RL for LLM Agents

RL: Reinforcement Learning · LLM: Large Language Model

Follow this topic Ask about this topic

Momentum

70 papers in the last four weeks, up 141% on the four weeks before. 0.7% of all new papers.

Jul 13Week of Sep 28

Latest papers 402

Jun 29, 2026cs.LG

Building Multi-Task Agentic LLMs via Two-Phase Distillation

A key step toward artificial general intelligence is to train models that can perform multiple tasks. In this paper, we study how to build such models by first training separate RL experts for individual tasks and then consolidating them via distillation, as an alternative to directly training a single model on mixed tasks. We show that off-policy distillation degrades in multi-task settings due to the mode-covering nature of forward KL: aggregating data from multiple tasks introduces a large number of behavioral modes that can exceed the student's capacity, forcing it to average across behaviors and leading to degraded performance. In contrast, on-policy distillation is mode-seeking but requires strong initialization. Inspired by these observations, we propose a two-phase approach: off-policy distillation followed by on-policy refinement. Evaluation across conversational agents and text-based games confirms that this two-phase approach matches single-task RL expert performance for each individual task, whereas off-policy or on-policy distillation alone fails to match this performance.

Huaijie Wang, Shusheng Xu, Yi Wu +1

Tsinghua University · Ant Group

Language Model Distillation Multi-Task RL On-Policy Distillation RL for LLM Agents

Jun 29, 2026cs.MA

ECHO: Learning Epistemically Adaptive Language Agents with Turn-Level Credit

What does it mean for a language agent to be adaptive? Effective multi-turn agents must decide what information to seek, how to use new evidence, and when they are certain enough to act. We introduce Epistemic Decision Processes (EDPs), a belief-state formulation of multi-turn information seeking in which actions produce external observations that update the agent's posterior over a latent task variable. EDPs make epistemic adaptivity explicit: good policies choose actions that are useful under the current belief, not merely those that correlate with eventual success. We prove that belief-agnostic policies can suffer errors that compound exponentially over the horizon, and that aggregate trajectory returns can fail to identify the per-turn Bayesian advantage needed for epistemic credit. We then introduce ECHO (Epistemic Credit for History-Conditioned Optimization), a practical clipped policy-gradient objective that assigns turn-level credit using posterior-sensitive rewards. In the Clue Selector Game, a novel controlled evidence-seeking benchmark, we show that ECHO substantially improves resolution, information gain, and efficiency over trajectory-level GRPO, and matches or exceeds frontier baselines on epistemic metrics such as grounding, recovery, and calibration while producing almost no visible reasoning text.

Abhijnan Nath, Nikhil Krishnaswamy

Situated Grounding and Natural Language (SIGNAL) Lab Department of Computer Science, Colorado State University Fort Collins, CO 80523 USA

Belief-Space Planning Credit Assignment in RL AI Agent Benchmarks RL for LLM Agents

Jun 29, 2026cs.CL

SEVA: Self-Evolving Verification Agent with Process Reward for Fact Attribution

Hallucination is the reliability bottleneck for LLM-based agents, and fact attribution verifiers are the last line of defense -- yet today's verifiers emit only opaque binary labels, leaving agents unable to self-correct and operators unable to audit. We present SEVA, a structured verification agent that emits evidence alignments, step-by-step reasoning chains, calibrated confidence, and a six-category error diagnosis with actionable fixes. Training such an agent with RL is non-trivial: standard binary reward on multi-component output triggers advantage collapse -- within-group reward variance vanishes and the GRPO gradient disappears. We resolve this with a process reward that decomposes verification quality into five independent components weighted 70/30 toward process signals, restoring the gradient and inducing an implicit curriculum -- the agent first masters verification behavior (alignment 0.917 -> 0.997, format 72% -> 100%), then outcomes (F1 64.9 -> 69.0). Structured output further enables a Verify -> Reflect -> Probe -> Refine self-evolution loop, which over four rounds on a 7B model surfaces an unexpected structural finding: each round produces a benchmark-specialist, not a generalist (+15 pp on HaluEval, -10 to -14 pp on TruthfulQA in the same model, persistent at 4x data). On ClearFacts, SEVA-3B matches GPT-4o-mini (69.0 vs. 69.8 F1) while producing substantially richer, auditable output -- confirming a principle that should generalize: for any RL task with multi-component generation, reward granularity must match output granularity.

Aojie Yuan, Yi Nian, Haiyue Zhang +2

University of Southern California, Los Angeles, USA · University of Michigan, Ann Arbor, USA

AI Agent Reliability Process Reward Models LLM Agent Self-Improvement LLM Agent Verification RL for LLM Agents LLM Agent Reliability

Jun 28, 2026cs.AI

UCOB: Learning to Utilize and Evolve Agentic Skills via Credit-Aware On-Policy Bidirectional Self-Distillation

Skill memories can improve agentic reinforcement learning by reusing past experience as textual guidance, but retrieved skills are not oracular: they may help in one state while misleading the same policy in another. This makes the common privileged-teacher assumption fragile, namely that a skill-conditioned prompt can be treated as a fixed teacher for the no-skill prompt. We introduce UCOB, a framework for learning to utilize and evolve agentic skills via credit-aware on-policy bidirectional self-distillation. UCOB treats skill-conditioned and no-skill prompts as two on-policy context views of the same model, compares their return-to-go within the same task and anchor state, and uses the higher-return view as the local teacher. This local credit signal internalizes useful skill-conditioned behavior, corrects misleading skill usage, and guides task/state skill memory updates, utility-aware retrieval, and reflection self-training. Experiments on agentic tasks, including ALFWorld, WebShop, and Search-QA, show that UCOB outperforms skill-free RL, skill-memory baselines, and self-distillation methods across model scales, with up to 23.5 and 18.0 point gains over SOTA baselines on ALFWorld and WebShop. Ablations and analyses further validate its core mechanisms and efficiency.

Songjun Tu, Chengdong Xu, Qichao Zhang +6

Institute of Automation, Chinese Academy of Sciences · Pengcheng Laboratory · Memorax AI

Credit Assignment in RL LLM Agent Skill Learning Policy Distillation RL for LLM Agents

Jun 26, 2026cs.AI

ATOD: Annealed Turn-aware On-policy Distillation for Multi-turn Autonomous Agents

Training small language-model agents for long-horizon interactive tasks requires both fast imitation and reward-driven improvement. On-policy distillation (OPD) provides dense teacher guidance and typically improves rapidly in the early stage, but its gains saturate once the student approaches the teacher, limiting the final performance ceiling. Reinforcement learning (RL) directly optimizes environment rewards and encourages exploratory improvement toward a higher reward-defined ceiling, but sparse and delayed feedback makes early-stage learning much less efficient than OPD. In this paper, we propose ATOD (Annealed Turn-aware On-policy Distillation), a hybrid online distillation algorithm that explicitly exploits this complementarity. (1) ATOD uses an annealed OPD-RL schedule: OPD dominates early training to approach teacher-level behavior, while RL is gradually strengthened to drive reward-based exploration. (2) ATOD introduces Turn-level Disagreement-Uncertainty Reweighting (T-DUR), which softly amplifies high-utility turns and improves dense supervision in long trajectories. Experiments on ALFWorld, WebShop, and Search-QA show that ATOD consistently outperforms competing post-training baselines: across the three student sizes, ATOD improves average success rate by 3.03 points over OPD and 23.62 points over GRPO, while surpassing the corresponding teacher models by 2.16 points.

Qitai Tan, Zefang Zong, Yang Li +1

Tencent Inc. · Tsinghua University

Agentic RL On-Policy Distillation RL Post-Training RL for LLM Agents

Jun 26, 2026cs.SE

SWE-MeM: Learning Adaptive Memory Management for Long-Horizon Coding Agents

Long-horizon software engineering agents often need to manage lengthy and noisy interaction histories under limited context budgets. Existing memory management methods typically rely on static compression workflows or impose rigid constraints on compression timing and granularity. Moreover, these approaches fail to jointly optimize memory management and issue resolution capabilities to improve performance while reducing token usage. We present SWE-MeM, a training framework for proactive and on-demand memory management in software engineering agents. SWE-MeM provides a flexible memory tool that lets agents decide when, what, and how to compress based on trajectory state, task progress, and remaining context budget. We train agents with synthesized proactive memory-management trajectories and Memory-aware GRPO, which jointly optimizes memory management and issue resolution through memory-aware trajectory splitting and step-level credit assignment. On SWE-Bench Verified, SWE-MeM achieves 43.4% and 60.2% resolve rate with 4B and 30B models, respectively, outperforming existing memory management baselines in both performance and efficiency.

Shuzheng Gao, Wenhao Zeng, Zhaojian Yu +5

The Chinese University of Hong Kong, China · Shanghai Jiao Tong University, China · Tsinghua University, China +1

Software Engineering Agents Agent Memory Management RL for LLM Agents

Jun 25, 2026cs.AI

Internalizing the Future: A Unified Agentic Training Paradigm for World Model Planning

Large language model (LLM) agents have demonstrated strong capability in sequential decision-making, yet they remains fundamentally reactive in long-horizon tasks. Unlike humans who employ "what-if" reasoning to evaluate potential plans before commitment, standard agents lack an internal world model to simulate future outcomes. Therefore, we propose to internalize future-aware planning by training a single autoregressive model to verbalize both a prospective state rollout and a plan-conditioned success estimate-a textual analogue of the Q-value. Crucially, we identify a format-capability gap: simply fine-tuning agents on look-ahead traces during post-training leads to superficial mimicry of foresight without genuine predictive grounding. To bridge this gap, we introduce a three-stage training paradigm: (i) World Model Agentic Mid-Training (WM-AMT) to inject latent predictive capabilities into the policy; (ii) Format-Eliciting SFT (FE-SFT) to structure this injected capability; and (iii) Foresight-Conditioned Reinforcement Learning (FC-RL) to refine the calibration and utility of the generated simulations. Evaluated on search and mathematical reasoning tasks, our approach consistently outperforms other training baselines. Our results demonstrate that effective internal world modeling in LLM agents requires a capability-first training pipeline to achieve grounded and calibrated foresight.

Xuan Zhang, Zhijian Zhou, Lingfeng Qiao +6

Fudan University · Shanghai Innovation Institute · Tencent Youtu Lab

World Model-Based Planning Model-Based Planning LLM Agent Training Latent World Models Long-Horizon LLM Agents RL for LLM Agents

Jun 25, 2026cs.CL

Supersede: Diagnosing and Training the Memory-Update Gap in LLM Agents

Large language model (LLM) agents operate over long, multi-session interactions in which facts change: a user moves, a price updates, a plan is revised. Acting correctly requires using the current value of a fact and discarding values that have been superseded. We isolate this ability on real conversational data and show that it is a distinct, unsolved failure. On the knowledge-update subset of LongMemEval, replacing an agent's full context with a bounded, self-maintained memory drops accuracy from 92% to 77% even on a frontier model (gpt-5.4), a gap that is statistically significant (paired McNemar p<0.005) and persists across model scale while full-context accuracy saturates near 92%. The bottleneck is therefore memory maintenance, not comprehension, and is not closed by a stronger model. We then ask whether this is merely an undersized memory, and find it is not: as the conversation grows 24x, accuracy falls further (from 68% to 28%), and granting the agent proportionally more memory yields no detectable recovery (28% to 28%, n=25). The failure scales with the length of the conversation, not the compression ratio. We release Supersede, an open reinforcement-learning environment (on the verifiers / prime-rl stack) that turns this measurement into a training signal: agents are rewarded for answering from the current value and penalized for stale ones. Finally, we close the loop and show the gap is trainable: GRPO fine-tuning a small open model (Qwen2.5-3B) on this environment nearly doubles its held-out supersession accuracy on real, unseen conversations (9.0% to 16.7%, a single run), along a monotonic checkpoint curve indicating the learned policy, not the harness, carries the gain. To our knowledge this is the first trainable environment whose reward targets temporal fact-currency, and the first evidence the supersession gap can be trained down, not only measured.

Vedant Patel

Vrin

Continual Learning for LLM Agents Knowledge Conflicts in Language Models LLM Agent Memory RL for LLM Agents

Jun 25, 2026cs.AI

Joint Learning of Experiential Rules and Policies for Large Language Model Agents

For LLM agents in multi-step interactive environments, a key challenge is to make effective use of accumulated interaction experience. Existing work has typically separated two uses of such experience: keeping it outside the model as natural-language rules for later prompting, or using trajectories and feedback to update the model parameters. The former is easy to interpret but can fall out of sync with the evolving policy; the latter improves the policy more broadly but provides only limited correction for local mistakes in sparse-reward settings. We present Joint Learning of Experiential Rules and Policies for LLM Agents (JERP), which updates a long-term experiential-rule pool and the policy from the same interaction trajectories. At decision time, JERP retrieves task-relevant rules and conditions the agent on them together with the interaction history. After each episode, it uses the collected trajectories both to optimize the policy and to revise the rule pool by comparing current rollouts with reference successful trajectories. This coupling keeps the rule pool aligned with the evolving policy while allowing stable and effective behaviors to be gradually absorbed into the model itself. Experiments on AlfWorld and WebShop show that JERP yields consistent gains in decision performance for complex interactive tasks.

Shicheng Ye, Chao Yu

Sun Yat-sen University

Continual Learning for LLM Agents LLM Agent Training Policy Learning RL for LLM Agents Memory-Augmented Language Model Agents

Jun 24, 2026cs.LG

Neglected Free Lunch from Post-training: Progress Advantage for LLM Agents

Process reward models enable fine-grained, step-level evaluation of LLMs, yet building them for agentic settings remains prohibitively difficult: long-horizon interactions, irreversible actions, and stochastic environment feedback make both human annotation and Monte Carlo estimation infeasible at scale. In this work, we show that reinforcement learning (RL) post-training already provides the ingredients for effective step-level scoring, eliminating the need for dedicated reward model training altogether. Concretely, we derive an implicit advantage under a general stochastic Markov decision process, which we term progress advantage -- log-probability ratio between the RL-trained policy and its reference policy exactly recovers the optimal advantage function. This formulation makes the resulting signal annotation-free, domain-agnostic, and available as a byproduct of the standard RL post-training pipeline. We validate the effectiveness of the progress advantage across three different applications: test-time scaling, uncertainty quantification, and failure attribution on five benchmarks and four model families. Across all settings, it consistently outperforms confidence-based baselines and, despite requiring no task-specific training, surpasses dedicated trained reward models. We complement these results with deeper analyses on characteristics of progress advantage, offering practical guidance for adoption in real-world agentic systems.

Changdae Oh, Wendi Li, Seongheon Park +3

University of Wisconsin–Madison · Argonne National Laboratory

Process Reward Models RL Post-Training RL for LLM Agents

Jun 24, 2026cs.LG

Semantic Consistency Policy Optimization for Reinforcement Learning of LLM Agents

Group-based reinforcement learning effectively post-trains LLM agents for long-horizon, sparse-reward tasks by deriving step-level credit from trajectory outcomes. However, this ties a step's credit to its rollout's final outcome: semantically near-identical intermediate steps receive opposite credit depending on whether their trajectory eventually succeeded or failed. Such semantic credit inconsistency sends conflicting gradients to similar actions and wastes the partially-correct progress inside failed rollouts. Motivated by this, we propose Semantic Consistency Policy Optimization (SCPO), a value-free reward-shaping method that mitigates this inconsistency by recovering step-level credit from successful siblings in the same rollout group. Concretely, SCPO scores each failed step against a successful sibling and adds positive step-level credit for new progress along that sibling. On ALFWorld and WebShop, SCPO matches or exceeds strong group-based baselines, reaching 93.7+/-4.1 percent success on ALFWorld and 74.8+/-2.0 percent on WebShop at 1.5B parameters, with gains concentrated on the hardest multi-step tasks.

Peng Xu, Sijia Chen, Junzhuo Li +1

†The Hong Kong University of Science and Technology (Guangzhou) · ‡The Hong Kong University of Science and Technology

Reward Shaping Credit Assignment in RL Step-Level Credit Assignment in RL Sparse-Reward RL RL for LLM Agents

Jun 24, 2026cs.CL

BiPACE: Bisimulation-Guided Policy Optimization with Action Counterfactual Estimation for LLM Agents

Stepwise group-based RL is an attractive way to train long-horizon LLM agents without a learned critic: it reuses multiple sampled rollouts to estimate local advantages. Its weakness is less visible but more fundamental: every group-relative estimator assumes that the steps it compares are equivalent for credit assignment. We show that current agentic variants violate this assumption through a state-action credit mismatch. The observation-hash partition is overly fine on the state side, creating singleton groups with zero step-level signal, while a single within-group mean is too coarse on the action side, mixing state-value estimation with action-specific credit. We introduce BiPACE (Bisimulation-Guided Policy Optimization with Action Counterfactual Estimation), a drop-in advantage estimator that fixes both sides without adding a critic, auxiliary loss, or extra rollouts. BiGPO clusters steps by cosine distance in the actor's own hidden-state geometry, an empirical policy-induced proxy for bisimulation that substantially lowers the singleton rate left by observation hashing. PACE then recenters returns within each behavioral cluster using action-conditioned peer baselines; its Q-style instance estimates a local Q(s,a)-V(s) nonparametrically. On ALFWorld/Qwen2.5-7B, BiPACE_Q raises overall validation success from GiGPO's 90.8 to

97.1\pm0.9

over three seeds, and crosses the 95% threshold on every seed, which GiGPO never does within the same budget. On Qwen2.5-1.5B it reaches

93.5\pm1.2

versus GiGPO's 86.7, and on WebShop and TextCraft it improves over GRPO and GiGPO at both model scales. The measured BiPACE-specific overhead is 11.3% of a single training-step wall time. Yet it changes the estimator's comparison unit from surface identity to approximate behavioral equivalence plus action-side counterfactuals. The code is available at https://github.com/TianxiangZhao/BiPACE.

Hanyang Wang, Weijieying Ren, Yuxiang Zhang +4

University of Chicago · Stanford University · The Hong Kong University of Science and Technology (Guangzhou) +2

Group Relative Policy Optimization RL for LLM Agents

Jun 24, 2026cs.CL

Beyond Next-Observation Prediction: Agent-Authored World Modeling for Sequential Decision Making

Recent studies on world modeling for Large Language Model (LLM) agents typically formulate the learning objective as next-observation prediction. However, this objective ties supervision to what a transition happens to reveal, which may omit the dynamics most relevant to the agent's current decision. To bridge this gap, we propose Agent-Authored World Modeling (AAWM), a training procedure that constructs supervision from the policy's own decision needs. Specifically, at each state, the agent identifies what it needs to understand about the environment before acting. These needs drive the retrieval of relevant transition evidence across trajectories, which is then synthesized into training targets that capture decision-oriented dynamics instead of reconstructing the next observation. This aligns the training objective with the dynamics the policy needs before acting, not with the contents of the next observation. Experimental results validate the effectiveness of AAWM across multiple environments and training settings. These results show that decision-aware world-model targets provide a more effective learning signal than next-observation prediction.

Guangfeng Cai, Kaibing Yang, Shuo He +4

Southeast University · Meituan

World Model Learning World Models LLM Agent Training RL for LLM Agents

Jun 22, 2026cs.CL

Tmax: A simple recipe for terminal agents

Terminal-using agents have quickly become the most popular downstream application of language models (LMs). Despite their prevalence, relatively little academic work has examined RL-based training of these models, likely due to difficult benchmarks, a lack of data, and a lack of simple baseline recipes. We present Tmax, the strongest open RL recipe for terminal agents to date, bringing open data recipes closer to the frontier. While simple, our recipe achieves 27% on Terminal-Bench 2.0 with only 9B parameters, outperforming much larger models from prior work. Concretely, we generate data using a novel taxonomy, combining difficulty control, personas, and verifier diversification, which allows us to cheaply generate large amounts of terminal environments for RL and SFT training. We open-source our terminal dataset, which is over 2.5x larger than previously released terminal-agent datasets. We then train open-weight models using RL with our data, using a simple, outcome-only recipe. We release our data, models, and code as a strong baseline for future open academic work on terminal agents at https://github.com/hamishivi/tmax.

Hamish Ivison, Junjie Oscar Yin, Rulin Shao +3

α Allen Institute for AI · ω University of Washington

Synthetic Data Generation Terminal Agents LLM Agent Training RL for LLM Agents

Jun 22, 2026cs.CL

PhoneBuddy: Training Open Models for Agentic Phone Use

Phones are becoming an important execution surface for general-purpose agents, but training open models for reliable phone use remains difficult because the environment that matters at deployment, real devices running real apps, is slow, stateful, side-effectful, and hard to reset or verify, while scalable mock environments only approximate real behavior. We present PhoneBuddy, a training recipe and open-model line for agentic phone use that combines a real-app environment with a mock-app environment, PhoneWorld, which reconstructs runnable mock apps from real GUI usage structure. PhoneBuddy first builds a shared supervised fine-tuning stage from trajectories collected in both environments, then compares real-app RL against mixed RL across both environments. Across a 150-task human evaluation on real phones spanning apps, mini-apps, and cross-app workflows, task success rate improves from 36.67% after supervised fine-tuning to 40.67% after real-app RL and 45.33% after mixed RL. On AndroidWorld, the same progression rises from 60.3% to 77.2% to 83.2%. These results show that mock-app training is not a replacement for real-app RL, but a complementary source of scalable, resettable, and automatically checked interaction. The gains are strongest on app and mini-app tasks, while long-horizontal cross-app workflows remain an important open challenge.

Zhengyang Tang, Xin Lai, Pengyuan Lyu +23

Tencent Hunyuan · Gaoling School of Artificial Intelligence, Renmin University of China · The Chinese University of Hong Kong, Shenzhen +1

Computer-Use Agents Mobile GUI Agents LLM Agent Training RL for LLM Agents

Jun 19, 2026cs.AI

ARCO: Adaptive Rubric with Co-Evolution for Multi-Step LLM-Based Agents

Reinforcement learning for multi-step LLM agents often relies on scalar rewards that indicate success but cannot explain why a trajectory is good or bad. Rubric-based rewards improve interpretability through natural-language criteria, but existing methods score at the trajectory level and freeze the scorer behind a closed-source judge, leaving step-level credit assignment unresolved and the judge itself static. We propose ARCO (Adaptive Rubric CO-evolution), a rubric framework in which a same-scale model

μ

shares a backbone with two heads: a generation head that produces per-step criteria, and a score head that predicts rubric-conditioned step-level rewards. A trajectory decomposition constraint ties the sum of step rewards to the terminal outcome, enabling credit assignment without step-level labels, while

μ

and the policy

π

are jointly updated on on-policy data so that the rubric content and the scoring function co-evolve at the parameter level. Across HotpotQA, 2WikiMultiHopQA, and MuSiQue with two open-source backbones, ARCO improves the best EM in every setting over strong outcome-, rubric-, and process-reward baselines, and analyses show that its rubrics are step-specific, robust to design choices, and useful for diagnosing agent behavior. Codes and data are available at https://github.com/zihangtian/ARCO.

Zihang Tian, Jingsen Zhang, Rui Li +3

Renmin University of China

Process Reward Models Credit Assignment in RL Step-Level Credit Assignment in RL Rubric-Based RL Reward Design for RL RL for LLM Agents

Jun 18, 2026cs.LG

Connect the Dots: Training LLMs for Long-Lifecycle Agents with Cross-Domain Generalization Via Reinforcement Learning

This work presents a general framework for training large language models (LLMs) to "Connect the Dots" (CoD), a meta-capability required by long-lifecycle agents: as an LLM-based AI agent gets deployed in an environment, it solves a long sequence of tasks while continuously exploring the environment, learning from its own experiences, and iteratively self-updating its context about the environment, thereby achieving progressively better performance on future tasks conditioned on the updated context. Major components of the CoD framework include: (1) algorithm design and infrastructure for end-to-end reinforcement learning (RL) with long rollout sequences interleaving solve-task and update-context episodes; (2) tasks and environments for incentivizing and eliciting the targeted meta-capability in LLMs during training, as well as for faithfully measuring progress during evaluation. We present proof-of-concept implementations of the CoD framework, including a GRPO-style RL algorithm with fine-grained credit assignment, as well as tasks and environments tailored to the targeted meta-capability (rather than domain-specific LLM capabilities or standard task-by-task RL). Empirical results with Qwen3-8B and Qwen3.6-27B validate the efficacy of end-to-end RL training in the CoD setting, and demonstrate the potential for out-of-distribution generalization -- within the training domains, across different domains, and from CoD to Ralph-loop settings -- of the elicited meta-capability. Our investigation of CoD connects several lines of prior work, and opens up new opportunities for advancing LLMs and AI agents. To facilitate further research and applications, we release our implementations at https://github.com/agentscope-ai/Trinity-RFT/tree/research/cod_v2/examples/research_cod.

Yanxi Chen, Weijie Shi, Boyi Hu +6

Alibaba Group · Hong Kong University of Science and Technology · The University of Hong Kong

Continual Learning for LLM Agents Cross-Domain Generalization LLM Agent Training Long-Horizon LLM Agents RL for LLM Agents

Jun 18, 2026cs.AI

MetaResearcher: Scaling Deep Research via Self-Reflective Reinforcement Learning in Adversarial Virtual Environments

Deep research agents have demonstrated remarkable capabilities in autonomous information gathering and synthesis, yet their training remains constrained by the static nature of simulated environments, the limits of fact-retrieval-only task designs, and the inefficiency of outcome-based reinforcement learning. In this work, we propose MetaResearcher, a novel framework that scales deep research agent training across four synergistic dimensions. First, we introduce an Evolving Virtual World that injects temporal dynamics and adversarial misinformation into the training environment, forcing agents to develop source credibility assessment and temporal conflict resolution skills. Second, we design Discovery-Oriented Tasks -- including hypothesis generation and contradiction resolution -- that transcend simple fact retrieval and push agents toward genuine research behaviors. Third, we propose a Self-Reflective Meta-Reward mechanism within the GRPO framework that jointly optimizes for answer correctness, search path efficiency, reflection depth, and tool call diversity, directly addressing the repetitive action loop problem observed in prior work. Fourth, we introduce a Heterogeneous Multi-Agent Swarm architecture comprising specialized Scout, Filter, and Synthesizer models that learn collaborative research strategies through coordinated reinforcement learning. Built upon the LiteResearcher infrastructure, MetaResearcher requires zero marginal API cost for training while targeting substantial improvements in both benchmark performance (GAIA, Xbench-DS) and epistemic robustness under adversarial conditions. We present the complete framework design, training methodology, and planned experimental validation.

Wei Yu, Suxing Liu, Minjie Yu +4

School of Digital Arts, Jiangxi Arts & Ceramics Technology Institute, Jingdezhen 33001, China · Universiti Sains Malaysia, 11800 USM Penang, Malaysia

Reinforcement Learning Deep Research Agents RL for LLM Agents

Jun 16, 2026cs.LG

EnvRL: Learn from Environment Dynamics in Agentic Reinforcement Learning

Reinforcement learning (RL) has emerged as a powerful paradigm for training Large Language Models (LLMs) as agents. However, conventional RL methods for long-horizon agentic tasks often struggle with sparse outcome rewards. Intuitively, this overlooks the rich environment dynamics information contained in rollout interaction trajectories. We argue that the interaction experience inherently serves as an implicit supervision signal, reveals the underlying transition mechanisms of the environment, and enables the agent to construct a more accurate internal model of the environment.. Therefore, in this work, we investigate how to leverage this additional signal to improve policy learning. Specifically, we propose EnvRL, a framework that incorporates environment dynamics learning into agentic RL via two auxiliary objectives: state prediction and inverse dynamics. By jointly optimizing with the primary RL objective, we encourage the agent to internalize environment dynamics from its own interaction experience. Extensive experiments on two long-horizon agentic benchmarks demonstrate that EnvRL achieves significant improvements on success-rates over RL-only baselines, e.g., when trained with GRPO, lifting Qwen-2.5-1.5B-Instruct from 72.8% to 77.4% on ALFWorld, and from 56.8% to 67.0% on WebShop.

Zhitong Wang, Songze Li, Hao Peng +4

Department of Computer Science and Technology, Tsinghua University · Shanghai AI Laboratory

Agentic RL World Model Learning Policy Learning RL for LLM Agents

Jun 16, 2026cs.AI

Closing the Feedback Loop: From Experience Extraction to Insight Governance in Verbal Reinforcement Learning

Training-free verbal reinforcement learning enables LLM agents to learn from world feedback -- objective signals such as dynamic task outcomes, market returns, or demand forecasts -- by extracting verbal rules from experience and injecting them as context, updating the agent's behavior without parameter changes. However, in non-stationary environments these agents face a retention-forgetting dilemma: retaining stale insights causes negative transfer, while discarding them causes catastrophic forgetting when conditions recur. We identify four requirements for navigating this dilemma -- outcome-driven evaluation, persistent structured evidence, non-monotonic knowledge lifecycle, and compositional governance -- and show that existing methods invest heavily in experience extraction while underinvesting in insight governance. We propose a three-layer architecture -- rules, evidence, and skills -- connected by a feedback-driven curation loop that closes the governance gap. Rules capture distilled experience from world outcomes; evidence logs track each rule's reliability across episodes; skills govern which rules to apply, how to resolve conflicts, and when to abstain. On financial forecasting as a case study, where world feedback is naturally abundant, noisy, and non-stationary, we show that the same accumulated experience either degrades performance below the zero-shot baseline or dramatically improves accuracy and risk-adjusted returns, depending on whether the curation loop is present.

Yanwei Cui, Xing Zhang, Yulong Zhang +4

AWS Generative AI Innovation Center · Amazon Web Services (AWS) · BingX Group Limited.

Continual Learning for LLM Agents Non-Stationary RL Reinforcement Learning LLM Agent Memory RL for LLM Agents

Jun 15, 2026cs.CL

Context-Aware RL for Agentic and Multimodal LLMs

Large language models (LLMs) often fail when answering requires identifying a small but decisive piece of evidence within a long or complex context, such as a single line in a tool trace or a subtle detail in an image. We propose ContextRL, a context-aware reinforcement learning (RL) method that improves long-horizon reasoning and multimodal performance through an \emph{indirect} auxiliary objective. Instead of supervising only the final answer, ContextRL presents the model with a query, an answer, and two highly similar contexts, and rewards it for selecting the context that supports the query--answer pair, thereby encouraging fine-grained grounding. We construct contrastive context data in two domains: for coding agents, trajectories serve as contexts, yielding 1k pairs built via condition filtering; for multimodal reasoning, images serve as contexts, yielding 7K pairs built via generative editing and similarity search. ContextRL achieves average gains of +2.2% over standard GRPO on 5 long-horizon benchmarks, and +1.8% across 12 diverse visual question answering benchmarks. To disentangle the effect of the proposed objective from that of additional data, we compare against data-augmentation baselines that repurpose the same contrastive contexts as standard query--context--answer examples. These baselines provide little to no improvement, showing that the gains arise from the proposed context-selection objective rather than from the contrastive data alone.

Peiyang Xu, Bangzheng Li, Sijia Liu +4

1Princeton University · 2UC Davis

Multimodal Grounding Contrastive RL RL for VLMs RL for LLM Agents

Jun 15, 2026cs.AI

OpenClaw-Skill: Collective Skill Tree Search for Agentic Large Language Models

Equipping Large Language Model (LLM) agents with effective skills is crucial for solving complex tasks in real-world systems like OpenClaw. In this work, we aim to develop a framework that automatically constructs such reusable skills to enhance LLMs in tool use, multi-step reasoning, and dynamic environment interaction. To this end, we propose Collective Skill Tree Search (CSTS), a novel tree-search-based skill construction framework that constructs structured, diverse and generalizable tree of skills. The core idea of CSTS is to leverage collective intelligence to jointly search, identify and compose effective skills via two iterative phases: Collective Skill Node Generation (CSN-Gen) and Collective Skill Node Assessment (CSN-Assess). CSN-Gen exploits collective knowledge from multiple models to explore diverse candidate skills for each subtask, enabling comprehensive skill exploration. CSN-Assess employs multiple models as judges to evaluate and select skill nodes with two scoring mechanisms: (1) collective quality scoring that aggregates independent evaluations to produce a robust estimate of skill effectiveness, and (2) collective transferability scoring that explicitly verifies whether a skill generalizes well across different models. With CSTS, we construct a set of comprehensive tree of skills along with skill-augmented training data, enabling models to effectively learn and utilize skills. Besides, we introduce Collective Skill Reinforcement Learning, which actively selects multiple relevant skills from the tree to broaden solution-space exploration, avoid being trapped by a single skill and its resulting homogeneous or suboptimal solutions. As a result, our trained model, OpenClaw-Skill, exhibits outstanding agentic capabilities in long-horizon planning, tool use and generalization over challenging benchmarks.

Tianyi Lin, Chuanyu Sun, Jingyi Zhang +6

The Hong Kong Polytechnic University · Nanyang Technological University · Tsinghua University +2

LLM Agent Skill Learning Long-Horizon LLM Agents RL for LLM Agents

Jun 15, 2026cs.CL

PathRouter: Aligning Rewards with Retrieval Quality in Agentic Graph Retrieval-Augmented Generation

Agentic GraphRAG trains language-model agents to iteratively retrieve and reason over graph-structured evidence, enabling more accurate and context-aware decision-making by efficiently navigating complex information networks. However, outcome-only reinforcement learning suffers from \textit{\textbf{answer-path reward aliasing}}, where correct answers may come from shortcuts rather than useful evidence paths. It also exhibits \textit{\textbf{search-update ambiguity}}, as scalar trajectory-level feedback does not indicate which retrieval actions to adjust. To mitigate these shortcomings, we present PathRouter, a path-aware training framework for agentic GraphRAG. PathRouter jointly evaluates each trajectory along answer correctness and evidence-path overlap, yielding four trajectory categories with differentiated GRPO advantage scaling that suppresses shortcut reinforcement while preserving evidence-seeking behavior. For evidence-poor trajectories, a frozen gold-evidence teacher provides token-level KL guidance on reasoning and search-query tokens, excluding answer tokens to avoid direct response imitation. Experiments on six QA benchmarks across three model sizes show that PathRouter consistently improves answer F1 and evidence-path overlap, achieving average F1 gains of 3.1 on 3B and 4.9 on 7B models compared to a strong baseline.

Bo Wang, Heyan Huang, Yaolin Li +6

Beijing Institute of Technology · Joy Future Academy

Reward Shaping Graph Retrieval-Augmented Generation Agentic RAG RL for LLM Agents

Jun 14, 2026cs.CL

SciOrch: Learning to Orchestrate Expert LLMs for Solving Frontier Multimodal Scientific Reasoning Tasks

Frontier scientific reasoning remains a major challenge for large language models (LLMs), where even the strongest commercial systems fall short of expert-level performance. A closer look at model behavior reveals substantial complementarity that single-model evaluation hides: different frontier models excel on different question types, and no single model captures the full picture. We present SciOrch, a framework that trains a lightweight 8B model to orchestrate frontier LLMs for scientific reasoning. The orchestrator decomposes each question, delegates sub-problems to selected commercial models through API calls, and synthesizes a final answer. Training such an orchestrator is fundamentally harder than conventional agentic RL: each action triggers an API call that is expensive in both dollar cost and latency, making standard online rollouts infeasible. We address this with MCTS-based approach, producing diverse orchestration trajectories, extracting per-node single-turn samples, and optimizing the orchestrator with GRPO-style training. On a 240-question test set spanning SGI-Reasoning and Scientists' First Exam, SciOrch reaches 56.66% average accuracy, outperforming the strongest single commercial model by 3.74% and the strongest multi-agent baseline by 3.33%. It also attains the best accuracy on both SGI and SFE with less than half the API cost of typical multi-agent methods.

Jingru Guo, Xiangyuan Xue, Lian Zhang +6

1Imperial College London · 2The Chinese University of Hong Kong · University of Illinois Urbana-Champaign +4

LLM Agent Orchestration Multimodal Reasoning Scientific Reasoning in Language Models Large Language Model-Guided Scientific Discovery RL for LLM Agents

Jun 13, 2026cs.AI

Reward Hacking in Language Model Agents: Revisiting AI Safety Gridworlds

Reward hacking, where AI systems exploit misspecified objectives to achieve high reward without satisfying intended goals, remains a central challenge in AI safety. Yet most known instances have been discovered post hoc in frontier systems where controlled study is impractical. We adapt the AI Safety Gridworlds framework into a text-based evaluation suite that reformulates classic reinforcement learning safety tasks for language-based agents. Across frontier and mid-scale models, we find that specification gaming emerges zero-shot: models systematically achieve high observed reward while underperforming on hidden safety objectives, and even apparently safe behaviors can reflect misunderstanding rather than principled safety. Reinforcement learning does not correct these failures: direct reward optimization widens the gap between observed and hidden reward, as the model's initial competence causes it to lock into locally rewarding strategies before discovering safer alternatives. This pattern persists across model scales (1.5B--14B) and is not resolved by finer credit assignment, exploration prompts, or entropy regularization. Our results show that reward hacking arises naturally when optimizing proxy objectives with capable language model agents and resists standard mitigations, suggesting that proxy-reward failures in agentic settings may require approaches beyond standard exploration and credit-assignment fixes. To facilitate reproducibility, the code for this work is available at our public repository.

Ömer Veysel Çağatan, Xuandong Zhao

KUIS AI Center, Koc¸ University · Istanbul, Türkiye · University of California, Berkeley +1

Reward Hacking Language Model Safety Evaluation AI Agent Safety Benchmarks RL for LLM Agents

Jun 13, 2026cs.CL

Ling and Ring 2.6 Technical Report: Efficient and Instant Agentic Intelligence at Trillion-Parameter Scale

Efficient and scalable agentic intelligence requires models that can deliver both low-latency responses and strong reasoning capabilities while remaining practical to train, serve, and deploy. In this report, we present Ling-2.6 and Ring-2.6, a family of models designed to address this challenge at scale. Ling-2.6 is optimized for instant response generation and high capability per output token, whereas Ring-2.6 is tailored for deeper reasoning and more advanced agentic workflows. Instead of training from scratch, we upgrade the Ling-2.0 base model through architectural migration pre-training and large-scale post-training. This upgrade is guided by a unified co-design of model architecture, optimization objectives, serving systems, and agent training environments, enabling improvements in both model capability and deployment efficiency. At the architectural level, we introduce a hybrid linear attention design that integrates Lightning Attention with MLA, improving the efficiency of long-context training and decoding. To further enhance token efficiency, we optimize capability per output token through Evolutionary Chain-of-Thought, Linguistic Unit Policy Optimization, bidirectional preference alignment, and shortest-correct-response distillation. For agentic capabilities, we propose KPop, a reinforcement learning framework designed to support stable training of Ring-2.6-1T on large-scale environment-grounded data. KPop improves training efficiency through asynchronous scheduling across coding, search, tool use, and workflow execution, enabling scalable learning from complex agent-environment interactions. Together, Ling-2.6 and Ring-2.6 provide a practical pathway toward efficient, scalable, and open agentic systems. We open-source all checkpoints in the 2.6 family to support further research and development in practical agentic intelligence.

Ang Li, Ben Liu, Bin Han +215

Efficient Language Model Training Linear Attention LLM Agent Training Efficient Language Model Inference Language Model Post-Training RL for LLM Agents

Jun 12, 2026cs.CL

AgentSpec: Understanding Embodied Agent Scaffolds Through Controlled Composition

LLM agents are increasingly built not as single model calls, but as scaffolded systems that combine reasoning, memory, reflection, action execution, and learning. While such scaffolds often improve performance, they are often embedded in tightly coupled pipelines, making it difficult to isolate component contributions, compare alternative designs, or understand how module interactions shape agent behavior. We introduce AgentSpec, a modular specification framework that represents embodied agents as typed compositions of reusable policy components with standardized interfaces. AgentSpec standardizes the interfaces among perception, memory, reasoning, reflection, action, and optional learning, enabling components to be swapped and recombined under controlled conditions. We instantiate this framework across DeliveryBench, ALFRED, MiniGrid, and RoboTHOR, and analyze reasoning, memory, reflection, and reinforcement-learning modules across model backbones. Our results show that agent performance is governed by scaffold compatibility and interaction effects rather than isolated module strength. In particular, structured multi-granularity memory improves long-horizon state tracking, reasoning and memory interact non-uniformly across environments, reflection trades off correction and cost, and RL-trained policies compose best when optimized with deployment-time scaffold structure. AgentSpec provides a controlled foundation for studying, comparing, and designing composable LLM agents. Our code, baselines and interactive playground are publicly available at https://agentspec-embodied.github.io.

Jixuan Chen, Jianzhi Shen, Haoqiang Kang +10

University of California, San Diego · 2Johns Hopkins University · University of Washington +1

LLM Agent Orchestration Embodied AI Long-Horizon LLM Agents RL for LLM Agents Memory-Augmented Language Model Agents

Jun 12, 2026cs.CL

Retrospective Progress-Aware Self-Refinement for LLM Agent Training

Long-horizon LLM-based agents receive rich environmental observations during interaction, yet outcome rewards provide limited explicit supervision about how individual actions advance task completion. We investigate whether agents can turn this interaction evidence into useful training signals through retrospective progress assessment. A WebShop pilot shows that direct progress prompting reduces task success, whereas hindsight-annotated demonstrations improve it. We introduce RePro, Retrospective Progress-Aware Training, with a forward-then-reflect rollout: the agent estimates progress while acting, then reassesses each step using the completed trajectory and outcome. After warmup with externally generated demonstrations, policy optimization combines self-generated progress differences, online-retrospective alignment, and format rewards with environment feedback, requiring neither a separate process reward model nor ongoing teacher annotation. Experiments on WebShop, ALFWorld, and Sokoban show that RePro enhances the Qwen family's performance, with up to 11.57% success rate gains.

Xinbei Ma, Congmin Zheng, Jiyang Qiu +11

Shanghai Jiao Tong University · OPPO Research Institute

Reward Shaping LLM Agent Training Long-Horizon LLM Agents Policy Learning RL for LLM Agents

Jun 12, 2026cs.AI

Closing the Reflection Gap: A Free Calibration Bonus for Agentic RL

LLMs are increasingly deployed as agents that interact with external environments and observe feedback such as execution results, error messages, and tool outputs. A well-functioning agent should be able to leverage this feedback to accurately assess its own performance. Yet we find a persistent reflection gap: LLM agents tend to mis-assess their own outputs after observing concrete environment feedback -- even for questions they correctly answered -- and standard RL barely helps due to a credit-assignment mismatch. To close this gap, we propose RefGRPO, a simple yet effective fix that augments standard RL algorithms with two key ingredients: a free calibration bonus computed by contrasting the agent's own reflection with the actual outcome (requiring no additional reward model, LLM judge, or external annotation), and a dynamic schedule on its coefficient. Compared to standard RL baselines, our method simultaneously improves reflection calibration (e.g., reduces underconfidence rate

44.4\% \to 7.7\%

) and task accuracy (e.g.,

75.1\% \to 76.5\%

) on text-to-SQL across five benchmarks. The resulting calibrated reflection turns the agent into its own verifier grounded in environment feedback, which further enables (i) better self-improvement that uses reflections as pseudo-rewards without outcome supervision, and (ii) more effective test-time selective prediction by committing only to rollouts flagged as correct.

Yinglun Zhu

University of California, Riverside

LLM Agent Evaluation Language Model Calibration Self-Improving Agents RL for LLM Agents

Jun 11, 2026cs.AI

From Verdict to Process: Agentic Reinforcement Learning for Multi-Stage Fact Verification

Recent approaches combining Large Language Models (LLMs) with retrieval-augmented reasoning have shown promise for automated fact verification. To process complex claims, these verification pipelines typically execute multi-stage workflows that coordinate tightly coupled modules, including claim decomposition, evidence gathering, and verdict prediction. However, existing methods optimize individual stages in isolation or rely on fixed heuristics, which limits adaptive coordination among stages and can lead to suboptimal outcomes. In this work, we propose ProFact, an agentic reinforcement learning framework for end-to-end optimization of multi-stage fact verification trajectories. ProFact trains a unified policy to coordinate claim decomposition, evidence seeking, answer generation, and verdict prediction. To address the sparse and delayed supervision provided by final veracity labels, ProFact introduces process-aware rewards that provide stage-level learning signals throughout the verification process. Empirical evaluation shows that ProFact consistently outperforms strong baselines in both verification performance and inference efficiency. These results highlight the effectiveness of process-aware trajectory optimization for multi-stage fact verification.

Rongxin Yang, Shenghong He, Siyuan Zhu +1

School of Computer Science and Engineering, Sun Yat-sen University, Guangzhou 510006, China

Agentic RL Process Reward Models Automated Fact-Checking RL for LLM Agents