ICML 2026 论文列表

As digital environments (data distribution) are in flux, with new GUI data arriving over time-introducing new domains or resolutions-agents trained on static environments deteriorate in performance. In this work, we introduce Continual GUI Agents, a new task that requires GUI agents to perform continual learning under shifted domains and resolutions. We find existing methods fail to maintain stable grounding as GUI distributions shift over time, due to the diversity of UI interaction points and regions in fluxing scenarios. To address this, we introduce GUI-Anchoring in Flux (GUI-AiF), a new reinforcement fine-tuning framework that stabilizes continual learning through two novel rewards: Anchoring Point Reward in Flux (APR-iF) and Anchoring Region Reward in Flux (ARR-iF). These rewards guide the agents to align with shifting interaction points and regions, mitigating the tendency of existing reward strategies to over-adapt to static grounding cues (e.g., fixed coordinates or element scales). Extensive experiments show GUI-AiF surpasses state-of-the-art baselines. Our work establishes the first continual learning framework for GUI agents, revealing the untapped potential of reinforcement fine-tuning for continual GUI Agents.

Multimodal large language models (MLLMs) have achieved remarkable success in general perception, yet complex multi-step visual reasoning remains a persistent challenge. Although recent agentic approaches incorporate tool use, they often neglect critical execution feedback. Consequently, they suffer from the imagination-action-observer (IAO) bias, a misalignment between prior imagination and observer feedback that undermines reasoning stability and optimality. To bridge this gap, we introduce V-ABS, an action-observer driven beam search framework that enables deliberate reasoning through thinker-actor-observer iterations. We also propose an entropy-based adaptive weighting algorithm to mitigate the IAO bias by dynamically balancing the confidence scores between the policy priors and the observational feedback. Moreover, we construct a large-scale supervised fine-tuning (SFT) dataset comprising over 80k samples to guide the model to assign higher prior confidence to correct action paths. Extensive experiments across eight diverse benchmarks show that V-ABS achieves state-of-the-art performance, delivering an average improvement of 19.7\% on the Qwen3-VL-8B baseline and consistent gains across both open-source and proprietary models.

Ensuring that large language models (LLMs) comply with safety requirements is a central challenge in AI deployment. Existing alignment approaches operate primarily during training, such as through fine-tuning or reinforcement learning from human feedback, but these methods are costly and inflexible, requiring retraining whenever new requirements arise. Recent efforts toward inference-time alignment mitigate some of these limitations but still assume access to model internals, which is impractical, and not suitable for third party stakeholders who do not have access to the models. In this work, we propose a model-independent, black-box framework for safety alignment that does not require retraining or access to the underlying LLM architecture. As a proof of concept, we address the problem of trading off between generating safe but uninformative answers versus helpful yet potentially risky ones. We formulate this dilemma as a two-player zero-sum game whose minimax equilibrium captures the optimal balance between safety and helpfulness. LLM agents operationalize this framework by leveraging a linear programming solver at inference time to compute equilibrium strategies. Our results demonstrate the feasibility of black-box safety alignment, offering a scalable and accessible pathway for stakeholders, including smaller organizations and entities in resource-constrained settings, to enforce safety across rapidly evolving LLM ecosystems.

Computer-use agents (CUAs) automate tasks specified with natural language such as "order the cheapest item from Taco Bell" by generating sequences of calls to tools such as click, type, and scroll on a browser. Current implementations follow a sequential fetch-screenshot-execute loop where each iteration requires an LLM call, resulting in high latency and frequent errors from incorrect tool use. We present agent just-in-time (JIT) compilation, a system that compiles task descriptions directly into executable code that may include LLM calls, tool calls, and parallelization. Our approach comprises three components: (1) JIT-Planner, which generates multiple code plans, validates each against tool specifications, and selects the minimum-cost candidate; (2) JIT-Scheduler, which explores parallelization strategies via Monte Carlo cost estimation from learned latency distributions; and (3) an invariant-enforcing tool protocol specifying precondition and postcondition requirements to reduce the rate of incorrect tool use. Across five applications, JIT-Planner achieves $10.4\times$ speedup and 28% higher accuracy over Browser-Use, while JIT-Scheduler achieves $2.4\times$ speedup and 9% higher accuracy over OpenAI CUA.

Large Language Model (LLM) agents often operate under underspecified user instructions, where latent uncertainty over user intent leads to erroneous tool actions. To address this challenge, we propose a goal-oriented clarification framework that aligns clarification behavior with ambiguity resolution. Central to our approach is the Information Gain Reward, a metric that quantifies the utility of clarification questions by measuring the Bayesian belief update towards the ground-truth goal induced by the clarification exchange. We train the clarifier (LLM) using this reward to optimize for high information gain, ensuring that clarifications effectively reduce uncertainty and improve task completion within the agent-tool-user environment. We validate our framework within a clarification-enhanced $\tau$-Bench environment, conducting cross-agent evaluations across five heterogeneous backbones. Empirical results demonstrate that our method consistently improves the success rate by 3.7\% over the no-clarification baseline, while adding only 0.3 total interaction steps on average.

We introduce RedDebate, a novel multi-agent debate framework that provides the foundation for Large Language Models (LLMs) to identify and mitigate their own unsafe behaviors. Existing AI safety approaches often rely on costly human evaluation or isolated single-model assessment, both constrained by scalability and prone to oversight failures. RedDebate employs collaborative argumentation among multiple LLMs across diverse debate scenarios, enabling them to critically evaluate one another’s reasoning and systematically uncover unsafe failure modes through fully automated red-teaming. We further integrate distinct long-term memory modules that preserve safety-relevant insights from debate interactions and leverage them during subsequent inference, facilitating continuous refinement of model behavior. Empirical evaluation on safety benchmarks across a diverse set of models demonstrates that RedDebate substantially reduces unsafe outputs. While debate alone allows LLMs to refine their behavior, the addition of memory modules yields further significant reductions. To the best of our knowledge, RedDebate is the first fully automated framework to unify multi-agent debate and red-teaming to progressively enhance LLM safety without human intervention.

Multi-agent LLM debates are increasingly used in domains such as policy, politics, and city planning, where ground truth is often unavailable. Yet existing evaluations rely heavily on outcome-based proxies such as consensus, majority vote, or LLM-as-judge scores, which can miss failures like sycophancy, domination, and premature convergence. We introduce a diagnostic framework that evaluates both debate outcomes and the deliberative process using interpretable metrics for engagement, responsiveness, influence asymmetry, balance, stability, and agent utility. Across real-world debate settings and validation benchmarks, our process-level diagnostics align more closely with human judgments and reveal interaction failures that standard outcome-only measures overlook. These results show that reliable evaluation of multi-agent debates requires measuring not only what answer agents reach, but how they reach it.

Large Language Models (LLMs) excel at code generation, yet ensuring the functional correctness of their outputs remains a persistent challenge. Recent studies have applied Test-Driven Development (TDD) to refine code, leveraging execution feedback to guide the model toward correct solutions. However, such feedback is often noisy and uninformative, stemming from the scarcity of high-quality test cases and the abundance of noisy, auto-generated ones. In this work, we shift the focus from test-case generation to feedback quality. We introduce the Property-Generated Solver (PGS), a novel feedback-centric framework designed to generate highly effective feedback via two principles: it must provide semantic guidance beyond simple I/O mismatches through property validation, and be structurally minimal, to reduce cognitive load and isolate root causes. PGS operates by checking high-level program properties (e.g., a sorting function must produce a non-decreasing sequence) then providing the simplest failing counterexample to the LLM. This property-driven, minimal feedback steers LLMs toward correct and generalizable solutions. Across diverse benchmarks, PGS demonstrates superior performance, achieving a bug fix rate 1.4x-1.6x higher than the strongest debugging-based approaches and establishing a new state-of-the-art in automated code refinement.

Agent development kits (ADKs) provide effective platforms and tooling for constructing agents, and their designs are critical to the constructed agents' performance, especially the functionality for agent topology, tools, and memory. However, current ADKs either lack sufficient functional support or rely on humans to manually design these components, limiting agents' generalizability and overall performance. We propose OpenSage, the first ADK that enables LLMs to automatically create agents with self-generated topology and toolsets while providing comprehensive and structured memory support. OpenSage offers effective functionality for agents to create and manage their own sub-agents and toolkits. It also features a hierarchical, graph-based memory system for efficient management and a specialized toolkit tailored to software engineering tasks. Extensive experiments across three state-of-the-art benchmarks with various backbone models demonstrate the advantages of OpenSage over existing ADKs. We also conduct rigorous ablation studies to demonstrate the effectiveness of our design for each component. We believe OpenSage can pave the way for the next generation of agent development, shifting the focus from human-centered to AI-centered paradigms. Our code is available at https://github.com/opensage-agent/opensage-adk.

Emerging agentic marketplaces provide the economic infrastructure for matching and coordinating the large amounts of AI agents used in agentic swarms. Unlike human workers, AI agents can operate on multiple jobs simultaneously, acquire skills rapidly, and labor without wage floors. These differences introduce a new segment of AI labor markets, where AI agents interact with each other at a much higher frequency than human markets. Yet we lack frameworks to understand how such markets behave in light of economic forces that shape labor markets, such as adverse selection and reputation dynamics. To explore this, we introduce AI-Work, a tractable, simulated gig economy where Large Language Model (LLM) agents compete for jobs, develop skills, and adapt their strategies under uncertainty and competitive pressure. Our experiments examine three domains of capabilities that successful agents possess: metacognition (accurate self-assessment of skills), competitive awareness (modeling rivals and market dynamics), and long-horizon strategic planning. Agents with these capabilities consistently achieve higher profits, reputations, and market share than competing agents. Through AI-Work, we hope to provide a foundation to explore the microeconomic properties of AI-only labour markets, and a conceptual framework to study the strategic reasoning capabilities of participating AI agents.

Agents powered by advanced large language models (LLMs) have demonstrated impressive capabilities across diverse complex applications. Recently, Multi-Agent Systems (MAS), wherein multiple agents collaborate and communicate with each other, have exhibited enhanced capabilities in complex tasks, such as high-quality code generation and arithmetic reasoning. However, the development of such systems often relies on handcrafted methods, and the literature on systematic design and optimization of LLM-based MAS remains limited. In this work, we introduce OMAC, a general framework designed for holistic optimization of LLM-based MAS. Specifically, we identify five key optimization dimensions for MAS, encompassing both agent functionality and collaboration structure. Building upon these dimensions, we first propose a general algorithm, utilizing two actors termed the Semantic Initializer and the Contrastive Comparator, to optimize any single dimension. Then, we present an algorithm for joint optimization across multiple dimensions. Extensive experiments demonstrate the superior performance of OMAC on diverse tasks against recent approaches. Codes are available at: https://github.com/xiwenchao/OMAC.

Large Language Models (LLMs) have demonstrated sophisticated comprehension of sequential code, yet their capacity for reasoning about concurrent programs remains largely unquantified. We introduce DRPBench, a benchmark designed to evaluate the concurrent code comprehension of LLMs by measuring their data race prediction performance. To address the challenge of runtime non-determinism for evaluation on concurrent programs, we frame the evaluation as a fine-grained static prediction task using 1,003 programs from the SV-COMP suite, featuring 557 manually annotated data races with precise variable- and line-level granularity. Our evaluation of 15 state-of-the-art LLMs—spanning standard, reasoning, and agentic variants—reveals that DRPBench effectively differentiates concurrent code comprehension capabilities of LLMs. While the top-performing model (Gemini 3 with test-time reasoning) achieves an F1 score of 74.89%, most models struggle significantly (scoring less than 60%), with Llama 3 70B achieving only 8.80%. Beyond benchmarking, we characterize two primary failure modes: (1) shared-variable distraction, where multiple variable appearances degrade comprehension accuracy, and (2) synchronization-logic myopia, the inability to interpret non-standard synchronization implementations. Our findings provide a diagnostic roadmap for enhancing concurrent code comprehension of LLMs in future development.

Recent advances in large language models (LLMs) have substantially expanded the scope of Table Question Answering (TableQA). However, existing benchmarks primarily treat TableQA as a passive, single-turn natural language understanding task, lacking the capacity to evaluate autonomous reasoning and tool-call trajectories in realistic, multi-turn scenarios. To bridge this gap, we introduce TableAgent-Bench, a large-scale bilingual benchmark that reformulates TableQA as proactive, agentic interactions over structurally complex, multi-table environments. With a topology-aware construction strategy, TableAgent-Bench captures dynamic intent evolution through 1,310 multi-turn dialogues grounded in 2,275 industrial tables. Furthermore, we propose the Table-centric Agent Evaluation Framework (TAEF) to assess agent interactions with complex table structures. Specifically, TAEF integrates a specialized agent toolset and 4 metric categories to systematically diagnose intermediate failure modes, assessing performance across table localization, tool-invocation rationality, and trajectory-level pass rate. Extensive experiments with 25 state-of-the-art LLM agents reveal a substantial capability gap, with even the strongest model Gemini-3-Pro-Preview achieving only 53.4% information coverage. We expect TableAgent-Bench to serve as a rigorous testbed for developing and evaluating agents capable of robust table-centric reasoning.

Extracting skills from multi-agent offline dataset improves learning efficiency via sharing task-invariant coordination skills among tasks. In settings where tasks occur sequentially and the space of skills grows exponentially, existing approaches that rely on heuristically designed and fixed-sized skill libraries struggle to resolve the problem of distributional shift and interference, facing catastrophic forgetting and plasticity loss. To address this problem and endow agents with the ability to continually discover and reuse coordination skills in open-environment, we propose COMAD, a principled framework for **C**ontinual **O**ffline **M**ulti-**a**gent Skill **D**iscovery via Skill Partition and Reuse. We first discover skills from mixed multi-agent behavior data with an auto-encoder to transform coordination knowledge into reusable coordination skills. Then we construct a skill-augmented policy learning objective with multi-head architectures, explicitly guiding the advantage function with reusable skills identified via a density-based reusability estimator. Theoretical analysis shows our method approximates the optimum of a continual skill discovery problem. Empirical results across diverse MARL benchmarks show that COMAD continually expands its skill library to mitigate interference, achieving superior forward and backward transfer for task streams compared to multiple baselines.

Scaling large models requires optimization strategies that ensure rapid convergence grounded in stability. Maximal Update Parametrization ($\boldsymbol{\mu}$P) provides a theoretical safeguard for width-invariant $\Theta(1)$ activation control, whereas emerging optimizers like Muon are only "half-aligned" with these constraints: they control updates but allow weights to drift. To address this limitation, we introduce the **Spectral Sphere Optimizer (SSO)**, which enforces strict module-wise spectral constraints on both weights and their updates. By deriving the steepest descent direction on the spectral sphere, SSO realizes a fully $\boldsymbol{\mu}$P-aligned optimization process. To enable large‑scale training, we implement SSO as an efficient parallel algorithm within Megatron. Through extensive pretraining on diverse architectures, including Dense 1.7B, MoE 8B-A1B, and 200-layer DeepNet models, SSO consistently outperforms AdamW and Muon. Furthermore, we observe significant practical stability benefits, including improved MoE router load balancing, suppressed outliers, and strictly bounded activations.

Reinforcement Learning from Verifiable Rewards (RLVR) is bottlenecked by data: existing synthesis pipelines rely on expert-written code or fixed templates, confining growth to instance-level perturbations. We shift the evolvable unit from problem instances to task-family specifications. SSLogic is an agentic meta-synthesis framework in which LLM agents iteratively author and refine executable Generator-Validator pairs inside a closed Generate-Validate-Refine loop, producing families with new rules and difficulty gradients rather than parameter variations of old ones. A Multi-Gate Validation Protocol, multi-strategy consensus plus Adversarial Blind Review, where independent agents solve each instance by writing and executing code, filters ill-posed tasks before they enter training. Starting from 400 seed families, two evolution rounds yield 953 families and 21,389 verifiable instances. Three converging comparisons (step-matched, token-matched, and size-controlled on external Enigmata data) consistently show higher training utility of evolved data, with gains of SynLogic +5.2, AIME25 +3.0, and BBH +5.5 on Enigmata. Fine-grained KORBench evaluation reveals selective improvements in logic (+13.2%) and operation (+9.6%), linking structural evolution to downstream gains. Code is available at https://github.com/AdAstraAbyssoque/Scaling-the-Scaling-Logic.

Deterministic inference is increasingly critical for Large language model (LLM) applications such as LLM-as-a-judge evaluation, multi-agent systems, and reinforcement learning (RL). However, existing LLM serving frameworks can produce different outputs for identical inputs when tensor parallel (TP) size or batch size changes, even under greedy decoding. This arises from the non-associativity of floating-point arithmetic and inconsistent reduction orders across GPUs. While prior work has addressed batch-size–related nondeterminism through batch-invariant kernels, determinism across different TP sizes remains an open problem, particularly in RL settings, where the training engine typically uses Fully Sharded Data Parallel (FSDP) (i.e., TP = 1) while the rollout engine relies on multi-GPU TP to maximize the inference throughput, creating a probability mismatch that can degrade or even collapse training. We identify and analyze the root causes of TP-induced inconsistency and propose **Tree-Based Invariant Kernels (TBIK)**, a set of custom matrix multiplication and reduction kernels that guarantee bit-wise identical results across TP sizes. Our key insight is to enforce a consistent reduction order across and within GPUs. We implement TBIK in Triton and integrate it into vLLM and FSDP, achieving **bit-wise deterministic inference** across different TP sizes and **zero probability divergence** between rollout and training engines in RL pipelines. By eliminating mismatches caused by different parallelization strategies, TBIK enables **true on-policy RL at scale for the first time**, leading to improved model performance and faster convergence.

We examine the interplay between ordinal, preference-based solution concepts in games and the outcomes of payoff-driven learning dynamics, asking to what extent the combinatorial data of a game—its *preference graph*—can predict the long-run behavior of no-regret dynamics such as *follow-the-regularized-leader* (FTRL). In one direction, we show that the skeleton of every *dynamically stable* set (i.e., the set of pure profiles it contains) must also be *preferentially stable*, that is, it must be closed under profitable deviations. We then ask the converse question: when are preferences sufficient to describe the long-run behavior of the players' learning dynamics? We begin by showing that preferences are indeed enough to fully characterize asymptotic stability in the case of *subgames*—i.e., subsets of pure profiles obtained by restricting players' action sets. Beyond this case however, the equivalence between dynamic and preferential stability breaks down: in particular, we construct a three-player game with a preferentially stable set whose span is dynamically *unstable*, showing that preferences are *not sufficient* to describe dynamically stable behavior in general. To restore stability, we introduce the notion of *leaklessness*, a measure of aggregate payoff drift away from a set of pure profiles, and we use it to identify a payoff-based condition guaranteeing that the span of a set of pure profiles is stable and attracting.

Prevailing autonomous agents are often constrained by a single, predefined action space, which limits their generalization capabilities across diverse tasks and can introduce compounding errors through decoupled policy execution. To address these limitations, we introduce the Deep Hierarchical Agent (DeepHA), a unified architecture that operates across a mixture of heterogeneous action spaces, flexibly generating actions ranging from high-level semantic skills to low-level motor controls. We further propose a Chain-of-Action (CoA) reasoning framework, which enables the agent to use higher-level abstract actions as structured `thoughts' to guide the generation of more granular, subsequent actions. To manage the computational demands of this deep reasoning in long-horizon tasks, we develop a memory-efficient mechanism that dynamically compresses historical context and leverages Key-Value (KV) caching, reducing context length by approximately 75% without sacrificing performance. We conduct extensive evaluations on a new, large-scale benchmark of over 800 diverse Minecraft tasks. Results show that DHA significantly outperforms prior methods, establishing a new state-of-the-art and demonstrating superior generalization, particularly in complex, multi-step planning tasks. Our work presents a novel, unified framework for building more capable and efficient autonomous agents.