative models, filters, and planners, enabling the study of arbitrary workflows ranging from fixed pipelines to agentic systems. We demonstrate this by conducting systematic experiments across a diverse range of systems and algorithms, finding that adaptive planning becomes more important to discovery efficiency as the search space scales.

LLM coding agents can generate working code, but their solutions often accumulate complexity, duplication, and architectural debt. Human developers address such issues through refactoring: behavior-preserving program transformations that improve structure and maintainability. We investigate whether agents (i) can execute refactorings reliably and (ii) identify the refactorings that human developers actually chose in real codebases. To this end, we construct CodeTaste, a benchmark mined from large multi-file open-source refactorings. To score solutions, we combine repository test suites that measure functional correctness with tailored static checks that verify removal of undesired and introduction of desired code patterns using dataflow reasoning. Our results show a clear gap: agents perform well at implementing refactorings that are specified in detail, but often fail to discover the human refactoring choices when given a focus area for changes. A propose-then-implement decomposition improves alignment, and selecting the best-aligned proposal before implementation can yield further gains. CodeTaste provides an evaluation target and a potential preference signal for aligning coding agents with human refactoring decisions in realistic codebases. We release the benchmark, leaderboard, and code.

Existing multimodal retrieval systems excel at semantic matching but implicitly assume that query-image relevance can be measured in isolation. This paradigm overlooks the rich dependencies inherent in realistic visual streams, where information is distributed across temporal sequences rather than confined to single snapshots. To bridge this gap, we introduce DeepImageSearch, a novel agentic paradigm that reformulates image retrieval as an autonomous exploration task. Models must plan and perform multi-step reasoning over raw visual histories to locate targets based on implicit contextual cues. We construct DISBench, a challenging benchmark built on interconnected visual data. To address the scalability challenge of creating context-dependent queries, we propose a human-model collaborative pipeline that employs vision-language models to mine latent spatiotemporal associations, effectively offloading intensive context discovery before human verification. Furthermore, we build a robust baseline using a modular agent framework equipped with fine-grained tools and a dual-memory system for long-horizon navigation. Extensive experiments demonstrate that DISBench poses significant challenges to state-of-the-art models, highlighting the necessity of incorporating agentic reasoning into next-generation retrieval systems.

World models enable learning policies via latent imagination, offering benefits such as history compression and sample efficiency. The primary challenge in applying world models to multi-agent tasks is that modeling multi-agent dynamics in latent space requires integrating information from different agents, often creating spurious correlations between their latent states. Existing methods either reconstruct the observation for each agent or employ communication to maintain correlation during execution, failing to learn disentangled latent states that are crucial for effective decentralized control. To address this, we present the Disentangled Multi-Agent World Model (DMAWM). It facilitates learning decentralized policies in the latent space through a novel architecture comprising independent agent modules and a shared environment module. During real-environment execution, agent modules independently process local information to form a factorized latent representation. The environment module is then trained to mirror the factorized structure generated by the agent modules, effectively disentangling individual latent states from the interaction dynamics. Consequently, imaginary rollouts generated by the environment module more faithfully simulate decentralized execution dynamics, facilitating the transfer of policies from imagination to decentralized execution. Empirically, DMAWM outperforms existing model-based and model-free approaches in convergence speed and final performance, with additional visualization demonstrating its efficacy in capturing agent interactions.

Integrating graph knowledge into Large Language Models (LLMs) via passive representation faces critical bottlenecks: limited context windows, unreliable numerical computation, and structural hallucinations. To solve this, we propose **GRASP** (**G**raph **R**easoning via **A**gentic **S**olving and **P**robing), shifting the paradigm from passive ingestion to proactive agentic exploration. By interleaving Neighbor Retrieval for on-demand probing with Code Interpreter as a deterministic solver, GRASP enables LLMs to autonomously navigate and compute over complex topologies. We employ a staged reinforcement learning strategy (GRPO) that transitions from visible tuning to a structure-blind environment, forcing the agent to develop genuine topological awareness. Evaluated on multi-domain graph reasoning benchmarks, our 4B model achieves a 53.06% average performance boost, surpassing SOTA baselines like DeepSeek-V3.2 and successfully generalizing to unseen tasks, with high potential for tackling sampling on million-node graphs and solving Hard-level LeetCode graph problems. Our implementation is open-sourced at https://github.com/PKU-ML/GRASP, with models hosted on Huggingface collection https://huggingface.co/collections/PKU-ML/grasp.