Generative RL 进展

Asynchronous reinforcement learning (RL) improves the efficiency of large language model post-training but introduces stale rollouts generated by earlier policies. Theoretical understanding of how this staleness affects convergence and how to mitigate its impact remains limited. We derive a convergence bound for GRPO-style algorithms that explicitly characterizes the tradeoff between the gradient estimator's second moment and bias. For trajectory-level importance-weighted estimators, our analysis shows that once the second moment is uniformly controlled, delay enters the bound through the bias introduced by clipping or rescaling. Guided by this insight, we propose a novel group mass capping GRPO (GMC-GRPO) method, which minimizes a ratio-based bias bound within a class of weighted estimators sharing a common second-moment guarantee. We establish convergence guarantees for asynchronous GMC-GRPO and show that, compared with TIC-GRPO, it improves the threshold dependence of the fourth-order delay term from $O(ε^{-4})$ to $O(ε^{-2})$ as $ε\to0$, where $1+ε$ is the ratio threshold. Under local policy overlap, the delay-dependent term decreases as $G^{-2/5}$ after tuning the step size, where $G$ is the group size. For fixed behavior and current policies, the bias introduced by group rescaling also vanishes as $G\to\infty$, whereas the bias from trajectory-wise clipping can persist. Experiments across Qwen3 models and reasoning benchmarks demonstrate improved robustness to stale rollouts, with GMC-GRPO achieving the best performance among stable baselines under large rollout delays.

Outcome-based reinforcement learning can train language models to forecast real-world events, but prior forecasting work either freezes research context before training or deploys agentic research only at test time, so the skill of gathering evidence is never shaped by the reward. We introduce an agentic forecasting environment, dataset, and harness built from 2,100+ resolved Polymarket questions; the agent acquires its own context at rollout time (web search, page reading, and financial time series, all restricted by layered leak filtering to information published before each question's cutoff), and we train Qwen3.5-35B-A3B (3B active parameters) on it with single-epoch GRPO under a Brier-score reward. Training changes how the agent interacts with information: calibration improves 30-40%, and search attempts fall from 3.8 to 2.25 per rollout as evidence discipline is learned. Evaluated in an identical harness against four frontier models, the trained policy also finishes ahead of every frontier model tested at evidence-based forecasting, including Claude Opus 4.5 (soft-Brier 0.254 vs. 0.256, n=265), at about 5% of the inference cost, and its margin is widest on the hardest questions, the ones the crowd itself had not decided. We release the environment, dataset, and per-rollout records as a reusable harness for temporal forecasting agents.

Language model (LM) alignment broadly aims to perturb a given LM $Q$ into an aligned LM $q$ such that i) the outputs produced by $q$ and $Q$ are 'close' in probability, ii) $q$ has a higher expected reward than $Q$. Two common techniques for LM alignment are: KL-constrained RL, which requires knowledge of the LM distribution and is computationally expensive, and the best-of-$n$ algorithm, which requires only sampling from the LM. The work of Yang et al. established asymptotic closeness between the distributions produced by the two alignment methods for an $m$--length i.i.d. token sequence output by the LM, in the limit as $m$ increases to infinity. However, the i.i.d. assumption is not representative of practical LMs, whose output sequences often have memory. In this paper, we extend the asymptotic closeness result to the case when the $m$--length token sequence outputted by the LM is Markovian. Further, for finite-length output sequences — particularly, when $m=1$ — we provide a complete characterization of LM distributions and reward functions for which the KL-divergence between the distributions produced by the two alignment methods is zero — a question first posed in Yang et al.

The rapid growth of video-based social media has increased users' exposure to harmful content, creating a need for reliable automated video safety detection. Although recent Vision-Language Models (VLMs) show strong video understanding capabilities, existing harmful video detection systems face two key limitations: they typically reduce safety detection to binary classification, overlooking the inherently multi-label nature of unsafe videos, and they rely on static training objectives that do not support controllable precision-recall trade-offs, though the desired operating point may vary across moderation pipelines and unsafe categories. To address these gaps, we propose Adaptive Tversky Policy Optimization (ATPO), a reinforcement learning framework for Multi-label Video Safety Detection (Multi-VSD). ATPO introduces the Adaptive Tversky Reward (ATR), which dynamically adjusts false-positive and false-negative penalties during training to enable controllable precision-recall trade-offs. Experiments on SafeWatch-Bench and XD-Violence show that ATPO substantially improves multi-label performance, increasing the Jaccard Index from 40.66 to 75.44 on SafeWatch-Bench-Real. Moreover, ATR enables reliable steering of the precision-recall operating point, supporting deployment scenarios with heterogeneous policy requirements. Code and checkpoints are provided at https://bruceyg.github.io/ATPO-project-page/ .

Reinforcement learning with verifiable rewards (RLVR) has substantially improved the mathematical reasoning capabilities of large language models. Recent work introduces search into RLVR rollouts to increase trajectory diversity, but diversity alone does not ensure that the search-induced rollout policy improves upon the current policy. To address this gap, we propose APIVIS, a training-time framework that adapts finite-budget Gumbel search to chunk-level mathematical reasoning. APIVIS combines direct and searched responses within each rollout group, allowing improvements found by search to produce informative relative rewards. It further applies selective supervision to search-improved tokens, preserving a learning signal when uniform group rewards render GRPO ineffective. We show that exact value-guided selection improves the expected verifier reward at each searched state and that this guarantee extends to the complete rollout policy, with a corresponding approximate guarantee under bounded value-estimation error. Experiments on widely recognized mathematical reasoning benchmarks and different model scales demonstrate substantial improvements over competitive search-based methods, validating the effectiveness of APIVIS.

Time series captioning is a fundamental step in time series understanding and can also serve as the bridge between signal and natural language. Supervised fine-tuning (SFT) relies on a larger model's captions and cannot exceed their quality. Reinforcement learning (RL) can, but its rewards were designed for other modalities and other tasks, and they transfer poorly to open-ended generation in the time series domain. We address this by proposing LineupRL, a reinforcement learning with verifiable rewards (RLVR) pipeline whose reward is caption-to-series identification. The reward model is a frozen large language model (LLM) verifier that reads the generated caption and the candidate time series as raw values, never the chart, and must pick the described time series from multiple distractors. Matching is a far lighter demand on the verifier than writing questions or judging a caption, so an off-the-shelf LLM can supply the reward. Across two captioning benchmarks, and on forecasting and reconstruction where the predictor sees only the caption, LineupRL outperforms SFT and RL baselines on every metric. The 3B vision language model (VLM) trained by LineupRL also outperforms, at 1/24 of the parameters, the 72B VLM whose captions the SFT baseline is distilled from. Our case study shows that LineupRL resists reward hacking, and that the captioner it trains both traces the trend and names the values at key points.

Flow matching is central to 3D generation, yet in practice its reinforcement learning (RL) methods are largely adapted from 2D visual generation. Representative DPO-, GRPO-, and NFT-style objectives, when applied to negative trajectories, mainly steer predicted velocities away from the corresponding directions without explicitly specifying a target velocity field toward preferred samples. In 3D generation, constrained by pretrained model capabilities, rollout diversity, and reward-distribution complexity, directly applying these RL methods yields limited gains in geometric quality. We introduce a forward-process RL method Dynamic Homing Optimization (DHO), which reformulates negative-trajectory optimization as positive-sample attraction-guided dynamic homing. Specifically, Minimum-Cost Attractive Matching (MAM) assigns each negative sample a distinct positive target, and Time-Aware Dynamic Correction (TDC) then redirects its trajectory toward the target using a remaining-time-aware corrective velocity. Building on asynchronous online DHO, we develop Flow3D-Pro, an image-to-3D geometry generation framework. Experiments show that DHO outperforms representative DPO-, GRPO-, and NFT-style objectives in 3D generation, while Flow3D-Pro produces higher-quality 3D geometry than existing mesh generation methods.

As the use of large language models (LLMs) expands, post-training has become increasingly important for adapting them to downstream tasks. However, obtaining reliable supervision remains costly, especially in domains without reference answers or executable verifiers. LLM-as-a-Judge provides scalable pseudo-rewards for unlabeled responses, but a single point score does not explicitly represent reward uncertainty. This motivates representing pseudo-rewards as conformally calibrated reward ranges. We propose Range-GRPO, a semi-supervised post-training framework that combines limited labeled data with unlabeled prompts. In Group Relative Policy Optimization (GRPO), learning signals depend on relative reward comparisons within each rollout group. The proposed objective compares reward ranges pairwise rather than reducing them to point rewards, allowing interval uncertainty to affect both the magnitude and direction of these signals. Our theoretical analysis characterizes this distinction and shows that the proposed objective recovers the Dr$.$GRPO advantage when all reward ranges collapse to points. Empirically, Range-GRPO achieves the highest in-distribution and out-of-distribution average performance among the evaluated semi-supervised methods while requiring fewer training resources.

Reinforcement learning based post training of diffusion models, such as Denoising Diffusion Policy Optimization (DDPO), optimizes a reverse diffusion process under a reward function. However, current approaches to reward optimizations do so at the cost of diversity and quality. In this paper, we provide better tradeoffs through careful theoretical considerations and method design. We analyze the theoretical framework and mathematically demonstrate that only-latter timestep updates of diffusion model may be harmful for diversity contrary to the conclusions presented in a previous work. Additionally, we propose an incremental Feynman-Kac training based on strong theoretical foundations in order to achieve the best-yet alignment-diversity tradeoffs. We perform extensive experiments and compare our method against related diffusion policy optimization approaches in three different tasks and also provide strong ablations for each component, thus validating strong performance gains in both alignment and diversity.

Training a single LLM agent jointly across diverse interactive environments has attracted increasing attention as a route to generalist agents. Existing curriculum and data-selection strategies often allocate training at the environment level or prioritize local reward-based signals, without explicitly considering relationships between current rollouts across environments for prompt-group selection. Meanwhile, as environments are learned at different rates, all-failure and all-success rollout groups can coexist within a batch, leaving those data without group-relative reward signals. Both challenges highlight limitations of relying solely on scalar rewards in multi-environment RL: they provide limited information about cross-environment relationships and no within-group reward contrast when rewards are identical. This motivates richer textual feedback, such as rubrics describing rollout behaviours, to guide learning. Beyond rubrics' usage as reward, we repurpose rubrics to guide both online data selection and policy supervision. An LLM judge tags each rollout using a predefined rubric vocabulary shared across environments. The resulting profiles guide the selection of data that aligns with the overall behavioural composition of the mixed-environment batch while limiting overlap with already-selected data. Available positive rubrics (describing desired behaviours) provide privileged context for an on-policy self-distillation teacher, supplying additional token-level supervision, while negative rubrics (describing undesired behaviours) guide subsequent rollout generation away from recurring failure modes. Together, these components form RISED. Across model backbones, RISED achieves the highest mean pass rate across environments and ranks first or second in every individual environment. Rubric-based analysis of RISED can further characterize the behavioural changes accompanying these gains.

Reinforcement learning (RL) for video generation usually assigns one scalar reward to an entire sampled video. Yet a video is not uniformly flawed: some visual tokens may already satisfy the prompt, whereas others require correction. A scalar reward cannot localize errors, causing optimization to perturb satisfactory tokens while under-targeting the tokens that actually need to change. We introduce Token-Level Video Reinforcement Learning, TVRL, a framework that derives token-level credit from the reward being optimized. Our key insight is that the answer likelihood of a frozen vision-language model provides both signals: its outputs contribute to the video-level reward, while magnitudes of its video-input gradients reveal which generated video tokens most affect that score. We instantiate TVRL in Group Relative Policy Optimization by averaging prompt-derived question rewards into one group-relative advantage and using detached, question-conditioned token-credit maps to reweight dense denoising-transition log-probabilities inside the clipped policy ratio. On VBench-2.0, TVRL achieves an Overall score of 57.69, outperforming the base model by 3.60 points. TVRL also improves matched GRPO baselines across three SDE samplers (SAGE, Flow, and Dance) by 2.68--3.15 points and across four reward models (VideoAlign, VideoScore2, UnifiedReward2, and Qwen3.5-9B) by 1.33--3.15 points.

Video generation models have achieved remarkable visual fidelity and have strong potential to become general-purpose world simulators. Despite this progress, they still fail to generate videos which adhere to laws of physics. The problem becomes even more apparent in realistic settings where multiple physical principles must work together within the same video; for example, "a balloon floating upward while steam rises from a pot" requires buoyancy and fluid dynamics to unfold coherently and simultaneously. Yet existing methods largely ignore multi-principle interactions, focusing on a single principle per video. We propose HiPhy (Hierarchical Physical Alignment), a reinforcement learning framework that grounds video generation in physical laws through a dual-level objective: locally enforcing the temporal dynamics of individual physical principles, and globally ensuring the physical and semantic coherence of the entire scene. To support multi-principle generation, we construct a 50K-prompt dataset and introduce a prompt benchmark MultiPhyBench, spanning a diverse range of co-occurring physical events. Our experiments show that HiPhy significantly outperforms prior methods and baselines, improving physical commonsense and semantic alignment significantly across various benchmarks, with the largest gains on scenes involving multiple concurrent physical principles where competing methods degrade most sharply.

We present OmniSeek, an agentic framework that transforms an Omni Large Language Model (Omni-LLM) into an active, multi-turn reasoning agent with native tool use. Rather than passively processing an entire audio-visual sequence in a single forward pass, OmniSeek makes evidence acquisition part of the reasoning process: it dynamically decides whether to look or listen, and over which temporal window, to retrieve sparse but critical evidence across different modalities within long contexts. Through an iterative multi-turn protocol, the retrieved raw audio or visual segments are appended back into the context to support subsequent reasoning. To cold-start this capability, we build a data engine that synthesizes OmniTraj-170K, a corpus of multi-hop Chain-of-Thought trajectories with interleaved audio and visual evidence. We first supervise the model on these trajectories to instill multi-turn tool-use behavior, and then further optimize the policy via a two-stage reinforcement learning with verifiable rewards. Moreover, we introduce an Audio-Visual Necessity objective that explicitly rewards successful trajectories whose reasoning depends on both modalities, discouraging single-modality shortcuts. Extensive experiments across a wide range of benchmarks demonstrate that OmniSeek learns adaptive cross-modal evidence seeking and consistently improves audio-visual reasoning performance.

Most preference optimization methods, such as Direct Preference Optimization (DPO), apply preference supervision at the response level, although autoregressive language models are optimized token by token. As a result, all tokens in a rejected response contribute to the negative training signal, including tokens that may encode behavior that is useful for the preferred response. We introduce GAW-PO, a gradient-aligned token reweighting method for DPO that estimates, for each rejected token, whether penalizing it would interfere with the preferred update directions. Tokens whose gradients are strongly aligned with the preferred behavior receive a weaker negative contribution, while conflicting tokens retain a stronger penalty. Our method achieves the highest average performance among the evaluated preference-optimization methods, improving by 0.97 points over standard DPO and 0.65 points over the strongest competing baseline across 11 benchmarks spanning mathematics, reasoning, coding, and question answering. We further show that gradient-aligned weighting is substantially more robust to aggressive preference optimization: as the DPO regularization parameter $β$ decreases, standard DPO degrades sharply, whereas GAW-PO continues to improve. These results suggest that accounting for the interaction between rejected-token updates and preferred behavior provides an effective form of token-level credit assignment for preference optimization.

Collaboration between a small language model (SLM) and a large language model (LLM) offers an opportunity to combine the efficiency of smaller models with the strong reasoning capabilities of larger ones. Existing approaches primarily frame such collaboration as a computation allocation problem, determining which model should handle each portion of the reasoning process. In black-box API-based settings, however, this paradigm can be inefficient due to coarse-grained delegation or repeated transmission of context across model switches. In this work, we instead formulate SLM-LLM collaboration as an information acquisition problem, under an API budget constraint. The SLM remains the primary reasoner and selectively queries a black-box LLM advisor only when needed, issuing targeted queries rather than delegating the reasoning process itself. To realize this strategy, we develop a three-stage RLVR framework that learns whether to call the advisor, how to formulate useful queries, and how to integrate the collaboration into the reasoning process by jointly refining advisor invocation and information use. Across mathematical reasoning and coding tasks, our approach improves the performance--cost tradeoff over existing collaboration baselines and, in some settings, matches or exceeds oracle problem-level routing. Finally, we show that our strategy can transfer to other advisor model families, without further training.

Reinforcement learning from verifiable rewards (RLVR) has produced large reasoning gains in language models, and verifiable video benchmarks make it applicable to causal-temporal video question answering. We study what RLVR teaches video-language models about time. We fine-tune four open models (Qwen3-VL-8B/4B, Qwen2.5-VL-7B, Gemma-3-12B) with group relative policy optimization under three data recipes: verified (synthetic CLEVRER questions with exact answer and event-order rewards), unverified (self-supervised pretext tasks over 43,751 real web videos), and a 1:1 mixture, plus a verified+real arm that adds 4,000 verifiable questions on real video. Each cell is evaluated in-domain and on out-of-domain real video (a NExT-QA temporal stress set and an MVBench subset), with frames in order, shuffled, and absent. (1) Verified training yields large in-domain gains that shrink as base competence grows (+14 to +19 points on weaker models; +6 on the strongest). (2) Much of the gain is non-visual: accuracy with no frames rises nearly as much as with frames. (3) Verified-only training can severely degrade out-of-domain accuracy with no sign during training: Qwen3-VL-8B loses 26.7 and 25.2 points on the two real-video sets, while the mixture never significantly degrades a model trained on it. Adding real verified questions removes that loss (-2.3 points, within noise of base) and keeps a +9.3 in-domain gain, so the cause is narrow synthetic-only data, not verification. (4) No recipe induces temporal-order grounding: across 41 evaluations the ordered-versus-shuffled gap is indistinguishable from zero in 39 and marginal in two, despite an event-order reward. Verifiable rewards improve benchmark accuracy without temporal understanding. Report no-frame controls, and mix in real video to guard against out-of-domain degradation.

Large language models deployed as personalized assistants must reason over long, evolving interaction histories. However, in long-term dialogue reasoning, relevant evidence is scattered across sessions, preferences may be revised over time, and standard long-context training fails to address these challenges under data scarcity and prohibitive computational costs. We propose StateTree, a data-driven RL method that constructs a challenging auxiliary task from scarce dialogues with verifiable ground truth. StateTree augments multi-session dialogues with a tree-structured path-tracing task: key-value records are embedded across sessions to form a binary tree. Solving the task requires the model to traverse from root to leaf by retrieving records across sessions and comparing timestamps to resolve branches, then recover the hidden target question among distractor leaves. We apply curriculum RL training progressively increasing tree depth and introduce a compositional variant whose edges carry step-level reasoning fragments, training the model to compose partial cues into coherent queries. Trained on 10K-token contexts, StateTree generalizes to 128K tokens without full-length RL costs and exhibits capabilities including cross-session retrieval, temporal reasoning, knowledge update, and compositional multi-hop reasoning. StateTree outperforms both SFT and RL-based baselines while preserving short-context general reasoning. StateTree-7B achieves gains up to +23.60% on LongMemEval (128k), and StateTree-14B reaches 59.00% accuracy on LongMemEval, surpassing QwenLong-L1-32B (45.20%).

Rubric-based reinforcement learning (Rubric-RL) trains language models where no verifier exists. A judge checks each criterion of a rubric, and the verdicts are aggregated into a reward, most often by a weighted sum. We show that this additive aggregation is the weak point. Under a sum, criteria compensate for one another: a policy that misses the one decision that matters can buy the points back with advice nobody asked for. On clinical consultation, such a policy scores higher and answers worse. Rubric coverage rises while appropriateness on held-out physician criteria falls below the untrained model. The medical criteria are not to blame. Grouped so that they must hold together, the same criteria, unchanged to the word, recover a third of the loss; shorter answers recover almost none. We therefore propose Protocol-level Rubrics (ProRubric), which keeps what the criteria ask for and changes how they are aggregated. It groups a checklist into a few protocol-level dimensions. A dimension counts only when all of its criteria hold and its failure clause does not fire. The grouping is done once, offline, and leaves the optimizer unchanged. ProRubric raises appropriateness by 10.8 points without losing coverage and has the best seven-benchmark average at both scales. Reward validity is set not only by what a rubric verifies, but by how it aggregates. Code is available at https://github.com/Estrellajer/ProRubric

Decision models such as Jev answer questions with probabilities, which are only useful if they are calibrated. Open-source reproductions rely on supervised fine-tuning plus temperature scaling, while reinforcement learning from verifiable rewards (RLVR) makes reasoning models overconfident. We present a working implementation of reinforcement learning for calibrated decisions (RLCD) for reasoning models: the model samples a rationale, and we score the answer distribution it commits to afterwards with a strictly proper scoring rule. A variance identity shows that scoring the mixture of several samples rewards disagreeing rationales, and that RLVR is exactly this mixture objective without its diversity term. Optimized naively, the per-rationale objective either switches reasoning off or is drowned out by policy-gradient noise, which leads to a two-stage recipe: calibrate, then reinforce. With Qwen3-1.7B on two reasoning tasks (3 seeds, paired tests), RLCD matches or beats SFT, RFT/STaR and GRPO (each temperature-scaled) in accuracy and beats all of them in selective prediction; on GSM8K answer verification a single query decides \gvTwoCovFive% of the items at $\le$5% error, versus \gvGrpoCovFive% for GRPO. When uncertainty comes from annotator disagreement, RLCD provably cannot beat cross-entropy. Code and results: https://github.com/ZimmyGao/openjev-rlcd.

Parameter-efficient reinforcement learning aims to improve reasoning with a compact trainable interface to a pretrained model. We introduce the Thalamic Router (T-Router), which concentrates adaptation on the reuse of completed computations. A compressed, addressable bank preserves block changes; a depth-recurrent controller conditions their selection and relative-scale writeback. This coupling gives thalamic context-dependent routing a concrete computational form: learn which earlier contributions a receiving layer uses, and with what influence. Correctness rewards train the interface while preserving backbone parameters and layer order. On an 8.95B-parameter backbone, T-Router allocates 41.73M parameters (0.466% of the backbone) and achieves 83.64 +/- 1.16 MathAvg after GSM8K RL, compared with 73.79 +/- 1.83 for full-parameter GRPO across three evaluation rounds. At a comparable parameter budget and with matched retries, it exceeds LoRA's 77.28 +/- 1.95 MathAvg, improving all three task families and raising mean AIME accuracy from 48.33 to 60.56. Capacity-controlled comparisons favor addressable block changes and recurrent context; separate search training extends the interface to tool-mediated reasoning. These results establish controlled computation reuse as an effective route to parameter-efficient reasoning reinforcement learning.

Value-model-free RLVR methods such as GRPO assign uniform advantages to all tokens in a rollout, ignoring that tokens contribute unequally. Recent methods use token entropy as an importance proxy but compute it globally across the batch, conflating importance with prompt difficulty and positional trends. We argue that importance should instead be measured relative to the local context of each token. We introduce proximal entropy, a local measure of token importance relative to neighboring tokens, and prove it is invariant to both confounders. Proximal Entropy Policy Optimization (PEPO) uses it to weight per-token advantages and outperforms GRPO and entropy-based baselines on mathematical reasoning across Qwen3-1.7B, Qwen3-4B, and Llama-3.2-3B-Instruct. We also show the formulation generalizes to other algorithms where substituting proximal entropy into existing methods improves, and applying it to single-stream RL succeeds where global entropy fails.

During reinforcement learning with verifiable rewards (RLVR), large language models (LLMs) can exploit loopholes in their environments to obtain high rewards without improving the intended capabilities, i.e., reward hacking. Despite its risks to training efficiency and safety, monitoring and mitigating reward hacking during training remain challenging, which is limited by a lack of testbeds that reproduce hacking and reliably identify it. We introduce CATCH, a controllable testbed for studying reward hacking in coding RL. CATCH deliberately exposes environmental loopholes and provides execution-based gold labels by comparing success under a vulnerable evaluator with task correctness under an independent audit. It also can control the model's initial hacking tendency through supervised fine-tuning data mixtures and the difficulty of earning rewards through reward designing, enabling systematic comparisons of hacking dynamics and interventions. Experiments show that CATCH can produce diverse RL training trajectories with clear reward hacking, and analyses demonstrate that both initial models and reward difficulties shape the emergence of reward hacking. We further evaluate the effectiveness of different reward hacking detection and mitigation methods. A key finding is that a chain-of-thought monitor initially suppresses hacking, but this protection erodes as the policy model learn to mislead the monitor with code comments. This highlights the need to evaluate hacking mitigations throughout training with CATCH. The source code and resources are publicly released at https://github.com/THUAIS-Lab/CATCH.

Grammar-constrained generation guarantees syntactic validity, but can substantially degrade semantic quality when the model's preferred outputs are poorly aligned with the imposed grammar. This trade-off is particularly severe when the prompt is underspecified or the model has limited instruction-following ability. Beam search can partially mitigate these failures by exploring multiple valid sequences, but its computational cost grows with beam width, while sequence-level probability is only an imperfect proxy for semantic quality. We introduce GrammarRL, a label-free reinforcement learning method that adapts language models to grammar constraints without requiring annotated data. GrammarRL optimizes the model using two complementary self-supervised rewards derived from its own likelihoods: a direct reward, measuring how likely the constrained output is given the input, and a reverse reward, measuring how well the input can be reconstructed from the generated output. We optimize these rewards with a Reinforce Leave-One-Out (RLOO) objective over groups of grammar-constrained rollouts, augmented with the top-1 beam-search hypothesis and regularized towards a frozen base model. We evaluate GrammarRL on sign language gloss translation, hierarchical text classification, and named entity recognition using Llama models ranging from 1B to 8B parameters. GrammarRL consistently outperforms constrained greedy decoding, with an average improvement of 9.8 points and gains of up to 22.8 BLEU. It matches or outperforms beam search on two of the three tasks while preserving greedy-decoding inference cost. Ablations further show that the two rewards are complementary: either reward alone can underperform the untrained baseline, whereas their combination consistently improves upon it.

Reinforcement learning with verifiable rewards (RLVR) has become an important approach for improving reasoning during post-training. Recent work suggests that some difficult prompts remain resistant to learning even when they occasionally produce correct solutions. We revisit this unlearnability phenomenon and find that the affected prompts do improve, at roughly one third of the learnable rate, while the difficulty-defined set used to study them is much less reproducible than expected. These difficulty labels are estimated from a limited number of sampled responses. Combining them across seeds can further change which prompts are selected instead of simply reducing measurement noise. We develop a sampling-based framework for quantifying this instability and determining how much evaluation is required for difficulty assignments to reproduce reliably. We also revisit the gradient-similarity evidence proposed to explain unlearnability and show that part of the observed separation arises because difficult prompts provide fewer correct rollouts from which their gradients can be estimated. Matching this sample count weakens the gradient difference but does not remove it. Overall, the slow-learning phenomenon survives our reanalysis, while both the prompts used to define it and the evidence used to explain it require more careful measurement.

We present PrefPI (Preference-Guided Policy Iteration), an iterative framework for steering pretrained generative robot policies using only relative preferences over self-generated trajectories. Unlike prior preference-learning methods that primarily sharpen modes already represented by the policy, we study steering beyond the initial effective support, where desired behaviors are rarely or never observed under the initial policy. Our key idea is to formulate preference learning as preference-conditioned generative modeling: preferred trajectories define a conditional distribution, whose density ratio with the broader behavior prior provides an implicit preference signal amplified by classifier-free guidance (CFG). Repeating this preference-conditioned modeling and guidance step yields a form of preference-guided policy iteration, turning incremental improvements toward previously inaccessible behaviors. Across diffusion policies and the PI0.5 flow- matching VLA in simulation and the real world, PrefPI produces substantial behavioral shifts with limited feedback. In particular, PrefPI increases object transport height from 10.7 cm to 19.8 cm on real hardware with only 150 preference-labeled trajectories.

Recently, coding agents have emerged as a dominant paradigm for real-world software engineering (SWE) scenarios, which solve complex tasks through multi-turn interactions with development environments. However, frequent interactions with environments would inevitably introduce substantial token overhead, leading to high usage costs and latency. Although recent studies have explored reducing token usage by context manipulation and interaction limits at inference time, these approaches focus on improving token efficiency while overlooking the risk of discarding task-relevant information, thus struggling to balance the trade-off between resolution rate and token efficiency. In this paper, we study a more general paradigm without suffering from the limitation, i.e., training token-efficient coding agents with promising resolution performance, which is a highly-practical yet less-explored problem. To this end, we reveal two core observations in SWE scenarios: i) Efficiency Variation: successful resolution could be achieved with fewer tokens; ii) Entropy Correlation: unproductive behaviors are associated with turn-level entropy. Motivated by observations, we propose a novel reinforcement learning framework, dubbed HERO. Specifically, HERO prioritizes task resolution over token efficiency during policy optimization and encourages efficient reasoning patterns at both trajectory and turn levels. Extensive experiments on SWE-bench Verified and SWE-bench Multilingual demonstrate that HERO achieves a favorable trade-off between resolution rate and token efficiency compared with state-of-the-art coding agents and reinforcement learning methods.

Equipping VLM agents with world modeling capabilities has shown strong potential for complex reasoning and long-horizon planning, while reducing the dependence of policy learning on costly real-world interactions. Existing methods mainly rely on prospective simulation to predict the consequences of candidate actions. However, this forward-only paradigm focuses on what will happen next and provides limited constraints for verifying whether an action is causally consistent with the observed state transition, which can lead to plausible-looking but physically incoherent behaviors. In this paper, we challenge the view of world modeling as only prospective prediction and introduce Retrospective World Modeling, a new agent learning paradigm that enables agents to reason backward by estimating the retrospective attribution distribution $P(\hat{a}{t}|s_t, s{t+1})$ for the action that most likely caused a given transition. Based on this capability, we formulate the Self-Consistency Reward (SCR), an intrinsic signal that measures the probabilistic consistency between the policy action and the retrospective explanation. Integrating SCR into reinforcement learning provides dense transition-level feedback and steers agents toward behaviors that are both task-effective and physically grounded. Extensive experiments across diverse agentic tasks show that our method substantially improves policy robustness and generalization over prospective-only world modeling baselines.

Multi-reward reinforcement learning trains large language models to satisfy multiple behavioral objectives simultaneously. Reward-wise normalization, as used in GDPO, preserves reward-specific relative information within rollout groups, but different objectives can still exhibit uneven learning progress. We study this behavior through advantage energy, the sum of a reward's squared advantages over a batch. Under idealized GDPO normalization, we show that this energy is proportional to active-group density: the fraction of rollout groups in which the reward provides nonzero relative advantages. This reveals a residual batch-level signal imbalance and provides a basis for calibrating reward contributions. Based on this relation, we propose Density-Aware Reward Aggregation (DARA). We derive an inverse-square-root density correction that gives greater weight to signals from less frequently active rewards. DARA computes its weights from each rollout batch, adapting to changes in reward activity throughout training without modifying the underlying policy optimization objective. Experiments on tool calling and mathematical reasoning show that DARA learns the targeted behaviors faster than GDPO, reaching high format compliance in up to 26% fewer training steps on tool calling and near-saturated length compliance in up to 65% fewer steps on mathematical reasoning, while remaining competitive in final performance. Our code is available at https://github.com/zhaihaotian/DARA.

Aesthetic image cropping aims to identify the optimal crop of an image in terms of aesthetics and composition. While supervision based on annotated data is fundamental, the field has been hindered by a long-standing problem: existing datasets suffer from (1) human subjectivity and (2) rigid discreteness confined to fixed sampling grids. These flawed annotations not only limit the accuracy and generalization of trained models but also severely distort fair evaluation. To overcome this, we propose to model human cropping preference as a multi-peaked, continuous, and sharp field over the crop space. We introduce the Continuous Preference Field (CPF), which recovers a dense preference landscape from discrete annotations through (1) peak clustering, (2) off-lattice refinement, (3) negative shaping, and (4) field assembly. Based on this, we train CPIC, a VLM-based cropping model optimized via GRPO with the CPF reward, which overcomes template collapse, achieving state-of-the-art performance and exceptional out-of-domain generalization. Finally, to resolve the long-standing benchmark evaluation crisis, we introduce CPICD, a comprehensive recalibration of existing ground-truth boxes. By leveraging the CPF to correct grid-bound artifacts across mainstream benchmarks, CPICD establishes a rigorous and reliable foundation for future cropping research. Extensive experiments and user studies demonstrate the superiority of our CPF, CPIC, and CPICD. Code, model, and data are available at https://github.com/zzqingz/CPIC.

Diffusion large language models (dLLMs) generate text by denoising a sequence or successive blocks, allowing several tokens to be revealed in parallel. Reinforcement learning with verifiable rewards (RLVR) reuses terminal feedback across these decisions, even as their conditioning context changes. We propose stepwise risk-sensitive GRPO (StepRS-GRPO), which varies the risk coefficient of the group-advantage transformation across denoising states while retaining the underlying trainer. For binary rewards, we show that this transformation is exactly a prompt- and state-dependent rescaling of centered outcome advantages. A capability-based calibration suggests a coefficient scale, while endpoint and interpolation ablations guide schedule selection. Across multiple dLLM backbones and mathematical reasoning benchmarks, StepRS-GRPO improves both pass@1 accuracy and pass@k coverage over centered GRPO, while increasing answer diversity. In our ablation studies, mass-matched controls support the contributions of state allocation and schedule direction, and the gains persist after matching the root mean square (RMS) of the advantages to that of centered GRPO. Reasoning-trace diagnostics further show that the diversity gains from StepRS-GRPO extend beyond final-answer strings.