Generative RL 进展

Historical street-view imagery records urban evolution, but uneven coverage leaves substantial gaps in historical records. Generating plausible past appearances requires restoring changed structures while preserving persistent scene content. We construct VIGOR-his, a decade-spanning cross-view dataset containing 43,653 location-level quadruplets across 11 cities on three continents. Its automated pipeline performs spatial pairing, consistency screening, change classification, and the generation and validation of satellite-based change descriptions and local editing instructions. Based on VIGOR-his, we propose CrossTimeEdit, a model that reformulates historical street-view generation as editing, using recent street views to constrain viewpoint and unchanged appearance and temporal satellite differences as change evidence. Starting from FLUX.2 [Klein] 4B, we train CrossTimeEdit through supervised fine-tuning (SFT) followed by online reinforcement learning (RL). We design three street-view editing criteria, namely Instruction Alignment (IA), Background Preservation (BP), and Quality and Physical Plausibility (QP), as both RL reward dimensions and evaluation metrics. We optimize this multi-reward objective using Within Group Relative Policy Optimization for flow-matching models (Flow-GRPO) with Group reward-Decoupled Normalization Policy Optimization (GDPO), which normalizes each reward dimension before aggregation. CrossTimeEdit improves overall performance across the three editing criteria by 17.12% over the pretrained baseline and outperforms cross-view generation models in scene consistency, visual realism, and perceptual quality. The implementation code, dataset, and model weights are available at https://luhanwen67.github.io/CrossTimeEdit-release/.

Vision-Language Models (VLMs) can generate rich video captions, yet often misidentify which person performs an action or which limb is involved, particularly across camera cuts. Improving these details requires evaluation and training that distinguish missing information from incorrect assertions. We introduce FlexBench, a benchmark spanning 3,105 shots and 18,161 evaluation queries, with human-verified identities and systematic per-person coverage of fine-grained limb actions and states. Its reference-derived checklists support automated assessment of complete captions in their person and shot contexts. Our Graded Physical Alignment score (GPA) awards credit for correct content and deducts points for incorrect or fabricated actions, making these errors explicit in the aggregate score. Building on this rubric, we propose Graded Margin Direct Preference Optimization (GM-DPO), which assigns stronger preference margins and greater training weight to more severe action errors. Across three VLM backbones, GM-DPO achieves the highest substantive-action and GPA scores among the evaluated preference objectives, improving GPA over DPO by 2.02-3.40 points. On Qwen3-8B, it reduces the weighted hallucination rate by 21.3% relative to DPO. These gains accompany sustained long-form output, improved shot structure, and competitive performance on three additional multimodal benchmarks.

Language-model agents are usually trained by reinforcement learning from one reward per episode, and privileged self-distillation enriches it by letting the same policy, given a skill, teach its skill-free self through token probabilities. However, we identify two phenomena that question this channel. Invisible Advantage: a skill in context lifts WebShop success from 42.2% to 56.2%, yet changes the probabilities of fewer than a quarter of the sampled tokens. Much to Align: a skill changes the hidden states of over 80% of response tokens, in a way that linear probes can trace back to the specific skill. To exploit this, we propose Privileged Representation On-policy Self-Distillation (PR-OPD). After a GRPO warm start, the policy writes a hindsight skill for each trajectory, re-reads its own responses with that skill as a stop-gradient teacher, and aligns its projected hidden states to the teacher's at every layer alongside the reward objective, with no external skill library, separate teacher, or inference overhead. On ALFWorld and WebShop with two backbones, PR-OPD achieves the best overall results in every setting, improving over GRPO by up to 4.7 points in ALFWorld success and 14.0 points in WebShop accuracy. Code is available at https://github.com/balibata/PR-OPD.

Reinforcement learning with verifiable rewards (RLVR) has become a standard paradigm for improving large language models (LLMs) on various tasks, yet its sparse outcome rewards lack token-level credit assignment for intermediate steps. To address this, on-policy self-distillation (OPSD) leverages a self-teacher with privileged context to provide additional dense learning signals. However, because the self-teacher is often overconfident and imposes excessive penalties on long reasoning trajectories, OPSD frequently struggles in practice. To mitigate this, we propose self-instructing policy optimization (SIPO) with a contrastive self-teacher to provide dense credit. At each iteration, SIPO samples multiple rollouts per prompt from the current policy, scores them with environment rewards, and constructs two teacher contexts for each rollout by pairing the reference answer with mistakes made within the group. The model then re-evaluates its own responses under both contexts, using the difference between the two teacher log-probabilities as token-level feedback, so that biases shared by both contexts are expected to largely cancel. The resulting objective yields a token-level advantage for every rollout: the reward still sets the main direction of each update while the self-teacher redistributes credit across tokens. Even in groups where every rollout fails and group-relative advantages vanish, SIPO still provides a learning signal. By preserving direct optimization of the task reward while providing dense, token-level feedback, this approach bridges reinforcement learning and on-policy self-distillation. Extensive experiments across multiple reasoning and code-generation benchmarks demonstrate that SIPO outperforms both RLVR and OPSD baselines without an external teacher or additional generation.

On-policy distillation (OPD) trains a student to match the teacher's next-token distributions on the student's own trajectories and has yielded substantial empirical gains. Generalized variants allow the student to surpass the teacher by extrapolating an implicit reward in output space. The language-model head, however, attenuates this change anisotropically: much of the change encoded in the teacher's hidden states reaches the logits at a small fraction of its weight, and the sampled-token log-probability ratios on which output-space extrapolation relies inject noise that the extrapolation amplifies, making training unstable. We observe that reinforcement learning (RL) shifts a model's internal representations relative to its base checkpoint, and that the direction of this shift can be measured at every layer. Motivated by this observation, we propose RIDE (RL-Induced Direction Extrapolation), which extrapolates the RL-induced change directly in representation space: at every layer and token position, RIDE computes the residual between the teacher and its pre-RL checkpoint and regresses the student's hidden states toward targets displaced beyond the teacher along this residual. Conditioned on a sampled trajectory, this regression is equivalent to maximizing a linear directional reward defined by the residual under a quadratic penalty centered at the teacher, which makes explicit how the objective moves the student along the RL-induced direction while limiting its deviation from the teacher. Across four base/RL-teacher pairs spanning different scales, architectures, and pre-training lineages, RIDE approaches or exceeds the RL-trained teacher on every pair and is the only method whose mean does so, and it consistently outperforms output-space extrapolation, which degrades the student whenever the teacher is close to its base. Project page: https://github.com/xixixixixxxx/RIDE.

Self-Distillation Fine-Tuning (SDFT) enables a language model to act as its own teacher: by conditioning on a demonstration, the model produces an implicit reward via pointwise mutual information, which guides on-policy learning without external supervision. However, SDFT operates at training time: it requires gradient updates and access to expert demonstrations, making it inapplicable at inference. We propose test-time self-distillation, a decoding-time method that extracts a steering signal from the self-distillation framework without any parameter updates, reward models, or training data. Our key insight is that counterfactual contexts, i.e. fixed textual templates that hypothetically prime the model for excellent versus poor reasoning, can substitute for the demonstration. The log-odds ratio of a candidate answer under these two counterfactual conditions defines a new reward signal. We derive the optimal KL-regularized policy under this reward, which takes the form of a Gibbs reweighting of the base distribution. Crucially, this reweighting is global: it cannot be decomposed into independent per-token operations without ignoring future trajectory quality. We therefore approximate the target distribution via beam search. Experiments on mathematical reasoning (MATH500), code generation (HumanEval), and graduate-level science QA (GPQA) across multiple model scales show that test-time self-distillation improves over standard sampling, low temperature, beam search and power sampling baselines on average, demonstrating that the self-distillation principle can be operationalized at inference time.

Reinforcement learning for long-horizon agents typically relies on sparse outcome-based rewards. This leads to a severe cold-start problem, as early-stage policies often fail to solve sampled tasks, leaving little useful reward signal for learning. To mitigate this problem, we use on-policy distillation (OPD) to provide token-level guidance on the student's own rollouts. We find that the benefit of this guidance depends on the performance gap between the teacher and the student. When the teacher substantially outperforms the student, distillation helps guide the student through the early training stage where outcome rewards provide little learning signal. As the gap narrows and eventually reverses, however, continued distillation becomes less beneficial and may hinder further improvement. Motivated by this observation, we propose Gap-Adaptive Teacher Scheduling (GATS), which augments the student's RL objective with an OPD term whose weight adapts to the teacher-student performance gap. Specifically, GATS gradually reduces teacher guidance as the student approaches the teacher's reference performance and withdraws it once that reference is reached. This enables GATS to leverage task-trained teachers smaller than the student, since teacher guidance is primarily needed during early training. Across ALFWorld, WebShop, and ScienceWorld with three Qwen2.5 teacher-student configurations, GATS achieves the highest average success rate among the compared methods in all three configurations, improving over reward-only GRPO by 4.37%-11.87% under matched student rollout budgets. Code is available at https://github.com/Ricardo-H/guide-then-let-go.

Explicit intermediate reasoning gives large language models (LLMs) a stronger problem-solving mode. We study learning from this think-mode advantage via on-policy distillation (OPD). OPD preserves student-generated trajectories and provides dense token-level teacher targets at student-visited prefixes. Privileged reasoning is used during distillation rather than student inference. Uniform ThinkOPD, a natural think-enabled OPD baseline, conditions a fixed teacher on one shared think trace and uniformly distills every sibling student response. Although its prefixes are on-policy, the trace need not follow a route compatible with every complete response: the same privileged trace can induce different teacher-student discrepancies even when responses reach the same outcome. We summarize this interaction with trace-response divergence (TRD) and introduce ThinkOPD, which routes supervision at the response level by combining group-relative reward gain with a TRD-based compatibility proxy. Final response weights are normalized within each rollout group. Across mathematical reasoning and code generation, ThinkOPD outperforms Uniform ThinkOPD in both same-model settings and both cross-model teacher-student pairs, and it exceeds representative rationale and self-distillation baselines in a controlled comparison. Controlled interventions show that outcome benefit and the TRD-based proxy provide complementary routing signals in this setting. Think-enabled OPD provides a controlled setting for studying how teacher advantage becomes transferable along student responses.

Despite rapid progress in video generation models, they still exhibit obvious motion deficiencies, often manifested as incorrect object motion. However, most existing video quality evaluations focus on aesthetic quality or text-video alignment. To address this gap, we study object-centric motion fidelity assessment, evaluating target objects along object consistency, motion continuity, and physical plausibility. To achieve this, we first introduce VidMotion, a diagnostic dataset of 6,879 videos with designated moving objects and fine-grained annotations including dimension-wise scores and failure causes. We further propose MotionInsight, a diagnostic evaluator that shifts assessment from implicit RGB-frame observation to explicit motion-space diagnosis. By constructing motion-aware representations, MotionInsight makes subtle motion deficiencies more observable. We also introduce motion-specific rewards during GRPO to transform observed motion into a diagnostic assessment. Experiments demonstrate that MotionInsight provides an effective basis for diagnosing object motion deficiencies, producing human-aligned scores along three dimensions and grounded explanations.

Visual reward models are essential for evaluating and improving visual generation models, yet existing approaches typically map task conditions and candidate outputs directly to scalar rewards, leaving implicit what should be evaluated for each individual case. We introduce Think Before You Score, a paradigm that explicitly determines what matters for each case before judging how well the candidate performs. Following this principle, we propose the Thinking Reward Model (TRM), which formulates case-adaptive rubrics, performs rubric-guided assessment, and produces fine-grained pointwise rewards. We further observe that conventional pairwise preference optimization can induce score polarization, and introduce Pairwise Dual-Group Relative Policy Optimization (PD-GRPO), which leverages pairwise supervision to improve reward discrimination while preserving fine-grained pointwise scoring. Extensive experiments on image generation and editing reward-modeling benchmarks demonstrate that TRM achieves state-of-the-art performance among open-source reward models while remaining highly competitive with proprietary alternatives. Moreover, using TRM as a reward for reinforcement learning consistently improves diverse visual generation models, demonstrating that its fine-grained, case-adaptive rewards translate into effective optimization signals for visual generation.

Power-sharpened sampling is an inference-time alternative to reinforcement-learning (RL) post-training for enhancing reasoning in large language models (LLMs). High-probability sequences are amplified under the base model without parameter updates or external rewards, avoiding the costly optimization and jagged generalization of RL. However, this approach faces a fundamental exploration--exploitation trade-off, as % strong sharpening restricts exploration, trapping samplers in plausible but incorrect reasoning trajectories, whereas weak sharpening leaves the answer distribution diffuse. To resolve this trade-off, we introduce Parallel Power Tempering (PPT), instantiating power-sharpened LLM sampling via parallel tempering. Running multiple interacting replicas in parallel at different sharpening levels allows lower-power replicas to explore diverse reasoning trajectories and higher-power chains to further exploit higher-likelihood responses favored by the sharpened target. Specifically, we tailor \method{} to inference-time sampling by mitigating a truncation bias, identified in prior power samplers, and investigate effective swap strategies under finite memory and compute budgets. Extensive experimentation shows that \method{} substantially improves single-chain power-sharpened sampling and outperforms RL-post-trained models, producing higher-quality reasoning traces and even achieving performance comparable to frontier models.

Fully asynchronous reinforcement learning (RL) improves resource utilization in large language model post-training by overlapping rollout generation with policy optimization, but it also introduces policy lag as trajectories are generated and queued while the trainer continues to update. We study how this lag accumulates over a trajectory's lifetime and how it can be controlled without sacrificing the wall-clock benefits of asynchronous execution. We decompose trajectory staleness into Generation Staleness, accumulated before rollout completion, and Waiting Staleness, accumulated after a completed trajectory enters the pool. Motivated by this decomposition, we introduce PACE (Pool-Aware Control of Effective Staleness). PACE converts excess pool occupancy into an adaptive rejection budget and ranks completed trajectories using an effective-staleness score that combines Waiting Staleness with prefix-aware Generation Staleness. This avoids penalizing long or interrupted rollouts solely because they span multiple policy versions. In single-turn mathematical reasoning, PACE improves the six-benchmark average validation accuracy by 18.7% over unfiltered asynchronous RL at the same wall-clock budget and matches synchronous RL performance with 47.1% less GPU time. PACE also improves validation performance in multi-turn tool-integrated reasoning, outperforming both synchronous and unfiltered asynchronous RL. Further experiments with the mixture-of-experts model and an alternative RL algorithm support its applicability across model architectures and training algorithms.

Post-training has been shown to significantly improve language models' performance on tasks with verifiable outcomes, including mathematical reasoning, software engineering, and computer use. However, whether the same approach can improve forecasting in financial markets is much less clear. Compared with tasks with verifiable outcomes, not only are realized returns noisy, but even what constitutes a relevant information set for making effective predictions is not obvious a priori: the model must decide which observations to gather and then commit to a numerical judgment before the outcome is known. We study this question in a chronological stock-price sandbox, where a language model gathers price, volume, relative-performance, and market-context evidence and predicts a future return. We post-train Qwen3-4B with supervised fine-tuning (SFT) on tool-use demonstrations, then proximal policy optimization (PPO) with a terminal reward given by the forecast score against the realized return. The resulting AURA-4B more than doubles the starting direction--magnitude score, from 20.94 to 43.31, and is comparable to frontier language models on this benchmark. Conditional magnitude agreement rises from 33.3 to 66.2, while directional accuracy changes from 62.9 to 65.4. SFT expands tool use, and PPO further increases the share of ranking and market-context queries. These results show that post-training can substantially improve financial forecasting performance, together with changes in how the model investigates the market, on this outcome-selected benchmark.

Agentic reinforcement learning (ARL) with verifiable rewards improves the ability of large language models (LLMs) to tackle knowledge-intensive tasks by learning to interleave search and reasoning. However, most existing ARL methods optimize only LLM-generated tokens and treat retrieved evidence as environment observations. This creates an information-credit gap: failures caused by missing or misleading evidence are attributed to the LLM policy rather than to the retriever, which motivates training the LLM and the retriever jointly. In this paper, we show that retrieval and LLM policy learning are order-sensitive: adapting the retriever before optimizing the policy yields a larger reward gain than the reverse order. To preserve this hierarchy while allowing both components to co-adapt, we formulate retrieval-augmented agentic RL as a bilevel optimization problem. To solve it efficiently, we introduce BRIDGE, a memory-efficient first-order bilevel method motivated by a loss-landscape analysis of the RL and retrieval objectives. Across seven open-domain QA benchmarks, BRIDGE achieves the highest average accuracy with both 3B and 7B backbones, improving the multi-hop average over the strongest baseline by 9.6 and 3.4 EM points, respectively. It also achieves the best averaged answer accuracy and reasoning quality across medical QA benchmarks.

Recent approaches to reinforcement learning (RL) post-training for large language models increasingly remove the critic to reduce training instability and memory overhead. Even where a critic is trained, it is discarded once training ends, although it has learned to predict outcomes. We revisit this trend and show that a pretrained critic's ability to predict future outcomes can make it a valuable asset for efficient long-horizon reasoning. First, we find that instability in critic-based RL for long chain-of-thought reasoning is largely an optimization artifact: keeping policy updates small and low in variance restores stable convergence. Second, a well-pretrained critic estimates the posterior probability of eventual success from later trajectory states and unfinished prefixes. Its predictions provide outcome-derived, dense, per-prefix learning signals that, during policy optimization, require neither completed rollouts, step-level annotations, nor external reward labels. Building on this insight, we introduce Reward-Free Policy Optimization (RFPO), which repurposes a single calibrated, frozen critic as a rollout-level reward, a value baseline for generalized advantage estimation, and a success forecaster for unfinished prefixes. We further show that binarizing the debiased score stops the policy from exploiting the critic's length bias. Binarized, RFPO matches supervised PPO without a single label in the training loop, while cutting compute and memory overhead. This makes RFPO well suited to long-horizon reasoning tasks, where outcomes arrive late and generation dominates cost: because rollouts can be rewarded before they finish, training no longer has to pay for waiting on every trajectory to complete. Our findings challenge the prevailing critic-free paradigm and establish critic-based, reward-free optimization as a scalable and computationally efficient path for LLM post-training.

Knowledge editing enables rapid updates of specific factual knowledge in large language models (LLMs) without full retraining. However, more realistic scenarios call for a lifelong framework that handles continual updates rather than one-off modifications. In such settings, existing editing methods often overfit to target prompts, significantly degrading both the generalization of the edited knowledge and the model's general capabilities. To address this issue, we propose GLIME (Generalizable Lifelong Model Editing), which combines knowledge editing with preference optimization over generation behavior. GLIME further incorporates replay-based editing and a gradient constraint to preserve previously edited knowledge. Experimental results show that GLIME significantly improves knowledge generalization in lifelong editing settings while maintaining both editing performance and general capabilities.

Safe reinforcement learning seeks policies that maximise task performance while satisfying safety constraints. In driving benchmarks, however, collision costs typically appear only at the time of collision, providing no advance warning of an approaching hazard. Frozen vision--language models can provide dense semantic feedback, yet it remains unclear whether their scores anticipate collisions and which component drives an observed safety improvement. Episodic cost can also favour policies that make little task progress. To address these gaps, we propose VLM-Safe-RL, a framework that integrates frozen CLIP signals into PPO-Lagrangian through reward shaping and an augmented multiplier update. On MetaDrive Hard, which combines the densest traffic with the largest map, the catastrophe rate falls from 31.6% to 19.4%. FormulaOne-L2 analysis finds no evidence that the CLIP signals anticipate collisions and shows that the VLM term has a negligible effect on the Lagrange multiplier. These findings show a conditional reduction in observed catastrophe rate without evidence of collision anticipation.

We identify a fundamental mismatch in empathetic reinforcement learning: support priorities evolve with the dialogue state, yet existing methods typically optimize predefined reward specifications that remain fixed across turns. To model these evolving support priorities, we organize empathetic support along cognitive, affective, and proactive empathy, and propose Context-Adaptive Rubric Evolution (CARE). At each turn, CARE generates a context-adaptive rubric by adjusting both the weights of these three empathy dimensions and their fine-grained evaluation criteria. The rubric generator is trained with turn-level rubric supervision and human preference data through supervised fine-tuning followed by preference-based reinforcement learning, and then serves as an adaptive reward interface for online empathetic RL. Integrated with both RLVER and MICA, CARE achieves state-of-the-art performance across SentientBench, EQBench3, and EMPA under three independent LLM judges. Notably, on EMPA, CARE improves EPM-Idx over the strongest baseline by at least 13 points under all three judges, including an increase from 28.11 to 83.54 under Gemini-2.5-Pro. Further analyses show that learned rubric priorities systematically vary across dialogue stages and user emotions, demonstrating that CARE adapts what is rewarded as support needs evolve.

Language model post-training is often bottlenecked by the need for human-collected preference data, which is expensive and difficult to scale. Reinforcement learning from AI feedback (RLAIF) style approaches that leverage pseudo labels offer an abundant alternative but introduce systematic biases that degrade downstream alignment. Recent general-purpose semi-supervised methods correct for teacher bias using a small set of human-labeled examples, but suffer from high variance especially when human annotations are scarce. To this end, we propose ABC-Align, leveraging abundant pseudo label signal to minimize variance and applying a lightweight, adaptive correction grounded in the human-labeled subset. The correction strength is tuned automatically during training using plug-in estimates of the relevant bias--variance quantities. On LLM alignment with RLHF, DPO, and GRPO where human feedback is scarce, we empirically demonstrate that ABC-Align achieves superior performance over prior semi-supervised baselines in a series of experiments on an increasing scale. Our code is available at https://github.com/SewoongLab/abc-align .

Scaling LLM-based optimization from textbook-scale instances to real-world, industrial tasks remains a critical open challenge. Existing approaches are predominantly evaluated on small, self-contained textual problems and often commit to a solver-integrated paradigm, limiting their ability to handle the scale and structural diversity of practical optimization workloads. In this work, we propose a practical framework for training open-source LLMs to tackle real-world, industrial-scale optimization. We first show empirically that solver-integrated reasoning, exact combinatorial algorithm, and heuristic search exhibit complementary strengths across different problem structures and scales. Motivated by this, we introduce Strategy-Diverse Reinforcement Learning (SDRL), which trains LLMs as adaptive optimization meta-solvers. SDRL leverages this complementarity through a correctness-gated hierarchical diversity reward that promotes robust exploration across varying strategies and within each strategy, effectively preventing premature strategy collapse. We further introduce a mixed-format training scheme that jointly supports both self-contained textual problems and file-grounded instances. Across comprehensive evaluations, our framework outperforms existing fine-tuned methods and frontier models including DeepSeek-V4-Pro and GPT-5.5, both on average across benchmarks and on industrial-scale optimization tasks.

Offline goal-conditioned reinforcement learning (GCRL) learns goal-directed policies from reward-free data, but in long-horizon tasks, goal-conditioned value functions often provide unstable guidance due to sparse rewards and discounting. Hierarchical methods partially mitigate this issue via subgoal decomposition; however, high-level decision-making still relies on noise-sensitive value estimates, leading to unstable behavior in complex environments. We address this limitation by proposing Diffusion Subgoal Planning (DSP), a diffusion-based framework for high-level subgoal generation. DSP casts high-level planning as guided generative inference over goal-conditioned subgoals and learns both conditional and unconditional flows, enabling classifier-free guidance to introduce a goal-directed bias at inference time. By removing explicit value-based guidance from high-level planning, DSP generates reachable and goal-directed subgoals through a generative model while retaining hierarchical execution. Experiments on offline GCRL benchmarks demonstrate that DSP outperforms prior methods on a range of navigation and manipulation tasks, with particularly strong performance in maze environments that require multi-step subgoal planning.

Reinforcement learning (RL) changes not only what language models say, but also how much they say, often increasing response length at the cost of token efficiency. Controlling this length growth is particularly challenging in open-ended RL because (i) response length is entangled with quality, (ii) open-ended tasks lack a natural success boundary for deciding when efficiency should be prioritized, and (iii) dense, graded rewards often yield small within-group quality margins, making quality-induced advantages especially sensitive to reward-level length shaping, which can perturb their magnitudes and even reverse their signs. We therefore adopt an asymmetric principle: quality should determine the direction of reinforcement, while length should only shape its magnitude. We instantiate this principle with Quality-Gated Length Advantage Shaping (QGLAS), which first computes advantages from quality rewards alone, then adds bounded bonuses only to shorter positive-advantage responses, leaving all other advantages unchanged. The bonus strength is further adapted to within-group quality separation, allowing conciseness to matter more when quality-favored responses are similar and less when their quality differences are clear. Across different model families, open-ended benchmarks, and reward sources, QGLAS consistently achieves a stronger quality--length trade-off than representative baselines. At approximately 30% compression, QGLAS retains 98.4--102.0% of the macro-average quality gains achieved by quality-only RL over the base model, compared with 68.3--75.5% for these baselines at comparable compression.

Reinforcement learning has become an effective approach to training language model agents, but sparse and delayed outcome rewards provide limited guidance for credit assignment across long interaction sequences. Recent work on on-policy self-distillation (OPSD) offers complementary supervision by evaluating a policy's sampled responses under privileged training-time context. However, our diagnostics show that positive average agreement between outcome and hindsight feedback coexists with substantial local disagreement, raising the question of how to allocate influence between them at each decision. We introduce UniOPSD (Unified On-Policy Self-Distillation), which unifies these feedback sources through adaptive local credit arbitration. UniOPSD constructs comparable credit estimates from environmental returns and successful-peer hindsight at shared interaction anchors. Historical agreement determines the global mixing level, while current signal availability and relative precision adjust each source's influence at individual decisions. The episode-level outcome contribution is retained, and bounded token modulation refines the fused step credit for policy optimization. With Qwen2.5-3B-Instruct and Qwen2.5-7B-Instruct, UniOPSD achieves ALFWorld success rates of $82.8%$ and $83.6%$, WebShop success rates of $75.0%$ and $82.0%$, and Search-QA aggregate accuracies of $45.3%$ and $49.8%$, respectively. On 3B WebShop, UniOPSD improves over SDAR by $7.0$ percentage points. Our code is available at https://github.com/Zenghuang-Fu/Uniopsd

Language models are increasingly used to sample from a specified distribution, for instance, to simulate survey respondents or generate synthetic data. Instruction-tuned models can state such a distribution correctly and still fail to sample from it. Prompting and changes to decoding reduce this mismatch only partly, which motivates training with policy optimization. Group relative policy optimization (GRPO) is a natural fit for this problem because it already samples a group of rollouts per prompt, and the group's empirical distribution can be compared with the target. However, scoring the group as a whole gives every rollout the same reward. Group-relative centering then sets all advantages to zero, and the model receives no learning signal. To give each rollout its own signal, we introduce the witness advantage, a per-rollout advantage derived from maximum mean discrepancy (MMD). It trains a model to match a target distribution over a finite set of outcomes. The MMD between the model's distribution and the target has a witness function that measures how over- or under-produced each outcome is. Each rollout's advantage estimates the negative witness at its outcome, so a rollout is rewarded for an outcome the group under-produces and penalized for one it over-produces. The witness advantage is computed in closed form from the group's outcome counts, and we use it as the reward in GRPO. On unseen target distributions, training with the witness advantage substantially reduces the total variation distance to the target while largely preserving the model's general capabilities.

Terminal utility evaluates a complete agentic workflow, but learning requires credit for the decisions within it. We introduce Attentive Search over Counterfactual Trees (ASCT), a framework that turns training-time multi-step search into local action credit. At actor-visited states, an auxiliary tree evaluates alternative legal actions from the same recoverable prefix. Its action-value table is centered by the frozen actor's probabilities and supplies credit for PPO on actor-sampled trajectories. This protocol connects counterfactual evaluation to policy learning while deploying the actor alone. Uniform, UCT, and cost-aware AgentUCT instantiate the framework. On HotpotQA agentic retrieval-augmented generation, all three improve mean held-out utility over trajectory-return PPO and workflow-adapted VinePPO. Across three seeds, ASCT-AgentUCT reaches 0.6187 utility versus 0.5939 for VinePPO, with gains in answer F1 and execution cost, and uses 50.3% fewer recorded auxiliary Qwen tokens. Transfer and component-description studies examine the learned policies beyond the training setting.

Reinforcement learning from verifiable rewards (RLVR) frequently reuses rollouts across multiple policy updates, increasing the mismatch between the current policy and the data-generating policy. We identify a sign-dependent gradient starvation problem in clipped policy optimization: clipping suppresses under-generated positive responses at the low-importance-weight tail while permitting severely over-generated negative responses to dominate the high-weight tail. To address this, we propose ReSPO (Reshaped Sequence Policy Optimization), which replaces clipping with a smooth, two-branch sequence-level kernel derived from an $α$-divergence variational objective and an exponential variance-control tilt. The positive branch preserves a nonzero gradient weight for under-generated positive responses, while the negative branch suppresses heavily over-generated negative responses. We demonstrate that ReSPO effectively learns from long positive reasoning trajectories during early training, even when accumulated policy drift relegates them to the low-importance-weight tail. On dense and MoE Qwen3 models, ReSPO accelerates early optimization, improves final training scores, and achieves higher held-out benchmark performance under a rollout reuse, validating our approach on importance-weight tail control in off-policy learning.

Rare-disease diagnosis is a long-tail reasoning problem: phenotypes are incomplete, individual disorders are sparsely documented, and relevant evidence is distributed across ontologies, gene annotations, and biomedical text. Language models consequently favor common conditions, miss rare candidates, or produce plausible but invalid names. We introduce RareDx, which couples controlled evidence use with knowledge-graph-grounded policy optimization. RareDx-Harness normalizes heterogeneous records into one ranked-diagnosis task and compares direct inference, static retrieval, adaptive tools, and structured phenotype-gene-disease reasoning over a shared knowledge layer. The training pipeline combines Top-10 post-training with RareDx-KGPO, our knowledge-graph-grounded policy optimization method. Its reward projects predictions into a canonical disease graph and integrates curated graded relevance, ontology proximity, biomedical similarity, and phenotype consistency. Vocabulary and output-budget constraints prevent dense partial credit from rewarding fabricated or overlong differentials. Across eight benchmarks, the complete RareDx system centered on Qwen3.5-9B reaches 38.34 macro Hit@10, 1.60 points above GPT-5.5 under the archived protocol; a disjoint validation-selection audit retains a 6.80-point routing gain over Direct on held-out cases. The 27B system reaches 23.53/36.56/40.76 at Hit@1/5/10. Controlled ablations show that retrieval is not uniformly helpful and that controlled routing is central to the gain. These results indicate that structured medical knowledge can turn a compact model into a competitive diagnostic ranker across heterogeneous long-tail settings in clinical practice.

In reinforcement learning with verifiable rewards (RLVR), imperfect verifiers can reward incorrect responses, creating opportunities for reward hacking. Using gradient flow with a fixed verifier, we characterize the conditions under which reward rises while correctness falls. We then show that the observations available during RLVR are, in general, insufficient to detect or identify accepted errors, or to guarantee their reduction without sacrificing correct responses. To address this limit, we construct a correction using additional feedback about correctness from audits. This correction achieves selective control: at the current policy, it lowers the probability of accepted errors and raises that of correct responses, provided it outweighs the pressure toward errors from verifier reward. Experiments with log linear and neural contextual bandits and with a language model support the analysis and show that selective control under partial auditing reduces accepted errors while increasing correctness.

People learn not only by repeating successful actions, but also by recounting and explaining their experiences, revising their understanding to guide future behavior. Can a language-model agent improve its future actions by training only on explanations of its own experience? We investigate this question by studying Retrospection-Only Fine-Tuning (ROFT), a minimal online procedure designed to isolate the effect of explanation-only training on subsequent behavior. The agent attempts a task, observes available feedback, generates a retrospective explanation, and is fine-tuned with a next-token prediction loss on the explanation tokens alone. The procedure uses neither an external teacher nor a reward-based policy update. In software-engineering experiments with Qwen3.5-4B, ROFT is trained on problems with mixed successful and unsuccessful base-model attempts. On held-out SWE-bench Verified and Pro, it reaches 49.2% and 26.8% solve rates after 20 updates without using a verifier, compared with GRPO's 48.0% and 25.3% after 40 updates in the evaluated runs, and makes faster early progress in training time and sampled attempts. It also learns to solve individual tasks on which all 64 sampled base-model attempts failed, showing that learning can begin without any initially successful trajectories. Behavioral analyses find that ROFT indirectly assigns credit to actions, encouraging good actions and discouraging incorrect ones. Moreover, prompting retrospections to emphasize more direct solutions yields shorter subsequent attempts even without an explicit length penalty. Together, these findings show that learning to explain can also improve learning to do, establishing self-generated retrospections as useful training targets and motivating further study of explanation-to-action transfer.

Recent work in visual question answering has shown that vision-language models can exhibit strong reasoning capabilities by translating visual inputs into textual representations. The effectiveness of this translation depends on how well visual details are retained; models need to surface and align both explicit and implicit knowledge sufficient to support reasoning, without introducing spurious assumptions. Existing methods that leverage detailed image captions introduce visual details unrelated to the reasoning task, inflating input token counts and increasing computational cost. To address these challenges, we propose VisKG, a reinforcement learning (RL) framework in which models learn to translate visual content into question-specific knowledge graph (KG) representations. This process filters out perceptual noise while preserving the entity-relation structure needed for chain-of-thought reasoning, following the principle of minimum sufficient information. To ensure stable RL post-training, VisKG adopts Group reward-Decoupled Normalization Policy Optimization (GDPO). In addition, we strengthen the supervision stage with negative rationale samples, exposing the model to incorrect reasoning paths before RL post-training. Experimental results across science, mathematics, and general visual understanding benchmarks show that VisKG achieves performance comparable to or better than baselines, while requiring fewer tokens than caption-based representations. Moreover, training VisKG with GDPO improves accuracy by 2% over its GRPO-trained counterpart on average. These results suggest that KG representations are a promising approach for supporting multi-step reasoning and open up future work on adaptively selecting the most suitable representation for a given task.