Reinforcement Learning with Verifiable Rewards

Linear interpolation between fine-tuned checkpoints has been shown to trace the Pareto front between competing objectives, but whether extrapolative weight averaging can extend such frontiers to new checkpoints useful at inference time, without additional RL training, remains unclear. We study this question in RL for competitive programming, where hidden unit tests under time and memory limits enforce both functional correctness and computational efficiency. Starting from a shared initialization, we train checkpoints under nested unit-test coverage: low-coverage rewards require passing smaller-input tests, while high-coverage rewards require passing progressively larger tests up to the full suite. This sweep reveals the emergence of a correctness-efficiency frontier: on hard problems, higher-coverage reward reduces optimization failures but increases correctness failures, leaving solve rate nearly unchanged. Interpolation between low- and high-coverage checkpoints recovers this frontier, while extrapolation extends it beyond the trained endpoints. Both the frontier and its extrapolative continuation appear across three inference settings, pure reasoning, tool use, and agentic coding, and across two model scales, 32B and 7B. At the problem level, moving along the frontier changes which problems are solved, making extrapolated checkpoints complementary policies in inference-time scaling. Ensembles with extrapolative weight averaging broaden coverage and improve pass@250 on LCB/hard by 3.3% over the best single checkpoint at matched sample budget. These results show that nested unit-test coverage in code RL induces a frontier that extrapolative weight averaging can navigate, extend, and exploit.

Reinforcement learning with verifiable rewards (RLVR) can yield large reasoning gains from very few training instances, yet its strong sensitivity to which instances are used makes data selection a central bottleneck. Most existing selection pipelines rely on training-time optimization signals and/or require access to verifiable rewards or ground-truth answers over large candidate pools, which is costly and often infeasible in specialized domains. We study RLVR data selection in a setting where selection must be performed before any RL training and without labels or reward evaluation on the full pool. We propose SHIFT, a one-shot, training-free selector based solely on inference-time hidden-state dynamics. For each candidate instance, SHIFT runs a single deterministic reasoning rollout and computes a reasoning-induced representation shift (RIRS) as the start-to-end hidden-state delta. SHIFT uses the RIRS magnitude as a lightweight proxy for instance utility and enforces coverage via a quality-weighted farthest-first CoreSet procedure in an RIRS-augmented feature space, producing compact subsets that scale to large unlabeled pools. Across mathematical reasoning and medical QA benchmarks under ultra-low budgets, SHIFT consistently outperforms training-free diversity and difficulty/uncertainty baselines, improving both in-domain accuracy and transfer to harder evaluation settings. Ablations show that RIRS-based coverage and quality-weighting contribute complementary gains, and analyses indicate that RIRS is not explained by simple input/output length statistics. Code is available at github.com/JianghaoWu/SHIFT.

Reinforcement Learning from Verifiable Rewards (RLVR) has improved language models in domains such as mathematics and code, where correctness can be checked automatically. However, many important tasks are only partially verifiable: prompts contain multiple requirements, responses may satisfy some but not all of them, or no single reference answer might exist. We introduce Soft-RLVR, a framework for reinforcement learning from decomposed, learned verification signals. Soft-RLVR converts each prompt into a checklist of atomic requirements, scores candidate responses item by item with an LLM verifier, and trains on the resulting soft reward. Checklist-based rewards turn sparse pass/fail supervision into a denser partial-credit signal, but they also introduce a tradeoff: averaging item-level judgments can reduce verifier noise, while partial credit can reward incomplete responses. We formalize this tradeoff and identify conditions under which checklist-based verification gives a more reliable RL training signal than holistic verification. We further introduce Soft-SVeRL, a self-verifying variant of Soft-RLVR in which the policy also acts as the verifier. We show that self-verification is prone to reward inflation from overly permissive self-judgments, and that explicit stabilization is needed to prevent this collapse. In a controlled instruction-following setting with rule-based ground-truth evaluation, checklist-based Soft-RLVR improves IFEval by up to 11.1 points using only learned verifier rewards. Our experiments further show that verifier quality and checklist quality both affect downstream RL outcomes, and that explicit stabilization is essential for effective self-verification.

While large language models have made significant progress in mathematical reasoning, they remain unreliable at judging the correctness of their own solutions. Existing approaches that equip models with self-verification typically treat solution generation and verification as two separate tasks, leading to substantially increased training time. In this paper, we propose FABSVer, which fuses these two tasks into a single generation pass, dramatically reducing training overhead while jointly optimizing both capabilities. We further identify a convergence bottleneck both theoretically and empirically: as training progresses, the reward reaches a plateau because the policy is constrained by a fixed reference model. To overcome this, we introduce Dynamic Reference Model Update (DRMU), which raises the reward ceiling and enables sustained reward growth. Extensive experiments on math benchmarks demonstrate that FABSVer achieves superior self-verification and reasoning performance across three model scales, while requiring only 51%--71% of the training time of existing methods. Analysis further reveals distinct learning phases in how models acquire self-verification, and that the gap between verify and answer rewards shrinks noticeably as model size increases.

Reinforcement Learning with Verifiable Reward (RLVR) is empirically shown to notably enhance the reasoning performance of large language models (LLMs), particularly in mathematics and programming. However, the mechanistic role of Sample Difficulty in RLVR remains poorly understood. In this paper, we investigate RLVR through the lens of difficulty-wise and one-sample analysis. We find that sample difficulty has a non-monotonic effect on RLVR: easy and medium-difficulty problems yield the strongest and most stable reasoning improvements, whereas overly hard problems often provide weak learning signals, induce degenerate behaviors such as answer repetition or skipping necessary computation, and can ultimately degrade the model's pre-existing capabilities. Beyond the obverse of response, we further analyze the model's internal feature dynamics using Temporal Sparse Autoencoders (T-SAE). Easy problems mainly reinforce direct-answer and basic-computation features while suppressing deliberative-reasoning features; hard problems activate reasoning-related features but become useful only when successful trajectories are sampled; medium-difficulty problems provide a more balanced signal, strengthening both computation and multi-step reasoning features. Motivated by these findings, we propose difficulty-adaptive strategies for hard-sample utilization, using backward-reasoning reformulation and T-SAE-guided training signals to improve reward density and credit assignment during RLVR. Overall, our results identify sample difficulty as a key factor governing both the optimization dynamics and representation evolution of RLVR.

Reinforcement learning with verifiable rewards (RLVR) trains reasoning models without labeled trajectories, using groups of verifier-scored rollouts to explore alternative reasoning paths. Limited rollout diversity is a central bottleneck, typically addressed through adjustments to temperature, prefixes, or rollout selection. We identify the first token of the response as a structurally distinct target for diversification, largely overlooked in prior work. We find that the first-token distribution is sharply concentrated and only weakly related to downstream correctness, as lower-probability candidates can yield similarly accurate responses. Diversifying the first token can therefore broaden the reasoning paths explored within each rollout group with little loss in response quality. Motivated by this observation, we introduce REFT (Rollout Exploration with First-Token Diversification), a lightweight modification to RLVR. REFT samples first tokens uniformly from the policy's top-$N$ candidates and allocates rollouts evenly across the sampled tokens, leaving the rest of the pipeline unchanged. We evaluate REFT on eight models spanning multiple architectures and sizes (0.5B-14B), with mathematical reasoning and code-generation tasks under GRPO and DAPO. Across these settings, REFT consistently improves Pass@1, Pass@8, and Pass@64. It also outperforms competing diversification methods at every evaluated budget, incurring the lowest rollout cost.

Reinforcement learning with verifiable rewards (RLVR) has become a key technique for en- hancing LLM reasoning, yet its data ineffi- ciency remains a major bottleneck. Existing methods address this problem only partially, each missing at least one of subset-level cov- erage, verifier signal use, or interpretability. To address this gap, we present IRDS (Inter- pretable RLVR Data Selection), which selects RLVR training instances on a sparse autoen- coder (SAE) cluster basis so the selection itself is auditable on recognizable problem motifs. To select instances the model both fails on and can still learn from, we introduce a verifier- coupled coverage objective on the SAE basis and solve it by greedy log-determinant max- imization. Experiments on three instruction- tuned models and six math reasoning bench- marks show that IRDS achieves the highest overall accuracy, exceeding the strongest base- line by +3.9/+4.0 pp on the two Qwen models and by +0.5 pp on Llama-3.1-8B, while run- ning an order of magnitude cheaper than the trajectory-based baseline.

Reinforcement Learning from Verifiable Rewards (RLVR) has emerged as the standard paradigm for improving reasoning capability of large language models, while Multi-Token Prediction (MTP) has been a widely adopted module in pretraining. Combining them is a natural approach, yet current RL practices detach MTP gradients because joint training degrades the performance. We revisit this failure from an optimization perspective. We show that the per-step effect of MTP on the RL objective can be decomposed into two terms: a first-order correlation and a second-order perturbation penalty. This decomposition unifies three MTP training regimes: Detach, Cross-Entropy loss, and Policy loss, and explains why each succeeds or fails. Further analysis of policy loss reveals that, although it aligns with intuition, performance still degrades: the correlation term decays while the quadratic penalty persists. Guided by the analysis, we propose Optimal Coefficient Calibration (OCC), an adaptive scheme that tracks the optimal coefficient online via a log-probability proxy at negligible cost. Across six competition-level mathematical reasoning benchmarks, OCC consistently matches or exceeds the detach baseline, delivering improved joint MTP-RL training performance.

Adaptive context compression is vital for scaling Large Language Models (LLMs) to complex, multi-turn agent tasks. However, rule-based compression methods may discard task-critical nuances, while Reinforcement Learning (RL) approaches usually struggle to balance information retention and token efficiency under the sparse rewards inherent to long-horizon workflows. To bridge this gap, we propose ZipRL, a novel adaptive compression framework tailored for Reinforcement Learning from Verifiable Rewards (RLVR). ZipRL features a multi-granularity compression mechanism for active, non-uniform information reduction, coupled with Hindsight Response Replay (HRR), a technique designed to densify training signals during RLVR optimization. Theoretically, we prove ZipRL's superior task-relevant utility over uniform methods. Concretely, ZipRL utilizes coarse-to-fine prompts for macro-compression and incorporates HRR into GRPO via generalized advantage reshaping. Multiple models of varying versions and parameter scales validate the effectiveness of our approach. Benchmarks on five agent tasks show ZipRL outperforms state-of-the-art approaches by 27.9% and 34.7% across Qwen3-4B and Qwen3-8B models, while maintaining exceptional token efficiency and robustness under extreme 256-turn extrapolation stress tests.

Visual captioning requires models to capture visual content faithfully while minimizing both omission and hallucination. As the dominant paradigm for captioning, MLLMs have achieved strong performance through scaling and high-quality data. Recently, RL has emerged as a key route to driving MLLMs toward higher precision and broader coverage, however, existing reward designs for captioning fail to provide fine-grained and reliable signals for factual verification, limiting their effectiveness. To address this, we propose VCap, a Witness-Adjudicator reward that pairs the reference caption (a witness) with the visual signal (an adjudicator). By explicitly verifying factual consistency between the reference and policy-generated captions grounded in the visual signal, VCap delivers a reward signal with hypergeometric-distribution-level precision for caption quality verification. This design enables effective learning even from imperfect references, facilitating weak-to-strong generalization in RL training. In our experiments, an 8B model trained with VCap outperforms open- and closed-source SOTA models on multiple image and video captioning benchmarks. Human evaluation further confirms its strong alignment with factual correctness. Additionally, VCap improves MLLM perceptual capability, generalizes across tasks, and surpasses best-of-N distillation, challenging prior assumptions about RLVR.

LLMs for code generation are commonly evaluated in repeated-sampling settings using Pass@k, where multiple candidate programs are executed against unit tests under a finite sampling budget. While recent verifier-based reinforcement learning (RLVR) methods improve executable correctness, how these objectives affect redundancy among sampled programs remains poorly understood. In this work, we study implementation-level redundancy in code generation using JPlag, a plagiarism-detection system for code. Across models and benchmarks, we show that correctness-only RLVR often concentrates generations around repeated implementations, whereas Pass@k-aware objectives maintain lower redundancy and improve larger-budget performance. Motivated by these observations, we augment RLVR with direct anti-redundancy rewards based on JPlag similarity. Across 3 models and 3 benchmarks, discouraging near-duplicate generations reliably improves finite-budget executable performance, often matching or outperforming specialized Pass@k-aware objectives.

Large Reasoning Models are typically trained via reinforcement learning from verifiable rewards (RLVR). However, existing approaches adopt fixed weights for positive and negative samples, and the conclusions hardly generalize to open-ended question answering (QA). In this paper, we systematically investigate the roles of positive and negative samples in reinforcement learning for open-ended QA. We propose a reward-mean-based strategy for distinguishing positive from negative samples, and observe that negative samples predominantly govern response diversity and the performance upper bound, whereas positive samples primarily determine response quality and convergence stability. Building on these observations, we propose EAPO, an Entropy-driven Adaptive Policy Optimization method that adaptively computes the weighting coefficients of positive samples based on the ratio of the current policy entropy to the initial entropy. During the entropy-decreasing phase, the weight assigned to positive samples is reduced to preserve exploration, whereas during the entropy-increasing phase it is amplified to reinforce stability, thereby mitigating entropy collapse. Experiments on two publicly available open-ended medical QA datasets demonstrate that EAPO consistently and substantially outperforms fixed-weight baselines in both response diversity and stability.

Reinforcement learning with verifiable rewards has become a standard recipe for improving the reasoning abilities of large language models. Existing algorithms face a tradeoff between computational efficiency and sample efficiency in value estimation and policy learning. We introduce BASIS, a critic-free post-training algorithm designed to address this tradeoff. At each online training step, BASIS samples only one rollout per prompt, but leverages rich information across prompts in the entire batch to improve value function estimation. Our experiments demonstrate that BASIS reduces MSE in value function estimation by 69% compared to REINFORCE++, a representative single-rollout baseline, and achieves lower MSE with one rollout than group mean estimators with 8 rollouts. This improvement in value estimation translates to better policy optimization: using substantially less training time, BASIS achieves performance close to multi-rollout GRPO-type baselines and often outperforms single-rollout REINFORCE-type baselines.

The proliferation of Reinforcement Learning from Verifiable Rewards (RLVR) datasets has exacerbated provenance collapse due to unclear lineage among existing datasets. To bridge this fragmented RLVR data landscape, we propose Atomic-source Tracing via Lineage-Aware Search (ATLAS), a systematic framework for tracing RLVR datasets back to their atomic sources, attributing over 99.7% of 1.45M instances to 20 atomic sources. Our analysis reveals that most RLVR datasets are variants of a small set of shared upstream sources, with few introducing genuinely new data, and many facing data contamination risks. These findings naturally motivate us to curate a new RLVR dataset, DAPO++, and to benchmark existing datasets from a lineage-aware perspective. To this end, we propose Source-level Counterfactual Attribution (SCA) as a guiding principle to curate a decontaminated training dataset with concentrated learning signals. Essentially, SCA measures a sample's marginal utility by comparing per-atomic-source RL checkpoints against a shared base model. Building upon these attribution signals, we further design a composite dataset quality score Q that strongly correlates with downstream RLVR performance. Experiments on Qwen3 series models verify that DAPO++ consistently improves performance on held-out benchmarks, while Q reliably predicts downstream RLVR training effectiveness. Our code and data is available at https://github.com/Celine-hxy/ATLAS.

Reinforcement learning with verifiable rewards (RLVR) has become central to post-training reasoning models, yet a key limitation of existing studies is their narrow view of the reasoning space: difficulty is treated as reasoning depth alone, and reward is concentrated on forward deductive state tracking. We instead characterize the reasoning space along two dimensions. Difficulty. Beyond reasoning depth, we study environment complexity, where models must identify the correct path amid distractors and interacting structures. Rewarded reasoning form. We consider four abilities core to real-world reasoning: deductive state tracking, abductive recovery of hidden events or facts, inductive rule induction, and analogical transfer. To disentangle these factors, we construct a synthetic knowledge-graph environment with controlled pre- and post-training distributions, where each instance varies along depth, complexity, and task family. Three findings emerge: joint depth-complexity coverage outperforms single-axis recipes; reasoning families respond non-uniformly, with abductive reasoning degrading outside the RL-covered region and task correlations clustering into deductive-abductive and inductive-analogy pairs; and uniform mixing outperforms staged curricula under a fixed budget. We also find that recent off-the-shelf models exhibit the same deductive-over-abductive asymmetry, suggesting that this gap is not merely an artifact of our controlled setup.

We present MobileGym, a browser-hosted, lightweight, fully controllable environment for everyday mobile use, targeting interaction fidelity without replicating proprietary backends. It enables two capabilities previously out of reach for everyday apps: verifiable outcome signals through deterministic state-based judging over structured JSON state, and scalable online RL through low-cost parallel rollouts. The full environment state is captured, configured, forked, and compared as structured JSON, and a single server can host hundreds of parallel instances, with about 400 MB memory per instance and about 3 s cold start. A layered state model and a declarative task-definition framework keep state programmability and task creation practical at scale, and a single programmatic judging mechanism delivers both deterministic evaluation verdicts and dense RL rewards. The accompanying MobileGym-Bench provides 416 parameterized task templates, including 256 test and 160 train templates, over 28 apps, with deterministic judges and a structured AnswerSheet protocol that avoids free-text matching failures. In a Sim-to-Real case study, GRPO on Qwen3-VL-4B-Instruct gains +12.8 percentage points on the 256-task test set, and on a 59-task real-device signal subset, real-device execution retains 95.1% of the simulation-side training gain. Project page: https://mobilegym.github.io.

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

Large Language Models (LLMs) have achieved remarkable advancements in reasoning capabilities empowered by Reinforcement Learning with Verifiable Rewards (RLVR). Nonetheless, RLVR intrinsically relies on ground-truth labels for reward computation, the acquisition of which is often prohibitively expensive in real-world scenarios. While unsupervised RLVR paradigms attempt to circumvent this by training on pseudo-labels, they are notoriously susceptible to training collapse. Moreover, different samples often exhibit varying annotation values. In this paper, we propose Reinforcement Learning with Active Verifiable Rewards (RLAVR), which actively acquires ground-truth labels for a small set of selected samples and integrates them with pseudo-labels, thereby stabilizing training dynamics and improving performance under limited annotation budgets. To identify valuable samples, we propose the Corrective Advantage Gap (CAG) metric and analyze the sample-level supervision value. Building on this, we introduce Correction-Aware Reliability Estimation for RLAVR (CARE), which translates the oracle CAG criterion into a practical pre-query acquisition policy to substantially improve training stability. Extensive experiments across diverse domains, model families, and model scales demonstrate the effectiveness and generality of our approach. Our code is available at https://github.com/Lumina04/CARE.

Reinforcement learning with verifiable rewards (RLVR) has driven breakthroughs in domains such as math, tool-use, and software engineering, yet its extension to computer-use agents (CUAs) has been bottlenecked by the scarcity of scalable training data with deterministic rewards. Constructing such data for CUAs requires consistent task instruction, executable environment, and verifiable reward. However, hand-curated benchmarks achieve high reward fidelity but cover few applications and LLM-as-judge-based datasets scale broadly but lack reliable verification. We present CUA-Gym, a scalable pipeline that co-generates task instructions, environment states, and reward functions. Concretely, a Generator agent constructs the initial and golden environment states, and a separate Discriminator agent writes the reward function from the task specification. An orchestrator agent drives the two through iterative rounds upon execution. Generated tuples then pass a final filter combining LLM majority voting and agent rollouts, ensuring quality beyond the per-task adversarial loop. To address the scarcity of training environments, we further synthesize CUA-Gym-Hub, a broad suite of high-fidelity mock web applications grounded in real-world software-use distributions, expanding the scale of CUA RLVR data by magnitude. Using this pipeline, we construct CUA-Gym, a dataset of 32,112 verified RLVR training tuples grounded in 110 environments. Trained with GSPO on CUA-Gym, our CUA-Gym-A3B and CUA-Gym-A17B achieve 62.1% and 72.6% on OSWorld-Verified, outperforming prior open-source CUAs at comparable scales, with performance scaling smoothly in both data volume and environment diversity. The same checkpoints also improve on the held-out WebArena benchmark, indicating transfer beyond the training environments. We will open-source the full synthesis pipeline, dataset, CUA-Gym-Hub environments, and models.

Contemporary reinforcement learning with verifiable reward methods post-train language models on multi-step reasoning by assigning a single outcome reward uniformly across all tokens in a trajectory. Such uniform assignment ignores which steps contributed to success or failure. Improving credit assignment can address this limitation by enabling targeted refinement of faulty reasoning steps, rather than updating entire trajectories uniformly. Resets are one such simple mechanism, enabling more precise credit assignment by returning to an intermediate state and resampling counterfactual continuations, so that outcome differences can be attributed to decisions made at that point. We propose two such methods: Random-Reset Policy Optimization (RRPO), where reset states are drawn randomly from reasoning steps, and Self-Reset Policy Optimization (SRPO), where the model self-localizes the erroneous step in an incorrect trajectory and resets there. We analyze these methods within the Conservative Policy Iteration (CPI) framework. Extending CPI with a credit-assignment oracle that targets improvable states yields provable improvements over random resets. Across models and reasoning benchmarks, SRPO consistently outperforms standard GRPO and RRPO by sampling multiple suffix continuations at a self-localized reset and learning from their rewards, using only the model itself with no external supervision.

Visual reasoning through reinforcement learning with verifiable rewards (RLVR) has achieved remarkable progress. However, when dealing with multi-source inputs, existing approaches tend to treat them as a mere accumulation of information, lacking explicit mechanisms to distinguish whether integrating additional sources yields information gain or introduces interference. Therefore, they struggle to effectively model dynamic interaction when integrating multiple sources, particularly when they differ significantly in physical properties and semantics, e.g., infrared and depth, leading to inferior performance to mono-source reasoning when a certain source holds the dominant signal. To address this issue, we propose MARS, a novel mono-anchored multi-source reasoning framework that models each visual modality as an independent information source. Specifically, by treating mono-source rewards as dynamic anchors, our method explicitly incorporates the information gain introduced by multi-source fusion into advantage normalization and adaptively emphasizes mutual promotion between sources while suppressing potential noise or conflicts during RLVR. From theoretical analysis, our method effectively quantifies information gain introduced by multi-source integration in gradient estimation, enabling consistent modality regulation. Empirical results also show impressive 3.2% and 4.9% performance gains on GRPO and DAPO across diverse datasets, confirming effectiveness of our method.

Reinforcement learning with verifiable rewards (RLVR) has become a core technique for post-training of Large Language Models (LLMs). While policy optimization is driven by all sampled tokens under a globally broadcast scalar reward, the heterogeneous policy behaviors exhibited along trajectories are largely overlooked without differentiation. Existing works address this by credit allocation, including token-level advantage reweighting, and selective token optimization, however, the allocation criterion are principally stagnant throughout training, limiting resilient policy evolution. In this work, we argue that \textit{when} learning signals are scheduled can be as important as \textit{where} they are allocated across tokens, and introduce the temporal dimension that scheduling the credit allocation criteria over the course of RLVR optimization. We find that prioritizing targeted tokens emphasized with specific policy behaviors, and gradually attenuating toward general optimization leads to more stable and efficient learning dynamics. Furthermore, we show that simple trajectory percentiles provide a natural perspective for distinguishing policy behaviors, and works effectively with temporal scheduling. Our analysis reveals that standard optimization substantially sacrifices policy entropy when simultaneously accommodating heterogeneous behaviors, whereas temporal scheduling yields healthier policy evolution dynamics. Experiments across mathematical and general reasoning benchmarks demonstrate consistent improvements, suggesting that temporal scheduling constitutes a promising optimization dimension.

Reinforcement learning with verifiable rewards (RLVR) has become a standard paradigm for post-training language models, but in practice, verifiers are rarely perfect. Recent theoretical work predicts that verifier noise affects the rate of learning but not its final outcome, implying that sufficient compute should close any gap induced by imperfect supervision. We test this prediction empirically by post-training Qwen2.5 (0.5B, 1.5B) with GRPO on GSM8K while injecting controlled false-positive and false-negative noise into the binary correctness signal, and varying rollouts per prompt as a compute axis. In practice, the gap in validation accuracy persists under substantial compute scaling, with returns to compute that are sharply diminishing. We further find a structural asymmetry where false negatives monotonically degrade performance more quickly than false positives. These findings suggest verifier quality and training compute are not interchangeable, and that reducing false negatives is a more effective lever than scaling compute alone.

Reinforcement learning with verifiable rewards (RLVR) has become a powerful paradigm for improving language models on reasoning-intensive tasks, but its effectiveness is often limited by exploration. For example, models often fail on hard problems, leaving little useful reward signal. External expert traces offer a natural source of guidance, yet they may also expose reward-relevant content along the critical path to the verifier target, such as final answers, intermediate values, executable implementations, or answer-related entities. This content can create an unintended reward hacking channel, allowing the policy to obtain reward by copying the trace rather than learning the underlying reasoning or agentic behavior. Existing guided-RL methods reduce this risk by using partial trajectories, but they mainly control how much expert information is shown heuristically rather than which parts should be hidden. To this end, we propose Semantic Masked Expert Policy Optimization (SMEPO), a fine-grained semantic masking strategy for expert-guided RLVR. Instead of truncating traces coarsely or revealing them unchanged, SMEPO masks reward-relevant semantic spans along the critical path while preserving the expert's decomposition, plan, and procedural structure. This turns hard problems from reasoning from scratch into a fill-in-the-blank process: the policy can follow the expert's problem-solving route, but must still reconstruct the missing values, code, or entities by itself. SMEPO is simple to apply and requires no changes to the reward function or RL objective. Across diverse domains, including math, code, and agentic search, SMEPO improves accuracy by up to 3.2 points over GRPO and reduces training time by up to 4.2x. The code is available at https://github.com/mit-han-lab/SMEPO.

Reinforcement learning with verifiable rewards can improve LLM reasoning, but learning remains sample-inefficient when terminal rewards are sparse. This has motivated a growing line of work on RL with textual feedback, where a critic model generates natural language feedback to guide a reasoning model (the actor), augmenting scalar rewards with richer learning signals. However, existing methods typically treat feedback as fixed or auxiliary, which misses a key property: feedback should not merely be correct, but should improve the policy (actor model) when provided in context. This motivates a paradigm of learnable textual feedback for RL. Yet the learnability and usefulness of feedback depend on the policy's ability to learn from it, making RL with learnable feedback an inherently bilevel problem. We formalize this coupling as a Stackelberg bilevel program and derive Bilevel Natural Language Actor-Critic (Bi-NAC), which jointly trains a critic to generate reward-improving feedback and an actor to exploit it. Across MATH-500, MBPP, and GPQA, Bi-NAC improves sample and parameter efficiency over RL and fixed-critic baselines: our 2B model outperforms the 3B GRPO baseline, achieving 46.6% versus 41.4% on MATH-500, while our 6B model surpasses the 7B GRPO baseline, achieving 49.3% versus 43.6% on GPQA.

Context or prompt-level reweighting has emerged as a central algorithmic lever in Reinforcement Learning with Verified Rewards (RLVR) for improving the reasoning capability of large language models, yet the principle determining what constitutes an optimal weighting remains poorly understood. We address this gap by formulating prompt reweighting as a functional derivative of a utility functional defined in the pass-rate function space, yielding a unified optimality framework that accommodates existing schemes, including REINFORCE and GRPO. Building on this optimality framework, we propose a distribution-aware prompt reweighting approach, called CurveRL, based on a quantile coordinate transform, in which the weight assigned to each prompt depends not on the absolute value of pass rates but on its rank and density to reflect the distributional structure of the pass rates in the learning dynamics. Extensive experiments across multiple benchmarks demonstrate that our proposed CurveRL consistently outperforms GRPO and other RLVR baselines. Our study identifies context-distribution control as a principled axis for analyzing and designing prompt-reweighted RLVR algorithms. The code is released in https://github.com/zhyzmath/CurveRL.

Reinforcement Learning with Verifiable Rewards (RLVR) has emerged as a central paradigm for scaling LLM reasoning, yet its optimization often suffers from training instability and suboptimal convergence. Through a systematic dissection of clipping-based GRPO-style objectives, we identify the rigid clipping decision induced by hard clipping as a key practical bottleneck in the studied RLVR setups. Specifically, our analysis suggests that informative signals can lie in the near-boundary region just beyond the clipping threshold, and are therefore discarded by the standard hard-clipping rule. Notably, once this bottleneck is precisely identified, even simple stochastic perturbations at the boundary can recover meaningful performance gains. Building on this finding, we propose Near-boundary Stochastic Rescue (NSR), a minimal, plug-and-play modification that stochastically retains these slightly out-of-bound tokens to recover lost signals. While NSR, via stochastic sampling, can be interpreted as inducing an implicit gradient decay in expectation, our ablations reveal that its stochastic, boundary-local rescue mechanism is consistently more effective than deterministic gradient decay. Validated by extensive experiments across model sizes from 7B to 30B and both dense and MoE architectures, as a plug-and-play solution, NSR substantially improves training stability and delivers consistent gains over strong baselines such as DAPO and GSPO.

Reinforcement Learning with Verifiable Rewards (RLVR) has become a promising paradigm for scaling reasoning capabilities of Large Language Models (LLMs). However, the sparsity of binary verifier rewards often leads to low efficiency and optimization instability. To stabilize training, existing methods typically impose token-level constraints relative to a reference policy. We identify that such constraints penalize deviations indiscriminately; this can flip verifier-determined direction when the policy attempts to outperform the reference, thereby suppressing gains. To resolve this, we propose One-Way Policy Optimization (OWPO), a method based on the principle of decoupling optimization direction from update magnitude. In OWPO, the verifier dictates the update direction, while the reference policy serves only to adjust the magnitude. Specifically, OWPO applies asymmetric reweighting: it performs Accelerated Alignment for inferior deviations (where the policy lags behind the reference) and Gain Locking for superior deviations (where the policy surpasses the reference). Furthermore, by incorporating iterative reference updates, OWPO creates a ``Ratchet Effect'' that continuously consolidates gains. Experimental results demonstrate that OWPO outperforms strong baselines, including DAPO, OPD, and MOPD, breaking the bottleneck of fixed priors to enable continuous self-evolution without reliance on external reference models.

Reinforcement learning from verifiable rewards (RLVR) has shown strong promise for LLM reasoning, but outcome-based RLVR remains inefficient on hard problems because correct final-answer rollouts are rare and sample-level credit assignment cannot use partial progress in failed attempts. We introduce SCRL (Subproblem Curriculum Reinforcement Learning), a curriculum RL framework that derives verifiable subproblems from reference reasoning chains and fixes the final subproblem as the original problem. This turns partial progress on hard problems into verifiable learning signals. Algorithmically, SCRL uses subproblem-level normalization, which normalizes rewards independently at each subproblem position and assigns the resulting advantages to the corresponding answer spans, enabling finer-grained credit assignment without external rubrics or reward models. Our analysis shows that subproblem curricula lift hard problems out of gradient dead zones, with larger relative gains as the original problem becomes harder. Across seven mathematical reasoning benchmarks, SCRL outperforms strong curriculum-learning baselines, improving average accuracy over GRPO by +4.1 points on Qwen3-4B-Base and +1.9 points on Qwen3-14B-Base. On AIME24, AIME25, and IMO-Bench, SCRL further improves pass@1 by +3.7 points and pass@64 by +4.6 points on Qwen3-4B-Base, indicating better exploration on hard reasoning problems.

Ranking systems used in decision-support settings should not only order candidates but also expose evidence that can be independently checked. We study evidence-certified candidate ranking: given an intent_id, a predefined plan skeleton, a window-local candidate roster, and text-derived candidate trajectories with span provenance, a system must output a Top-K list together with doc_id:span evidence certificates whose cited spans are sufficient to recover the decision. We instantiate this task on MAVEN-ERE and RAMS with fixed upstream extraction, window-local randomized candidate identifiers, skeleton-aligned trajectory supervision, hard negatives, and audit references. We introduce Evidence-Coupled Policy Optimization (ECPO), a listwise policy-optimization objective whose action is the joint object of ranking and evidence certificate. ECPO first learns an interpretable trajectory reward from skeleton alignment, argument consistency, and optional graph features; it then optimizes a constrained policy with three coupled rewards: listwise ranking utility, span-level certificate validity, and an evidence-cycle reward computed by a label-free deterministic verifier that reconstructs candidate support from claim-stripped cited spans. This reframes the goal from maximizing ordinary NDCG alone to maximizing CertNDCG and decision-evidence coupling. The evaluation compares ECPO against zero-shot, SFT, and GRPO policies, RM-only scoring with deterministic evidence attachment, grammar/JSON-constrained decoding, validator retry, best-of-N RM selection, and post-hoc evidence rationalization under closed-roster, predicted-roster, and hybrid-roster settings.