RL for LLM Agents

RL-based post-training for reasoning models is increasingly bottlenecked by repeated fresh rollout generation, particularly in agentic settings where environment interaction dominates wall-clock cost. Replay can reduce this burden by reusing past trajectories, but existing methods typically embed it within larger training pipelines involving exploration, experience restructuring, or mixed-policy optimization. This makes replay's own contribution difficult to isolate. We ask a focused question: how far can principled replay selection alone go? We introduce Headroom-Drift Replay, a group-level replay control primitive for GRPO that separates reuse into two decisions. Headroom ranks stored groups by remaining learning value, while Drift gates them by compatibility with the current policy. The fresh on-policy stream remains unchanged, and the method adds no auxiliary generation or training machinery. Across mathematical reasoning, multimodal reasoning, and Agentic Search benchmarks, this single intervention outperforms naive replay and matches or exceeds broader replay methods on Avg Mean@32. In Agentic Search, where environment interaction dominates cost, it delivers comparable quality at materially lower wall-clock time.

Audio enhancement in real-world scenarios involves complex distortion couplings and requires personalized enhancement. Existing solutions struggle to address both simultaneously. To improve robustness and enable autonomous operation in such scenarios, we propose StrixAE, an agent based on a multimodal large language model (MLLM). StrixAE leverages the MLLM as a controller to coordinate multiple audio enhancement and personalization models. To further enhance system robustness, reduce artifacts, and improve generalization across diverse real-world scenarios, StrixAE is trained through a two-stage process: first, CoT supervised fine-tuning on AcoustBench to ground basic reasoning and tool invocation; second, Audio Perception Reinforcement Learning (APRL), a reward design specifically tailored for audio restoration pipelines that jointly optimizes format validity, structural coherence, and perceptual quality. Unlike generic RL fine-tuning, APRL introduces structured rewards that enforce executable pipelines and logical section ordering, enabling the agent to produce reliable, interpretable enhancement plans without hallucinated tools. Based on real-world test datasets, our proposed method outperforms most existing open-source and proprietary solutions, achieving state-of-the-art performance across multiple perceptual metrics and demonstrating strong generalization robustness.

Graph-based policy optimization improves credit assignment for long-horizon LLM agents by organizing rollout trajectories into state-transition graphs. However, existing methods construct graphs independently within each policy update, discarding transitions discovered by earlier policies and limiting advantage estimation to small, batch-local rollout groups. We propose \emph{Temporal Instance-Graph Policy Optimization} (TIGPO), which extends graph-based credit assignment across policy updates. TIGPO maintains a persistent transition graph for each task, allowing valid transitions discovered by different policy versions to jointly determine credit for current rollouts. To actively reconnect current exploration with historical experience, TIGPO allocates a fixed rollout budget between Exploration slots for ordinary task sampling and Revisit slots for delayed reattempts of previously explored tasks. For each revisit, TIGPO pairs the current rollout group with its corresponding earlier Exploration group to construct a cross-temporal reference. The enlarged reference is designed to stabilize relative advantage estimation under small rollout groups, while comparison on the same task directly captures policy improvement across training stages. Historical transitions and scores serve only as structural and detached statistical references and are never replayed in the policy loss. Experiments on ALFWorld and WebShop demonstrate that TIGPO consistently outperforms prior group-based and graph-based policy optimization methods.

Large language model (LLM) agents for long-horizon interactive tasks typically follow a ReAct-style protocol, issuing one primitive action per LLM round. While this enables frequent replanning, it is inefficient for long-horizon tasks where many rounds are spent on routine action sequences. A natural alternative is to let the agent emit variable-length action chunks. However, naively training such policies with standard reinforcement learning fails: the agent either collapses to single-action behavior or over-commits to excessively long sequences. Both failures share a common root cause: the inability to learn chunk boundaries. We propose SPACE, which addresses this challenge by distilling chunk-boundary supervision from trajectory-induced programmatic skills. We induce two-level programmatic skills from successful trajectories, where subskill boundaries serve as direct chunk-boundary supervision. This temporal structure is then distilled into a primitive-chunk policy via hybrid on-/off-policy optimization with chunk-aware credit assignment. Experiments on ALFWorld and ScienceWorld show that SPACE improves success rates by 7.0%-31.3% over the strongest baseline in each setting while reducing average LLM decision rounds by up to 78.9%.

Natural language is emerging as a primary feedback channel for improving language agents, capable of conveying intent, preferences, and causal structure in forms interpretable by both humans and modern language models. We call this paradigm Verbal Reinforcement Learning (VRL) and offer the first unified account of it. We organize the field around a single axis, \textit{when} verbal feedback takes effect in an agent's lifecycle and \textit{what} it modifies, yielding three pillars: (1) \textbf{Language as Grounding Signal}, where language defines the task itself by specifying goals, states, and reward structures; (2) \textbf{Language as Deliberative Feedback}, where natural language guides reasoning at test time without the need to update model parameters; (3) \textbf{Language as Learning Signal}, where language-based feedback shapes model parameters through training. Within each pillar, we synthesize representative work, distinguish key subcategories of approaches, and outline the distinct role language plays in shaping agent behavior. Together, this taxonomy shows how verbal reinforcement is reshaping agent development, while also defining the challenges and opportunities for building more capable and aligned agents.

Training open-ended agents via reinforcement learning (RL) is hindered by the lack of verifiable gold answers and scalable rubrics. Moreover, even near the model's capability boundary, long-horizon open-ended agentic tasks often yield brittle and unstable rewards, resulting in weak or noisy rollout contrast that obscures fine-grained optimization signals for group-based policy learning. To address these challenges, we propose ARISE-RL, a novel full-cycle self-evolution framework that couples a task/rubric Generator and a reasoning Solver through rubric-mediated co-evolution. The Generator grounds tool-related rubric criteria in real tool observations and is rewarded for producing valid, intermediate-difficulty tasks aligned with the Solver's evolving capability boundary. The Solver, in turn, learns from fine-grained rubric satisfaction signals through multi-step reasoning and tool use. We further introduce Reward-Gated Self-Evolution Distillation (RG-SED), which selectively distills a memory-augmented variant of the same policy back into itself only when the memory yields empirical reward improvement, thereby reducing distribution mismatch and avoiding blind imitation of noisy guidance. Finally, to support rigorous evaluation, we present ECR-Bench, an expert-calibrated rubric benchmark suite covering single-tool deep research and multi-tool travel planning. Extensive experiments demonstrate that ARISE-RL consistently achieves robust and stable overall state-of-the-art performance across all evaluated benchmarks.

Vision-and-Language Navigation (VLN) requires an agent to navigate through unseen 3D environments according to natural-language instructions. Explicit reasoning can improve instruction understanding and semantic grounding, but autoregressive generation at every step accumulates large decision-stage latency over multi-step navigation. We propose VerNav, a verifier-first framework for low-latency LLM-based VLN. The verifier reduces decision-stage latency by replacing per-step autoregressive generation with batched action verification, while an entropy-based adaptive generator is invoked only for uncertain decisions to produce compact state evidence. To further improve navigation performance with the verifier, we introduce a two-stage alignment scheme: (i) VPO improves local action-preference alignment in static verifier training, and (ii) step-level reinforcement fine-tuning provides dense progress rewards over multi-step navigation rollouts during dynamic task execution. Experiments on the Room-to-Room (R2R) benchmark show that the verifier-only decision path of VerNav achieves competitive navigation performance among representative LLM-based VLN agents while reducing average decision-stage LLM latency per step by more than $10\times$ compared with autoregressive methods.

While reinforcement learning has enabled LLM-based search agents to invoke external tools, existing methods train under fixed budgets and cannot adapt when constraints vary at deployment. We propose AnySearch, a framework that enables a single policy to perform budget-aware search under any budget constraint through a training scaffold and curriculum reinforcement learning. In the first phase, we train the agent with explicit budget state injection and structured reasoning prompts that guide efficient allocation under linearly decaying budgets. In the second phase, the scaffold is removed and the agent learns to operate autonomously under adaptively sampled budget constraints, matching inference conditions. Both phases are optimized with a composite reward that couples answer accuracy with budget efficiency through absolute and relative signals, where an adaptive weight amplifies the efficiency signal for high-accuracy queries and attenuates it for low-accuracy ones. Extensive experiments on seven general and multi-hop QA benchmarks show that our method outperforms baselines across all budget scales, generalizes to unseen constraints beyond the training range, and achieves superior tool productivity without excessive token overhead. Our code is available at https://github.com/xwsun01/AnySearch.

How do LLM agents come to both understand environments they act in and master tasks set within them? Through controlled experiments combining world-model training (next-state prediction) and policy training (reward maximization), we investigate this question. We dissect the resulting models through their additive parameter updates. Geometrically, we find effective world-model updates are low-rank and share an input-feature subspace with policy updates while writing to nearly orthogonal output directions, whether trained separately or sequentially. However, we find that, in projection interventions, the sequential update induces more robustness than separate policy RL when removing the world model's leading input directions, suggesting that it has learned alternative input pathways. Behaviorally, we find the sequentially trained agent explores a wider range of states and actions. Based on this, we ask: does policy training preserve world knowledge as well as it could? We probe this with training-free merging built on the geometrically motivated input basis plus an online world-model loss during policy RL, and show both improve over the untreated baseline. Our findings suggest world knowledge and task-directed ability can be learned in geometrically complementary forms, and that future post-training pipelines should consider how best to engineer the interface between them.

Flat retrieval-augmented generation treats a corpus as a bag of chunks, discarding document hierarchy and cross document structure. We introduce SearchWiki, a harness framework that synthesizes a corpus into a hierarchical, typed, navigable wiki and trains an agent, WikiResearcher-9B, to retrieve information through multi-turn tool use. The wiki organizes knowledge into three layers - document overviews, cross- document topic pages, and page-level source records; enabling progressive refinement of retrieval when initial lookup misses. We optimize the agent's navigation policy with on-policy reinforcement learning with a multi-component reward function balancing answer correctness, retrieval quality and trajectory efficiency. Evaluation on ViDoRe-V3 (8 domains), FinanceBench, and memory benchmarks (LoCoMo, LongMemEval, PersonaMem-v2) shows that WikiResearcher- 9B which is our RL-tuned Qwen 9B model, significantly outperforms same-size untrained baselines and exceeds or matches larger external models. SearchWiki paired with WikiResearcher-9B demonstrates that learned navigation over structured corpora is a superior alternative to flat retrieval.

Localizing issue-relevant code regions is a critical step in automated software engineering. However, due to their reliance on sparse trajectory-level signals, existing methods cannot identify which per-turn actions are effective and often discover correct code regions during exploration but fail to commit them. To address these limitations, we propose an action-aware reinforcement learning method that combines a per-turn reward sequence rewarding both the discovery and commitment of gold code regions with an action-level advantage estimation scheme that isolates each action's credit by grouping turns sharing the same exploration context. Extensive evaluations show that our method improves the average F1 over the state-of-the-art (SOTA) by 1.58% on SWE-Bench Verified and 8.55% on SWE-Bench Pro, with our 4B model outperforming baselines up to 8x larger. Our code is available at https://github.com/donian00/A2Agent.

Large language model agents increasingly solve long-horizon tasks through multi-agent harnesses in which a central agent coordinates specialized sub-agents, tools, and environments. Training the central policy in such a harness raises two challenges. First, an action label is a low-cardinality decision, whereas its args form a high-dimensional conditional sequence; optimizing both with a shared sequence-level signal can produce conflicting gradients. Second, dynamic scheduling creates interdependent sessions with branches, parallel calls, and rewritten contexts, which cannot be faithfully reduced to one flat token sequence. We introduce Harness-RL, a structured reinforcement learning framework that combines Conflict-Aware Policy Optimization (CAPO) with interface-level black-box trajectory construction. The black-box component captures Interface Call Records, builds per-session prefix trees, and aligns outcome and process rewards with trainable tokens. CAPO uses forward activations to identify parameter partitions associated with action and args tokens, then routes their policy gradients to the corresponding subspaces. Harness-RL supports both central-only and joint multi-agent training. Across seven multi-hop question answering and agentic retrieval benchmarks, it reaches average F1 scores of 42.93 and 47.79 with Qwen2.5-1.5B and Qwen2.5-3B, respectively, while ablations validate the contribution of CAPO and favor central-only optimization in the evaluated setting. Our code is available at https://github.com/jiangxinke/Harness-RL.

Retrieval-Augmented Generation (RAG) improves the factuality of large language models (LLMs), yet existing RAG systems often struggle with complex, multi-step reasoning that requires adaptive retrieval and continuous revision of intermediate contexts. Recent reinforcement learning (RL)-based agentic RAG methods partially alleviate this issue, but typically rely on coarse-grained action spaces and trajectory-level rewards, resulting in weak reward assignment and a bias toward short-horizon, stereotyped reasoning template. To address, we propose AgenticRag-R1, a RL framework that deeply integrates reasoning, retrieval, and memory via a memory stack and fine-grained action space, supported by hierarchical action-aware rewards and an information-aware trajectory rejection strategy to enable effective long-horizon learning. Experiments across a diverse set of multi-hop, open-domain, and agentic reasoning benchmarks, spanning multiple backbone model sizes, demonstrate that AgenticRag-R1 consistently outperforms strong baselines. Moreover, AgenticRag-R1 learns more robust, interpretable, and memory-aware reasoning behaviors, highlighting the effect of fine-grained action modeling and information-aware optimization for long-horizon reasoning. Our code is anonymous available at https://github.com/jiangxinke/Harness-RL/tree/AgenticRAG-R1-Whitebox.

Terminal agents are a compelling application of large language models (LLMs), with the potential to integrate deeply into users' daily workflows. Reinforcement learning (RL) is a key technique for improving their capabilities, making scalable training environments a central challenge. Since public real-user interaction data are scarce, synthetic environments provide a practical alternative, but often suffer from domain gaps and limited fidelity, leading to poor generalization. Existing work mainly scales the quantity and diversity of synthetic environments, while reward-signal quality and the mechanisms governing generalization remain under-explored. We study how RL improves terminal agents and propose the Agentic Compositional Generalization hypothesis: rather than teaching new domain-specific skills from scratch, RL primarily shapes high-level decision-making behaviors that compose and route low-level skills acquired during pre-training and supervised fine-tuning (SFT). This account is consistent with our empirical results and suggests that verifier quality, which determines which behaviors are reinforced, is more important than simply increasing environment quantity or diversity. Motivated by this insight, we propose River, a simple training recipe that improves reward quality by filtering low-quality environments and augmenting outcome rewards with process-level behavior regularization. Using this recipe, our RL-trained agent achieves the best performance among evaluated open-source RL-trained 8B models across four terminal-agent benchmarks. River also generalizes across model families, scales, agent harnesses, and RL objectives. Using fewer than 30% of the TMax training environments, River improves RL gains by 106% and 30% on average for models ranging from 2B to 27B on Terminal-Bench-Lite and Terminal-Bench-v2.1, respectively.

Tool-Integrated Reasoning (TIR) is a fundamental capability for LLM agents to solve complex tasks by interacting with external tools iteratively. Reinforcement Learning (RL) has become the dominant paradigm for enabling this capability. However, existing approaches typically assign uniform trajectory-level advantages and treat all correct tool calls equally, ignoring the varying difficulty and learning value across trajectories and reasoning steps. This can lead to imprecise learning signals that do not adequately distinguish between trivial and challenging tool-use patterns. To address this limitation, we propose HiDiffTIR, a Hierarchical Difficulty-aware policy optimization framework for multi-turn TIR. HiDiffTIR performs difficulty-aware credit assignment at both trajectory and turn levels, enabling the policy to focus on more informative trajectories and harder reasoning steps. Notably, this fine-grained optimization is achieved without additional supervision, relying solely on group-level statistics derived from standard RL rollouts. Extensive experiments on three tool-using benchmarks demonstrate that HiDiffTIR consistently improves multi-turn TIR performance and tool invocation accuracy over strong RL baselines, highlighting the necessity of difficulty-aware credit assignment for effective policy optimization in tool-integrated LLM agents.

Agentic reinforcement learning (RL) has emerged as an important post-training approach for enhancing the capabilities of Large Language Models (LLMs). However, existing methods face a trade-off between policy performance and resource efficiency. Conventional Proximal Policy Optimization (PPO) implementations incur substantial memory overhead from a separate critic, whereas critic-free group-relative methods require multiple rollouts and face potential learning bottlenecks on long-horizon tasks. In this work, we propose Single-rollout Autoregressive Policy Optimization (SAPO), an efficient PPO-style framework that unifies policy optimization and value learning within a single causal language model. SAPO exploits the autoregressive structure of LLMs to sequentially generate action and value estimation at distinct causal boundaries with shared parameters, and then jointly optimizes the PPO objectives and an auxiliary on-policy SARSA objective with turn-level generalized advantage estimation, where the latter is designed to facilitate value learning. Extensive experiments on ALFWorld and WebShop with Qwen2.5-1.5B/7B and Qwen3-14B demonstrate that SAPO reduces peak GPU memory usage by 23.1% and per-iteration runtime by 24.8% over strong PPO baseline, while matching or slightly improving task success rate. Our experiments also show that SAPO outperforms Group Relative Policy Optimization (GRPO) and recent cutting-edge variants in both task success and training stability.

Continuous self-improvement requires an ever-expanding pool of self-generated, diverse, adaptive goals. For language agents, existing training environment pools (hand-curated, statically synthesized, or frozen-verifier) keep the goal distribution fixed as the learner scales. We introduce SPADE (Self-Play in Adaptive Synthetic Executable Environments), a self-play RL framework in which a single LLM plays two roles: an Environment Designer that writes complete, long-horizon training environments as executable code with an OpenAI Gym-style reset()/step() interface, and a Reasoning Agent that learns to act in them. Each is a stateful, multi-turn environment (state transitions, reward functions, and verification code), so one interface spans reasoning problems and multi-step agentic tool use. The Reasoning Agent's regret is estimated using the gap between its reward with and without privileged hints; in optimizing this regret signal the Environment Designer learns to target environments at the edge of the agent's capabilities while keeping them feasible. Through extensive experimentation, we find several components critical to success: grounding the Environment Designer on documents sampled from a large pretraining corpus, and giving it an accumulated environment memory. Scaling to 30B-parameter models, SPADE improves over the strongest fixed-environment baseline by +5.3 on average across eight held-out math, science, code, and reasoning benchmarks, and lifts the tool-use setting by +5.7 on BFCL-v4 multi-turn and +13.9 on ACEBench-Agent; on the games setting, the margin over the strongest baseline grows with model scale. By making environment design itself a learnable component, SPADE takes a concrete step toward open-ended self-improvement.

AI-powered digital avatar streamers must answer product questions, engage viewers, and execute marketing strategies in real time, demanding low latency, frequent strategy updates, and accurate yet effective responses. Evolvable Harnesses, whose Skills, Hooks, prompts, and tools can be updated independently of model weights, enable rapid iteration but expose a trade-off: large models adapt zero-shot yet are too slow, whereas compact models meet latency targets but overfit to fixed Harness configurations. We propose Harness-Aware Training (HAT), which trains compact models to adapt to changing Harnesses. Its key component, Harness-State Augmentation (HSA), applies task-preserving transformations to Skill identifiers and content, tool schemas, prompt structures, and Hook functions. Training proceeds in three stages: HSA-SFT learns reasoning and tool use from strong-model trajectories across diverse environments; General On-Policy Distillation restores generalization lost during SFT; and HSA-RL improves robustness to changing Harnesses through reinforcement learning in augmented environments. Across four evaluation sets, HAT achieves 94.8 on Live-Stream QA (base: 80.3; strongest general LLM: 93.0) and 94.6 on Harness-Variant QA (base: 75.4). Unlike Fixed-Harness SFT, which lowers IFEval by 7.7 points from the base model, HAT avoids this regression and reaches 83.5. On one NVIDIA H20 GPU, the optimized system delivers P50 and P95 latencies of 3.4 s and 8.1 s. Deployed in Taobao Live's digital-avatar service, it also yields positive online A/B test results for item-page views.

Deep search agents operate over trajectories spanning dozens of steps, yet standard reinforcement learning provides only a single outcome reward per trajectory, which is far too sparse for effective credit assignment. On-policy self-distillation (OPSD) addresses this by using the model's own logits as dense token-level teachers, but extending it to search agents introduces a fundamental tension: the teacher, having access to privileged information such as the correct answer, produces a distribution that differs systematically from the student's exploration-based reasoning, and naive distillation causes the student to inherit this information asymmetry rather than learn better search strategies. We resolve this tension through two contributions. First, we construct Evidence Anchors, which are concise, step-level evidence snippets extracted from the web, as privileged information that captures key reasoning steps without revealing the entire answer path. Second, we propose Step-Level Self-Distilled Policy Optimization (SSPO), which converts teacher-student disagreement into step-level advantage weights within GRPO, applied exclusively to incorrect trajectories. This design decouples what to update from how much to update: the outcome reward determines the direction of policy change, while the teacher modulates its magnitude at each step. Correct trajectories are left untouched, preserving their diversity. On Qwen3-8B, SSPO consistently outperforms GRPO across BrowseComp, GAIA, and FRAMES, surpassing or matching GRPO trained with twice as many gradient steps while adding only about 5 percent overhead per step from a single additional forward pass.

Automating empirical research is a long-standing direction of AI. Recent automatic research (AutoResearch) agents bring this goal within reach, as modern LLMs show the capability to independently implement solutions and learn from the execution outcomes. Behind these gains, post-training (especially RL) plays a central role. In this paper, we identify a fundamental tension when scaling RL for these agents: the two components of every AutoResearch trajectory (agent generation and environment execution) scale in very different manners, since all generation shares compute through batching, while each execution occupies its exclusive sandbox and real machine time. As a result, the environment execution dominates the training cost and becomes the bottleneck as trajectories grow. To resolve this tension, we propose World Model RL (WMRL), which replaces environment execution with a world model to remove this bottleneck. Additionally, the world model can be imperfect, as its rewards are corrupted by bias and noise. Therefore, we further equip WMRL with two mitigations, Online Debiasing and Inverse-Variance Denoising, which offset the bias and suppress the noise respectively. Theoretically, we prove that both mitigations of WMRL strictly improve the convergence guarantee. Empirically, WMRL accelerates training by 3-4x on various tasks at different agent scales, while exceeding the performance of standard RL baselines. Moreover, our post-trained 4B and 9B agents outperform much larger open-weight agents of 48B and 120B on held-out benchmarks. Beyond AutoResearch, WMRL also transfers to post-training embodied VLA policies, which demonstrates the generalizability of our method.

Large language model agents increasingly rely on long-horizon reasoning to solve complex tasks involving planning, tool use, and memory. A critical capability in such settings is reflection: assessing trajectory progress, identifying missing evidence and unreliable intermediate states, and deciding whether to continue, revise, or abandon the current branch. Learning effective reflection, however, is challenging because reflection is performed locally within the current branch, whereas its utility can only be determined by its contribution to the final trajectory outcome. This local-global mismatch makes outcome-based reinforcement learning provide only local, sparse and delayed supervision for reflective decisions. To solve these, we propose LoongReflect, a training framework that formulates reflection as a memory-control policy. The agent operates over a reversible trajectory tree using explicit reflect and backtrack actions. Reflection consolidates verified facts, missing evidence, and branch-specific risks into working memory, while backtracking removes an unreliable branch from the active context and preserves a concise corrective lesson. To learn this policy, LoongReflect combines two complementary signals through a look-ahead, extragradient-style coordination mechanism. A fast channel distills globally informed reflective behavior from a privileged teacher, with supervision restricted to reflection and backtracking tokens. A slow channel optimizes complete trajectories using outcome-based GRPO, aligning local control decisions with final task success. Experiments on multi-hop retrieval-augmented generation and mathematical reasoning benchmarks demonstrate consistent improvements over outcome-only reinforcement learning and self-distillation baselines.

Large language model-based shopping agents are increasingly deployed in real-world e-commerce platforms, generating massive amounts of user interaction logs that provide valuable supervision for improving these agents. However, existing approaches primarily rely on offline training signals, such as user-item interactions or synthetic preference data, while largely overlooking the rich supervision contained in users' natural conversational feedback. Moreover, the available online feedback is heterogeneous, sparse, and noisy, making it difficult to transform into reliable learning signals automatically. To address these challenges, we propose LOFA, a framework that enables shopping agents to learn directly from real online interaction logs without human annotation. LOFA combines reinforcement learning over verifiable purchase outcomes with feedback-aware on-policy distillation, which identifies users'in-dialogue directives and converts them into dense token-level supervision. These complementary objectives capture both collaborative behavioral patterns and user-specific preferences. Extensive experiments on real-world e-commerce logs demonstrate that LOFA consistently improves recommendation quality, response helpfulness, and user-satisfaction alignment over strong baselines, highlighting the effectiveness of learning shopping agents from real online user feedback.

We study whether the localized numerical operations and integrative judgments of financial analysis benefit from the same form of LLM specialization. Larix maps a 16-lens European listed-real-estate analysis framework to eight lens-aligned specialists; we compare a frontier LLM under monolithic versus specialist-decomposed prompting while holding the model, source evidence, task instructions, output schema, and scoring fixed. Across 19 firms spanning seven regulatory wrappers, decomposition improves the numerical-task aggregate by 15.8 percentage points but does not reliably improve, and can reduce, performance on judgment tasks, a pattern stable across four frozen-template dispatches; a single-agent control given the complete framework does not reproduce the numerical gain. Post-training Qwen3.5-9B with GRPO using task-aligned structured rewards then raises the development-split score by 12.0 points and the judgment aggregate by 14.2 points, with gains on all four sub-ceiling tasks; the gains transfer to unseen firms (+15.2 points overall; +40.4 on covenant stress) and to unseen regulatory wrappers (+4.3), with positive transfer on all three anti-memorization splits. Prompt-level decomposition thus improves modular numerical execution, whereas targeted parameter adaptation improves integrative financial judgment.

Agent skills represent a standardized format for packaging procedural knowledge and domain expertise, serving within agent harness systems as an essential mechanism to continually constrain a language model's behavior space for repeatable, high-quality task execution. However, because strong closed-source models entail high inference costs, current popular agent harnesses, such as Codex and OpenClaw, remain prohibitively expensive when deploying these skills to accomplish real-world tasks. The rapid capability enhancement of open-source models deployable on consumer-grade GPUs presents a compelling opportunity to drastically reduce these costs by leveraging skill-based behavioral constraints. Nevertheless, automatically generating effective skills tailored specifically for such compact models remains a significant practical challenge. To address this, we propose SKILLER, a natural-language-driven reinforcement learning framework designed to automatically generate executor-specific skills for small models, which employs a strong model as the actor and critic, treats the small-model agent system as the environment, and propagates all reinforcement learning signals entirely via natural language. Extensive experimental evaluations across five relevant benchmarks using Qwen3.5-9B and Qwen3.5-4B demonstrate that SKILLER outperforms three open-source and one closed-source skill generation or evolution methods, achieving absolute gains ranging from 4.3 to 20.4 percentage points for the 9B model and 1.8 to 13.3 points for the 4B model, while remarkably matching the performance of strong closed-source models on single-skill tasks in SkillsBench. The project is available at https://github.com/DANG-ai/SKILLER.

Reinforcement learning (RL) for large language models is moving toward multi-turn agentic workloads, where rollout tasks repeatedly pause for external environments, resume with growing contexts, and finish at highly variable times. In this setting, RL training goodput, measured by training throughput, matters more than raw GPU occupancy: GPU waiting and repeated prefill recomputation are pure overhead. We present TideRL, a readiness-aware elastic RL system with Continuous Task Batching, Resource-Aware Ref-Actor Pipelining, and Elastic Resource Scaling. CTB preserves useful rollout state, $\textrm{RA}^2\textrm{P}$ selects between decoupled streaming and colocated aggregation from the ready backlog and arrival interval, and ERS moves ranks between rollout and training using the same readiness signals. Across text-only and multi-modal agentic workloads, TideRL improves RL training goodput by up to 5.6$\times$ over synchronous baselines and over 33% over asynchronous baselines, while reaching similar task performance. It also improves KV cache hit rate by 1.58$\times$, reduces per-step training time by up to 44.3%, and cuts total waiting time by up to 77.6%.

Long-horizon tool-using agents must reason over user goals, domain policies, tool calls, simulator state, and delayed verifiable rewards. Reinforcement learning (RL) is a natural fit for this setting, but multi-turn on-policy rollouts create long contexts, while model-specific attention layers may require custom masks and learned sink normalization. We present SINKFLEX-RL, a modular training system for RL in dual-control tool-use environments. The system combines a Gymnasium-compatible environment wrapper, a VERL-style rollout dataflow, group-relative policy optimization without a separate value model, and a sink-aware FlexAttention path designed to preserve model-specific sink scaling under causal and sliding-window masks. In a preliminary Tau2Bench retail run, validation reward (mean@1) rises from 0.25 early in training to $0.44$ later in the observed training window, while training-score and trajectory-reward proxies also trend upward. In a fixed-configuration memory benchmark, the optimized attention path reduces peak VRAM from 28.06GB to 22.52GB at 4096 tokens, a $19.7\%$ reduction, and runs the measured 8192-token configuration using $25.53$~GB where the eager baseline runs out of memory. These results illustrate the value of integrating environment interfaces, RL dataflow, and attention-kernel design for memory-feasible long-horizon agent training.

External natural-language skills provide large language model (LLM) agents with reusable and editable guidance for solving complex tasks. Yet their effectiveness depends not only on skill quality, but also on whether the policy can translate the provided guidance into appropriate actions. However, methods specifically designed to improve this skill-utilization ability remain largely underexplored. In practice, skill-based agents are commonly trained with reinforcement learning objectives centered on task-level rewards, which offer limited supervision and struggle to capture subtle differences in how effectively the policy uses the provided skills. We propose BCSD (Bidirectional Context Self-Distillation), a framework that combines self-distillation with reinforcement learning to train LLM agents to use external skills more effectively. Unlike prior self-distillation methods that rely on a single privileged context, BCSD evaluates each trajectory from two complementary skill-context views. The augmented view introduces higher-level Meta-Skill guidance, while the reduced view prunes general guidance to highlight task-specific skills. Their complementary token-level signals are combined to rescale the RL advantage. Experiments on ALFWorld and WebShop demonstrate that BCSD achieves the strongest overall performance across model scales, enabling agents to utilize external skills more effectively. Ablation studies further verify the complementary contributions of the augmented and reduced context views. Code will be released to ensure full reproducibility.

Large language model-based multi-agent systems have recently shown strong potential for complex, long-horizon tasks. However, existing methods mainly rely on coarse prompt-level differentiation without parameter adaptation for diverse subtasks, resulting in insufficient inter-agent heterogeneity and limited specialized capability that bottleneck performance on tasks with complex requirements. To address this, we introduce a Task-Oriented Multi-Agent System with Mixture of Role-Subtask Experts (MoRSE) that distinguishes agents with (role, subtask)-conditional specialization at both the task structure and parameter levels. To make agents' responsibility explicit at the task structure level, we formulate a task-oriented multi-agent system that decomposes each task into a dependency-aware Directed Acyclic Graph of subtasks and assigns each agent a specific (role, subtask), introducing task-level specialization across collaborating agents. Additionally, to address the diverse role and subtask parameter adaptation demands, we propose a dynamic Mixture of (role, subtask) LoRA Experts module with a prototype-based semantic router for subtasks, augmenting agents with parameter-level specialization on a shared LLM substrate cost-effectively. Then, to co-optimize experts and router stably under sparse task rewards, we further propose a hierarchical group-relative policy optimization with two-layer credit assignment that isolates expert updates from the cross-route variance introduced by routing decisions, disentangling expert quality from routing quality. Experiments on code-generation benchmarks across three backbones demonstrate the effectiveness of our approach, with improvements in both whole-task and step-wise performance, and the gains from trained specialization generalize across held-out task categories and domains.

Multi-step language-model agents repeatedly process growing interaction histories, leading to substantial context costs. Vision--text compression reduces these costs by rendering history as images, but the resulting modality shift creates a marked capability gap. Through controlled evaluations of history recovery, matched-state decisions, and complete trajectories, we show that this gap cannot be explained by OCR quality alone. Visual-history agents exhibit systematic drift in action selection, query formulation, stopping, and evidence use, revealing an agentic policy gap. We introduce \textbf{CAPS}, a two-stage \textbf{C}ross-modal \textbf{A}gentic \textbf{P}olicy \textbf{S}elf-distillation framework that uses the same model's stronger text-history policy to supervise its visual-history counterpart. Offline trajectory self-distillation transfers successful text-policy behavior to visual-history inputs, while online policy self-distillation provides dense supervision on states visited by the visual-history policy during reinforcement learning. On SearchQA, CAPS improves over AgentOCR by 5.0% and 3.4% with 3B and 7B backbones, respectively. On full-history ALFWorld, the corresponding gains are 15.6% and 14.5%. Across settings, CAPS reduces average memory-context cost by up to 63.3% and peak cost by up to 83.4% relative to matched text-history policies. These results show that explicit cross-modal policy self-distillation can preserve agent capability under vision--text compression. Our code will be made publicly available in a future release.

Embodied agents using LLM-based planners often struggle with physical hallucinations, poor generalization to long-horizon tasks, and lack of environmental awareness. We propose GraphThink, a novel framework that integrates a task graph to provide structured knowledge for robust planning and a scene graph to maintain environmental memory for event-driven replanning. Specifically, the task graph guides LLM thinking through contextual prompting and iterative refinement, effectively mitigating planning hallucinations. Furthermore, within the GRPO framework, the task graph offers delicate reward design to train the LLM planner, enhancing long-horizon planning capabilities and improving generalization. Finally, an event-driven replanning module, powered by the scene graph, enables closed-loop environment awareness and error correction. GraphThink achieves state-of-the-art performance on the ALFRED benchmark. In particular, our high-level planner surpasses leading API-based LLMs on both the validation set and held-out long-horizon tasks, underscoring its robust zero-shot and few-shot capabilities. Additional evaluations further demonstrate strong out-of-distribution generalization to novel tasks and environments.