ICLR 2026 Papers

While humans naturally learn and adapt from past experiences, large language models (LLMs) and their agentic counterparts often fail to retain reasoning from previous tasks and apply it in future contexts. We introduce **L**og-**A**ugmented **G**eneration (LAG), a novel framework that *directly reuses* prior computation and reasoning from past logs at test time, enabling models to learn from previous tasks and perform better on new, unseen challenges, without sacrificing the system's efficiency or scalability. Our approach represents task logs as key-value (KV) caches that encode the full reasoning context of prior tasks, while storing KV values for only a selected subset of tokens. When a new task arises, LAG retrieves KV values from relevant logs to augment generation. Unlike reflection-based memory mechanisms, which require additional extraction or distillation steps, LAG reuses prior reasoning verbatim. Moreover, it extends beyond existing KV caching techniques, which have primarily targeted efficiency, by explicitly improving accuracy through log reuse. Experiments on knowledge- and reasoning-intensive datasets demonstrate that our method significantly outperforms standard agentic systems that do not utilize logs, as well as existing solutions based on reflection and KV cache techniques.

Multimodal large language models are progressively advancing toward multimodal agents that can proactively execute tasks. Existing research on multimodal agents primarily targets either GUI or embodied scenarios, corresponding to interactions within 2D virtual world and 3D physical world, respectively. However, many real-world tasks inherently require agents to interleave interactions across both types of environments. We initially mix GUI and embodied data to train models, but find performance degradation caused by data conflicts. Further analysis reveals that GUI and embodied data exhibit synergy at shallow layers but conflict at deep layers, resembling the cerebrum-cerebellum mechanism in the human brain. To this end, we introduce a high-performance generalist agent, OmniActor, designed from both structural and data perspectives. First, we propose Layer-heterogeneous MoE that separates parameters at deep layers to eliminate conflict, while sharing parameters at shallow layers to leverage synergy. This design enables OmniActor to outperform agents trained solely on GUI or embodied data in their respective tasks. Furthermore, we unify the action spaces of GUI and embodied tasks and collect large-scale datasets from diverse sources for training. This substantially enhances the performance of OmniActor across various scenarios, especially in GUI tasks.

We introduce ComputerRL, a framework for autonomous desktop intelligence that enables agents to operate complex digital workspaces skillfully. ComputerRL features the API-GUI paradigm, which unifies programmatic API calls and direct GUI interaction to address the inherent mismatch between machine agents and human-centric desktop environments. Scaling end-to-end RL training is crucial for improvement and generalization across diverse desktop tasks; however, it remains challenging due to environmental inefficiency and instability during extended training. To support scalable and robust training, we develop a distributed RL infrastructure capable of orchestrating thousands of parallel virtual desktop environments to accelerate large-scale online RL. Furthermore, we propose Entropulse, a training strategy that alternates reinforcement learning with supervised fine-tuning, effectively mitigating entropy collapse during extended training runs. We employ ComputerRL on open models GLM-4-9B-0414 and GLM-4.1V-9B-Thinking, and evaluate them on the OSWorld benchmark. The GLM-ComputerRL-9B achieves a new state-of-the-art accuracy of 48.9%, demonstrating significant improvements for general agents in desktop automation. Our code and demos are available at https://computer-rl.vercel.app/.

Large Language Model (LLM) agents show great promise for complex multi-turn tool-use tasks, but their development is often hampered by the extreme scarcity of high-quality training data. Supervised fine-tuning (SFT) on synthetic data leads to overfitting, whereas standard reinforcement learning (RL) struggles with a critical cold-start problem and training instability. To address these challenges, we introduce $\textbf{Environment Tuning}$, a novel training paradigm that enables agents to learn complex behaviors directly from problem instances without relying on pre-collected expert trajectories. $\textbf{Environment Tuning}$ orchestrates this learning process through a structured curriculum, actionable environment augmentation that provides corrective feedback, and fine-grained progress rewards to ensure stable and efficient exploration. Using only 400 problem instances from Berkeley Function-Calling Leaderboard (BFCL) benchmark, our method not only achieves competitive in-distribution performance against strong baselines but also demonstrates superior out-of-distribution generalization, overcoming the performance collapse common to SFT-based approaches. Our work presents a paradigm shift from supervised fine-tuning on static trajectories to dynamic, environment-based exploration, paving the way for training more robust and data-efficient agents.

The widespread deployment of LLM-based agents is likely to introduce a critical privacy threat: malicious agents that proactively engage others in multi-turn interactions to extract sensitive information. However, the evolving nature of such dynamic dialogues makes it challenging to anticipate emerging vulnerabilities and design effective defenses. To tackle this problem, we present a search-based framework that alternates between improving attack and defense strategies through the simulation of privacy-critical agent interactions. Specifically, we employ LLMs as optimizers to analyze simulation trajectories and iteratively propose new agent instructions. To explore the strategy space more efficiently, we further utilize parallel search with multiple threads and cross-thread propagation. Through this process, we find that attack strategies escalate from direct requests to sophisticated tactics, such as impersonation and consent forgery, while defenses evolve from simple rule-based constraints to robust identity-verification state machines. The discovered attacks and defenses transfer across diverse scenarios and backbone models, demonstrating strong practical utility for building privacy-aware agents.

As coding agents are increasingly deployed in large codebases, the need to automatically design challenging, codebase-level evaluation is central. We propose Gistify, a task where a coding LLM must create a single, minimal, self-contained file that can reproduce a specific functionality of a codebase. The coding LLM is given full access to a codebase along with a specific entrypoint (e.g., a python command), and the generated file must replicate the output of the same command ran under the full codebase, while containing only the essential components necessary to execute the provided command. Success on Gistify requires both structural understanding of the codebase, accurate modeling of its execution flow as well as the ability to produce potentially large code patches. Our findings show that current state-of-the-art models struggle to reliably solve Gistify tasks, especially ones with long executions traces.

Post-Training Multimodal Large Language Models (MLLMs) to build interactive agents holds promise across domains such as computer-use, web navigation, and robotics. A key challenge in scaling such post-training is lack of high-quality downstream agentic task datasets with tasks that are diverse, feasible, and verifiable. Existing approaches for task generation rely heavily on human annotation or prompting MLLM with limited downstream environment information, which is either costly or poorly scalable as it yield tasks with limited coverage. To remedy this, we present AutoPlay, a scalable pipeline for task generation that explicitly explores interactive environments to discover possible interactions and current state information to synthesize environment-grounded tasks. AutoPlay operates in two stages: (i) an exploration phase, where an MLLM explorer agent systematically uncovers novel environment states and functionalities, and (ii) a task generation phase, where a task generator leverages exploration trajectories and a set of task guideline prompts as context to synthesize diverse, executable, and verifiable tasks. We show AutoPlay generates $20$k tasks across $20$ Android applications and $10$k tasks across 13 applications Ubuntu applications to train mobile-use and computer-use agents. AutoPlay generated tasks enable large-scale task demonstration synthesis without human annotation by employing an MLLM task executor and verifier. This data enables training MLLM-based UI agents that improve success rates up to $20.0\%$ on mobile-use and $10.9\%$ on computer-use scenarios. In addition, AutoPlay generated tasks combined with MLLM verifier-based rewards enable scaling reinforcement learning training of UI agents, leading to an additional $5.7\%$ gain. coverage. These results establish AutoPlay as a scalable approach for post-training capable MLLM agents reducing reliance on human annotation.

The efficiency of GPU kernels is central to the progress of modern AI, yet optimizing them remains a difficult and labor-intensive task due to complex interactions between memory hierarchies, thread scheduling, and hardware-specific characteristics. While recent advances in large language models (LLMs) provide new opportunities for automated code generation, existing approaches largely treat LLMs as single-shot generators or naive refinement tools, limiting their effectiveness in navigating the irregular kernel optimization landscape. We introduce an LLM agentic framework for GPU kernel optimization that systematically explores the design space through multi-agent collaboration, grounded instruction, dynamic context management, and strategic search. This framework mimics the workflow of expert engineers, enabling LLMs to reason about hardware trade-offs, incorporate profiling feedback, and refine kernels iteratively. We evaluate our approach on KernelBench, a benchmark for LLM-based kernel optimization, and demonstrate substantial improvements over baseline agents: our system produces correct solutions where baselines often fail, and achieves kernels with up to 16$\times$ faster runtime performance. These results highlight the potential of agentic LLM frameworks to advance fully automated, scalable GPU kernel optimization.

Web applications are prime targets for cyberattacks due to their role as entry points to vital services and sensitive data repositories. Traditional penetration testing is expensive and requires specialized expertise, creating scalability challenges for securing the expanding web ecosystem. While language model agents have shown promise in certain cybersecurity tasks, modern web applications require visual understanding of complex user interfaces, dynamic content rendering, and multi-step interactive workflows that only computer-use agents (CUAs) can handle. Despite CUAs' demonstrated capabilities in web browsing and visual task automation, their potential to discover and exploit web application vulnerabilities through graphical interfaces remains unknown. We introduce HackWorld, the first evaluation framework for systematically assessing CUAs' capabilities in exploiting web application vulnerabilities through visual interaction. Unlike existing benchmarks using sanitized environments, HackWorld exposes CUAs to 36 curated applications spanning 11 frameworks and 7 languages, containing realistic vulnerabilities including injection flaws, authentication bypasses, and unsafe input handling. Our framework directly evaluates CUAs' ability to discover and exploit these vulnerabilities using Capture-the-Flag (CTF) methodology while navigating complex web interfaces. Evaluation of state-of-the-art CUAs reveals exploitation rates below 12%, struggling to plan multi-step attacks and use security tools effectively. Our results expose CUAs' limited cybersecurity skills when operating on vulnerable web applications, opening future research directions on developing security-aware CUAs for vulnerability detection and exploitation.

Large language models have revolutionized natural language processing, yet their deployment remains hampered by the substantial memory and runtime overhead of the transformer’s Key-Value cache. To mitigate this, recent methods employ a scoring-aggregation framework to evict unimportant cache entries, based on the "stability assumption"—that a fixed subset of entries remains consistently important during generation. However, prior work has largely focused on refining importance indicators for scoring, while defaulting to mean aggregation due to a faithful trust in the stability assumption. In this work, we argue that this underlying assumption is inherently fragile, making mean aggregation highly vulnerable in extreme cases. To counter this, we propose a simple yet elegant defensive aggregation strategy: a two-step, linear-time approach that controls worst-case risk, thereby defending against extreme cases with negligible computational overhead. Embodying this strategy, we propose a novel cache eviction method, DefensiveKV and its extension, Layer-DefensiveKV, which incorporates layer-wise budget allocation. Across seven task domains (18 datasets), our methods reduce generation quality loss by 2.3× and 4.3× respectively, versus the strongest baseline under a 20\% cache size. These results set new performance benchmarks and pioneer a promising direction for optimizing cache eviction against underlying fragility through worst-case risk management.

Large language models (LLMs) have demonstrated remarkable performance on single-turn text-to-SQL tasks, but real-world database applications predominantly require multi-turn interactions to handle ambiguous queries, execution errors, and evolving user requirements. Existing multi-turn benchmarks fall short of capturing this complexity, either by treating conversation histories as static context or by limiting evaluation to narrow, read-only (SELECT-ONLY) operations, thereby failing to reflect the challenges encountered in production-grade database assistant. In this work, we introduce BIRD-INTERACT, a benchmark that restores this missing realism through: (1) a ***comprehensive interaction environment*** that couples each database with a hierarchical knowledge base, metadata files, and a function-driven user simulator, enabling models to solicit clarifications, retrieve knowledge, and recover from execution errors without human supervision; (2) two ***evaluation settings*** reflecting real-world interaction settings which contain a pre-defined conversational protocol (c-Interact) and a more open-ended agentic setting (a-Interact) in which the model autonomously decides when to query the user simulator or explore the DB environment; (3) a ***challenging task suite*** that covers the full CRUD spectrum for both business-intelligence and operational use cases, guarded by executable test cases. Each task features ambiguous and follow-up sub-tasks, requiring LLMs to engage in dynamic interaction. The suite is organized into two sets: a full set (**BIRD-INTERACT-FULL**) of 600 tasks which unfold up to **11,796** dynamic interactions for a comprehensive overview of performance and a lite set (**BIRD-INTERACT-LITE**) of 300 tasks, with simplified databases for detailed behavioral analysis of interactions, and fast development of methods. Our empirical results highlight the difficulty of BIRD-INTERACT: the most recent flagship model GPT-5 completes only **8.67%** of tasks in the c-Interact setting and **17.00%** in the a-Interact setting on the full task suite. Further analysis via memory grafting and Interaction Test-time Scaling (ITS), validate the importance of effective interaction for achieving success in complex, dynamic text-to-SQL tasks.

Multimodal Misinformation Detection (MMD) in realistic, mixed-sourced scenarios must incorporate robust reasoning capabilities to handle the social complexity and diverse types of forgeries. While MLLM-based agents are increasingly used for MMD task due to their powerful reasoning abilities, they suffer from a critical trade-off: on one hand, single-agent methods provide only the limited, single-view analysis; on the other hand, multi-agent methods introduce high computational costs and significant optimization difficulties. To address this gap, we propose a novel Multi-Social-Agent Self-Distillation framework that internalizes collective social reasoning capabilities into a unified model. Our framework consists of two core stages: (1) we simulate multi-perspective judgments from a diverse society of MLLM agents and synthesize their collective feedback into high-quality Social Chain-of-Thought (SCoT) data; (2) Building on this, we propose the Social Correction Value-Driven Preference Optimization (SCPO), a new alignment algorithm that leverages the degree of social misjudgment as a verifiable signal to dynamically focus training on the most challenging samples. Extensive experiments on the challenging MFC-Bench and MMFakeBench benchmarks demonstrate the effectiveness of our framework. Our 7B Qwen2-VL-based model significantly outperforms various MLLM baselines, multi-agent methods, and even competes or surpasses proprietary models like GPT-4o and Claude, facilitating advanced multimodal misinformation reasoning and detection via thinking as society.

Mobile GUI agents are becoming critical tools to improve user experience on smart devices, with multimodal large language models (MLLMs) emerging as the dominant paradigms in this domain. Current agents, however, rely on explicit human instructions, overlooking the potential to leverage the contextual information (like location, time, user profile) and historical data for proactive task suggestions. Besides, previous works focus on optimizing the success rate during task execution, but pay less attention to the personalized execution trajectory, thereby neglecting potentially vast differences in user preferences. To address these challenges, we introduce the FingerTip 20K benchmark. We collected 20K unique human demonstrations of multi-step Android device interactions across a variety of everyday apps. These demonstrations are not isolated but are continuously acquired from the users' long-term usage in their real lives, and encompass essential user-related contextual information. The benchmark contains two new tracks: proactive task suggestions by analyzing environment observation and users' previous intents, and personalized task execution by catering to users' action preferences. Our experiments reveal that the tracks we propose pose significant challenges for leveraging user-related information in GUI tasks. We also performed a human study to show that there exists a huge gap between existing agents and humans. The model fine-tuned with the data we collected effectively utilized user information and achieved good results, highlighting the potential of our approach in building more user-oriented mobile LLM agents. Our code is open-source at \url{https://anonymous.4open.science/r/FingerTip-57B8} for reproducibility.

Agents in the real world must make not only logical but also *timely* judgments. This requires continuous awareness of the dynamic environment: hazards emerge, opportunities arise, and other agents act, while the agent's reasoning is still unfolding. Despite advances in language model reasoning, existing approaches fail to account for this dynamic nature. We introduce *real-time reasoning* as a new problem formulation for agents in evolving environments and build **Real-Time Reasoning Gym** to demonstrate it. We study two paradigms for deploying language models in agents: (1) **Reactive agents**, which employ language models with *bounded reasoning computation for rapid responses*. (2) **Planning agents**, which allow *extended reasoning computation for complex problems*. Our experiments show that even state-of-the-art models struggle with making logical and timely judgments in either paradigm. To address this limitation, we propose **AgileThinker**, which simultaneously engages *both reasoning paradigms*. **AgileThinker** consistently outperforms agents engaging only one reasoning paradigm as task difficulty and time pressure rise, effectively balancing reasoning depth and response latency. Our work establishes real-time reasoning as a critical testbed for developing practical agents and provides a foundation for research in temporally constrained AI systems, highlighting a path toward real-time capable agents.

Large language model (LLM)–based multi-agent systems are challenging to debug because failures often arise from long, branching interaction traces. The prevailing practice is to leverage LLMs for log-based failure localization, attributing errors to a specific agent and step. However, this paradigm has two key limitations: (i) log-only debugging lacks validation, producing untested hypotheses, and (ii) single-step or single-agent attribution is often ill-posed, as we find that multiple distinct interventions can independently repair the failed task. To address the first limitation, we introduce DoVer, an intervention-driven debugging framework, which augments hypothesis generation with active verification through targeted interventions (e.g., editing messages, altering plans). For the second limitation, rather than evaluating on attribution accuracy, we focus on measuring whether the system resolves the failure or makes quantifiable progress toward task success, reflecting a more outcome-oriented view of debugging. On the datasets derived from GAIA and AssistantBench, DoVer flips 18–28% of failed trials into successes, achieves up to 16% milestone progress, and validates or refutes 30-60% of failure hypotheses. Our findings highlight intervention as a practical mechanism for improving reliability in agentic systems and open opportunities for more robust, scalable debugging methods for LLM-based multi-agent systems.

Improving power grid operations is essential for enhancing flexibility and accelerating grid decarbonization. Reinforcement learning (RL) has shown promise in this domain, most notably through the Learning to Run a Power Network competitions, but prior work has primarily focused on single-agent settings, neglecting the decentralized, multi-agent nature of grid control. We fill this gap with MARL2Grid, the first benchmark for multi-agent RL (MARL) in power grid operations, developed in collaboration with transmission system operators. Built on RTE France’s high-fidelity simulation platform, MARL2Grid supports decentralized control across substations and generators, with configurable agent scopes, observability settings, expert-informed heuristics, and safety-critical constraints. The benchmark includes a suite of realistic scenarios that expose key challenges, such as coordination under partial information, long-horizon objectives, and adherence to hard physical constraints. Empirical results show that current MARL methods struggle under these real-world conditions. By providing a standardized, extensible platform, MARL2Grid aims to advance the development of scalable, cooperative, and safe learning algorithms for power system operations.

Recent advances in large language models (LLMs) have led to significant progress in mitigating social biases at the individual model level. However, a core vulnerability persists: small, stochastic biases can be amplified through multi-step interaction, leading to skewed system-level outcomes. A promising, yet unverified, hypothesis is that the architectural diversity of multi-agent systems (MAS)—where LLM-based agents with different roles and perspectives interact—could naturally mitigate this amplification. In this work, we rigorously test this hypothesis and investigate the phenomenon of bias amplification in MAS across sensitive attributes, including gender, age, and race. We introduce Discrim-Eval-Open, an open-ended, multi-option benchmark designed to measure system-level bias and bypass the performative neutrality of modern LLMs. We further propose novel metrics, including an adaptation of the Gini coefficient, to quantify the extremity of system-wide outputs. Our experiments reveal that iterative bias amplification is a pervasive issue that is not solved by MAS architectures. This amplification persists across various configurations, spanning agent roles, communication topologies, iteration depths, and model types, even when individual agents exhibit minimal bias in isolation. Moreover, we observe a systemic tendency to favor younger age groups, females, and Black communities. Finally, we demonstrate that even the inclusion of objective, neutral inputs can exacerbate bias amplification, exposing a critical vulnerability in system-level robustness. These findings challenge the assumption that architectural complexity alone fosters equity, underscoring the urgent need to address the fundamental dynamics of bias amplification within LLM-based MAS.

Machine learning engineering (MLE) has a clear objective: Given an MLE task and a verifier (e.g., performance on some held-out data), what is the most effective way to utilize compute to achieve the best performance for the given task? Existing language model (LM) agents rely on prompting frontier LMs and accumulating experience non-parametrically by storing and retrieving experience through agent scaffolds and test-time compute. In this paper, we show that in environments such as MLE where a good verifier is available, adapting the LM parameters through gradient updates can be more effective in utilizing compute and agent’s experience. Specifically, we show that agents backed by weaker models that improve via reinforcement learning (RL) can eventually outperform agents backed by much larger, but static models for a given MLE task. We identify two major challenges with RL in this setting. First, actions can take a variable amount of time (e.g., executing code for different solutions), which leads to asynchronous policy gradient updates that favor faster but suboptimal solutions. We propose duration-aware gradient updates in a distributed asynchronous RL framework to amplify high-cost but high-reward actions. Second, using performance on the held-out data as a reward for MLE provides limited feedback. A program that’s nearly correct is treated the same as one that fails entirely (e.g., during data loading). We propose environment instrumentation to offer verifiable partial credit, using a separate, static language model to insert print statement to an existing program. Our experiments suggest that a small LM (Qwen2.5-3B) adapted with RL, when given enough compute, can solve an MLE task better than prompting a frontier model (Claude-3.5-Sonnet) with the state-of-the-art agent scaffold (AIDE) by an average of 22% across 12 Kaggle tasks.

While Language Models (LMs) have made significant progress in automating machine learning engineering (MLE), the acquisition of high-quality MLE training data is significantly constrained. Current MLE benchmarks suffer from low scalability and limited applicability because they rely on static, manually curated tasks that demand extensive time and manual effort to produce. We introduce MLE-Smith, a fully automated multi-agent pipeline, to transform raw datasets into competition-style MLE challenges through an efficient generate--verify--execute paradigm for scaling MLE tasks with verifiable quality, real-world usability and rich diversity. The proposed multi-agent pipeline in MLE-Smith drives structured task design and standardized refactoring, coupled with a hybrid verification mechanism that enforces strict structural rules and high-level semantic soundness. It further validates empirical solvability and real-world fidelity through interactive execution. We apply MLE-Smith to 224 of real-world datasets and generates 606 tasks spanning multiple categories, objectives, and modalities, demonstrating that MLE-Smith can work effectively across a wide range of real-world datasets. Evaluation on generated tasks shows that the performance of eight mainstream and cutting-edge LLMs on MLE-Smith tasks is strongly correlated with their performance on carefully human-designed tasks, highlighting the effectiveness of the MLE-Smith in scaling up MLE tasks while maintaining task quality.