Generative RL 进展

Rubric-based reinforcement learning (RL) provides interpretable rewards for aligning large language models (LLMs) by evaluating responses against query-specific evaluation criteria. To construct rubrics at scale, a straightforward approach to LLM-based rubric generation is to prompt an LLM to generate a rubric directly from the query. However, rubrics directly generated by LLMs are vulnerable to reward hacking, since omitted or underspecified criteria allow the policy to obtain high rubric rewards with low-quality responses. Existing LLM-based rubric generation methods improve the granularity and coverage of the generated criteria but do not proactively guard against reward hacking. To address this limitation, we propose RubricArmor, an adversarial framework that exposes and mitigates potential reward hacking at the rubric generation stage before it occurs in subsequent RL. Specifically, RubricArmor performs adversarial evolution, in which an attack step and a repair step alternate over multiple rounds. The attack step simulates the reward hacking of the policy by constructing adversarial responses that satisfy the current rubric but fail to properly complete the task. The repair step then revises the rubric to detect the response defects exposed by the attack step while preserving other valid criteria. Extensive experiments demonstrate that RubricArmor outperforms competitive rubric generation baselines and translates into more effective downstream rubric-based RL.

The next generation of multimodal research agents must reason over long-lived research histories rather than short model completions. During a single task, an agent may repeatedly search the web, inspect visual evidence, revisit earlier hypotheses, and accumulate tens of thousands of tokens of multimodal context. Despite this trend, online RL for multimodal research agents remains largely confined to shorter contexts and interaction horizons. We push online RL training to 128k context and 75+ tool-interaction turns. To our knowledge, this is the first online multimodal deep-research RL study trained at 128k context, and the first trained with a 75 tool-turn horizon. Scaling to this regime exposes several practical limitations of conventional RL training. Early in training, weak policies make poor use of large interaction budgets, causing expensive rollouts with little reward improvement. Later, policy entropy can collapse before performance has saturated, prematurely ending useful learning. We introduce Long-MDR, a three-component training recipe designed specifically for this setting: On-Policy Distillation Warmup, Progressive Horizon Expansion, and Entropy-Triggered Rescue. Together, these techniques improve both the learning efficiency and stability of long-horizon RL, enabling continued gains in a regime where direct training is slow and costly. At a 50-turn evaluation budget, our RL-trained Long-MDR-9B ranks first on five of six benchmarks among the compared 7B-9B agents.

Generative reward models (GRMs) are important for LLM optimization. Unlike scalar reward models, GRMs generate natural-language critiques alongside preference judgments, providing finer-grained evaluation signals. Their effectiveness depends heavily on critique reliability. However, existing GRM training typically uses final preference correctness as outcome supervision. Because the preference outcome space is highly constrained, unreliable critiques can still yield correct outcomes and thus be reinforced. Recent work leverages human critiques for process supervision, but such critiques are scarce and are often reduced to scalar rewards, leaving their fine-grained evaluative information underutilized. We argue that evaluative criteria learned from human critiques can be generalized to broader outcome-only preference data. To this end, we propose EnGRICH, a GRM training framework that pairs the GRM with a training-time MetaCritic learned from a small set of human critiques. MetaCritic constructs response-specific rubrics and uses them to evaluate the evidence coverage and correctness of generated critiques. The resulting signals provide both process rewards for fine-grained credit assignment and structured guidance for exploring better critiques. During GRM training, MetaCritic is further optimized to generalize human-grounded evaluative criteria to outcome-only data. At inference, the trained GRM operates independently. Experiments across seven reward-model benchmarks show that EnGRICH consistently improves over competitive baselines, while further analyses validate the effectiveness of its core mechanisms.

Reinforcement learning with verifiable rewards scales multimodal reasoning, but an outcome reward says how much a trajectory is worth, not how that value should be spread over the decisions that produced it. We introduce Residual Visual Credit Optimization (RVCO), which treats token credit as a conserved routing problem. A controlled visual intervention yields a per-token evidence response; robust within-trajectory coordinates remove incidental scale; and a budgeted entropic router distributes a fixed amount of sequence utility according to perceptual dependence. A residual support path guarantees positive credit at every valid position, and an analytic correction restores the prescribed credit mass exactly. The resulting field is selective, bounded, full-support, and invariant to response-local score shifts, and recovers hard token selection as a limiting case. Across four model families and seven reasoning benchmarks, RVCO improves accuracy over strong RLVR baselines while maintaining late-stage optimization stability, corruption robustness, and competitive training cost. Rewards, rollouts, and the group-relative advantage estimator are unchanged; only the geometry of token-level credit differs.

Aligning large language models (LLMs) with human values is important for safe, efficient, and beneficial AI deployment. However, human values are multifaceted: helpfulness, harmlessness and humor trade off against one another, and different users want different trade-offs. Multi-objective alignment (MOA) addresses this by training a policy that can provide any point of the Pareto front, but existing methods either train one model per preference, interpolate a few separately aligned experts post hoc, or train a single conditioned model without considering which preferences it should be trained on. Since the hard regions of the preference simplex depend on the objectives at hand, existing methods leave them under-trained and do not get the most out of a single model. Therefore, we propose MAESTRO (Multi-objective Alignment via End-to-end STeering and Robust Optimization), which formulates MOA as a minimax problem over preference distributions and trains a single prompt-conditioned policy end-to-end with RL against an adversarial preference distribution: a Dirichlet distribution updated by online mirror descent toward the preferences the current policy serves worst, rather than on a fixed one. On HH-RLHF, BeaverTails and a summarization task, with up to three objectives, MAESTRO attains the best Pareto front on most tasks in a single training run, at the lowest training cost among the compared methods. The largest margins appear in the hard regions that a fixed preference distribution leaves under-trained, confirming that a single prompt-conditioned model is capable of covering the objective trade-offs on its own.

On-policy self-distillation (OPSD) provides denser token-level supervision and better computational efficiency than Reinforcement Learning with Verifiable Rewards (RLVR). However, this denser supervision may introduce substantial noise and training instability. Existing improvements often rely on high-variance per-token statistics and introduce extra hyperparameters and trade-offs. Based on the advantage formulation in RLVR, we analyze the OPSD objective from the same perspective, incorporating outcome correctness signals. We find that vanilla OPSD imposes insufficient penalties and excessive rewards on incorrect trajectories because it applies a fixed divergence objective regardless of outcome correctness. Furthermore, the reliability of teacher supervision is associated with both trajectory outcome and the cumulative average teacher entropy along the rollout. Based on these observations, we propose Outcome-Guided On-Policy Self-Distillation (OG-OPSD), which dynamically adapts both the divergence objective and distillation position according to binary outcome rewards and the cumulative average teacher entropy. Extensive experiments show that OG-OPSD consistently improves the performance of vanilla OPSD and multiple strong baselines in mathematical reasoning, multimodal reasoning, and out-of-distribution tasks across Qwen3 models at 1.7B, 4B, and 8B scales, as well as Qwen3-VL-2B.

On-policy distillation (OPD) improves the reasoning capabilities of small language models through token-level teacher supervision on student-generated trajectories. Yet can teachers that excel at solving problems independently also guide student reasoning effectively? Prior work shows that when student prefixes follow reasoning paths that differ from the teacher's own or contain errors, teachers can be less accurate when continuing from these prefixes than when solving problems independently. To this end, we propose Prep-OPD, which uses reinforcement learning (RL) before distillation to train the teacher to adapt to the student's existing reasoning state and correct course when errors arise. Training optimizes teacher continuations from fixed student prefixes using final-answer correctness as the reward. The prepared teacher then trains the student through trajectory guidance and token-level supervision. We evaluate Prep-OPD on eight mathematical reasoning benchmarks, using Qwen3-4B-Instruct-2507 as the teacher and Qwen3-0.6B and Qwen3-1.7B as students. With the 4B teacher and 1.7B student, Prep-OPD improves average accuracy over standard OPD and the strongest baseline, Relay-OPD, by 8.28 and 2.30 percentage points, respectively. Controlled experiments further show that teacher RL conditioned on student-generated prefixes yields higher student accuracy than problem-start teacher RL with and without handoff on Qwen3-1.7B. Reusing the same prepared teacher also improves Qwen3-0.6B.

Pretrained world models can generate diverse environments, yet users often want to explore a particular scene specified by their own video. This requires learning the scene's visual identity while retaining the quality of action-conditioned generation. We introduce Personalized World Models (PWM), a framework for customizing interactive world models from short scene videos through online reinforcement learning. In PWM, the support trajectory and its associated controls provide reward feedback on continuations sampled from the current policy. In the GRPO instantiation, group-relative optimization updates a compact LoRA adapter using a unified reward for scene appearance, visual continuity, and motion, while base-policy anchoring regularizes changes to the pretrained generation prior of a frozen Yume-5B backbone. The same adaptation procedure is applied across real and rendered environments. We also instantiate PWM with DiffusionNFT as an alternative reward-guided optimization method for learning the scene-specific adapter. We also introduce PWM-Bench, comprising 150 customization tasks across Indoor, Outdoor, and Gaming, with paired evaluation on held-out continuations. The GRPO and DiffusionNFT instantiations of PWM improve customization over native Yume in 71.3% and 65.3% of the evaluated scenes, respectively, with positive mean gains across all three domains. For the GRPO instantiation, matched SFT comparisons further demonstrate higher mean customization gains and better mean image-quality scores in every domain, while retaining frame-level visual quality close to the pretrained model.

Preference alignment has become a standard practice for text-to-image diffusion models. Direct Preference Optimization (DPO) simplifies this process by eliminating explicit reward modeling. Its diffusion variant, Diffusion-DPO, has become a widely adopted baseline. Diffusion-DPO essentially encourages the likelihood of preferred samples while suppressing dispreferred ones. In this paper, we revisit DPO-style alignment methods for diffusion models from the perspective of the manifold hypothesis. Under this view, natural images concentrate near a low-dimensional manifold embedded in the high-dimensional ambient space, whereas DPO directly optimizes preference distributions in the full space without accounting for this geometric structure. This creates a mismatch in the optimization dynamics: it suppresses geometry-preserving tangential updates, while insufficiently restricting hazardous normal-direction updates. This mismatch gradually degrades image quality and diversity. To address this issue, we propose Anisotropic Geometry-Aware Preference Optimization (APO), which replaces the uniform Euclidean treatment of prediction errors with a geometry-aware anisotropic metric derived from the reference model. Concretely, APO adaptively strengthens regularization in directions where the reference denoising function is highly sensitive, while relaxing constraints in directions that permit safe semantic adjustment. This recalibrates preference optimization according to the local manifold geometry, and maintains the original manifold structure. Experiments show that APO achieves strong performance and an average win rate exceeding 60% against various existing alignment methods across diverse benchmarks. It requires significantly fewer training steps than prior methods, and preserves generation diversity throughout training.

Diffusion-based dataset distillation (DD) suffers from a fundamental objective mismatch: likelihood-driven diffusion models prioritize density approximation over the discriminative decision boundaries required for downstream tasks. Beyond semantic mismatch, relying solely on density also leads to geometric coverage loss, where generated samples collapse into a few high-density modes and fail to cover the manifold's structural diversity. We propose Manifold-Guided Policy Optimization (MGPO), which reformulates DD as a multi-objective reinforcement learning problem and achieves Dual-Space Alignment via a pixel-space discriminative reward and a latent-space geometric reward guided by a class-wise Minimum Spanning Tree (MST). The discriminative reward enforces class separability, while the MST-based geometric reward encourages generated latents to cover a sparse geometric skeleton of each class, jointly addressing both failure modes. We further provide an idealized analysis that motivates the MST-based reward, including a Hausdorff approximation bound and a subsampling bound independent of the dataset size. The reward-modular design extends to structured tasks such as object detection and segmentation by substituting the frozen task reward model. Extensive experiments show MGPO consistently outperforms existing methods, including a +8.0% mIoU gain on segmentation under low-budget settings.

Deep search agents tackle complex knowledge tasks through iterative retrieval, multi-hop reasoning, and evidence synthesis across multiple sources. Existing approaches typically assume relatively stable retrieval systems and operate over short-horizon tool interaction. However, when retrieval is sensitive to query formulation, even a semantically appropriate query may fail to surface critical evidence because of mismatched entity names, aliases, or keyword combinations. Recovering from such failures requires repeated query reformulation and longer interaction trajectories. This setting poses a distinct training challenge, as the policy must sustain long-horizon query exploration while managing an expanding volume of retrieved content. We propose LexiHorizon, a framework for training search agents over long horizons that expands the trajectory context budget, manages accumulated retrieval content using a window over recent tool observations while preserving the reasoning history, and introduces an outcome-gated search-effort reward that provides a bounded bonus for tool invocations to trajectories with nonzero answer reward. Experiments on XBench, WebWalkerQA, and BrowseComp-ZH show that the resulting 9B model consistently outperforms both its base model and MiroThinker-1.7-mini, with maximum absolute gains of 8.7 and 23.8 percentage points, respectively. These results suggest that combining an extended context budget with reasoning-preserving context management benefits long-horizon deep search agents.

Large language model (LLM) agents solve complex tasks through multi-step interactions with external tools. These interactions often contain recurring local tool sequences. Treating such sequences as composite "Skills" can shorten tool-use trajectories and reduce repeated low-level decisions. However, when atomic tools and composite skills coexist, skill use becomes a policy problem: the agent must decide whether the current state requires atomic fine control or skill-level abstraction. In this paper, we argue that effective skill use should be studied as adaptive tool granularity selection. The most direct training signal for this problem is to compare the consequences of atomic and skill choices available from the same state. Based on this view, we propose CIPO, a Counterfactual Imagination Policy Optimization framework for adaptive tool granularity. CIPO constructs executable skills through budget-constrained mining of successful tool-use trajectories and trains granularity decisions with counterfactual branch rollouts. For each base rollout, CIPO branches at the first eligible granularity decision and replaces the chosen action with a feasible atomic or skill alternative. The paired outcome difference serves as a supplementary reward for policy optimization. Experiments across multiple benchmarks and model backbones show that CIPO improves task success and decision efficiency over baselines. Further analyses show that CIPO learns effective skill use by improving the choice between atomic tools and composite skills based on the current state, without simply increasing skill frequency.

Large language models can intervene in reinforcement learning through both reward design and action selection, yet aggregate performance offers an incomplete account of what these interventions actually do. Similar returns can conceal different learning mechanisms, while plausible rewards can induce undesirable behavior. We introduce LocusRL, a diagnostic framework that connects controlled reward-policy comparisons with audits of reward judgments, signal delivery, optimization objectives, and executed actions. The framework traces performance differences to testable explanations and checks targeted corrections through executable rules and counterfactual replay. Across two evaluation batches covering ten Connect Four training seeds, we uncover seed-dependent reversals in intervention effects and show how tracing actual updates changes their interpretation: historical Qwen training operates through reward-weighted teacher-action likelihood. A separate matched three-seed reward-direction experiment distinguishes sensitivity to a learning signal from its usefulness. With terminal rewards held fixed, a sign-reversed dense oracle yields a 2.8% aggregate win rate, compared with 57.2% for terminal-only training and 46.7% for the positive dense oracle. Thus, a reward can strongly influence learning without improving performance. At the decision level, counterfactual replay verifies a winning correction to a diagnosed action error. Complementary experiments in Leduc and reward-validation studies in Goofspiel extend the analysis to imperfect-information settings, revealing how reference-label definitions and validation-data exposure affect intervention assessment. Together, these findings show why evaluating LLM interventions requires tracing how their outputs become learning signals and actions. LocusRL turns aggregate outcomes into actionable diagnoses and verifiable corrections.

Correction-based offline preference pipelines commonly treat model failures only as rejected responses under the original prompt. This supervision is incomplete for boundary failures: responses that violate the given instruction yet coherently satisfy a nearby intent or constraint setting. We introduce Bidirectional Preference Synthesis (BPS), a data-construction method for standard Direct Preference Optimization (DPO) that makes this missing prompt dependence explicit. For each validated boundary failure, BPS keeps the conventional forward pair under the original prompt and adds a reverse pair under a synthesized achieved prompt, so the same response is rejected where it is wrong and chosen where it is right, without changing the DPO objective, training a reward model, or requiring online sampling. On Qwen3-4B-Instruct-2507, BPS preserves original-side pairwise ranking while raising achieved-side ranking accuracy from 6.8% to 62.3% on held-out crossed anchors, with a similar shift under a Kimi-K2.6 cross-teacher probe. A blind human audit supports the intended reverse preference direction, and downstream evaluations show the clearest separation from Forward-DPO in multilingual multi-turn instruction following, with consistent capability-retention patterns on agentic, tool-use, and code checks.

Diffusion and flow models provide expressive policy classes for online reinforcement learning (RL), enabling multimodal behaviors and improved performance. However, training these policies remains challenging: the critic specifies the desired policy as an unnormalized Boltzmann density but does not provide direct samples from it. Many existing methods rely on importance sampling to construct training signals, which can suffer from high variance, increasing computational cost and destabilizing training. We propose Score-Calibrated Flow (SCF), a simple and efficient algorithm for training generative models to sample from unnormalized densities without importance sampling or backpropagation through the sampling trajectory. We learn the desired flow by enforcing self-consistency, bypassing target posterior mean estimation. By jointly exploiting the prescribed target score and the structure of flow matching, we establish these self-consistency requirements as score-calibrated optimality conditions, first for the terminal density and then for the trainable velocity field. We prove that their unique solutions are, respectively, the target density and the ideal flow model that conditional flow matching (CFM) would recover if target samples were available. We formulate the velocity condition as a fixed-point equation and exploit its conditional-expectation structure to construct a stop-gradient objective for enforcing it. The resulting training procedure retains the scalable sample-interpolate-regress structure of CFM despite the absence of target samples, using endpoints generated by the current flow. For online RL, the critic gradient supplies the target score at the generated actions, yielding a direct approach to actor training. Experiments on RL benchmarks demonstrate that SCF matches or improves upon state-of-the-art generative-policy baselines, while substantially reducing training time.

LLM-based long-horizon agentic post-training is often bottlenecked by rollout generation: trajectories span many interaction turns, completion times vary substantially, and synchronous update barriers leave faster workers waiting for stragglers. Asynchronous reinforcement learning which has been adopted in LLM post-training addresses this inefficiency by consuming trajectories as they arrive, but introduces policy lag and off-policy optimization. Evolution strategies (ES) offer a backpropagation-free alternative for LLM post-training, yet it relies on a larger number of rollouts and existing practices have remained largely synchronous. In this short-form paper, we introduce bounded-staleness asynchronous ES and demonstrate it on Endless Terminals benchmark using Qwen2.5-7B-Instruct. Across three evaluation seeds, natural Async-1 matches synchronous ES, achieving 25.9% versus 25.4% held-out success. Controlled schedules that delay 10% of each update cohort by four or eight policy updates reduce success by only 1.6 and 3.1 percentage points, respectively, without explicit off-policy correction. GRPO performs better overall, reaching 29.0% held-out success, but importantly our results show that ES tolerates moderate policy staleness with limited degradation, opening possibilities for future improvement of ES-based post-training with asynchronous algorithms. To the best of our knowledge, we are the first to demonstrate the effectiveness of sync and async ES on a multi-turn terminal style agentic coding task.

Reinforcement learning for embodied control remains constrained by the difficulty of reward specification. Although recent large language model (LLM)-based methods can synthesize reward functions from natural-language descriptions, they often fail to capture subtle behavioral properties that humans care about, such as natural gait, posture, and movement style. This limitation arises because many desired behaviors are easier to recognize visually than to encode in a reward function. We introduce Reward Optimization via Observable Trees (ROOT), a framework for discovering reward functions that align learned policies with user-specified embodied behaviors. Rather than relying solely on scalar training statistics, ROOT casts reward design as an observation-guided search over a persistent experiment tree that stores reward programs, trained policies, and rollout observations, together with behavioral insights distilled by a video-language model, to diagnose behavioral failures and guide subsequent reward refinements. We evaluate ROOT on seven tasks across four embodiments: simulated Hopper, HalfCheetah, Ant, Unitree Go2, and as well as the real-world Unitree Go2. ROOT produces behaviors that better align with user intent than those generated by existing LLM-based reward-generation methods, achieving up to 86.8% locomotion-completeness accuracy and improving Vid-LLM behavioral alignment from 3.56/5 to 4.14/5, a 16.5% improvement over baselines. Human evaluations further support these results, with ROOT preferred in 51-63% of pairwise comparisons.

On-policy distillation (OPD) has emerged as a widely used paradigm for post-training large language models, reducing the train--test mismatch of conventional distillation by supervising the student on its own generated trajectories. However, existing OPD objectives remain largely token-local and outcome-agnostic, optimizing teacher--student agreement at each prefix despite reasoning quality being determined at the trajectory level. Reinforcement learning with verifiable rewards (RLVR), particularly Group Relative Policy Optimization (GRPO), provides complementary outcome-level supervision but suffers from sparse rewards and coarse credit assignment. We show that OPD and RLVR exhibit complementary blind spots: teacher signals provide dense local guidance but are weakly aligned with rollout correctness, whereas group-relative rewards capture task success but provide coarse token-level credit and vanish on all-failure groups. We introduce DiffGate, an outcome-gated objective that combines GRPO with selective, bounded teacher guidance. Teacher supervision is applied only to failed trajectories, scaled by group difficulty, and smoothly bounded to prevent extreme teacher--student discrepancies from dominating optimization. The verifier therefore determines which trajectories receive teacher guidance, while the teacher provides dense token-level update directions within those trajectories. Across Qwen3-0.6B and Qwen3-1.7B students, DiffGate improves code avg@8 over matched GRPO by $+1.7$ and $+1.8$ points and pass@8 by $+1.6$ and $+5.7$ points, respectively. On mathematics, avg@8 remains within $0.5$ points of GRPO while pass@8 improves by $+1.1$ and $+3.9$ points. Overall, DiffGate improves pass@8 across all four model--domain settings, demonstrating improved solution coverage under our evaluation protocol.

Reward-guided image editing at test time seeks to improve a specified reward while preserving source content and visual plausibility. Many existing approaches optimize candidates through pretrained generation processes, making repeated adjustment depend on costly large-model execution and, in some cases, backbone backpropagation. We develop a theoretical framework that jointly accounts for reward, source preservation, and pretrained-prior preferences, allowing the desired output distribution to be specified separately from the dynamics used to realize it. Based on this framework, we introduce FASTER, which trains a small network for each source and objective to perform inexpensive editing, while pretrained and reward models provide feedback on candidate outputs. By reusing each candidate and its feedback across multiple small-network updates, FASTER reduces repeated sampling and supervision queries without placing the pretrained generative backbone inside the inner optimization loop. On SD3, FASTER leads all four target metrics and several validation metrics among the evaluated methods. Compared with the evaluated baseline that optimizes controls along pretrained generation trajectories, FASTER achieves editing-time speedups of up to \({6.91\times}\) on Stable Diffusion 3 and \({24.14\times}\) on Stable Diffusion 1.5.

Although reinforcement learning (RL) post-training repairs the localized segmental errors of zero-shot text-to-speech (TTS), arriving at a working recipe still relies on tedious manual tuning, and whether LLM agents can take over this research pipeline is unclear. We investigate this question with AgenticTTS-Forge, a collaborative workflow that structures human guidance and agentic execution around a shared workspace, applied to CosyVoice2-0.5B. To measure what the agent automates, we audit its trajectory stage by stage against the published recipe. To measure what it exploits, we score its policies with held-out observers hidden from the agent. Our results show that the agent recovers an underspecified recipe, improves it, and, when gains stall, surveys the literature unprompted and pivots from the LM carrier to the flow carrier, halving Bad cases. However, its autonomy exposes three traps across the data, proxy, and algorithm axes: the held-out set leaks through a channel the contract never reads, a self-shaped reward inflates the proxy where it is scored, and separately tuned policies do not compose additively. These findings show that the binding constraint is measurement rather than reasoning, and can inform the design of harnesses whose contracts read every channel the agent does.

Reinforcement learning with verifiable rewards (RLVR) has been shown to improve the reasoning capability of large language models (LLMs) across diverse reasoning tasks. However, group-based RLVR methods, such as GRPO, assign a uniform advantage to all tokens within rollouts of the same outcome. While existing works refine credit assignment of GRPO based on local signals such as token locations or entropy, they often fail to capture the global semantic novelty of a reasoning behavior relative to the current policy. In this work, we propose a hierarchical credit assignment approach for group-based RLVR methods, called HarA, which identifies and encourages semantically novel reasoning behaviors during RLVR. HarA represents each sampled rollout as a distribution over the hidden states and locations of tokens, and computes the Fused Gromov-Wasserstein (FGW) barycenters of all rollouts with the same outcome, capturing the internal reasoning patterns in the latent space under the current policy. The semantic novelty of a reasoning element can then be measured by its contribution to the FGW distance between the current rollout and the barycenter. While solving the FGW formulation is expensive, we introduce an anchor-guided linearization that turns it into a Wasserstein formulation solvable via the Sinkhorn algorithm efficiently. By reweighing token-level advantage of group-based RLVR methods based on the novelty signals, HarA highlights novel reasoning behaviors at flexible granularities to encourage fine-grained LLM exploration. Extensive experiments across three group-based RLVR methods show that our plug-and-play method effectively enhances the exploration of LLMs, outperforming existing methods across diverse reasoning benchmarks.

Recent studies on reinforcement learning (RL) report seemingly conflicting evidence about large language model (LLM) reasoning. Training on mathematics can improve performance in other domains, yet gains in Pass@1 can coincide with lower Pass@$N$ than the base model. This raises a fundamental question: does RL expand an LLM's reasoning boundary, or merely reweight its existing reasoning space? We revisit these phenomena across Qwen and Gemma model families, showing both cross-domain gains and forgetting, while coverage at large sampling budgets increases on some tasks and decreases on others. Detailed analysis of solution traces before and after RL indicates a shift in the reasoning strategies the model employs, motivating a two-stage autoregressive policy model that separates strategy selection from problem-specific execution. Within this framework, we prove how RL's implicit bias reshapes strategy preferences, allowing gains on some tasks while suppressing strategies required by others. This mechanism can also broaden or narrow coverage at a given sampling budget even without expanding strategy support. We further provide theoretical justifications for log-sigmoid and log-linear scaling laws in RL compute, and evaluate their predictive power. Together, these results connect changes in strategy selection to cross-domain transfer, reasoning coverage, and compute scaling.

Reinforcement learning with verifiable rewards improves reasoning, while the allocation of learning signal shapes which solutions remain accessible under repeated sampling. Group-relative objectives assign equal advantages to equally rewarded responses, making aggregate credit proportional to sampled mode frequency. We introduce Exploration-Preserving Policy Optimization (ExPPO), a lightweight advantage-shaping rule that redistributes credit using prompt-relative, length-normalized response surprisal and prompt pass rate. ExPPO combines bounded shaping with shared normalization to preserve verifier polarity and approximately maintain each prompt group's total absolute sequence-advantage mass. Our analysis characterizes response-level credit allocation alongside sampled mode updates, deriving local conditions for gains in entropy and correct-mode discovery. Experiments show improved in-domain and out-of-domain reasoning coverage, higher aggregate response accuracy, and strong coverage at large sampling budgets. A controlled multi-answer evaluation further demonstrates increased correct-mode yield and gains in diversity among verified-correct responses. Code is available at https://github.com/jinhangzhan/ExPPO

Developing socially intelligent AI remains heavily dependent on human-annotated data, limiting the scale and breadth of social understanding models can acquire. Methods that derive training signals from unlabeled data offer a path beyond this dependence, but social predictions lack the verification oracles available in mathematics and coding. Moreover, core social targets such as affect, intent, preference, and pragmatic meaning are often ambiguous. The same behavior can support multiple plausible interpretations, making it difficult to verify which is best supported. To address this challenge, we introduce Reinforcement Learning with Comparative Evidence (RLCE), a reinforcement learning method that learns social understanding from unlabeled training data without constructing rewards from ground-truth annotations. Given distinct answers in a rollout group, RLCE constructs evidence tests that identify observable evidence favoring an answer over another, validates these tests against the input sample, and aggregates test outcomes to determine the best-supported interpretation. Tests are regenerated as the policy produces new answers, enabling them to evolve with the policy. Across four benchmarks spanning affect, pragmatics, communicative intent, and preference, RLCE attains the strongest performance among seven methods that use no ground-truth training labels for rewards, including consensus, policy LLM-judge verification, multimodal co-evolution, and rubric-based rewards. Gains over the strongest baseline reach up to +18.93 points. Analyses further show that RLCE exhibits a larger share of reward variation between correct and incorrect predictions than compared rubric methods, can overturn erroneous policy-derived preferences, and benefits from pairwise test construction, compositional test aggregation, and on-policy test evolution.

Large Audio Language Models (LALMs) are prone to hallucinating and over-relying on text priors when simultaneously presented with audio and text inputs. To mitigate these hallucinations, we propose utilizing the multimodal Direct Preference Optimization (mDPO) objective, which forces the model to ground its generation in the acoustic input by contrasting intact and distorted audio counterparts. We extend this preference learning framework to the audio domain by applying a variety of acoustic perturbations. Evaluating the Qwen2-Audio backbone across the DCASE 2025 Challenge and AH Existence datasets, we demonstrate that extending mDPO to LALMs significantly enhances temporal reasoning in the complex DCASE dataset, and improves performance on basic existence verification in the AH benchmark. We identify temporal reversal, frequency masking, and random noise as the most effective perturbations. Ultimately, our approach achieves an absolute accuracy improvement of 14.0% on the DCASE 2025 dataset and 27.4% on AH Existence.

Inverse materials design is a long-standing goal of computational materials discovery. Generative models for crystalline materials are typically trained to match the distribution of a structure database, while nothing in their training objective points them at specific design goals such as targeted properties. We use group-relative policy optimization (GRPO) to align a generative model based on stochastic interpolants and discrete flow matching with general black-box reward functions through reinforcement learning. Atom types are generated by a discrete flow and the policy gradient of our generalization of GRPO directly acts on the likelihoods of the atom-type transitions, which differentiates our work from previous reinforcement-learning approaches for diffusion and flow-based generative models of crystalline materials. We introduce a reward function that raises the yield of metastable, unique and novel structures (mSUN) from 13.4% for the pretrained model to 45.5% for the reinforced model, as evaluated by a community benchmark. Our reward also improves the performance of a reinforcement learning framework for crystalline materials based on latent denoising diffusion models. At the same time, we find that directly reinforcing atom-type transition likelihoods enables reward exploitation that has to be prevented with explicit guards. The same analysis also exposes a gap in the community metric. Single-element structures in distinct packings are counted as metastable, unique and novel materials and inflate mSUN without yielding any new compounds. A stability claim is only as good as its reference hull. We report every result split by the number of reference phases behind it and argue that benchmarks should do the same.

Building LLMs that behave well socially, not merely correctly, requires Building LLMs that behave well socially, not merely correctly, requires more than producing locally helpful responses. A socially competent agent must infer users' unstated goals, respect their preferences, and adapt as the conversation unfolds. These behaviors are inherently multi-turn and social, making them hard to optimize: real interaction data is scarce, and user preferences are typically latent rather than directly observable. To address these challenges, we build on a persona-driven social simulation environment (consisting of a persona library, LLM-based user simulators, and a user-satisfaction scoring system ranging from [0, 1]), to introduce preference-batched GRPO (PB-GRPO), a post-training algorithm that learns socially adaptive policies from conversation-level feedback. Compared to vanilla GRPO, PB-GRPO computes advantages using a normalization estimated across a bucket of users with similar preferences, stabilizing training across a diverse social population. Empirical evidence shows that PB-GRPO improves models' social behavior over strong reinforcement learning baselines in our simulated environment.

Methods for training language models on stale samples are judged by comparisons against importance-corrected baselines. We show that details of the experimental harness can reverse the observed ranking of methods, and we introduce PTH (Probe The Harness), a set of checks that makes the harness visible. Our case is a comparison between SAN, a behaviour-free method, and truncated importance sampling (TIS) on verl and in a single-GPU trainer, in which SAN first finished ahead in both stacks. Four details of the harness changed this comparison: the PPO ratio was taken against the learner's own recomputed probabilities, the data seed did not reach the TIS arm, the replay queue reused its first batch for 33 updates, and two loss normalisers differed from their description. In each case the logged quantity looked consistent with a working setup, while the quantity that defines the comparison went unchecked. With the harness checked, TIS matches SAN on verl, and in the trainer TIS learns steadily while SAN keeps a margin. We contribute the signature of each detail and its effect on the comparison, reference results for TIS and uncorrected GRPO under sampler lag, and the PTH checklist.

Improving joint audio-visual reasoning in Omni Large Language Models typically incurs substantial data construction and training costs. Our diagnostics reveal multi-hop reasoning difficulties despite correct answers to all corresponding single-hop questions and suggest partial decoupling in the local optimization of perception and reasoning objectives. This motivates post-training with different emphases on these capabilities. Text-only reasoning training yields gains across data sources, model scales, and families. With the best-performing text-only configuration, supervised fine-tuning followed by reinforcement learning (RL) raises Qwen2.5-Omni-7B's geometric mean of nine reasoning scores by 25.83% over the base model, outperforming the complete native audio-visual route with 56.6% fewer GPU-hours. Training on data synthesized entirely by a text-only LLM raises this geometric mean by 21.01% without audio-visual data in construction or training. However, text-only training degrades perception. We therefore propose a text-centric post-training paradigm: text-only training provides the main reasoning optimization, and reduced-data native audio-visual RL then refines perception. Refinement uses about 90% fewer input tokens than full-data audio-visual RL, restores perception above the base level, and retains 93.5% of the best-performing text-only pipeline's reasoning gain.

Rubric-based reinforcement learning extends reward-driven optimization to open-ended tasks by assigning partial credit to individual response requirements. However, rubric judges can assign a high criterion score even when the information or action it requires is absent from the response, a failure mode we term Vacuous Credit. Such awards persist after the required information is removed and can reverse the sign of a response's GRPO advantage. To address this problem, we introduce MetaRubric, which alternates evidence-aware policy optimization with response-guided rubric adaptation. We construct counterfactual counterparts by changing one task-relevant fact in each prompt. During policy optimization, credit is assigned only when the response contains sufficient evidence to satisfy the required rubric criterion. After each policy-optimization stage, current policy responses guide revisions to original and counterfactual criteria while preserving the meaning of the original prompt's initial rubric as interpreted under each prompt's facts. We also adapt criterion weights at stage boundaries to better address observed policy errors. Across multiple backbones, MetaRubric improves PubMedQA accuracy by 6.00--20.40 percentage points over static-judge GRPO, with further gains on HealthBench-Hard and two multimodal medical benchmarks.