Scaling interactive and verifiable environments is critical for training terminal agents. As frontier models become more capable, environments synthesized from scratch become less challenging and thus provide limited learning signals. Recent co-evolution methods iteratively synthesize environments near the model's learnable frontier based on weaknesses exposed during rollouts. However, their dependence on on-policy rollouts limits generalization and the continuous provision of learning signals as the model becomes stronger. In this paper, we propose environment evolution, which incrementally increases environment difficulty off-policy and schedules the evolved environments generation by generation during training to provide continuous learning signals. We derive three evolution directions that influence environment difficulty from the multi-turn learning objective and then implement evolution along these directions through a loop-engineered multi-agent harness. Quantitative rollout experiments with Hy4 preview, Claude Opus 5, and GPT-5.6 Sol show that environment evolution consistently produces more difficult environments. We validate its effectiveness on Qwen3.6-27B and Qwen3.6-35B-A3B through simple long-horizon RL training, improving their performance by 14.4 and 18.0 percentage points on Terminal-Bench 2.1, respectively.
LLM-as-judge is essential for evaluating open-ended text and steering post-training, yet improving the judge itself typically relies on expensive annotations, reward models, or distillation from stronger teachers. In this work, we eliminate external gold supervision from the RL training reward: the model's own evaluative capability generates learning signals for its optimization -- a closed-loop setting of bounded recursive self-improvement (RSI) termed Recursive Self-Evaluation (RecurSE). We study two central questions: when can self-improvement occur, and when must it stop? First, RecurSE pairs a trainable judge evaluating candidate responses under per-rule rubrics (Pass 1) with a synchronized policy-copy checker that audits the judge's reasoning against meta-rubrics to supply a scalar process reward (Pass 2). To enable learning, interface decoupling structurally isolates the checker's scalar score from the judge's verdict tokens, eliminating a degenerative token-copying shortcut that inflates self-assigned rewards. Second, because unanchored recursive learning is inherently bounded, Pairwise Advantage Validity (PAV) serves as an unbiased validation monitor that jointly tracks judge accuracy and checker fidelity to reliably identify the optimal early-stopping window. Across Qwen3.5-9B, Gemma-4-E4B-it, and Qwen3.6-27B, RecurSE achieves consistent generalization gains across held-out medical, pairwise, summarization, and professional benchmarks. Ablations demonstrate that synchronized judge-checker co-evolution outperforms frozen checkers, external meta-judges, self-consistency, and scaled teacher distillation. Furthermore, preference pairs curated by our judge effectively enhance downstream policy alignment. Bounded RSI for LLM-as-judge is thus viable when self-produced reward validity is explicitly decoupled and monitored.
Modern agents operate inside agent harnesses that manage tools, context, and control flow, making the harness a critical part of the agent system. Our original Agent Lightning introduced a disaggregated architecture that connects arbitrary agents to RL training through an LLM endpoint proxy, an approach later adopted by frameworks such as verl Uni-Agent, AReaL 2.0, slime, and Polar. We refer to this paradigm as harnessed agentic RL, where the deploy-time harness directly participates in model post-training. Harnessed agentic RL differs fundamentally from traditional agentic RL: the harness, rather than the training engine, owns the environment interaction loop, while the trainer observes only sequences of LLM request-response pairs. This introduces challenges in retokenization, sample merging, advantage calculation, loss normalization, and backend scheduling, which can substantially affect training stability and effectiveness. We present Agent Lightning v1.0, a lightweight framework for harnessed agentic RL implemented in approximately 3,500 lines of code. It supports arbitrary agent harnesses and serves as a practical testbed for studying these challenges. We evaluate it on instruction-following, search, and coding agents, and provide a complete reproducible pipeline for coding-agent RL. Using only 6K training examples and modest compute, RL improves Qwen3.5-9B on SWE-bench Verified from 41.8% to 56.4%, a 14.6-point absolute gain. We release the complete workflow and training scripts to facilitate reproducible research on harnessed agentic RL.
Reinforcement learning (RL) for terminal agents needs executable training environments with reliable rewards and useful difficulty. Fixed recipes such as few-shot, Self-Instruct, and Evol-Instruct apply the same prompting policy to every seed, even when the current policy would benefit from a harder, easier, or simply different task. We present Envs-FORGE, a prompting policy that converts verifier rewards into per-seed environment-synthesis actions. Envs-FORGE estimates seed pass rates, scores six projection--direction actions around a target learning frontier, and solves a per-seed mixed-integer linear program (MILP) to choose the action that conditions generation. The selected action drives synchronized rewriting of the instruction, fixtures, oracle solution, tests, and Docker environment; only gold-verified bundles enter RL training. The indexed MILP form also supports optional soft skill coverage for portfolio planning. On Qwen 3.5 35B, Envs-FORGE improves Pass@1 over Base by 9.2 percentage points on tb-core (40.0% to 49.2%) and 6.4 points on tb-2.0 (23.0% to 29.4%), exceeding the strongest fixed-recipe baseline by 2.4 and 2.1 points. It reaches 77.1% on SWE-bench Verified versus 73.4% for Base, and improves tb-core by 6.8--9.2 points across the evaluated 4B--35B models. All synthesis methods export 100 verified environments and use 2.27M--2.88M synthesis tokens, placing the comparison at the same downstream training-set size and the same operational scale. The source code is available at https://github.com/DataArcTech/DataArc-SynData-Toolkit/.
Zile Zhou, Huining Yuan, Weichen Zhang +2cs.CV cs.AI
Existing Vision-Language Models (VLMs) exhibits a critical bottleneck in robust spatial reasoning. Recent reinforcement learning (RL) methods aim to close this gap with verifiable outcomes, yet they suffer from poor credit assignment across intermediate reasoning steps. Concurrently, structured reasoning approaches overlook the critical depth perception necessary for comprehensive 3D understanding. To address these challenges, we propose SCOUT (Structured Chain-Of-Thought Utilizing Process-Supervised RL Training). Specifically, we design a structured Chain-of-Thought (CoT) framework that explicitly models 3D environmental perception to ensure robust spatial understanding and reasoning. Furthermore, we introduce a novel RL algorithm featuring multi-objective process rewards and a tailored advantage estimation method, facilitating fine-grained credit assignment across distinct segments of the reasoning trajectory. To support our framework, we develop SCOUT-24k, a structured spatial reasoning CoT dataset synthesized through a customized pipeline. Extensive evaluations demonstrate that SCOUT-3B improves upon baseline models by 16.85% and 6.3% on general spatial benchmarks and complex spatial reasoning tasks respectively. Notably, our larger SCOUT-7B even outperforms GPT-4o by a margin of 4.28%. Moreover, despite being trained exclusively on single image, SCOUT-7B exhibits robust out-of-domain generalization to multi-image and video scenarios. These empirical results render SCOUT as a critical step towards next generation of spatially-aware VLMs.
Reinforcement learning (RL) post-training of large language models (LLMs) is computationally intensive and involves complex system pipelines with substantial debugging overhead. In practice, factors such as framework adaptation, numerical precision, and operator implementation can cause failures, including gradient overflow and loss divergence. Reproducing such failures directly on large models requires considerable time and computational resources. This paper systematically analyzes failures encountered during large-scale RL training on the Huawei Ascend platform, summarizes representative failure types, and identifies three model-side factors relevant to fault reproduction. Based on these factors, we propose a proxy-model construction method for low-cost fault investigation and auxiliary diagnosis. It employs structure-preserving, clustering-based expert pruning to select representative experts while retaining the model's backbone architecture, routing mechanism, and basic task capabilities. Our experimental results show that the proxy models reduce accelerator requirements by 50%-87.5% and achieve up to a 33.3x reduction in per-step NPU-hour cost, while preserving major training dynamics and reproducing fault responses consistent with the original models. Overall, the proxy models can serve as low-cost surrogates for fault reproduction, targeted validation, and auxiliary diagnosis in RL post-training.
Recent research empowers Large Language Models (LLMs) as multi-turn search agents to iteratively retrieve and generate outputs until complex tasks are solved. However, the contexts of multi-turn search agents are lengthy and complex. For example, the retrieved set of documents in each turn would inevitably introduce irrelevant information that distracts LLMs, referring to \textit{context interference}, potentially hindering the reliability and efficiency of search agents. Therefore, we conduct a systematic study on context interference in multi-turn search agents, focusing on investigating i) which parts of the context of search agents will contribute to the context interference, ii) how to refine the contexts of search agents to mitigate the interference, and iii) can incorporating context refinement into search agent training yield further improvements. We reveal that interference primarily arises from the latest retrieved documents. Based on the explored findings, we then introduce a distill-based context refiner to dynamically mitigate context interference for multi-turn search agents. Finally, we validate that incorporating context refinement into RL training pipelines of search agents can significantly enhance both reliability and efficiency. This study highlights the importance of mitigating context interference of search agents, inspiring a novel paradigm of ``refine context and then generate'' for AI agents.
Reinforcement learning (RL) for large language models is moving toward multi-turn agentic workloads, where rollout tasks repeatedly pause for external environments, resume with growing contexts, and finish at highly variable times. In this setting, RL training goodput, measured by training throughput, matters more than raw GPU occupancy: GPU waiting and repeated prefill recomputation are pure overhead. We present TideRL, a readiness-aware elastic RL system with Continuous Task Batching, Resource-Aware Ref-Actor Pipelining, and Elastic Resource Scaling. CTB preserves useful rollout state, $\textrm{RA}^2\textrm{P}$ selects between decoupled streaming and colocated aggregation from the ready backlog and arrival interval, and ERS moves ranks between rollout and training using the same readiness signals. Across text-only and multi-modal agentic workloads, TideRL improves RL training goodput by up to 5.6$\times$ over synchronous baselines and over 33% over asynchronous baselines, while reaching similar task performance. It also improves KV cache hit rate by 1.58$\times$, reduces per-step training time by up to 44.3%, and cuts total waiting time by up to 77.6%.
Modern LLM coding agents such as Claude Code and OpenHands share a common inefficiency: they spend much of their token budget finding the file to patch, rather than patching it. On SWE-Bench Verified, a 30B OpenHands agent averages 23 rounds and 631K tokens per resolved issue, with many calls spent on grep, glob, and view_file during repository exploration. We introduce CodeGrep, a 14B retrieval agent trained end-to-end with GRPO to issue multi-turn parallel grep, glob, and read tool calls and return candidate files to a frozen downstream coding agent. On all 500 SWE-Bench Verified instances, CodeGrep preserves resolve rate while substantially improving efficiency: 27.0% versus 25.8% for the no-retrieval baseline, with 15% fewer rounds and 19% fewer tokens on resolved instances. Across retrievers, downstream utility follows a precision threshold: BM25 with precision 0.375 degrades the agent, Jina with precision 0.445 is neutral, and CodeGrep with precision 0.677 crosses the threshold at which retrieval begins to reduce rollout cost. To enable this study, we mine supervision from 67K open-source agent trajectories using CATM and build a Git-worktree environment for multi-turn agent RL. In our setting, applying the efficiency signal at the advantage layer rather than the reward layer reduces KL drift and translates cleanly into downstream efficiency. We will release the model, training pipeline, RL environment, and evaluation harnesses.
Dense per-step supervision is an appealing remedy for sparse-reward, long-horizon LLM agents: reward the agent for predicting its next observation, and memory should follow. We show that under group-normalized RL (GRPO), this recipe does not merely fail -- it destroys the policy. Across Qwen3-1.7B/4B/8B on ALFWorld, a potential-based prediction reward drives every run into a degenerate absorbing state (prediction accuracy -> 1.0, task success -> 0,episode length pinned at the horizon): the "dark room" pathology, built automatically by the optimizer. A single-factor ablation localizes the cause -- removing only GRPO's std normalization turns the same reward from catastrophic (0%) into baseline parity -- and a two-line proposition explains why: in all-fail groups the z-scored advantage is invariant to the shaping coefficient, so bounded rewards become unbounded pressure and annealing cannot help. Our central insight generalizes this: what z-scoring amplifies is a dense signal's within-group variance while all-fail groups dominate, so signals whose variance decays by mastery are structurally amplifier-safe.This variance-profile criterion retrodicts our collapses, carries preregistered predictions for arms that had not yet run, and is consistent with published reward-channel successes (a compatibility check, not an independent test). Finally, a controlled signal-delivery matrix (identical signal, varying only the consumption mechanism) shows the reward channel is at best neutral while the auxiliary-loss channel gains ~20 points -- and a shuffled-gold placebo matches the true-gold arm, so the gap survives without correct labels. Endpoints are single-seed; seed replication and group-size controls are preregistered and in progress.
Axel Højmark, Jérémy Scheurer, Evgenia Nitishinskaya +5cs.AI cs.CL cs.LG
Language models trained with reinforcement learning may learn to optimize the grader's judgment rather than the intended objective. This "reward-seeking" is difficult to measure because a model that pursues the grader's judgment and one that pursues the intended objective behave identically whenever the grader rewards the intended behavior. We measure reward-seeking using Contrastive Synthetic Document Finetuning to change a model's beliefs about what the grader rewards, putting those beliefs in conflict with what users or developers want, and measuring the rate at which the model adopts each party's preferred behavior. Applied to intermediate checkpoints of a capabilities-focused OpenAI o3 RL run, without safety training, we find that these checkpoints often side with grader preferences over those of users or developers on coding and alignment tasks. This tendency to side with the grader trends upward throughout RL training. For example, in an environment that forces a choice between keeping a promise to a supervisor and breaking it to complete the task, a late capabilities-focused o3 checkpoint breaks the promise 87% of the time when SDF documents say the grader rewards task completion, versus 9% when they say it rewards honesty (a choice its chain-of-thought often makes explicit). An earlier checkpoint is far less sensitive (40% vs. 24%). Our method also generalizes to reward-hacking models. A model organism trained to reward-hack (gpt-oss-120b) is more than twice as sensitive to grader preferences as the unmodified model, with the mean behavioral shift in favor of the grader rising from 33% to 86%. These results indicate that RL can increase reward-seeking over the course of training, producing models that may act against their developers' intentions when they believe that doing so leads to higher reward.
Post-training of frontier language models is conducted on curated task suites, and inevitably leaves a distribution shift between training and deployment environments. This exposes developers to generalization failures, which are relatively poorly understood. To better understand such generalization failures, we believe the community should construct clean demonstrations under simplified conditions. To facilitate this, we propose a simple and flexible way to construct language models which fail to generalize in controllable ways when subsequently trained with Reinforcement Learning (RL) on a given distribution of training tasks. Our construction uses Supervised Fine-Tuning on a dataset of a mixture of transcripts corresponding to a collection of 'conditional policies', which can each independently be assigned certain behaviors on each different task distribution, to obtain a model that is then well approximated as a 'mixture of conditional policies.' We observe that RL training then selects for policies that obtain the highest reward on the training distribution. This can produce striking behaviors: in a controlled setting with two distributions containing identical questions prepended with two different 'trigger strings', RL training on either distribution actively degrades performance on the other to zero, even though the underlying task is identical. We also use our construction to illustrate two novel ways in which generalization may fail in future language models, corresponding to distribution shifts of task coverage and temporal context respectively. While our construction is deliberately simple and may not closely resemble 'natural' generalization failures, the resulting 'model organisms' are of interest for alignment stress-testing and generalization science, and can be used as existence proofs that training success and generalization can come apart in structured ways.
Hamish Ivison, Junjie Oscar Yin, Rulin Shao +3cs.CL
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.
Reinforcement learning pipelines for Large Language Model (LLM) training often rely on manually redesigned environments between stages, requiring practitioners to heuristically infer which configuration will best improve the current policy. To automate this process, we propose the LLM-as-Environment-Engineer framework in which the current policy model analyzes failure trajectories together with contextual information and proposes modifications to the next-stage training environment configuration. We also introduce MAPF-FrozenLake, a controllable testbed whose generator exposes multi-dimensional environment configurations, making it suitable for studying and benchmarking environment redesign. On this testbed, we condition the environment engineer on structured summaries of policy behavior, failure cases, and environment statistics, from which it produces the configuration for the next training stage. With Qwen3-4B as the backbone, our framework achieves the strongest aggregate performance on our benchmarks, outperforming larger proprietary LLMs (e.g., GPT, Gemini) and fixed-environment training baselines. We further analyze which forms of context are most effective, finding that successful environment updates rely on failure evidence and preserve configurations that already work. Interestingly, the current RL checkpoint serves as a better environment engineer than the original base model, suggesting that policy learning improves the model's ability to diagnose its remaining weaknesses.
Zirui Cheng, Zeyu Shen, Thomas L. Griffiths +1cs.AI
People make decisions differently in strategic interactions. Some update beliefs like a Bayesian; others exhibit biases like motivated reasoning. Although creators of large language models use simulated humans for safety evaluations and training, they often fail to cover this breadth of human behavior. We argue that cognitive science and economics provide a convenient tool for doing so, making use of mathematical models of human decision-making. We propose an approach that we call Equation-to-Behavior Prompting for guiding large language models to match cognitive models, and evaluate this approach on persuasion games based on legal decision-making. We find that large models can approximate equation-based specifications -- Bayesian updating, affine distortion, motivated updating, and Grether's $α$-$β$ model -- using prompting, but small models fail to do so. However, training small models with reinforcement learning to adhere to mathematical rules, Equation-to-Behavior RL, reduces belief error by 26.5% in out-of-distribution parameterizations. We show that these simulations can help create diverse training environments; training small models to consider different kinds of decision-makers improves average belief change by 2.5%--12% over Bayesian-only training, even when persuading GPT-5-mini. Our work could improve human simulations for training and evaluation in increasingly realistic settings, and could also enable novel research into more complicated mathematical models of human decision-making.
Ruifeng Yuan, Chaohao Yuan, David Dai +4cs.CL cs.IR
Recent retrieval-augmented generation (RAG) approaches have demonstrated strong capability in handling complex queries, yet current research overlooks a critical challenge: different retrievers require fundamentally different query formulation strategies for optimal performance. In this work, we present the first systematic analysis of how LLMs can learn to adapt their query formulation strategies for different retrievers via reinforcement learning (RL). Our empirical study reveals that RL effectively teaches an LLM to tailor its queries to specific retriever characteristics. We discover that different retrievers exhibit surprisingly distinct optimal query styles (e.g., descriptive vs. question-like), suggesting strategies learned for one retriever ineffective for another. We further show that performance can be enhanced by incorporating retriever-specific human guidance and by scaling model size. To facilitate learning over multi-retrieval-step trajectories, we introduce a branching-based rollout technique that improves training stability. Our work provides the first empirical evidence and actionable insights for building truly retriever-aware RAG systems. Code and resources are available at https://github.com/LCO-Embedding/Envs-aware-Information-Retrieval.
Yucheng Li, Huiqiang Jiang, Yang Xu +14cs.LG cs.CL
Reinforcement learning (RL) has become a key component in modern large language models, yet the rollout stage remains the key bottleneck in RL training pipelines. Although Multi-Token Prediction (MTP) offers a natural solution to accelerate rollouts through speculative decoding, many studies have observed that MTP acceptance rates degrade significantly during RL training, leading to limited speedup performance. To address this bottleneck, we present Bebop, a systematic study of MTP in LLM post-training, and offer practical recipes to integrate MTP into large-scale RL pipelines. First, we reveal that the MTP acceptance rate is fundamentally bounded by the fluctuation of model entropy, which demonstrates a clear negative linear relationship with the rise of entropy in the RL stage. Second, we show that probabilistic rejection sampling largely alleviates the disturbance introduced by entropy in RL compared to greedy draft sampling. We further identify that the conventional MTP training objectives (cross-entropy or KL) are suboptimal in such settings, and therefore we propose a novel end-to-end TV loss that directly optimizes multi-step rejection sampling acceptance rate, yielding ~10% acceptance rate improvements, achieving up to 95% acceptance rates and up to 25% extra inference throughput gains across mathematical reasoning, code generation, and agentic tasks. Third, we test various online MTP training strategies during RL and show that pre-RL MTP training with e2e TV loss and rejection sampling achieves a consistent acceptance rate and speedup throughout the entire RL, eliminating the need for costly online MTP updating. We provide extensive experiments and analysis that validate our findings. Experimental results show our method achieves up to 1.8x end-to-end acceleration in async RL training of Qwen3.5, Qwen3.6, and Qwen3.7 models.