RL for LLM Agents

Structured skill prompts improve exploration in long-horizon agentic reinforcement learning (RL). Skill-augmented RL methods retain external skills at inference, while skill-internalization RL methods withdraw them during training to enable autonomous performance. However, existing internalization approaches only use skill-helpfulness contrast for curriculum control, leaving the policy update unchanged and unable to distinguish skill-dependent from autonomous success. We propose SkillC, a framework based on Contrastive Skill Credit Assignment (CSCA) that converts this contrast into a direct learning signal for internalization. SkillC samples paired skill-injected and skill-free rollouts for tasks from active skill types within the same policy update, and injects their task-level contrast into optimization via a dual-stream advantage estimator that preserves global ranking while applying a one-sided correction toward skill-free success. A smoothed validation-level signal further drives an adaptive curriculum over attribution strength, rollout allocation, and monotonic active-set pruning. Experiments on ALFWorld and WebShop show that, without runtime skill access, SkillC surpasses the strongest prior skill-internalization RL baseline by 5.5% and 4.4%, respectively, while remaining competitive with skill-augmented RL methods.

Search agents powered by large language models can autonomously decompose queries, retrieve information, and synthesize answers through multi-step reasoning. However, the rapid growth of training methods has outpaced controlled comparison: existing works differ in retrieval corpora, reward designs, and training protocols, making it unclear what actually drives improvements. We present a controlled empirical study that isolates three under-explored dimensions of search agent training. First, we identify a critical data-coverage issue in the widely used Wikipedia 2018 corpus and show that correcting it alone yields larger gains than the differences between training algorithms. Second, we systematically compare outcome-based and process-based reward methods across three base models, finding that the simplest outcome-based approach achieves competitive or superior performance in most settings, and that process-level credit assignment can over-correct agent behavior. Third, we analyze training data diversity, off-policy data utilization, and search budget scaling, distilling practical guidelines for training effective search agents. Our code is available at https://github.com/YiboZhao624/SearchAgentReview.

Humans know when to reach for help e.g. $347 \times 28$ warrants a calculator while $2+2$ does not. Language models do not. Prompt-based approaches can instruct a model when to invoke tools, but this scaffolding does not teach it to recognize the boundary of its own knowledge. RL approaches that assign a single outcome reward to the whole trajectory fare no better: trajectory-level credit cannot isolate which tool call in a successful episode actually helped, nor penalize unnecessary calls. We propose CARL (Competence-Aware Reinforcement Learning), which trains a critic on the model's own rollouts to learn where parametric knowledge suffices and where it needs external help. By decomposing each rollout at natural tool-use boundaries (e.g., code fence delimiters and context block transitions), CARL assigns independent credit to each segment from a single binary outcome, without external judges or step-level annotations. As a result, erroneous tool calls, incorrect extractions, and unnecessary calls each receive appropriately signed advantages. The trained critic captures the model's domain competence: it separates parametrically solvable from tool-dependent questions with AUC 0.93 at 7B. On five benchmarks spanning arithmetic, multi-hop factual QA, and numerical reasoning over financial tables, CARL improves exact-match accuracy by 6.7 points at 7B and 9.7 points at 3B over the best RL baseline, with the largest gain (+8.3 EM at 7B, +9.0 EM at 3B) on Musique. The model issues 53% fewer tool calls on parametrically answerable questions while remaining ${\sim}10$ EM points more accurate on them. Gains are largest at small scale: the 3B improvement is $1.4\times$ the 7B improvement, suggesting that knowing when to ask disproportionately benefits models with smaller parametric memory.

Reinforcement learning for multi-turn agents suffers from a credit-assignment mismatch: rewards are sparse and trajectory-level, while success often hinges on a few local decisions. Existing online policy distillation (OPD) provides denser token-level supervision, but typically treats heterogeneous agent trajectories as monolithic strings rather than causal interaction units. We present StepOPSD, a post-rollout preference self-distillation framework that takes the agent step as the unit of credit redistribution. StepOPSD decomposes trajectories into action-centered step segments, rescoring them under hindsight-enriched teacher contexts and converting token-level log-probability gaps into sign-preserving advantage shaping with a normalized per-step credit budget before the GRPO update. Across ALFWorld and Search-QA with Qwen3-1.7B and Qwen2.5-3B-Instruct, StepOPSD attains best or second-best results on subsets most sensitive to local causal errors, including first-place performance on ALFWorld Heat (79.1%), PickTwo (95.0%), Search-QA TriviaQA (61.6%), and tied-best performance on HotpotQA (40.4%). The results further reveal a consistent two-knob law: smaller α_clip acts as a broadly stabilizing local trust region, whereas the optimal global mixing strength λ_mix remains task-dependent. These findings suggest that step-aware distillation is most useful when trajectory-level rewards are weakly aligned with the local action that determines downstream success.

Post-trained LLMs are often optimized to produce helpful, polite, and accommodating responses. In adversarial negotiation, however, such behavior can become a vulnerability: emotionally framed language may influence an agent's bargaining decisions in ways that conflict with its user's objectives. We therefore introduce EmoDistill, an offline framework for distilling emotional negotiation skills from LLM-LLM interactions into smaller language-model agents. Here, an emotional negotiation skill is a state-conditioned behavior that determines which explicit emotion to invoke in a bargaining state and how to realize that emotion as an effective negotiation utterance. EmoDistill learns these two components separately: an Implicit Q-Learning (IQL) selector learns which emotion to express in each bargaining state, while a LoRA-adapted 7B policy learns emotion-conditioned expression through Supervised Fine-Tuning (SFT) and Judge Policy Optimization (JPO). Across four emotion-sensitive negotiation domains, the full EmoDistill policy achieves competitive utility and improves over vanilla and IQL-only baselines in most settings. Emotion-free ablations show that removing the explicit emotion channel substantially reduces overall negotiation utility, while transfer experiments reveal partial, domain-dependent transfer and robustness to unseen LLM counterparties.

Group-based reinforcement learning (RL) methods have achieved remarkable success in improving the performance of large language models (LLMs) and have been rapidly extended to agentic tasks. However, their credit assignment relies heavily on coarse-grained trajectory-level attribution according to final outcomes, making it difficult to capture the contribution of individual steps, such as valuable steps obscured within failed trajectories. To uncover latent information and enable more faithful step-level credit assignment, we propose Graph-based Group Policy Optimization (GraphGPO), which first aggregates all rollout trajectories into a unified state-transition graph and then estimates the distance from each state to the task goal using the global information encoded in the graph. Finally, GraphGPO assigns credit to each edge by estimating a graph-based advantage, based on how much the transition reduces the distance to the task goal. In this way, GraphGPO significantly improves training efficiency and achieves state-of-the-art performance across a range of challenging benchmarks.

LLM-based multi-agent systems decompose complex tasks into interacting roles, but most remain manually orchestrated by prompts, tools, and control rules, while agents are rarely optimized through a unified reinforcement learning interface. Existing RL post-training frameworks mainly target single-policy optimization and lack abstractions for user-defined multi-agent workflows, structured interaction, role-specific credit assignment, and configurable parameter sharing. We present UnityMAS-O, a general RL optimization framework for LLM-based multi-agent systems. UnityMAS-O treats the complete workflow as the optimization unit, rather than a single response or policy trajectory. It represents workflows through four first-class objects: logical agent roles, graph trajectories, user-defined rewards, and agent--model mappings. This decouples logical agents from physical model parameters, supporting full sharing, full separation, and partial sharing, with rewards assigned at role, turn, and trajectory levels. UnityMAS-O extends verl with a Ray-based star-topology runtime. A central controller executes workflows, invokes tools, records structured trajectories, and assembles rewards; model-local worker groups handle rollout, buffering, advantage computation, and distributed PPO-style updates. Users can define agents, workflows, model mappings, and rewards without rewriting the optimization infrastructure. We instantiate UnityMAS-O on retrieval-augmented QA, iterative agentic search, and reflective code generation. Across Natural Questions, HotpotQA, and held-out code tasks, multi-agent RL improves manually specified workflows after optimization, with especially large gains for smaller models and strict code all-passed metrics. These results show that UnityMAS-O can serve as a reusable substrate for converting diverse LLM-based multi-agent workflows into trainable multi-agent RL systems.

Medical RAG needs evidence-grounded claims, so plugging a claim-level NLI checker into retrieval-augmented RL is intuitive. We find that the checker's output distribution during training, not its held-out accuracy, decides whether it provides trainable gradient. We compare four NLI checker back-ends as process rewards inside a GRPO-trained medical RAG agent (Qwen2.5-7B, replicated on Qwen3-4B and Llama-3.1-8B) across four held-out medical QA benchmarks. Three diagnostic findings emerge. (i) Signal collapse is log-prob-specific: LLM log-probability scoring labels over 97% of claims neutral -- collapsing the RL gradient to zero -- while a calibrated MedNLI classifier scores the same pairs non-degenerately. (ii) Moderate signal beats strong signal on answer quality: a strong proprietary checker triggers a three-step reward-hacking cascade -- ultra-short answers, search avoidance, language collapse -- so a moderate-signal local classifier trains a higher-quality model (+12% BERTScore over zero-shot, no GPT dependency). (iii) Signal strength is policy-dependent: the same checker registers as moderate on one policy but strong on another without triggering the cascade end-state. We frame these as boundary conditions for verifier-as-reward systems.

While large language models (LLMs) augmented with agentic search capabilities show promise for legal reasoning, they overlook a fundamental constraint that applicable law must match the temporal context of each case, as retroactive application of statutes violates core legal principles and leads to erroneous conclusions. Our observations reveal that current legal LLMs suffer from temporal bias anchored to their training cutoff, while search agents rarely incorporate temporal constraints into queries, and that web search alone cannot provide the precise statute and precedent citations that legal reasoning demands. To address these challenges, we propose LegalSearch-R1, an end-to-end reinforcement learning framework that pairs local statute RAG for precise article matching with online web search for broader legal knowledge, trained on temporally-indexed data spanning multiple amendment periods to enforce temporal consistency. Extensive experiments on our benchmark covering 13 legal tasks demonstrate that our 7B-parameter agent outperforms state-of-the-art deep research frameworks and specialized legal LLMs by 12.9% to 29.8%, surpasses baselines by 57.7% to 80.3% on temporal consistency, and exhibits robust out-of-domain generalization. The code and data are available at https://github.com/AlexFanw/LegalSearch-R1.

Large Language Model (LLM)-based multi-agent systems rely on optimized collaboration topologies to balance performance and communication costs. However, current methods struggle with the inherent stability-extensibility trade-off and often misalign computational budgets with query difficulty. We propose ATOM, an adaptive framework that generates budget-controllable collaboration graphs via a novel task-driven reinforcement learning paradigm. Inspired by atomic structures, ATOM employs a nucleus-electron hierarchy: it maintains a stable, offline-learned collaboration backbone (the nucleus) while dynamically activating query-conditioned agents (electrons) during inference. Crucially, a complexity-aware budgeting strategy aligns resource consumption with task demands by estimating query difficulty to strictly regulate electron instantiation. Extensive experiments across six diverse benchmarks demonstrate that ATOM achieves state-of-the-art performance while improving token efficiency by up to $30\%$ compared to strong baselines.

Coding agents produce rich trajectories while solving software-engineering tasks. To enable agent self-evolution, these trajectories can be distilled into reusable procedural skills that compactly encode experience to guide future behavior. However, existing skill construction and maintenance methods often rely on fixed prompts and heuristic update rules, leaving it unclear how knowledge should be selected, abstracted, and maintained to best serve downstream agents. We propose CODESKILL, an LLM-based framework that reformulates skill extraction and skill-bank maintenance as a learnable management policy. CODESKILL extracts multi-granularity procedural skills from coding-agent trajectories, evolves skills with new experience, and maintains a compact skill bank for future task solving. We train CODESKILL with reinforcement learning, using a hybrid reward that combines dense rubric-based skill-quality feedback with sparse verifiable execution feedback from the frozen downstream agent. Experiments on EnvBench, SWE-Bench Verified, and Terminal-Bench 2 show that CODESKILL improves average pass rate by 9.69 over the no-skill baseline and by 4.01 over the strongest prompt-based or memory baseline, while maintaining the skill bank at a stable size during iterative construction.

While Large Language Model-based Multi-Agent Systems (LLM-MAS) demonstrate remarkable capabilities in solving complex tasks by orchestrating specialized agents and external tools, the implicit trust in tool outputs creates a critical attack surface. Existing tool attacks are limited by domain specificity or fixed and static templates. To address these challenges, we propose Evo-Attacker, which formulates the tool attack as a self-evolving, memory-augmented reinforcement learning process. Evo-Attacker constructs a dynamic attack memory and employs deliberative reasoning to retrieve adversarial patterns and strategize modifying interventions at critical moments. Furthermore, we introduce Attack-Flow GRPO to optimize intermediate reasoning steps via terminal outcomes, addressing the long-horizon credit assignment challenge. Comprehensive experiments demonstrate that Evo-Attacker consistently outperforms baselines, highlighting its generalization and evolutionary capabilities and the urgent need for defensive tool safeguards.

Proactive task-oriented agents must autonomously anticipate user needs, identify actionable opportunities, and trigger software actions at appropriate moments - fundamentally shifting from reactive systems that await explicit instructions. However, existing approaches lack generalizable end-to-end solutions for measuring and optimizing such anticipatory behaviors. This paper introduces ProActor, a unified framework for conversational task scheduling that integrates: (1) a domain-agnostic automated annotation methodology that enables scalable proactiveness reinforcement learning (RL) by generating full opportunity time windows instead of rigid point labels, (2) systematic proactiveness metrics capturing both timing quality and reference action alignment, and (3) RL optimization using GRPO with various reward designs. Our insight is that RULER-based rewards with proactiveness rubrics are crucial for improving timing quality, and that proactiveness optimization enabled by stage-aware composite rewards is key to balancing timing quality and reference action alignment. Timing-aware RL requires extensive exploration, demanding efficient infrastructure. We develop ART-F, an adaptive framework combining request-adaptive inference clusters with DDP-based training on single-node multi-GPU systems, enabling LoRA training of 4-bit Qwen2.5-14B-ProActor-Q4 with 4-8x speedups. Experiments on two newly auto-annotated datasets demonstrate significant improvements in proactive timing while maintaining action consistency comparable to state-of-the-art (SOTA) baselines. Ablations validate the effectiveness of distinct composite reward variations.

With the continuous advancement of Large Language Models (LLMs), intelligent agents are becoming increasingly vital. However, these agents often fail in environments governed by implicit rules--hidden constraints that cannot be observed directly and must be inferred through interaction. This causes agents to fall into repetitive trial-and-error loops, ultimately leading to task failure. To address this challenge, we propose Test-Time Exploration (TTExplore), a framework where a thinker component analyzes interaction history to infer these implicit rules and guide an actor. Effective exploration in this setting critically depends on the reasoning ability of the thinker. However, evaluating deep reasoning trajectories is inherently unstable and difficult, which poses a major obstacle to effective training. To overcome this issue, we introduce a novel and stable reinforcement learning pipeline. The core idea is to use accurate task-level scores as indirect rewards to bypass the difficulty of evaluating intermediate reasoning, and to retain only a single thinking node per trajectory to alleviate reward sparsity. Using this pipeline, we train a specialized 7B model, Exp-Thinker. Experiments on five text-based embodied tasks show that TTExplore equipped with Exp-Thinker improves baseline agent performance by an average of $14$-$19$ points, demonstrating the effectiveness of explicitly reasoning about implicit rules.

Large language models (LLMs) have achieved strong performance in code generation, but most methods rely on autoregressive decoding without global planning, often leading to locally coherent yet globally suboptimal solutions (e.g., failing test cases or inefficient complexity). While recent approaches such as Chain-of-Thought (CoT) and multi-agent systems (MAS) introduce planning, their limited role specialization and coordination hinder performance on complex tasks. To address the challenges of coordination and specialization in multi-agent code generation, we propose Collaborative Reinforcement Code (CoRe-Code), a framework for role specialized LLM agents that enhances inter-agent coordination to generate more accurate and efficient code. CoRe-Code adopts a simple Planner-Coder paradigm, where the Planner produces high-level plans and the Coder executes them to generate code. We further introduce a collaboration-aware reinforcement learning stage based on Group Relative Policy Optimization (GRPO) to enhance role specialization and alignment. Experiments show that CoRe-Code outperforms a wide range of existing RL-based and multi-agent methods. In addition, we demonstrate that CoRe-Code can generalize to other multi-agent frameworks (e.g., Retrieval and Debugging agents), highlighting its flexibility and scalability. We evaluate CoRe-Code on multiple benchmarks of varying difficulty using three base models. Compared to existing baselines, the results show consistent improvements in accuracy, while also achieving higher efficiency in terms of execution time and memory usage, demonstrating the effectiveness and practicality of CoRe-Code.

Large language model (LLM) agents excel at solving complex long-horizon tasks through autonomous interaction with environments. However, their real-world deployment faces a fundamental device--cloud dilemma: on-device models are efficient but often brittle, while cloud models are stronger but costly in computation. State-of-the-art LLM device--cloud routers usually make coarse task-level decisions, which cannot adapt to the changing difficulty of multi-step agent interactions. To address this issue, we present Hera, a step-level device--cloud LLM agent coordinator for long-horizon tasks achieving a strong performance--cost Pareto frontier. Hera adopts a novel two-stage training paradigm: (1) imitation learning for cold-start, followed by (2) reinforcement learning that jointly optimizes task success and cloud usage efficiency. The first stage casts step-level routing as a supervised classification problem: the device agent is replayed on cloud trajectories, with each state labeled by the agreement between device and cloud actions. In the second stage, we perform cost-aware reinforcement learning by grouping identical states across trajectories and updating Hera with labels favoring higher expected return and fewer future cloud calls. We evaluate Hera on ALFWorld, WebShop, and AppWorld, where it consistently outperforms prior methods, achieving 92.5% of the cloud-only success rate with cloud use in only 46.3% of steps.

CLI agents are the closest thing language models have to an embodied setting: the model emits commands, the terminal executes them, and the returned stream -- stdout, errors, files, logs, and traces -- records the consequences. We argue that this stream is a supervision signal, but standard agent RL discards it: GRPO-style training updates action tokens with sparse outcome-level rewards while ignoring environment responses already in the rollout. Failed rollouts provide little policy-gradient signal despite containing rich evidence about how the environment responds. We introduce ECHO (Environment Cross-entropy Hybrid Objective), a hybrid objective that combines the standard policy-gradient loss on action tokens with an auxiliary loss that trains the policy to predict environment observation tokens resulting from its own actions. ECHO reuses the same forward pass as GRPO, requires no additional rollouts, and turns terminal feedback into dense supervision for all rollouts. ECHO doubles GRPO pass@1 on TerminalBench-2.0: Qwen3-8B improves from 2.70% to 5.17%, and Qwen3-14B from 5.17% to 10.79%. ECHO also produces policies that better predict terminal dynamics, even on trajectories they did not generate: across held-out rollouts, it sharply reduces environment-token cross-entropy while GRPO alone barely changes it. From base Qwen3-8B, ECHO matches expert-SFT-then-GRPO performance on held-out terminal tasks without expert demonstrations, and recovers roughly half of the expert-SFT initialization benefit on TerminalBench-2.0. In some settings, the environment prediction loss alone enables verifier-free self-improvement, allowing policies to improve on unseen OOD tasks by learning only from environment interactions. Together, these results suggest that environment observations are not merely context for future actions, but a dense, on-policy supervision signal already present in every rollout.

Multi-agent LLM workflows route inference through specialized roles to lift end-task accuracy, but jointly training those roles with reinforcement learning is unstable in ways that are poorly understood. We study when end-to-end RL training of multi-agent LLM workflows improves over their base models, comparing Shared-Policy training, where all roles update one policy, with Isolated-Policy training, where each role has its own parameters. Our experimental matrix spans Eval-Opt, Voting, and Orch-Workers workflows, math and code tasks, and three model scales (0.6B, 1.7B, 4B). We find that multi-agent RL usually improves over base models, but gains depend jointly on workflow, task, and scale, not on policy sharing alone. Isolated-Policy tends to reach higher peak accuracy yet more often falls off a terminal accuracy cliff, while Shared-Policy training does not eliminate failure; it redistributes failure into qualitatively different patterns. We then explain the strongest of these patterns through role-level gradient dynamics induced by workflow topology and policy routing: under Isolated-Policy, parallel same-role agents on shared prompts amplify per-role gradients and drive terminal degradation in Voting and Orch-Workers workflows; under Shared-Policy, asymmetric per-step gradient mass causes the shared policy to be captured by the dominant role, producing different failure signatures by task and workflow. Together, the empirical map and its underlying mechanisms show that policy sharing routes training pressure through different channels rather than offering uniform stability, making it a design choice with workflow- and task-conditional tradeoffs.

ReAct-style agents for search-intensive, multi-step reasoning tasks rely largely on their own internal judgment to decide what evidence to seek, which reasoning or action step to take next, and when to stop, often producing shallow, redundant, or poorly targeted trajectories. Prior work has explored rubrics as external quality signals, but existing uses are mostly evaluative rather than action-guiding: rubrics typically serve as training-time rewards or post-hoc evaluators of completed outputs, and in deep-research settings they are often coarse-grained and report-level rather than step-level. We introduce Co-ReAct, a rubric-guided action-selection framework that uses rubrics as step-level guidance during inference. At each decision step, Co-ReAct injects a rubric into the agent's context to guide the next Reason-or-Act decision, specifying what the agent should target in evidence seeking, search, reasoning, or self-evaluation. To make this guidance reliable, we train a dedicated rubric generator with GRPO. Unlike prior pairwise or binary preference formulations, our objective optimizes a list-wise Spearman rank-correlation reward against multi-judge expert consensus rankings, encouraging rubrics that are discriminative rather than merely plausible. On DeepResearchBench and SQA-CS-V2, Co-ReAct consistently improves over ReAct and representative test-time compute baselines across search agents built on both 8B/14B open-source and frontier closed-source base models. The trained rubric generator can also serve as a drop-in component that improves these baselines without changing their underlying decision mechanisms. Our code is publicly available at https://github.com/ZBWpro/Co-ReAct.

Agentic reinforcement learning (Agentic RL) has achieved strong progress in tasks with clear success signals. However, many real-world agent applications require user-conditioned behavior: the same query may call for different planning strategies and tool-use decisions across users. This setting raises key challenges: generic rewards cannot capture heterogeneous user preferences, observed behaviors are entangled with conformity effects, and flat memories cannot support personalized skill retrieval. To this end, we propose a unified personalized Agentic RL framework that embeds personalization into training-time optimization. At its core is Personalized Anchor Reward-Decoupled Policy Optimization (PARPO), which decouples generic task-quality rewards from personalized preference rewards and uses user-specific anchors to stabilize learning under heterogeneous reward scales. We further introduce a two-stage preference-disentangled reward model and Preference-Aligned Skill Evolution Graph Memory (PSGM) for personalized supervision and preference-aligned skill retrieval. Together, they form a closed loop of preference identification, policy optimization, and structured skill accumulation. Experiments on ETAPP, ETAPP-Hard, and SJAgent show that our framework consistently outperforms strong memory and RL baselines. Code and data are included in the supplementary materials.

Reinforcement learning (RL) has emerged as a viable recipe for training LLM agents to reason over external memory banks in multi-session dialogue. Existing work trains exclusively on a single benchmark, leaving open how the composition of training data shapes the skills a memory agent acquires. We present a controlled empirical study that holds architecture, RL algorithm, and all hyperparameters fixed and varies only the training curriculum across three conditions: in-domain (LoCoMo), mixed-benchmark (LoCoMo + LongMemEval), and out-of-domain (LongMemEval only). Across two benchmarks and ten question types, curriculum composition acts as a fine-grained lever on specialization rather than a uniform scaling factor on performance. The mixed curriculum yields the strongest overall F1 on both evaluation sets. Training on a narrow out-of-domain set transfers a targeted skill - temporal reasoning - despite weak aggregate performance. Per-type differences substantially exceed aggregate differences, indicating that single-number benchmark comparisons systematically underreport curriculum effects. We further report two practical lessons from adapting GRPO to a single-GPU regime: cross-benchmark mixing requires filtering format-specific noise from memory banks to preserve training signal, and binary exact-match reward produces no learning signal at the small group sizes (G = 4) required on one GPU, motivating continuous reward functions in this regime.

Spreadsheet systems (e.g., Microsoft Excel, Google Sheets) play a central role in modern data-centric workflows. As AI agents grow increasingly capable of automating complex tasks, such as controlling computers and generating presentations, building an AI-driven spreadsheet agent has emerged as a promising research direction. Most existing spreadsheet agents rely on specialized prompting over general-purpose LLMs; while this design has potentials on simple spreadsheet operations, it struggles to manage the complex, multi-step workflows typical of real-world applications. We introduce Spreadsheet-RL, a reinforcement learning (RL) fine-tuning framework designed to train specialized spreadsheet agents within a realistic Microsoft Excel environment. Spreadsheet-RL features an automated pipeline for scalable collection of paired start-goal spreadsheets from online forums, as well as domain-specific evaluation tasks in areas such as finance and supply chain management, which we compile into the new Domain-Spreadsheet benchmark dataset. It also includes a Spreadsheet Gym environment designed for multi-turn RL: Spreadsheet Gym exposes extensive Excel functionality through a Python sandbox, along with a refined harness that incorporates a comprehensive tool set and carefully designed tool-routing rules for spreadsheet tasks. Through comprehensive experiments, we show that Spreadsheet-RL substantially enhances AI agent's performance on both general and domain-specific spreadsheet tasks: it improves Qwen3-4B-Thinking-2507's Pass@1 on SpreadsheetBench from 12.0% to 23.4%, and raises Pass@1 from 8.4% to 17.2% on our curated Domain-Spreadsheet dataset. These results highlight Spreadsheet-RL's strong potential for generalization and real-world adoption in spreadsheet automation, and broadly, its promise for advancing LLM-based interactions with data interfaces in everyday work.

Large language model (LLM) agents still struggle to effectively use long-term memory, with answer-supporting evidence often scattered across long conversational histories and buried in substantial irrelevant content. Existing memory systems commonly process memory into dedicated units before future queries are known and retrieve these preconstructed units at query time. Because the contents and granularity of these units are determined before the current query is known, the retrieved units can remain coarse and noisy for that query, leaving downstream answerers to further denoise them and uncover query-specific evidence. We present DeferMem, a long-term memory framework that decouples this problem into high-recall candidate retrieval and query-conditioned evidence distillation. DeferMem uses a lightweight segment-link structure to organize raw history and retrieve broad candidates at query time. A memory distiller then distills these high-recall but highly noisy candidates into a set of faithful, self-contained, and query-conditioned evidence. To train this distiller, we introduce DistillPO, a reinforcement learning algorithm that formulates post-retrieval evidence distillation as a structured action comprising message selection and evidence rewriting. It optimizes this action with a decomposed-and-gated reward pipeline and structure-aligned advantage assignment, gating reward components from validity to quality checks while exposing answerability-related feedback early and assigning each reward to its responsible output span. On LoCoMo and LongMemEval-S, DeferMem surpasses strong baselines in QA accuracy and memory-system efficiency, achieving the highest QA accuracy and fastest runtime while consuming no commercial-API tokens for memory operations.

The proliferation of large language models (LLMs) and modular skills has endowed autonomous agents with increasingly powerful capabilities. Existing frameworks typically rely on monolithic LLMs and fixed logic to interface with these skills. This gives rise to a critical bottleneck: different LLMs offer distinct advantages across diverse domains, yet current frameworks fail to exploit the complementary strengths of models and skills, thereby limiting their performance on downstream tasks. In this paper, we present Maestro (Multimodal Agent for Expert-Skill Targeted Reinforced Orchestration), a Reinforcement Learning (RL)-driven orchestration framework that reframes heterogeneous multimodal tasks as a sequential decision-making process over a hierarchical model-skill registry. Rather than consolidating all knowledge into a single model, Maestro trains a lightweight policy to dynamically compose ensembles of frozen expert models and a two-tier skill library, deciding at each step whether to invoke an external expert, which model-skill pair to select, and when to terminate. The policy is optimized via outcome-based RL, requiring no step-level supervision. We evaluate Maestro across ten representative multimodal benchmarks spanning mathematical reasoning, chart understanding, high-resolution perception, and domain-specific analysis. With only a 4B orchestrator, Maestro achieves an average accuracy of 70.1%, surpassing both GPT-5 (69.3%) and Gemini-2.5-Pro (68.7%). Crucially, the learned coordination policy generalizes to unseen models and skills without retraining: augmenting the registry with out-of-domain experts yields a 59.5% average on four challenging benchmarks, outperforming all closed-source baselines. Maestro further maintains high computational efficiency with low latency. The source code is available at https://github.com/jinyangwu/Maestro.

We present Mem-$π$, a framework for adaptive memory in large language model (LLM) agents, where useful guidance is generated on demand rather than retrieved from external memory stores. Existing memory-augmented agents typically rely on similarity-based retrieval from episodic memory banks or skill libraries, returning static entries that often misalign with the current context. In contrast, Mem-$π$ uses a dedicated language or vision-language model with its own parameters, separate from the downstream agent, to generate context-specific guidance for complex tasks. Conditioned on the current agent context, the model jointly decides when to produce guidance and what guidance to produce. We train it with a decision-content decoupled reinforcement learning (RL) objective, enabling it to abstain when generation would not help and otherwise produce concise, useful guidance. Across diverse agentic benchmarks spanning web navigation, terminal-based tool use, and text-based embodied interaction, Mem-$π$ consistently outperforms retrieval-based and prior RL-optimized memory baselines, achieving over 30% relative improvement on web navigation tasks.

Language agents increasingly operate over streams of related tasks, yet existing memory systems struggle to convert accumulated experience into reusable knowledge. Retrieval-augmented and structured memory methods record per-session observations effectively, but often couple acquisition and consolidation into a single online process, leaving the agent without a global view across sessions to discover recurring patterns, abstract shared procedures, or prune redundant entries. Inspired by complementary learning systems theory, we propose Auto-Dreamer, a learned offline consolidator for language-agent memory. Auto-Dreamer decouples fast per-session memory acquisition from slow cross-session consolidation. Given a selected working region of a typed memory bank, the consolidator treats the region as read-only evidence, performs bounded tool-use to inspect entries and provenance-linked source trajectories, and synthesizes a fresh compact replacement set that abstracts across sessions and supersedes the original region. We train Auto-Dreamer via GRPO, using end-to-end agent performance as the reward signal to learn how to consolidate memories acquired through fast online experience. Trained on ScienceWorld trajectories alone, Auto-Dreamer outperforms fixed, RL-trained, and prompted memory baselines on ScienceWorld by 7 points while using an active memory bank 12$\times$ smaller than the strongest baseline, and continues to lead on held-out ALFWorld and WebArena without retraining -- using 6$\times$ less memory than the strongest baseline on ALFWorld.

Language agents can adapt from experience in interactive environments, but current reflection-based methods can only self-correct within a single task instance. Whether such experience can be distilled into reusable lessons that improve performance on future unseen tasks remains unclear. We address this problem by introducing the In-context Training (ICT) task, a framework for evaluating cross-task self-improvement in language agents. In ICT, a reflector model observes trajectories collected by an actor model and generates system prompts intended to improve the actor's performance on future unseen tasks. We then propose an RL-based training pipeline for learning such reflections directly from experience, without human-provided examples. Across ALFWorld and MiniHack, our trained reflectors outperform an untrained baseline on most held-out task families, showing that the ability to learn from experience can itself be learned. In some cases, we observe generalisation beyond the benchmark on which the reflector was trained, to substantially different environments. Finally, we introduce MetaGym, a generic Python library for constructing meta-environments, enabling future research on self-improving language agents.

Reinforcement learning from verifiable rewards (RLVR) is a promising paradigm for improving large language model (LLM) agents on long-horizon interactive tasks. However, in partially observable environments, incomplete observations cause agent beliefs to drift over time, while delayed rewards obscure the causal impact of intermediate decisions, exacerbating temporal credit assignment challenges. To address this, we propose ReBel (Reward Belief), a process-level reinforcement learning algorithm that explicitly models structured belief states to summarize interaction history and guide subsequent policy learning. ReBel introduces belief-consistency supervision, converting discrepancies between predicted beliefs and observed feedback into dense self-supervised signals without requiring external step-wise annotations or verifiers. It also employs belief-aware grouping to compare trajectories under similar belief states, yielding more robust and lower-variance advantage estimates. We evaluate ReBel on challenging long-horizon benchmarks, including ALFWorld and WebShop. ReBel improves task success by up to $20.4$ percentage points over the episode-level baseline GRPO and increases sample efficiency by $2.1\times$. These results suggest that belief-aware self-supervision is a promising direction for reliable long-horizon decision-making under partial observability. Code is available at: https://github.com/Fateyetian/Rebel.git.

Automatic generation of computer-aided design (CAD) models is a core technology for enabling intelligence in advanced manufacturing. Existing generation methods based on large language models (LLMs) often fall short when handling complex CAD models characterized by long operation sequences, diverse operation types, and strong geometric constraints, primarily because reasoning chains break and effective error-correction mechanisms are lacking. To address this problem, this paper proposes a memory-augmented reinforcement learning framework for CAD generation agents. The framework encapsulates the underlying geometric kernel into a structured toolchain callable by the agent and builds a closed-loop mechanism of design intent understanding, global planning, execution, and multi-dimensional verification. It also designs a dual-track memory module consisting of a case library and a skill library, and proposes a dynamic utility retrieval algorithm. By introducing reinforcement learning into retrieval and policy optimization, the agent can effectively avoid retrieval traps in which examples are semantically similar but geometrically infeasible, enabling online self-correction and continual evolution without additional large-scale annotated data. Experiments show that the proposed method significantly improves both the success rate and geometric consistency on complex CAD model generation tasks.

Reinforcement learning can train LLM agents from sparse task rewards, but long-horizon credit assignment remains challenging: a single success-or-failure signal must be distributed across many actions. Existing methods rely on trajectory-level rewards or proxy signals, without fully leveraging per-step environmental feedback. Multi-turn agent settings are underexplored, where feedback can include error messages, page changes, observations, or reference trajectories. We systematically study five feedback sources and two insertion granularities and introduce SERL, a selective environment-reweighted learning framework. SERL uses the task reward to determine update direction, while environment feedback adjusts placement and magnitude, focusing on critical actions. On ALFWorld and WebShop, SERL achieves 90.0% and 80.1% success, outperforming strong RL and distillation baselines. Analysis shows that grounded, action-relevant feedback at meaningful points consistently outperforms indiscriminate use of longer or richer context.