RL Post-Training

Scaling test-time compute has emerged as a powerful mechanism for enhancing Large Language Model (LLM) performance. However, standard post-training paradigms, Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL), optimize the likelihood of individual samples under a base policy, creating a misalignment with test time procedures that rely on aggregated or filtered outputs. In this work, we propose Compute Aligned Training, which aligns training objectives with test-time strategies. By conceptualizing inference strategies as operators on the base policy, we derive new loss functions that maximize performance when said strategies are applied. We instantiate such loss functions for SFT and RL across common test time strategies. Finally, we provide empirical evidence that this training method substantially improves test time scaling over standard training.

Product mapping, the task of deciding whether two e-commerce listings refer to the same product, is a core problem for price monitoring and channel visibility. In real marketplaces, however, sellers frequently inject promotional keywords, platform-specific tags, and bundle descriptions into titles, causing the same product to appear under many different names. Recent LLM-based and multi-agent frameworks improve robustness and interpretability on such hard cases, but they often rely on expensive external APIs, repeated retrieval, and complex inference-time orchestration, making large-scale deployment costly and difficult in privacy-sensitive enterprise settings. To address these issues, we present EPM-RL, a reinforcement-learning-based framework for building an accurate and efficient on-premise e-commerce product mapping model. Our central idea is to distill high-cost agentic reasoning into a trainable in-house model. Starting from a curated set of product pairs with LLM-generated rationales and human verification, we first perform parameter-efficient fine-tuning (PEFT) on a small student model using structured reasoning outputs. We then further optimize the model with Reinforcement Learning (RL) using an agent-based reward that jointly evaluates output-format compliance, label correctness, reasoning--preference scores from specially designed judge models. Preliminary results show that EPM-RL consistently improves over PEFT-only training and offers a stronger quality--cost trade-off than commercial API-based baselines, while enabling private deployment and lower operational cost. These findings suggest that reinforcement learning can turn product mapping from a high-latency agentic pipeline into a scalable, inspectable, and production-ready in-house system.

We present JigsawRL, a cost-efficient framework that explores Pipeline Multiplexing as a new dimension of RL parallelism. JigsawRL decomposes each pipeline into a Sub-Stage Graph that exposes the intra-stage and inter-worker imbalance hidden by stage-level systems. On this abstraction, JigsawRL resolves multiplexing interference through dynamic resource allocation, eliminates fragmented utilization by migrating long-tail rollouts across workers, and formulates their coordination as a graph scheduling problem solved with a look-ahead heuristic. On 4-64 H100/A100 GPUs across different agentic RL pipelines and models, JigsawRL achieves up to 1.85x throughput over Verl on synchronous RL, 1.54x over StreamRL and AReaL on asynchronous RL, and supports heterogeneous pipelines with moderate latency trade-off.

Reward models in RLHF are trained to score only the final token of a response - a choice that discards rich signal from every intermediate position and produces models whose token-level outputs are noise. We argue this is a missed opportunity: a well-trained reward model's output at any token should represent the conditional expectation of the final reward given the response so far. We introduce Temporally Coherent Reward Modeling (TCRM), which induces this property via two regularization terms on top of the standard Bradley-Terry loss, with minimizers provably equal to conditional expectations. The regularizers correspond to Monte Carlo and TD value-learning objectives, establishing a direct connection to RL value functions. TCRM requires zero changes to architecture, data, or inference, yet unlocks three capabilities from one principle: interpretable token-level reward trajectories (middle-token pairwise accuracy improved from 50% to 88.9%, final-token accuracy preserved); state-of-the-art PRM performance on ProcessBench (44.9% average F1) among models trained only on outcome data; and unified reward/value modeling in PPO, reducing peak GPU memory by 27% and step time by 19% with matching LLM quality.

Reinforcement Learning from Verifiable Rewards (RLVR) on chain-of-thought reasoning has become a standard part of language model post-training recipes. A common assumption is that the reasoning chains trained through RLVR reliably represent how a model gets to its answer. In this paper, we develop two metrics for critically examining this assumption: Causal Importance of Reasoning (CIR), which measures the cumulative effect of reasoning tokens on the final answer, and Sufficiency of Reasoning (SR), which measures whether a verifier can arrive at an unambiguous answer based on the reasoning alone. Through experiments with the Qwen2.5 model series and ReasoningGym tasks, we find that: (1) while RLVR does improve task accuracy, it does not reliably improve CIR or SR, calling the role of reasoning in model performance into question; (2) a small amount of SFT before RLVR can be a remedy for low CIR and SR; and (3) CIR and SR can be improved even without SFT by applying auxiliary CIR/SR rewards on top of the outcome-based reward. This joint reward matches the accuracy of RLVR while also leading to causally important and sufficient reasoning. These results show that RLVR does not always lead models to rely on reasoning in the way that is commonly thought, but this issue can be remedied with simple modifications to the post-training procedure.

Offline reinforcement learning (RL) can learn effective policies from fixed datasets, but deployment objectives may change after training, and in many applications the trained actor cannot be retrained because of data, cost, or governance constraints. We study deployment-time adaptation for frozen offline actors using Product-of-Experts (PoE) composition with a goal-conditioned prior. Our main practical finding is graceful degradation rather than universal performance gain: under degraded or random priors, precision-weighted composition remains anchored to the frozen actor, while additive and prior-only adaptation collapse, and a KL-budget selector often recovers a near-oracle operating point. We also make explicit a closed-form identity in the frozen-actor setting: for diagonal-Gaussian actors and priors, PoE with coefficient alpha yields the same deterministic policy as KL-regularized adaptation with beta = alpha / (1 - alpha), with posterior covariances differing only by a global scalar factor. Empirically, across four D4RL environments (3,900 MuJoCo episodes), we observe a 4/5/3 HELP/FROZEN/HURT split. Extending the analysis to six harder cells and two AntMaze diagnostics reveals an actor-competence ceiling: medium-expert remains HURT in all 9 cells at every tested alpha, while AntMaze with a behavior-cloned frozen actor yields zero success for all composition rules. Overall, PoE and KL-regularized adaptation are best viewed as a single actor-anchored safety mechanism for deployment-time steering.

Reinforcement Learning (RL) post-training has become the standard for aligning generative models with human preferences, yet most methods rely on a single scalar reward. When multiple criteria matter, the prevailing practice of ``early scalarization'' collapses rewards into a fixed weighted sum. This commits the model to a single trade-off point at training time, providing no inference-time control over inherently conflicting goals -- such as prompt adherence versus source fidelity in image editing. We introduce ParetoSlider, a multi-objective RL (MORL) framework that trains a single diffusion model to approximate the entire Pareto front. By training the model with continuously varying preference weights as a conditioning signal, we enable users to navigate optimal trade-offs at inference time without retraining or maintaining multiple checkpoints. We evaluate ParetoSlider across three state-of-the-art flow-matching backbones: SD3.5, FluxKontext, and LTX-2. Our single preference-conditioned model matches or exceeds the performance of baselines trained separately for fixed reward trade-offs, while uniquely providing fine-grained control over competing generative goals.

Reinforcement learning (RL) for LLM post-training faces a fundamental design choice: whether to use a learned critic as a baseline for policy optimization. Classical theory favors critic-based methods such as PPO for variance reduction, yet critic-free alternatives like GRPO have gained widespread adoption due to their simplicity and competitive performance. We show that in sparse-reward settings, a learned critic can inject estimation noise that exceeds the state signal it captures, increasing rather than reducing advantage variance. By casting baseline selection as a Kalman filtering problem, we unify PPO and GRPO as two extremes of the Kalman gain and prove that explained variance (EV), computable from a single training batch, identifies the exact boundary: positive EV indicates the critic reduces variance, while zero or negative EV signals that it inflates variance. Building on this insight, we propose Explained Variance Policy Optimization (EVPO), which monitors batch-level EV at each training step and adaptively switches between critic-based and batch-mean advantage estimation, provably achieving no greater variance than the better of the two at every step. Across four tasks spanning classical control, agentic interaction, and mathematical reasoning, EVPO consistently outperforms both PPO and GRPO regardless of which fixed baseline is stronger on a given task. Further analysis confirms that the adaptive gating tracks critic maturation over training and that the theoretically derived zero threshold is empirically optimal.

Reinforcement Learning (RL) enhances LLM reasoning, yet a paradox emerges as models scale: strong base models saturate standard benchmarks (e.g., MATH), yielding correct but homogeneous solutions. In such environments, the lack of failure cases causes the advantage signal in group-relative algorithms (e.g., GRPO) to vanish, driving policies into mode collapse. To address this, we propose Constrained Uniform Top-K Sampling (CUTS), a parameter-free decoding strategy enforcing structure-preserving exploration. Unlike standard sampling that follows model biases, CUTS flattens the local optimization landscape by sampling uniformly from constrained high-confidence candidates. We integrate this into Mixed-CUTS, a training framework synergizing exploitative and exploratory rollouts to amplify intra-group advantage variance. Experiments on Qwen3 models demonstrate that our approach prevents policy degeneration and significantly boosts out-of-domain generalization. Notably, Mixed-CUTS improves Pass@1 accuracy on the challenging AIME25 benchmark by up to 15.1% over standard GRPO, validating that maintaining diversity within the semantic manifold is critical for rigorous reasoning.

Exploration is a cornerstone of learning from experience: it enables agents to find solutions to complex problems, generalize to novel ones, and scale performance with test-time compute. In this paper, we present a framework for post-training language models (LMs) that explicitly encourages optimistic exploration and promotes a synergy between exploration and exploitation. The central idea is to train the LM to generate sets of responses that are collectively accurate under the reward function and exploratory in their reasoning strategies. We first develop a general recipe for optimizing LMs with set reinforcement learning (set RL) under arbitrary objective functions, showing how standard RL algorithms can be adapted to this setting through a modification to the advantage computation. We then propose Polychromic Exploratory Policy Optimization (Poly-EPO), which instantiates this framework with an objective that explicitly synergizes exploration and exploitation. Across a range of reasoning benchmarks, we show that Poly-EPO improves generalization, as evidenced by higher pass@$k$ coverage, preserves greater diversity in model generations, and effectively scales with test-time compute.

Misaligned artificial agents might resist shutdown. One proposed solution is to train agents to lack preferences between different-length trajectories. The Discounted Reward for Same-Length Trajectories (DReST) reward function does this by penalizing agents for repeatedly choosing same-length trajectories, and thus incentivizes agents to (1) choose stochastically between different trajectory-lengths (be NEUTRAL about trajectory-lengths), and (2) pursue goals effectively conditional on each trajectory-length (be USEFUL). In this paper, we use DReST to train deep RL agents and to fine-tune four LLMs (Qwen3-14B, Gemma 4 12B, Granite 4.2 8B, and gpt-oss-20b) to be NEUTRAL and USEFUL. We find that these DReST models generalize to being NEUTRAL and USEFUL in unseen contexts at test time. Indeed, DReST RL agents achieve 11% (PPO) and 17% (A2C) higher USEFULNESS on our test set than default agents, while DReST LLMs achieve high NEUTRALITY while remaining near-maximally USEFUL. We also test our LLMs in an out-of-distribution setting where they can pay costs to influence when shutdown occurs. Relative to default fine-tuning, DReST fine-tuning lowers the share of answers that pay costs to influence shutdown in all four LLMs, by 5 to 20 percentage points, and by 17 to 27 points where the benefit of influencing shutdown is small. Our results thus provide some early evidence that DReST could be used to train more advanced agents to be useful and shutdownable.

Reinforcement learning (RL) has emerged as a powerful post-training paradigm for enhancing the reasoning capabilities of large language models (LLMs). However, reinforcement learning for LLMs faces substantial data scarcity challenges, including the limited availability of high-quality external supervision and the constrained volume of model-generated experience. These limitations make data-efficient reinforcement learning a critical research direction. In this survey, we present the first systematic review of reinforcement learning for LLMs under data scarcity. We propose a bottom-up hierarchical framework built around three complementary perspectives: the data-centric perspective, the training-centric perspective, and the framework-centric perspective. We develop a taxonomy of existing methods, summarize representative approaches in each category, and analyze their strengths and limitations. Our taxonomy aims to provide a clear conceptual foundation for understanding the design space of data-efficient RL for LLMs and to guide researchers working in this emerging area. We hope this survey offers a comprehensive roadmap for future research and inspires new directions toward more efficient and scalable reinforcement learning post-training for LLMs.

Reinforcement learning (RL) has emerged as a promising paradigm for training reasoning-oriented models by leveraging rule-based reward signals. However, RL training typically tends to improve single-sample success rates (i.e., Pass@1) while offering limited exploration of diverse reasoning trajectories, which is crucial for multi-sample performance (i.e., Pass@k). Our preliminary analysis reveals that this limitation stems from a fundamental squeezing effect, whereby probability mass is excessively concentrated on a narrow subset of high-reward trajectories, restricting genuine exploration and constraining attainable performance under RL training. To address this issue, in this work, we propose Steering Probability Squeezing (SPS), a training paradigm that interleaves conventional RL with inverse reinforcement learning (IRL). SPS treats on-policy rollouts as demonstrations and employs IRL to explicitly reshape the induced trajectory distribution, thereby enhancing exploration without introducing external supervision. Experiments on five commonly used reasoning benchmarks demonstrate that SPS can enable better exploration and improve Pass@k. Beyond algorithmic contributions, we provide an analysis of RL learning dynamics and identify an empirical upper bound on Pass@k, shedding light on intrinsic exploration limits in RL-based reasoning models. Our findings suggest that alternating between RL and IRL offers an effective pathway toward extending the exploration capacity of reasoning-oriented large language models.

Reinforcement Learning (RL) has achieved impressive success in post-training Large Language Models (LLMs) and Vision-Language Models (VLMs), with on-policy algorithms such as PPO, GRPO, and REINFORCE++ serving as the dominant paradigm. However, these methods discard all collected trajectories after a single gradient update, resulting in poor sample efficiency, particularly wasteful for agentic tasks where multi-turn environment interactions are expensive. While Experience Replay drives sample efficiency in classic RL by allowing agents to reuse past trajectories and prioritize informative ones, directly applying Prioritized Experience Replay (PER) to LLMs fails. The rapid policy evolution of billion-parameter models renders stored priorities stale, causing old high-priority trajectories to dominate sampling long after they have become uninformative. We propose Freshness-Aware PER, which addresses this priority staleness problem by augmenting any PER-based priority with a multiplicative exponential age decay grounded in effective sample size analysis. To the best of our knowledge, Freshness-Aware PER is the first work to successfully apply PER to LLM/VLM reinforcement learning. We evaluate on eight multi-step agentic, reasoning, and math competition tasks with 0.5B, 3B, and 7B models. Freshness-Aware PER significantly outperforms on-policy baselines, achieving +46% on NQ Search, +367% on Sokoban, and +133% on VLM FrozenLake, while standard PER without age decay consistently degrades performance. Our code is publicly available at https://github.com/Vision-CAIR/Freshness-Aware-PER.

Frontier models have demonstrated exceptional capabilities following the integration of task-reward-based reinforcement learning (RL) into their training pipelines, enabling systems to evolve from pure reasoning models into sophisticated agents. However, debate persists regarding whether RL genuinely instills new skills within a base model or merely sharpens its existing distribution to elicit latent capabilities. To address this dichotomy, we present an explicit comparison between distribution sharpening and task-reward-based learning, utilizing RL as a tool to implement both paradigms. Our analysis reveals the inherent limitations of distribution sharpening, demonstrating from first principles how and why the optima can be unfavorable and the approach fundamentally unstable. Furthermore, our experiments using Llama-3.2-3B-Instruct, Qwen2.5-3B-Instruct and Qwen3-4B-Instruct-2507 on math datasets confirm that sharpening yields limited gains, whereas incorporating task-based reward signal can greatly help achieve robust performance improvements and stable learning.

Current post-training methodologies for adapting Large Vision-Language Models (LVLMs) generally fall into two paradigms: Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL). Despite their prevalence, both approaches suffer from inefficiencies when applied in isolation. SFT forces the model's generation along a single expert trajectory, often inducing catastrophic forgetting of general multimodal capabilities due to distributional shifts. Conversely, RL explores multiple generated trajectories but frequently encounters optimization collapse - a cold-start problem where an unaligned model fails to spontaneously sample any domain-valid trajectories in sparse-reward visual tasks. In this paper, we propose Supervised Group Relative Policy Optimization (S-GRPO), a unified post-training framework that integrates the guidance of imitation learning into the multi-trajectory exploration of preference optimization. Tailored for direct-generation visual tasks, S-GRPO introduces Conditional Ground-Truth Trajectory Injection (CGI). When a binary verifier detects a complete exploratory failure within a sampled group of trajectories, CGI injects the verified ground-truth trajectory into the candidate pool. By assigning a deterministic maximal reward to this injected anchor, S-GRPO enforces a positive signal within the group-relative advantage estimation. This mechanism reformulates the supervised learning objective as a high-advantage component of the policy gradient, compelling the model to dynamically balance between exploiting the expert trajectory and exploring novel visual concepts. Theoretical analysis and empirical results demonstrate that S-GRPO gracefully bridges the gap between SFT and RL, drastically accelerates convergence, and achieves superior domain adaptation while preserving the base model's general-purpose capabilities.

Reinforcement Learning with Verifiable Rewards (RLVR) is widely used for post-training Large Language Models, but practical verifiers can make errors. We study how the rate and structure of reward noise affect RLVR in code generation, with a preliminary scientific-reasoning check. In multi-seed Qwen3 8B experiments on MBPP, mean validation reward over two fixed late evaluations is within 1 percentage point of the clean baseline at the tested resampled group-rollout noise rates through 20%, and within about 2 points at 30%. Confidence intervals allow larger losses; these point estimates do not establish a general tolerance threshold. We also examine full-program pass@k, four controlled noise structures, two model-based verifiers, and policy models from three families spanning 4B-9B parameters. We derive conditional advantage distributions for symmetric and asymmetric group noise, including retained format penalties, and show why clipping limits simple gradient-scaling arguments. The analysis identifies information preserved by whole-group corruption and limits on interpreting our asymmetric sweep as a precision-recall comparison. Overall, the results indicate that imperfect verification can support effective RLVR in the tested settings, while aggregate error rates alone do not characterize the learning signal.

Building general-purpose reasoning models using reinforcement learning with verifiable rewards (RLVR) across diverse domains has been widely adopted by frontier open-weight models. However, their training recipes and domain mixtures are often not disclosed. Joint optimization across domains poses significant challenges: domains vary widely in rollout length, problem difficulty, and sample efficiency. Further, models with long chain-of-thought traces increase inference cost and latency, making efficiency critical for practical deployment. We present Apriel-Reasoner, trained with a reproducible multi-domain RL post-training recipe on Apriel-Base, a 15B-parameter open-weight LLM, across five domains using public datasets: mathematics, code generation, instruction following, logical puzzles, and function calling. We introduce adaptive domain sampling that preserves target completed-rollout ratios despite heterogeneous rollout dynamics, and a difficulty-aware length penalty that, at no additional training overhead, encourages longer reasoning for difficult problems and shorter traces for easy ones. Trained with a strict 16K-token output budget, Apriel-Reasoner remains effective at a 32K output budget and improves over Apriel-Base on AIME 2025, GPQA, MMLU-Pro, and LiveCodeBench while producing 30-50% shorter reasoning traces. Among the evaluated open-weight models of similar scale, it improves the accuracy-token tradeoff using a fully public-data recipe.

Reinforcement learning has become a central post-training paradigm for improving LLM and agent capabilities. Yet existing RL post-training methods share a common blind spot: they construct local learning signals from sampled trajectories, rewards, or feedback-conditioned targets, then update the policy without explicitly verifying whether the resulting policy outperforms its predecessor. Optimizing these local signals does not necessarily produce a better policy, while finite sampling, generation stochasticity and feedback noise can further widen this gap. We argue that the missing ingredient is policy improvement feedback: the ability to measure progress across policy iterations. We introduce Policy Improvement Reinforcement Learning (PIRL), which formulates inter-iteration performance gain as an explicit objective structurally aligned with final task performance. Building on PIRL, we propose Policy Improvement Policy Optimization (PIPO), a plug-in closed-loop framework that verifies the previous update against a sliding-window historical performance anchor. PIPO uses this improvement feedback to modulate the local learning signal of the base policy optimization algorithm, reinforcing updates associated with measured progress and suppressing those associated with performance drops. We provide theoretical evidence that PIPO locally aligns policy updates with the PIRL improvement objective. Experiments on mathematical reasoning, code, tool-use, and self-distillation settings show that PIPO yields consistent gains across PPO, group-relative, and self-distillation policy optimization families.

Reinforcement learning (RL) has become a standard technique for post-training diffusion-based image synthesis models, as it enables learning from reward signals to explicitly improve desirable aspects such as image quality and prompt alignment. In this paper, we propose an online RL variant that reduces the variance in the model updates by sampling paired trajectories and pulling the flow velocity in the direction of the more favorable image. Unlike existing methods that treat each sampling step as a separate policy action, we consider the entire sampling process as a single action. We experiment with both high-quality vision language models and off-the-shelf quality metrics for rewards, and evaluate the outputs using a broad set of metrics. Our method converges faster and yields higher output quality and prompt alignment than previous approaches.

Post-training with reinforcement learning (RL) has recently shown strong promise for advancing multimodal agents beyond supervised imitation. However, RL remains limited by poor data efficiency, particularly in settings where interaction data are scarce and quickly become outdated. To address this challenge, GIPO (Gaussian Importance sampling Policy Optimization) is proposed as a policy optimization objective based on truncated importance sampling, replacing hard clipping with a log-ratio-based Gaussian trust weight to softly damp extreme importance ratios while maintaining non-zero gradients. Theoretical analysis shows that GIPO introduces an implicit, tunable constraint on the update magnitude, while concentration bounds guarantee robustness and stability under finite-sample estimation. Experimental results show that GIPO achieves state-of-the-art performance among clipping-based baselines across a wide range of replay buffer sizes, from near on-policy to highly stale data, while exhibiting superior bias--variance trade-off, high training stability and improved sample efficiency. Code is available at https://github.com/distanceLu/GIPO.

Despite their sophisticated general-purpose capabilities, Large Language Models (LLMs) often fail to align with diverse individual preferences because standard post-training methods, like Reinforcement Learning with Human Feedback (RLHF), optimize for a single, global objective. While Group Relative Policy Optimization (GRPO) is a widely adopted on-policy reinforcement learning framework, its group-based normalization implicitly assumes that all samples are exchangeable, inheriting this limitation in personalized settings. This assumption conflates distinct user reward distributions and systematically biases learning toward dominant preferences while suppressing minority signals. To address this, we introduce Personalized GRPO (P-GRPO), a novel alignment framework that decouples advantage estimation from immediate batch statistics. By normalizing advantages against preference-group-specific reward histories rather than the concurrent generation group, P-GRPO preserves the contrastive signal necessary for learning distinct preferences. We evaluate P-GRPO across diverse tasks and find that it consistently achieves faster convergence and higher rewards than standard GRPO, thereby enhancing its ability to recover and align with heterogeneous preference signals. Our results demonstrate that accounting for reward heterogeneity at the optimization level is essential for building models that faithfully align with diverse human preferences without sacrificing general capabilities.

Reinforcement Learning (RL) for Large Language Models (LLMs) often suffers from training collapse in long-horizon tasks due to exploding gradient variance. To mitigate this, a baseline is commonly introduced for advantage computation; however, traditional value models remain difficult to optimize, and standard group-based baselines overlook sequence heterogeneity. Although classic optimal baseline theory can achieve global variance reduction, it neglects token heterogeneity and requires prohibitive gradient-based computation. In this work, we derive the Optimal Token Baseline (OTB) from first principles, proving that gradient updates should be weighted inversely to their cumulative gradient norm. To ensure efficiency, we propose the Logit-Gradient Proxy that approximates the gradient norm using only forward-pass probabilities. Our method achieves training stability and matches the performance of large group sizes ($N=32$) with only $N=4$, reducing token consumption by over 65% across single-turn and tool-integrated reasoning tasks.

RLVR has become a standard recipe for training LLMs on reasoning tasks with verifiable outcomes, but when rollout generation dominates the cost, efficiency hinges on which prompts are sampled and when. In practice, prompt pools are often static or only weakly coupled to policy progress, so uniform sampling fails to track the moving capability frontier and wastes rollouts on regions that are already solved or still unreachable. Prior methods improve efficiency via filtering, curricula, adaptive rollout allocation, or teacher guidance, but they often assume a fixed pool, which does not support stable on-policy pool growth, or they introduce additional teacher cost and latency. In this work, we propose HeaPA (Heap Sampling and On-Policy Query Augmentation), which maintains a bounded, evolving pool, tracks the frontier with heap-based boundary sampling, grows the pool via on-policy augmentation under lightweight asynchronous validation, and stabilizes correlated queries via topology-aware pool statistics re-estimation and controlled reinsertion. Across two training corpora, two training recipes, and seven benchmarks, HeaPA consistently improves accuracy and reaches target performance with fewer computations at comparable wall-clock time. Analyses attribute the gains to frontier-focused sampling and on-policy pool growth, with more pronounced improvements at mid-to-large model scales. Our training code is publicly available at https://github.com/horizon-llm/HeaPA.

On-policy reinforcement learning (RL) for language model post-training suffers from a fundamental tension: as training progresses, policy entropy collapses and sampling diversity diminishes, causing the model to ``forget'' its own earlier exploratory capacity. While off-policy data can restore diversity, existing methods mix entire trajectories at the sequence level, introducing severe policy mismatch and training instability. We argue that the core question is not \emph{whether} to use off-policy data, but \emph{where} in the sequence it should appear. Based on this insight, we propose \textbf{Thinking Seeds}, a token-level mix-policy framework that uses the model's own historical checkpoints as off-policy prefixes, providing diverse starting points for reasoning, while the critical continuation is generated on-policy to preserve gradient quality. Through token-level importance ratios, Thinking Seeds effectively leverages historical diversity without compromising training stability. Extensive experiments across models and mathematical reasoning benchmarks demonstrate that Thinking Seeds consistently outperforms standard on-policy training and existing off-policy extensions. Our analysis reveals that the method maintains higher effective entropy, reduces gradient loss from clipping, and expands the explorable solution space, clarifying how position-aware mix-policy modeling improves both exploration and final performance in LLM RL.

Existing reinforcement learning methods for LLM reasoning implicitly assume that the policy generating training trajectories should coincide with the one producing inference responses. We argue that this is a misleading inductive bias: the optimization-optimal trajectory distribution favors informative gradients, whereas the inference-optimal response distribution emphasizes accuracy and consistency. Forcing both into a single policy entangles their gradients and suppresses exploration. We propose R$^2$PO (Residual Rollout Policy Optimization), which attaches a lightweight Residual Rollout-Head atop the policy to decouple training trajectories from inference responses, diversifying rollouts during training while keeping inference generation intact. Experiments show that R$^2$PO consistently outperforms baselines, with average accuracy gains of 3.4% on MATH-500 and 1.3% on APPS, alongside more diverse rollouts and reduced length bias. Our code is available at https://github.com/RRPO-ARR/Code.

Reinforcement learning (RL) algorithms such as PPO and GRPO are widely used to train large language models (LLMs) for multi-turn agentic tasks. However, in off-policy training pipelines, these methods can exhibit unstable optimization dynamics and are prone to perfor- mance collapse. Through empirical analysis, we identify two fundamental sources of instability in this setting: (1) a granularity mismatch between token-level policy optimization and turn- structured interactions, and (2) high-variance and unreliable gradient updates induced by off- policy importance sampling and inaccurate advantage estimation. To address these challenges, we propose SORL, Stabilizing Off-Policy Reinforcement Learning for Long-Horizon Agent Train- ing. SORL introduces mechanisms that align policy optimization with the structure of multi- turn interactions and adaptively suppress unreliable off-policy updates, yielding more conserva- tive and robust learning dynamics. Within this framework, we instantiate two stabilized algo- rithms: SO-PPO and SO-GRPO. Both algorithms are designed to mitigate gradient variance and prevent optimization collapse without requiring careful early stopping or heuristic tuning. We evaluate SO-PPO and SO-GRPO on benchmarks spanning open-domain QA, multi-hop QA, and medical multiple-choice QA, and further assess their transfer to asynchronous RL for mathe- matical reasoning by training on DAPO-Math-17k and validating on AIME-2024. These results demonstrate that SORL provides a practical, scalable, and general framework for stabilizing re- inforcement learning in multi-turn LLM agent training and asynchronous RL for mathematical reasoning.

Adapting language models (LMs) to new tasks via post-training carries the risk of degrading existing capabilities -- a phenomenon classically known as catastrophic forgetting. In this paper, toward identifying guidelines for mitigating this phenomenon, we systematically compare the forgetting patterns of two widely adopted post-training methods: supervised fine-tuning (SFT) and reinforcement learning (RL). Our experiments reveal a consistent trend across LM families (Llama, Qwen) and tasks (instruction following, general knowledge, and arithmetic reasoning): RL leads to less forgetting than SFT while achieving comparable or higher target task performance. To investigate the cause for this difference, we consider a simplified setting in which the LM is modeled as a mixture of two distributions, one corresponding to prior knowledge and the other to the target task. We identify that the mode-seeking nature of RL, which stems from its use of on-policy data, enables keeping prior knowledge intact when learning the target task. We then verify this insight by demonstrating that the use on-policy data underlies the robustness of RL to forgetting in practical settings, as opposed to other algorithmic choices such as the KL regularization or advantage estimation. Lastly, as a practical implication, our results highlight the potential of mitigating forgetting using approximately on-policy data, which can be substantially more efficient to obtain than fully on-policy data.

Reinforcement learning (RL) promises to expand the capabilities of language models, but it is unclear if current RL techniques promote the discovery of novel behaviors, or simply sharpen those already present in the base model. In this paper, we investigate the value of deliberate exploration -- explicitly incentivizing the model to discover novel and diverse behaviors -- and aim to understand how the knowledge in pre-trained models can guide this search. Our main finding is that exploration with a simple, principled, representation-based bonus derived from the pre-trained language model's hidden states significantly improves diversity and pass@k rates -- both for post-training, and in a novel inference-time scaling setting we introduce. For inference-time, exploration with representation-based diversity improves efficiency, consistently improving pass@k rates across a variety of models and reasoning tasks. For example, for Qwen-2.5-14b-Instruct we obtain over 50% improvement in verifier efficiency on almost all tasks. For post-training, we show that integrating this exploration strategy into an RL pipeline improves reasoning performance over that of the initial model and over standard RL post-training. For example, on AIME 2024, our post-trained Qwen-2.5-7b-Instruct's pass@80 matches the pass@256 of GRPO on the same model, demonstrating a 3x improvement in test-time sample efficiency. Overall, our findings suggest that deliberate exploration -- with the right notion of diversity -- is a practical path toward discovery of new behaviors beyond sharpening.

Today's generative models thrive with large amounts of supervised data and informative reward functions characterizing the quality of the generation. They work under the assumptions that the supervised data provides knowledge to pre-train the model, and the reward function provides dense information about how to further improve the generation quality and correctness. However, in the hardest instances of important problems, two problems arise: (1) the base generative model attains a near-zero reward signal, and (2) calls to the reward oracle are expensive. This setting poses a fundamentally different learning challenge than standard reward-based post-training. To address this, we propose BaNEL (Bayesian Negative Evidence Learning), an algorithm that post-trains the model using failed attempts only, while minimizing the number of reward evaluations (NREs). Our method is based on the idea that the problem of learning regularities underlying failures can be cast as another, in-loop generative modeling problem. We then leverage this model to assess whether new data resembles previously seen failures and steer the generation away from them. We show that BaNEL can improve model performance without observing a single successful sample on several sparse-reward tasks, outperforming existing novelty-bonus approaches by up to several orders of magnitude in success rate, while using fewer reward evaluations.