RL Post-Training

On-policy (self) distillation (OPSD) is increasingly adopted for language-model post-training. It strengthens the teacher with privileged information but can induce a privilege illusion: the student learns privilege-dependent behavior it cannot reproduce from its inference-time context, yet behaves as if the training-time privileged information remained available, ultimately degrading performance. In this paper, we identify information asymmetry between the privileged teacher and the student at inference as the root cause of this failure in OPSD. To resolve this asymmetry, we propose Dual-Anchored Policy Distillation (DAPD), a unified framework with two levels of anchoring. Dual-Path Anchoring (DPA) introduces a self-conditioned bridge and aligns reference and rollout behavior along two matched-information paths, preventing privilege-dependent behavior from being transferred to the inference-time student. Dual-Source Anchoring (DSA) applies these paths in both reference-to-rollout and rollout-to-reference directions, reducing reliance on privileged reference guidance while preserving correctness supervision. Extensive experiments show that DAPD significantly alleviates privilege illusion, outperforming OPSD on Qwen3-4B by +2.00 points on average across tasks. Notably, its gains persist across scales, reaching +2.69 at 4B and +2.78 at 32B.

Compact web agents can reduce deployment cost, but training them poses challenges in both data collection and post-SFT reinforcement learning (RL). Successful trajectories are expensive to collect and often contain inefficient detours. After supervised fine-tuning (SFT), full trajectory corpora are dominated by routine states; moreover, when group-relative RL is applied to web actions, inadequately designed action-level rewards can yield weak or misleading relative updates, while groups rejected as unsuitable for such updates receive no fallback learning signal. We present RMSWeb, a three-part recipe for Qwen3-VL-Instruct at 8B and 32B. Reflection-conditioned retries increase collection yield and shorten successful trajectories; failure-mode mining concentrates offline RL on critical states exposed by the SFT policy; and Salvage-DS combines an action-semantic polarized reward, contrast-and-competence-gated dynamic sampling, and an action-only anchor for rejected groups. Policies trained with reflection-collected data use up to 19.7% fewer action steps on solved tasks. On WebVoyager, Online-Mind2Web, and WebTailBench, RMSWeb improves over SFT by 2.4-7.0 points at 8B and 1.2-7.7 points at 32B. Our 8B model also achieves the strongest reported Online-Mind2Web result among similarly sized open-weight models in our comparison and a leading reported accuracy-cost trade-off on WebVoyager and WebTailBench, with the caveat that external evaluation protocols differ.

Error-penalized scoring rules ($+1$ for a correct answer, $-λ$ for a wrong one, $0$ for abstaining) are increasingly prescribed against hallucination: a rational agent facing such a rule answers exactly when its correctness probability exceeds Chow's threshold $t^\ast=λ/(1+λ)$. We prove that a KL-anchored gradient learner can do the opposite. When abstention is a discrete action, the reward gradient and the anchor's restoring force are throttled by the same gate-saturation factor and die together: under explicit conditions (among them, blanket answering loses score in expectation and prompts share a bounded readout) the model drifts toward refusing everything, its mean training reward rising to zero like $1/t$ in training time $t$, so the curve reads as improvement while coverage collapses. The advantage estimator compounds the failure: in its sparse-answer regime, group normalization silently replaces every designed penalty with an effective penalty of one, moving the learned threshold from $t^\ast$ to $1/2$. The repair is structural: train a mandatory confidence report with a strictly proper score plus a correctness reward, and abstain only at deployment by thresholding the report. The always-emitted report has no gate to saturate, so no shared factor can kill its reward gradient and its anchor together, and its calibrated optimum is attracting. Simulations confirm every prediction, and experiments on language models at two scales confirm the mechanism live: the rule silences questions the models demonstrably still solve within ten optimizer steps, an ablation isolates the cause, and report-level training raises coverage, accuracy, and calibration together.

For reinforcement learning (RL) in self-contained environments, a policy can get rewards by exploiting environment-specific regularities (tool schemas, grader parsing, task templates) rather than by acquiring transferable skill, and an in-distribution holdout shares those regularities. We argue that the discriminating question is behavioral, namely how a trained agent acts, and that cross-benchmark transfer is the right place to look for it. We post-train an open-weight mixture-of-experts model (Qwen3.5-122B-A10B) on 363 long-horizon Model Context Protocol (MCP) tasks across 27 categories, using a two-stage SFT-then-RL pipeline. Toolathlon performance informed the initial base-family and SFT-teacher choices, but no external-benchmark task or grader entered training and no external score informed the reward, training hyperparameters, trained-checkpoint selection, or stopping. At greedy pass@1, the trained model improves over the base on five reported external evaluations: Toolathlon (+9.6 pp), $τ^2$-Bench (+5.3 pp), BFCL-V4 (+3.5 pp), SWE-Bench Pro (+5.8 pp), and Terminal-Bench 2 (+2.8 pp). Both software-engineering benchmarks improve despite the training collection containing no software-engineering tasks. An exploratory paired-trajectory analysis identifies four recurring behavioral differences (more careful local-goal formation, building goal-relevant working state, keeping parent goals stable through local repairs, and verifying completion) that appear in analogous forms across office workflows and code. These results provide descriptive evidence that long-horizon multi-tool post-training can change ways of working that transfer beyond its training domain.

Pretraining navigation diffusion policies rely on large-scale expert demonstrations. These data are typically generated by a fully-informed oracle planner suited to a single nominal robot. This limits the policy's generalization to diverse embodiments and challenging scenarios (e.g., escaping dead ends or detouring long obstacles) that demand diverse local reactive behaviors with only onboard local observations. Post-training the policy with reinforcement learning (RL) offers a principled remedy. However, previous RL for diffusion approaches lead to only marginal improvements. This is because the intractable likelihood of diffusion policies renders policy gradients unstable in addition to inefficient policy exploration. To address these challenges, we propose a data-efficient diffusion RL post-training framework - GQRM (Group Q-score Reweighted Matching). Our framework introduces two complementary designs: (i) a self-bootstrapped exploration strategy with behavior perturbation that preserves the pretrained policy prior, and (ii) a group Q-score normalization mechanism that computes per-trajectory values on each state for efficient reweighted score matching. By conducting distributed online RL training across heterogeneous embodiments, the resulting fine-tuned policy, X-NavDP, achieves state-of-the-art cross-embodiment visual navigation performance, improving the overall success rate from 61.20% to 84.28% in simulation and 10% to 65% in real-world hard cases. The code and model are publicly available at https://yty-sky.github.io/x-navdp-project-page.

Supervised fine-tuning (SFT) can equip large language models (LLMs) with domain knowledge for high-performance computing (HPC) tasks such as data race detection and benchmark question answering. However, knowledge alone does not guarantee task-appropriate behavior: the same SFT model that correctly classifies 88.65% of C/C++ data race samples produces verbose, imprecise answers to factual queries, with 65.9% of MLPerf responses exceeding 40 characters. Reinforcement learning (RL) post-training addresses this gap by optimizing for task-specific rewards rather than token-level imitation. Yet HPC tasks exhibit extreme heterogeneity, with binary classification, factual QA, and semantic generation differing by 58x in answer length, spanning three distinct reward distributions, and showing widely varying SFT accuracy. This makes uniform-weight RL methods such as GRPO suboptimal. We propose HARGO, Heterogeneity-Aware Reward-Guided Optimization, which introduces per-response importance weighting via confidence-modulated advantage: computing a discrimination signal from group-level reward contrast and a confidence signal from reference model log-probabilities, then modulating the advantage before computing per-response weights, without requiring task-type labels. Across four HPC tasks and nine methods, HARGO achieves the best performance on all three primary metrics: WinRate 54.62%, Data Race F1 91.30%, and PLP Similarity 0.8558. Ablation confirms complementary contributions from both signals. HARGO establishes the best overall alignment quality among compared methods for heterogeneous HPC tasks.

Recent advances in post-training Large Language Models (LLMs) increasingly rely on Reinforcement Learning with Verifiable Rewards (RLVR) or On-Policy Self-Distillation (OPSD). While OPSD provides dense, logit-level supervision, it inherently suffers from exposure bias due to the privileged information of the self-teacher. In multi-turn agentic settings, this leads to reasoning route convergence and the loss of clear optimization directions. To tackle these challenges, we introduce Contrastive Reinforced Policy Optimization (CRPO), which reformulates agentic OPSD from a contrastive learning perspective. By leveraging predictive entropy to distinguish between positive positions (reflective exploration) and negative positions (exposure bias), CRPO conducts group-wise contrast to preserve reliable, fine-grained optimization signals. Extensive evaluations across 13 challenging reasoning and deep-search benchmarks demonstrate that CRPO consistently outperforms existing reinforcement learning and self-distillation baselines, significantly enhancing training stability and generalization in long-horizon interactions.

Recently, Reinforcement Learning (RL) has emerged as a crucial paradigm for the post-training of Large Language Model (LLM) agents. However, existing methods predominantly rely on sparse task rewards for policy optimization, failing to fully exploit another class of inherently dense supervisory signals naturally present during online interaction: environmental feedback following action execution. Recent theoretical studies suggest that generalization in multi-step, goal-oriented tasks hinges on predictive knowledge of environmental consequences. Inspired by this, we propose TAPO: Transition-Aware Policy Optimization for LLM Agents, a unified training framework that alternates between policy optimization and transition supervision. Beyond standard RL updates, TAPO repurposes rollout data to apply action-conditioned next-observation prediction supervision on a shared backbone model. This approach enhances the model's sensitivity to environmental transition dynamics and action consequences while concurrently optimizing the policy. It serves as a computationally lightweight, plug-and-play enhancement module for existing agent RL algorithms, requiring no additional expert data, extra sampling costs, or inference-time overhead. We conduct systematic experiments on WebShop and ALFWorld, integrating foundation models of various scales with different policy optimization algorithms. Empirical results demonstrate that TAPO consistently improves task performance over pure policy optimization baselines.

Group Relative Policy Optimization (GRPO) has become a standard reinforcement learning method for post-training language models. Recent work shows that GRPO can reduce the base model's reasoning capacity and underperform it in Pass@k when k is large, indicating reduced coverage of reasoning paths. We find that this reduction is associated with GRPO concentrating on responses that the base model already generates with high probability. We trace this concentration to two mechanisms in the GRPO update. At the response level, high-probability responses dominate the group gradient through repeated occurrence. At the token level, GRPO's importance ratio scales gradients, further reinforcing tokens that become more likely under the current policy. We propose ReCo, a reweighting method that addresses both effects. Response contributions are normalized by their expected occurrence within the rollout group, and the token-level importance ratio is replaced with a variance-based ratio that gives larger update scale to non-saturated decision points where alternative token choices remain plausible. Across Qwen2.5-Math-1.5B/7B and Llama-3.1-8B-Instruct on five mathematical reasoning benchmarks, ReCo improves Pass@k for large values of k and is comparable to GRPO for small values of k.

Code models are increasingly trained with execution feedback, but most training signals still stop at correctness. This leaves an important gap for systems code: two programs can pass the same tests while differing greatly in runtime. We study how to train code agents to prefer faster correct implementations, rather than treating efficiency only as an evaluation metric. The key difficulty is that runtime is a fragile reward. It is meaningful only after a program is correct, varies across tasks, and gives little guidance when most sampled programs fail to compile or run. We propose \textbf{RLPF}, reinforcement learning from performance feedback, which turns execution outcomes into a staged reward. Failed programs are ordered by execution progress, while correct programs are ranked by their relative improvement from the baseline toward the expert reference. This gives useful feedback before correctness and performance-sensitive feedback after correctness. Fine-tuning Qwen3-32B with RLPF on PerfCodeBench raises correct-and-runnable solutions from $11.1\%$ to $54.6\%$ and improves relative efficiency from $8.1\%$ to $38.6\%$. The trained model becomes competitive with stronger open-weight systems, and its optimization behavior transfers modestly to EffiBench-X. Additional studies show that model-generated references provide useful but weaker supervision, and that the full composite reward is more reliable than correctness-only or runtime-only baselines. These results suggest that code agents can be trained not only to pass tests, but also to optimize the programs they write.

We present, to our knowledge, the first end-to-end FP4 RL post-training, in which both the rollout and training policies, including their forward and backward passes, operate at 4-bit precision. A systematic study reveals that the dominant source of degradation in FP4 RL is not training-side quantization error but rollout activation quantization: outliers stretch the dynamic range so far that a large number of activation values underflow to zero under FP4. Counterintuitively, restoring the training policy to higher precision while keeping the rollout in FP4 makes accuracy worse than full FP4 baseline, exposing rollout-training mismatch as the principal failure mode and ruling out standard pretraining-style fixes. We address this with Rollout Residual Quantization (Rollout-ResQ): a single residual correction term constrained to a hardware-friendly sparsity pattern, added only to the FP4 rollout matmul -- a lightweight correction that recovers most of the precision lost to outlier-driven underflow without inflating the rollout's compute footprint. On Qwen2.5-3B and Qwen2.5-Math-7B, Rollout-ResQ paired with the HiFloat4 (HiF4) format -- whose three-level hierarchical scaling preserves resolution under FP4's tight 4-bit budget -- closes the accuracy gap to BF16 from 4.9% to 1.1%, bringing fully quantized FP4 RL within striking distance of full precision. Applied to the open-standard MXFP4, the same recipe narrows the gap from 13.6% to 5.3%, revealing that FP4 format choice is a key factor that determines the ceiling on recoverable accuracy. Together, these results establish HiF4 as the enabling format for end-to-end FP4 RL post-training, and Rollout-ResQ as the activation-side mechanism that makes the gap to BF16 closable.

Reinforcement-learning post-training dominates modern language-model development, yet its power behavior on GPU hardware has not been characterized, and datacenters manage GPU power with workload-blind mechanisms, static caps and reactive throttling, that slow hardware indiscriminately. We instrument GRPO training with half-second power telemetry at 7B, 14B, and 72B scales on one to four A100s (380,000+ samples), and train a PPO meta-controller that adapts the workload's own generation parameters to measured power. Against the full 500-step 7B trace, the controller cuts power-limit violations by 89.8% while increasing token output by 18.1% and energy efficiency by 26.2% (tokens per MWh). Deployed live at 72B, the same controller family yields replicated null results, diagnosed as the group-size actuator losing authority under model sharding. An actuator-authority sweep shows the same parameters applied as generation concurrency retain 17-22% power authority, isolating an occupancy-versus-volume principle; a controller rebuilt on that actuator controls a live 72B rollout-generation workload across three replications: 35.7% more output than a static safe baseline at 2.27 +/- 1.08% budget violations, 87.2% fewer violations than uncontrolled operation, and the best mean throughput and energy per token among constrained controllers, with an adaptive threshold rule matching it in one of three operating conditions. Under realistic measurement windows the original 72B transients fall from 23.6% at half-second resolution to 1.6% at 30 s and zero at 5 min; a composed 16-GPU fleet shows zero violations at 30 s and longer, with peak demand at 50-56% of nameplate. For this fleet mix, roughly twofold oversubscription of nameplate appears feasible, subject to operator validation. We quantify the economic and carbon consequences and specify a low-cost operator pilot.

Asynchronous reinforcement learning (RL) accelerates large language model (LLM) post-training by overlapping rollout generation with policy optimization, but the resulting stale, off-policy data destabilizes optimization and can cause policy collapse. Existing methods gate tokens by ratio magnitude alone, applying one threshold at every position. We show that the ratio's natural scale is set by token entropy, so deviations from mid-trajectory weight updates stay within this scale and carry genuine exploration. We further identify an overlooked low-entropy regime that breaks this scaling, where a near-zero probability amplifies train--inference mismatch into noise far beyond what the local entropy admits. A magnitude threshold admits this noise and discards the exploration. We therefore propose the Entropy-Normalized Trust Region (ENTR). Across long-horizon agentic tasks and mathematical reasoning benchmarks, ENTR outperforms existing asynchronous methods. It improves avg@1 on BrowseComp-Plus by $6.9\%$ over the strongest baseline, trains stably up to $30$ policy versions of staleness, and matches synchronous GRPO at a $2.6\times$ speedup.

Model merging plays a crucial role in consolidating multiple specialized models into a single, unified model, especially in the era of large language models (LLMs). Recent research has primarily focused on developing strategies to enhance merging performance with the trained models, while the impact of training paradigms, such as supervised fine-tuning (SFT) and reinforcement learning (RL), on the effectiveness of model merging remains underexplored. In this study, we systematically explore the merging behavior of RL-trained LLMs compared to those trained with traditional SFT. Through comprehensive evaluations across five representative tasks, we find that RL significantly reduces task conflicts and results in less performance degradation after merging, making RL-trained models particularly well-suited for this process. To unearth the reasons behind the superior suitability of RL for model merging, we conduct extensive empirical experiments and theoretical analyses. Our findings highlight three key factors: (1) On-policy training data in RL control the gradient updates in a smaller magnitude, reducing the risk of overwriting existing knowledge for other tasks in the model. (2) The RL optimization objective, which favors ``\textit{enough is as good as a feast}", progressively reduces the magnitude and the number of conflict parameter updates as the model converges. (3) Joint optimization of positive and negative examples in RL steers the model towards an unbiased task-specific parameter subspace, ensuring robust performance while further preventing parameter conflicts.

Recent large reasoning models often develop long chain-of-thought responses during reinforcement learning (RL), resulting in high inference latency and deployment cost. Existing methods for response length control typically rely on explicit length penalties or additional control modules, which require careful tuning and may compromise reasoning quality. We propose Quadrant-weighted Sampling for Length-aware Policy Optimization (QLPO), a simple resampling-based variant of GRPO that introduces implicit length control without modifying the reward function. QLPO first over-generates candidate responses and then resamples the training group by preserving the empirical correct/incorrect ratio while favoring short correct responses and long incorrect responses. This reshapes the training distribution and implicitly encourages shorter model outputs. Across models ranging from 1.5B to 32B parameters, including both base models and strong reasoning models, QLPO consistently improves the accuracy-length trade-off. It reduces response length by 30% to 70% while preserving reasoning performance. These results suggest that structured resampling provides an effective and robust approach to efficient reasoning.

Closing the gap between benchmark performance and reliable real-world operation remains a central challenge for Vision-Language-Action (VLA) humanoid robots, which must handle execution errors, distribution shifts, and environmental variability. This paper presents DEED (Data-Efficient Post-Training and Experience-Driven Learning), a systems-level approach evaluated on a supermarket chip-restocking task using a Unitree G1-Edu humanoid robot and the GR00T N1.6 foundation model. DEED comprises three key components: (1) a data-efficient post-training pipeline with control-frequency alignment, data curation, task-relevant visual highlighting, and reduced VLA dependence; (2) a real-world study of experience-driven refinement, adapted from RECAP via a text-based advantage prefix and a vision-language value function; and (3) a latent-space analysis tool for studying in- and out-of-distribution behavior. Our results suggest that bridging the lab-to-store gap is primarily a systems integration challenge rather than an architectural one: careful data design and targeted post-training can transform a policy that fails under naive fine-tuning into a competent real-world system using only a single GPU.

LLM preference alignment aims to optimize models toward human preferences across diverse user instructions. Reinforcement learning has become a major post-training approach for this goal, but existing proxy rewards are often outcome-level, mainly evaluating the final response while providing limited guidance for the reasoning trajectory. This can make credit assignment coarse when multiple responses receive similar final scores, leaving trajectory-level preferences under-specified. To address this limitation, we propose Thinking Checklist Reward (TCR), a process-oriented reward for RL-based preference alignment. TCR converts preference pairs into sample-specific thinking checklists and uses them to evaluate whether the generated reasoning trace addresses the preference-implied considerations. To reduce overlap with outcome-level supervision, TCR further introduces an exponential moving average (EMA) residual formulation to isolate a complementary thinking surplus beyond what is predictable from the outcome reward. Experiments on five models from three model families show that TCR consistently improves alignment performance across diverse benchmarks, with ablations further validating the importance of EMA-based residual formulation and sample-specific checklist supervision.

Reinforcement learning with verifiable rewards (RLVR) is rapidly advancing the reasoning capabilities of language models, yet the optimization layer that converts reward feedback into weight-space updates remains poorly understood. Building on our prior analysis (Zhu et al., 2025), we study this missing layer through the singular structure of model weights and identify spectral inheritance: RLVR can reuse the base model's weight spectra while acquiring new behavior through changes in the associated input and output singular frames. We operationalize spectral inheritance as Isospectral Optimization (ISO), an RLVR-native, fixed-spectrum optimization framework with complementary offline and online instantiations. Offline, ISO-Merger combines the frame changes of shared-base specialists into a single fixed-spectrum model, requiring no post-merge data, rollouts, gradient updates, or on-policy distillation (OPD). It recovers complementary specialist capabilities and achieves the strongest aggregate performance among the compared data-free merging methods. Online, ISO-Optimizer applies a chosen base optimizer, including AdamW and Muon, to the frame variables while keeping the base spectra fixed. Across reasoning and coding tasks ranging from 1.5B to 8B parameters, ISO-Optimizer improves accuracy in the reported runs and reaches matched scores with substantially fewer training steps. On Qwen3-8B-Base, AdamW reaches an aggregate accuracy of 0.495 after 270 training steps. ISO-AdamW reaches the same accuracy after only 100 training steps and improves further to 0.509 after 210 training steps. Together, ISO offers a concrete answer to RLVR's missing optimization layer: rather than inheriting pre-training optimization wholesale, design post-training around the structure of reward-driven adaptation: inherit the spectrum, optimize the frames.

Large-scale online reinforcement learning (RL) is the predominant means of eliciting advanced abilities including long-term reasoning and agentic tool use in large language models (LLMs). However, continuing to scale it across vast task domains of interest remains challenging in both computational infrastructure and cost, especially when considering RL as merely a one-off learning stage. Recently, a widely used technique for distilling knowledge across various domains and training stages, multi-teacher on-policy distillation (MOPD), helps to decouple the RL stage, saving costs, while maintaining generality across vast domains. Nonetheless, similar to online RL, MOPD requires coupled inference and backward passes, which continues to limit its scalability and computational efficiency. To address these challenges, we propose REGEN: Replay-recycling for Expert-to-Generalist Distillation with Offline RL. Instead of distilling from multiple teacher models, REGEN trains a generalist by simply recycling the replay memory -- the free by-product of the teachers' specialized RL training -- and employing offline RL algorithms. REGEN completely decouples the rollout sampling from the backward training process and thus greatly reduces the training cost. Across mathematical reasoning, code generation, and instruction following, REGEN matches the accuracy of MOPD at substantially lower cost. It potentially turns online RL into a data synthesis process instead of a one-off learning stage, and can be extended to large-scale post-training without requiring heavy computational load. Code is available at https://github.com/yunjie-sysu/REGEN.

Asynchronous reinforcement learning improves throughput by decoupling rollout generation from optimization, but the resulting staleness is an inevitable byproduct, compounded jointly by policy lag, engine delays, and mixture-of-experts routing. From a trust-region perspective, this mismatch is critical: in the finite-horizon improvement bound, training-inference divergence governs the approximation error, whereas PPO clipping only gates sampled outward updates and therefore acts as a sampled surrogate rather than a full-policy constraint. As a result, the high-staleness update can remain weakly controlled in exactly the asynchronous regime where stale rollouts matter most. We introduce the Staleness-Adaptive Trust Region (SAT), which uses the detached sampled log-ratio as a practical staleness proxy, identifies the high-mismatch tail within each batch through Staleness-based kernel function scaling, and contracts only the sign-selected endpoint of the nominal PPO interval using Effective contraction factors. This design preserves the baseline behavior on ordinary tokens, while making the update more conservative exactly on newly intercepted outward bands. We evaluate SAT in a fully decoupled asynchronous reinforcement learning setup built on Qwen3-30B-A3B-Base, leveraging SGLang as the inference engine and Megatron as the training pipeline. In this setting, SAT-GSPO w/ R3 attains the best observed AIME24 avg@8, reaching 35.83 at lag 1 and 34.79 at lag 8, while SAT-GSPO reaches 34.17 at lag 1. More broadly, the results indicate that aligning the clip interval with observed staleness heterogeneity is an effective way to stabilize the reported asynchronous regime.

Group Relative Policy Optimization (GRPO) has shown strong effectiveness in reinforcement learning from verifiable feedback, where sampled rollouts can be compared within a group using task-provided correctness signals. However, extending group-relative optimization beyond verifiable settings is challenging because success in many tasks is not captured by a single correctness criterion. We propose \textbf{Reference-Relative Policy Optimization (RRPO)}, which generalizes GRPO by replacing direct correctness-based advantage construction with reference-relative contrastive comparisons. RRPO first uses \emph{stratified conditional rollouts} to construct positive and negative anchor sets, and then trains a metric projection head with a set-contrastive objective to compare candidate rollouts against these anchors. The resulting alignment scores directly define contrastive advantages: during policy optimization, the projection head is frozen, and the scores are centered within each rollout group in a standard group-relative objective. We evaluate RRPO using anchor-based contrastive advantages throughout policy optimization, without relying on task ground-truth verifiers. Across verifiable reasoning, open-ended generation, and post-SFT settings, RRPO remains competitive with verifier-based optimization, improves over weakly supervised baselines, and provides additional gains after supervised fine-tuning.

PPO and the GRPO baseline studied here use clipped surrogate objectives whose favorable-direction saturation introduces an abrupt change in the scalar objective's derivative. We ask whether Output Reset (OR), a smooth one-sided saturation rule, offers a useful alternative for large language model post-training. PPO-OR and GRPO-OR replace the clipped policy term with an OR squared-margin loss in rollout-relative token log-ratio space; the advantage sign determines the update direction, and a token contributes zero direct OR residual after crossing the favorable margin. We compare PPO-clip with PPO-OR under generalized advantage estimation (GAE), and GRPO with GRPO-OR under group-relative advantages, using \texttt{Llama-3.2-1B-Instruct} on Anthropic \texttt{hh-rlhf} with one shared reward model and three seeds per method. Under GAE, PPO-OR has a mean final training-time reward-model score $0.305$ higher than PPO-clip, with a larger observed across-seed spread. Under group-relative advantages, GRPO-OR does not have a higher mean score, but shows a smaller observed spread, a near-zero terminal OR residual, and a declining overshoot fraction, while the matched GRPO clipped-objective trace remains variable. Both group-relative methods exhibit substantially larger rollout-to-current log-ratio displacement than the GAE methods, and OR does not consistently reduce it. Thus, OR changes optimization behavior in both matched comparisons, but the observed reward effect differs between them. At $G=2$, the GRPO-OR diagnostics do not translate into a reward-score gain. Whether larger groups change this outcome remains open. The reported scores are training-time reward-model measurements, not held-out human-preference performance.

Reinforcement learning (RL) on open-ended tasks compresses an LLM's rubric-based evaluation into a scalar reward, discarding rich textual feedback and conflating responses with distinct quality profiles. We propose Experiential Learning (EL), which repurposes the feedback model from an LLM-as-a-Judge into an LLM-as-a-Coach. The coach distills its assessment of each on-policy response into transferable experiential knowledge, which conditions a teacher model and is internalized by the policy through on-policy context distillation. Compared with scalar rewards, this higher-bandwidth feedback channel provides dense supervision and preserves fine-grained preferences among high-quality responses. Across two policy families, with feedback from the policy itself or a proprietary model, EL consistently outperforms rubric-based RL on held-out and unseen open-ended tasks. Notably, EL generalizes better beyond the training distribution, and mitigates reward hacking. These findings establish experiential knowledge as a richer and more generalizable learning signal for post-training on non-verifiable tasks.

Large language model (LLM) post-training is essential for improving reasoning, adaptation, and alignment. Existing methods mainly follow two paradigms: reinforcement learning (RL) and on-policy distillation (OPD). However, RL relies on coarse-grained outcome supervision, resulting in difficult credit assignment and limited capability to acquire new knowledge. OPD, meanwhile, unconditionally matches teacher logits through KL divergence, which creates a dilemma: similar teachers provide little new knowledge, while substantially different teachers often yield ineffective guidance, largely restricting OPD to within-family distillation. We propose Distilled Reinforcement Learning (Distilled RL), which integrates teacher supervision into the RL objective to provide fine-grained guidance, selectively transfer new knowledge and avoid unconditional imitation. Distilled RL contains three components: reverse importance sampling with clipping, negative sample reset, and sequence-level geometric normalization. Through a concise and interpretable case study, we demonstrate that Distilled RL can effectively transfer previously unavailable knowledge from a teacher model to a student model. Extensive experiments across both within-family and cross-family distillation settings show that Distilled RL substantially outperforms standard RL and OPD in terms of both pass@1 and pass@k. Our code is available at https://github.com/597358816/Distilled-RL.

Muon is competitive with AdamW in large-scale pre-training, but its operating regime in reinforcement-learning post-training remains unclear. We map this regime on ALFWorld, a sparse-reward agentic benchmark, using three group-based objectives and Qwen2.5 models from 0.5B to 3B. Under a shared KL and clipping recipe, matched optimizer comparisons and AdamW rate controls trace the usable step-size range. AdamW responds non-monotonically to rate, whereas fan-in Muon remains stable at a more aggressive effective step: at $3 \times 10^{-5}$ it improves late success over an AdamW $10^{-6}$ baseline after correction across rate-metric tests. Its normalized-AUC effect is directionally positive but less uniform; the heuristic-matched lower-rate effect is less consistent, and tuned AdamW nearly matches high-rate Muon at 3B GraphGPO. High-rate Muon applies $3.53 \times$ AdamW's hidden-matrix update RMS; a full-budget RMS-matched control removes the late-success gain. Together, these results identify a recipe-level operating regime in which fan-in Muon supports a more aggressive stable effective step under shared KL and clipping: the margin is largest when optimization headroom remains and contracts near saturation, after AdamW tuning, or under magnitude matching. The scale-matched control ties this spectral effect to Muon's scale convention rather than establishing a universal optimizer ranking. Code is available at https://github.com/x66ccff/verl-muon.

Reinforcement learning (RL) has become central to improving large language models (LLMs) on complex reasoning tasks, yet RL post-training is largely studied in isolation from the pretraining that precedes it. As a result, two basic questions remain open: (1) how do pretraining choices (model size, data) shape the returns to RL compute, and (2) what does RL actually do to the model? These questions are difficult to study in the standard LLM setting: pretraining corpora are vast and uncontrolled, making it hard to attribute behaviors to pretraining versus RL, and systematic compute sweeps across both stages are prohibitively expensive. To address these challenges, we use chess as a controlled testbed for studying reasoning across the full pretraining-to-post-training pipeline. We follow the standard LLM training pipeline by pretraining language models from 5M to 1B parameters on human chess games, supervised fine-tuning on synthetic reasoning traces, and running RL on chess puzzles with verifiable rewards. Using this framework, we find that the post-RL performance at given RL compute level is well-predicted from the pretraining loss, and slope of the RL reward curves improves approximately linearly with the pretraining tokens. Beyond scaling, we find that RL does not simply sharpen the SFT policy: on easy puzzles it amplifies correct moves the SFT policy already preferred, while on hard puzzles it surfaces correct moves that were nearly absent under SFT. We further test whether our findings transfer beyond chess by training a 1B language model on math-domain text, where the same predictive pattern emerges: longer-pretrained checkpoints reach higher post-RL performance and improve faster under RL. In sum, we provide a quantitative account of the pretraining-to-RL interface and a controlled testbed for studying the science of reasoning across the full pretraining-to-post-training pipeline.

A growing gap separates inference context lengths from RL post-training: inference systems are approaching million-token contexts, while post-training workloads often remain at 256K tokens or below and rely on length generalization at deployment. The gap is especially important for AI agents, whose observations, tool outputs, documents, and prior decisions accumulate over long trajectories. LongStraw is an architecture-aware execution stack for million-token RL post-training under a fixed GPU budget, instantiated with Group Relative Policy Optimization (GRPO). It evaluates the shared prompt without autograd, retains only model-specific state needed by later tokens, and replays short response branches one at a time, reducing the live training graph at the cost of additional replay time. We implement it for the hybrid recurrent and full-attention Qwen3.6-27B and the compressed-attention mixture-of-experts GLM-5.2. On eight H20 GPUs, LongStraw completes grouped Qwen scoring and response backward at 2.1M positions for groups of 2 and 8; increasing the group size adds only 0.21 GB of peak allocated memory, while a separate stress test reaches 4.46M positions. On 32 H20 GPUs, we validate the end-to-end LongStraw execution path for a 2.1M-token prompt across all 78 layers of GLM-5.2. These experiments establish execution capacity rather than complete training correctness because the captured prompt state is detached and some distributed forward and gradient composition paths remain incomplete.

Multi-turn agents solve complex tasks through extended sequences of tool interactions before producing a final answer, making credit assignment a fundamental challenge during post-training. Outcome rewards provide reliable supervision for short-horizon reasoning, but become sparse and high-variance as trajectories grow to tens or hundreds of tool calls. They can also be misleading: a failed rollout may contain many useful actions that move the agent closer to the goal, yet outcome-only training assigns them the same negative advantage as the eventual mistake. We propose TRACE (Turn-level Reward Assignment via Credit Estimation), a dense credit-assignment method for agentic reinforcement learning. TRACE represents rollouts as state transitions at tool-call boundaries, obtains gold-answer log-probabilities from a frozen reference model, transforms them into log-ratio state values, and derives per-action rewards as Temporal-Difference changes in those values. This requires no additional critic or process-label training, and its one-step log-ratio TD component telescopes across redundant tool calls. On long-horizon complex search, TRACE substantially improves base-model tool-use ability using pure RL, without a cold-start supervised fine-tuning stage, an agentic mid-training stage, or training on live-web data. On the closed-web BrowseComp-Plus benchmark, it raises Qwen3-4B from $7.2$ to $35.6$ and Qwen3-30B-A3B from $8.4$ to $42.6$. The learned search behavior also transfers to open-web benchmarks, and the learning curves show earlier improvement and faster convergence during RL training.

Large language models have made strong reasoning gains through supervised fine-tuning, reinforcement learning, and on-policy distillation, yet these post-training methods are usually evaluated only by final-answer accuracy. We study how they reshape confidence during reasoning. We introduce a three-stage calibration framework that evaluates confidence before, during, and after chain-of-thought generation, corresponding to difficulty estimation, early termination, and answer aggregation. Through a controlled comparison on mathematical reasoning benchmarks, we find that OPD provides the most useful pre-reasoning confidence, SFT gives the strongest online signal for early stopping, and RL produces the most reliable trace-level signal for aggregation. We further show that confidence reliability is position-dependent: RL confidence becomes informative after a path-commitment phase, while OPD confidence is useful early but can become inversely calibrated later. Based on this observation, we propose PosConf, a position-aware confidence strategy that uses confidence only from reliable relative-position intervals. PosConf improves RL answer aggregation by 6.1 points over majority voting and consistently improves OPD early stopping under tight token budgets, with gains up to 4.3 points by avoiding its later inverse-calibration region, showing that \emph{confidence in reasoning models should be used both stage-wise and position-awarely}. Our code is available at https://github.com/EIT-NLP/Post-Training-Calibration.

Generative Flow Networks (GFlowNets) offer a promising alternative to reward-maximizing reinforcement learning (RL) for large reasoning models, encouraging diverse reasoning paths by matching reward distributions rather than collapsing to dominant modes. Recent work shows promise on math and code, but scaling GFlowNet-style RL to modern post-training pipelines remains difficult: as model size, rollout horizon, reward noise, and distributed-systems complexity grow together, a learned prompt-conditional partition function becomes a source of gradient instability and engineering overhead rather than a useful normalizer. Through systematic analysis, we find that the learned partition function, previously treated as essential, can be replaced by an in-batch Monte Carlo estimate computed from the rollout group already required for training. We propose GFlowRL, a streamlined GFlowNet-style RL algorithm that removes the auxiliary partition network entirely while preserving the reward-distribution-matching objective, completed by two stabilizers: importance-sampling correction for rollout/trainer drift and asymmetric flow-gap clipping for outlier residuals. GFlowRL exceeds all counterparts on math, code, and adversarial red-teaming benchmarks, reaching a Codeforces rating of 2048 at the 14B scale (within 25 Elo of o3-mini) and attaining the highest average ASR@1 on AdvBench and HarmBench, outperforming the previous SOTA multi-turn attacker in a regime where FlowRL, a prior GFlowNet-style method, diverges. The same recipe transfers to all evaluated MoE configurations up to 235B parameters, where FlowRL again fails to converge. To our knowledge, GFlowRL is the first GFlowNet-style RL algorithm to scale stably across both dense and sparse architectures. Code will be at: https://github.com/microsoft/gflowrl