Generative RL 进展

Why does removing reference-policy KL regularization sometimes improve group policy optimization? This motivates studying how reference-policy information should enter group-relative updates. We analyze seven potential failure modes in the interactions between KL and rewards: residual KL updates after reward clipping, after gradient cancellation, and in groups with identical rewards; KL growth with response length and an imbalance in its relative contribution; KL concentration on a small number of tokens; and sampling noise when k1 is incorporated into rewards. We propose Zero-Sum Calibrated Policy Optimization (ZCPO), which uses relative drift measured by conditional KL to calibrate within-group reward coefficients and integrates them into the base surrogate. Mathematical reasoning experiments and ablations support this design's effectiveness in our settings.

Foundation models supply reinforcement learning (RL) with priors that mitigate its longstanding weaknesses in sample efficiency and transfer, but their token-by-token generation makes queries sequential and costly. Jev, a recently released decision model, generates nothing and returns calibrated, typed answers in a single forward pass. Existing work studies foundation models in RL either as models to be trained or as generators to be prompted, and Jev belongs to neither category, having so far served only as a black box in single domains. How well such a model decides on its own in RL environments, and how it can improve RL as a component of training, therefore remain unaddressed. To this end, in this paper we first examine the requirements that the objects of an RL system place on the answers they consume, and establish that Jev can fulfill all of them except the cardinal use of a value function. The remaining objects form positions that admit several roles each. We then construct algorithms that employ Jev at three of these positions, as a reference policy, an exploration judge, and a replay rater, to improve sample efficiency, exploration, and learning performance. Across nine MiniGrid tasks and three Atari games, training with Jev outperforms a standard RL learner, including where the learner makes no progress alone, while the model itself remains untrained. To our knowledge, we present the first use of Jev within the RL learning process and establish a frozen decision model as a usable component of RL training, inviting further exploration of how Jev and other advanced decision models can improve RL.

Reinforcement learning (RL) is the central training paradigm for advancing large foundation models towards self-improvement. This report introduces the MiMo-V2.6 series, an omni-modal family that pushes the frontier of model intelligence by scaling RL compute. Prior to RL, we conduct mid-training on a broad multimodal corpus to provide ample exploration space, and build a solid infrastructure on the pretrained hybrid-SWA architecture to support subsequent scale-up. We scale RL compute along three dimensions: (1) larger batches and higher throughput, with an asynchronous training that consumes 1,568 samples and 2.7-3.7B tokens per step at context lengths of up to 1M; (2) more diverse and complex environments, spanning code, general, visual, and cyber domains under a mixture of agent harnesses; and (3) more grader compute, via groupwise agentic grading that yields more accurate reward signals for long-horizon tasks and steers the model towards shorter, more token-efficient solutions. To keep training stable at scale, we freeze the MoE router and establish a multi-layer defense against reward hacking. We further build infrastructure for mixed-task agentic RL, including a unified trajectory representation, high-concurrency multi-framework rollout, decoupled control and data planes, and training-inference consistency. We open-source the training dynamics, RL environments, and RL framework to facilitate reproduction and further research on scaled RL and model self-improvement.

Long-form video understanding often involves multiple questions about different aspects of the same recording. Yet existing video agents typically process each question through an isolated tool-use trajectory. This repeatedly restarts video exploration and memory construction, missing opportunities to acquire evidence jointly and progressively build a shared understanding that supports the complete question set. We introduce VAMR (Video Agent for Multi-Question Reasoning), which coordinates all questions about a video through one shared tool-use trajectory. At each round, a persistent policy model can invoke tools for one or more unresolved questions and submit answers for questions with sufficient evidence. Question-conditioned visual perception retrieves fine-grained clues for several questions in one call, while layered multi-question memory integrates reusable context into a shared video story and preserves separate evidence for individual questions. After supervised fine-tuning initializes this interaction protocol, we propose question-horizon policy optimization (\qhpo) to optimize shared trajectories in which questions progress and finish at different rounds. Specifically, a question-level critic estimates the value of each active question, while round alignment maps each question advantage to the rounds that directly serve it before the aligned advantages are aggregated to optimize the shared actor. Across LVBench, Video-Holmes, and LongVideoBench, VAMR achieves the highest accuracy overall and the fewest reasoning rounds among iterative methods. On LVBench, it reaches 62.1% accuracy, exceeding VideoARM by 4.3 points while reducing reasoning rounds and processed frames by 85.9% and 61.4%.

Large language models (LLMs) are commonly post-trained against LLM judges and other cheap surrogates because the true reward, such as human preference, is too expensive to query at scale. This practice often leads to reward hacking, where reinforcement learning against a miscalibrated surrogate leads to undesirable side effects. In this work, we study a setting in which a small number $n$ of model outputs are annotated with ground-truth labels (e.g., from expert review) and used to recalibrate the LLM judge before optimizing against it. Prior approaches to judge recalibration are costly or heuristic, and it is known that on-policy sampling fails when the surrogate is miscalibrated on a rare set of outputs. In this work, we propose envelope sampling, a theoretically-grounded method for judge recalibration that seeks to minimize an upper bound on the regret of the post-trained model under the assumption that the human reward and re-calibrated reward lie in an $L^2$ ball around the judge. We give practical algorithms to sample from the envelope by rejection or by fine-tuning against a modified reward, and experiments on clinical note generation and on a controlled sycophancy task show that recalibrating on envelope samples mitigates reward hacking where recalibrating on base-model samples does not.

Language models (LMs) are commonly trained with Reinforcement Learning with Verifiable Rewards (RLVR) to enhance their reasoning capabilities. However, since RLVR does not explicitly account for calibration during training, it can lead to severe calibration degradation, including overconfidence. Recent calibration-aware training methods for LMs, which incorporate objectives for uncertainty estimation into training, improve calibration but still exhibit overconfidence under distribution shift, while sacrificing reasoning performance. To this end, we propose RL-ARC, a calibration-aware training framework that jointly leverages reasoning confidence and answer confidence. Specifically, RL-ARC leverages reasoning confidence as an auxiliary signal for calibrating answer confidence, applying it as reasoning-guided regularization for correct cases and as an overconfidence penalty for incorrect cases. Comprehensive results across ID and OOD settings show that, beyond improving calibration, RL-ARC enables reasoning models to adaptively estimate confidence based on the given question without substantially sacrificing reasoning performance, thereby highlighting the importance of reasoning confidence for training reliable reasoning models.

A verified reference solution provides a correct trajectory for training a reasoning model. Alternatively, a prefix of the reference can guide the model in generating a trajectory of its own. How much reference guidance should we provide? We study this question through prefix continuation, where the model continues from a reference prefix and keeps the resulting trajectory if it passes verification, falling back to the reference otherwise. Since both procedures produce correct trajectories, we compare their distributions with the ideal distribution, the model's own distribution conditioned on successful verification. For one continuation, we derive the KL divergence in closed form, which, up to a bounded term, decreases with the product of the probability of generating a different correct trajectory and the reference surprisal, the negative log probability of the reference suffix given the prefix. Since a longer prefix tends to raise the former but lowers the latter, continuation success alone does not determine the preferred amount of guidance. From this analysis, we learn a prefix selector shared across training questions from continuation outcomes, without estimating success probabilities or additional generation. The resulting Adaptive Reference Guidance (ARG) constructs correct trajectories within a fixed generation budget, and we apply it to all-failure groups in Group Relative Policy Optimization (GRPO). Experiments on Qwen3-4B and Qwen3-8B across five mathematical reasoning benchmarks show that ARG achieves the highest aggregate pass@12 among the evaluated methods with competitive average sampled accuracy.

A language model (LM) can usually answer the same question in more than one way, but reinforcement learning with verifiable rewards (RLVR) is indifferent to which correct answer a model produces. A solution will earn the same reward whether it is the thousandth copy of a familiar answer or one the model has never produced before. Yet, there is potential value in having the model retain multiple correct solutions as it is trained. For instance, multiple modes may give users a choice and provide problem-solving strategies that improve overall model performance. Here, we introduce ModeBench, a benchmark of multi-solution tasks in which the verifier returns both correctness and mode discovered. We then use ModeBench to measure how solution diversity changes under RLVR post-training. We find that RLVR post-training concentrates probability onto fewer correct modes even as accuracy holds or improves, and moreover, that frontier models are already highly concentrated. We then introduce our solution, Re:Max, which stores one verified example per discovered mode in a replay buffer and trains on those stored modes uniformly. A solution found once is, therefore, practiced as often as one found repeatedly. Across three model scales, two RL objectives, and harder task constructions, replay improves both how often a policy succeeds and how many different ways it can succeed.

Large Language Models (LLMs) have shown strong reasoning capabilities when fine-tuned with reinforcement learning (RL), particularly through Group Relative Policy Optimization (GRPO). However, existing GRPO methods assume centralized access to training data, which may not hold in practice due to privacy or regulatory constraints. To this end, we propose Fed-GRPO, a federated GRPO training framework that addresses these privacy constraints by enabling collaborative reasoning training without sharing raw data, which leverages the reward statistics naturally produced during GRPO training as zero-cost signals to guide aggregation, local training, and communication. Fed-GRPO contains three reward-signal-driven mechanisms: (i) signal-weighted aggregation that weights clients by their reward standard deviation, prioritizing clients with stronger learning signals; (ii) global reward calibration that re-weights per-prompt objectives based on the local-global reward gap, steering each client toward its relative weaknesses; and (iii) adaptive sparse communication that allocates bandwidth based on the informativeness of each client's update. Extensive experiments on mathematical reasoning tasks demonstrate that Fed-GRPO achieves the best performance among all federated methods, clearly outperforms FedAvg and approaches centralized training performance, while losslessly reducing communication by $32\times$ and supporting up to $621\times$ compression under tight bandwidth budgets with only graceful accuracy degradation. Our code is available at https://github.com/HKU-HealthAI/Fed-GRPO.

Few-shot learning aims to recognize novel categories from limited labeled examples. Recent studies incorporate textual semantics to compensate for limited visual observations and improve class representations. However, high image-text agreement may reflect both intrinsic object properties and incidental context, making support prototypes susceptible to contextual contamination. To address this problem, we propose DVLA-RL++, which extends DVLA-RL with complementary semantic purification (CSP) and counterfactual reinforcement-learning gating (CRG). Specifically, CSP generates intrinsic and nuisance descriptions from labeled supports and compares their agreement with each support token. An ambiguity-dependent rejection margin guides sparse evidence allocation, while an intrinsic semantic anchor fills the unassigned mass to provide a fallback when visual evidence is unreliable. CRG learns layer-wise semantic fusion strengths using a reward that balances recognition performance and nuisance exposure. An independently executed reference trajectory on the same episode provides a paired learning signal. Theoretical analysis relates retained evidence and anchor quality to prototype stability and establishes conditions for unbiased on-policy gradient estimation. Experiments on standard, fine-grained, and cross-domain benchmarks show state-of-the-art accuracy, with an average gain of 1.4% over DVLA-RL. The project page is available at https://peacelwh.github.io/TPAMI27-DVLA-RLpp/.

Vertical-domain few-shot classification remains challenging for small language models, as limited supervision makes it difficult to acquire domain-specific decision knowledge. On-Policy Distillation (OPD) can improve teacher-guided adaptation by supervising student-generated rollouts, while GRPO-based reinforcement learning can further refine downstream predictions. However, existing KD-to-RL pipelines typically rely on globally fixed transition schedules, ignoring that different samples may require different amounts of teacher-guided acquisition before reward-driven refinement. We propose PIVOT (Perplexity-Informed Transition Optimization), a dynamic transition framework that routes samples between OPD and GRPO according to teacher-evaluated sequence perplexity. PIVOT moves low-perplexity samples to GRPO for reward-driven refinement while keeping high-perplexity samples under OPD for continued domain knowledge acquisition. Experiments on Banking77 and HWU64 show that PIVOT consistently outperforms continued OPD and globally synchronized OPD$\rightarrow$GRPO baselines under the same number of post-warm-up student optimization steps, achieving stronger downstream performance and more stable training dynamics.

Reinforcement learning with verifiable rewards (RLVR) and on-policy distillation (OPD) have become two main paradigms for post-training reasoning models. RLVR gives each response a single outcome label, leaving the steps inside it without separate credit. OPD provides token-level guidance at student-visited prefixes, but its pointwise signal does not directly reflect the pattern of teacher--student disagreement across the vocabulary. Dense, unbounded log-ratio supervision can amplify the teacher's influence, yet a strong solver is not necessarily a suitable guide when the student's solution paths depart from the teacher's. We propose Residual Advantage (\RA{}), which treats the teacher--student probability residual as a bounded one-step reward, subtracts the corresponding state value under the student policy to form a standard advantage, and centers the result within each response before adding it to the verifier advantage. The guidance term has zero mean within each response, so the verifier advantage remains the response's mean label and the teacher only redistributes credit among the steps within it. \CoRA{} further updates a teacher LoRA with verifier advantages on the same scored student batch and uses the updated teacher in the next iteration's residual, adapting guidance to the student's attempts. With Qwen3-1.7B-Base and Qwen3-4B-Base students and a Qwen3-8B teacher, \RA{} combined with GRPO or REINFORCE++ improves the underlying sequence-advantage algorithm in all 24 comparisons on three mathematical benchmarks, raising macro Avg@8 by 1.7--3.6 points and Pass@8 by 3.9--6.3 points. Both combinations surpass teacher-only OPD, and \CoRA{} adds a further 1.0--1.5 Avg@8 points.

Large foundation models have accelerated progress toward general-purpose agents that interact with humans and other agents through language and multimodal signals. However, robust multi-agent decision-making requires reasoning about what other agents know, intend, and are likely to do under partial observability. Current agentic systems often operate through prompt design, memory, or end-to-end behavioral shaping, but typically do not learn an explicit partner-state representation that can be reused as a decision variable across tasks. We introduce mental-model-enabled agents, a framework that equips an agent with a latent mental model of its counterpart, allowing it to infer hidden beliefs, intentions, and likely reactions from the observed history and use these inferences to guide action selection. Our method learns an amortized recursive Theory-of-Mind representation, with first- and second-order mental-state structure, jointly with a belief-conditioned reward model that evaluates candidate actions relative to the inferred partner state. A policy is then learned under this belief-aware signal, yielding an agent that can act independently at inference time while retaining the benefits of explicit partner modeling. We evaluate the same framework on both language-only and multimodal benchmarks. Across these settings, explicit mental-state modeling consistently improves interaction quality and Theory-of-Mind performance over base agentic systems, showing that structured partner modeling is a useful inductive bias for general multi-agent systems. Our code is publicly available at https://github.com/hananshafi/Mental-Models

LLM agents increasingly improve at inference time by accumulating experience in context rather than by updating parameters. This process is often described as in-context reinforcement learning (ICRL). Whether in-context learning (ICL) can actually play the role of RL, however, has not been tested. We study this question in its simplest form, direct ICRL, where the model conditions directly on raw trajectory-reward pairs, and ask whether the reward acts as a learning signal. Through controlled experiments on four benchmarks across six models, we find that the reward is read, but its effect is small: flipping, randomizing, or removing the reward leaves the improvement curve almost unchanged, and this holds even under meta-prompts that explicitly instruct the model to explore, exploit, or reason over rewards. Trajectories drive improvement, but not through their semantic content: shuffled or corrupted trajectories work as well as real ones. These patterns closely mirror those known in ICL, suggesting that direct ICRL is better understood as a special case of ICL than as inference-time RL. This reframing has implications for agent memory design: ICL factors such as input distribution and demonstrations may matter more than RL elements such as reward shaping and exploration.

Reinforcement learning (RL) methods such as GRPO substantially improve large language model reasoning but often suffer from policy entropy collapse: the loss of sampling diversity weakens exploration and limits further improvement. Existing methods address this issue either through algorithm-level interventions, such as reward modification and entropy/KL regularization, or through token-level reweighting. We investigate a complementary perspective: entropy collapse can also be mitigated by changing which generated rollouts contribute to policy updates. Under the same sampling budget, not all rollouts contribute positively to an update, and selectively excluding some can improve learning. To address this, we propose GRPODropout: before the standard update, we use a simple strategy that selectively removes a small number of high-probability positive-advantage rollouts and recenters the retained advantages. To motivate this design, we develop a rollout-level theoretical analysis that guides method design and threshold selection. The method changes only rollout usage, and adds negligible computational overhead. Experiments show higher accuracy than original GRPO and higher actor entropy while using fewer rollout samples for updates, illustrating "less is more." This work provides insight into RL rollout usage: removing some rollouts can improve performance. Code is available at https://github.com/hexuandeng/GRPODropout/.

We introduce Conversational Voice Aesthetic Model, a speech large language model for describing the voice aesthetics of real or synthetic speech responses in natural conversational contexts. Given a context and a response speech, CVAM describes salient moments that characterize the voice and predicts nine categorical attributes spanning gender, pitch, pacing, emotion, and delivery. The key challenge lies in perceptual fields such as emotion and delivery, which are inherently subjective and lack definitive ground truth. Therefore, we collect ~10 human annotations for each of 3k real and synthetic responses derived from the CANDOR corpus. CVAM is supervised finetuned on synthesized aesthetic descriptions and labels, then optimized with Group Relative Policy Optimization on human judgments. Experiments show that CVAM better agrees with human listeners than Gemini 3.1 Pro and open-source speech LLMs, and outperforms single-human-vs.-rest agreement. Together, we demonstrate the importance of grounding voice aesthetics in human perception and propose a principled framework for human alignment.

Vision-Language (VL) reasoning requires a model to both extract relevant and accurate information from an image (visual reasoning, VR), and to infer the answer from it (language reasoning, LR). Reinforcement learning with verifiable rewards typically trains both through a single chain-of-thought with a final-answer reward. This gives every CoT token the same sequence-level advantage, failing to distinguish capability specific errors. We propose SPLIT-RL, a staged post-training approach that trains VR and LR in disjoint phases. Because a group's rollouts differ along one capability at a time, the group-relative advantage isolates it, and each phase is optimized using phase-specific reward. We further introduce Claim-Level Advantage (CLA-GRPO), which decomposes VR-phase rollouts into atomic visual claims and provides a fine-grained advantage at claim level based on visual-type group formation. Although trained in two phases, trained policy is evaluated like GRPO model, with a single CoT call at inference time. Under this protocol, SPLIT-RL improves average accuracy over GRPO by 1.4-6.1 points across Qwen3-VL models from 2B to 30B-A3B and InternVL3.5-8B. Evaluating each capability using an oracle based diagnostic shows that answer-only GRPO leaves perception unchanged, whereas SPLIT-RL improves both VR and LR.

Reinforcement learning with verifiable rewards (RLVR) has become a standard recipe for post-training vision-language models (VLMs), but it typically assumes a static training environment. As the actor improves, fixed tasks drift out of its learning frontier: many become trivial, others remain unsolvable; and the learning signal collapses. We argue that VLM post-training should evolve the visual environment alongside the actor, not just the actor itself. We propose VICO, a co-evolutionary framework in which an actor and an Environment-as-Rewriter (EnvRewriter) are trained jointly: the EnvRewriter edits verifiable image-side structures, such as scene graphs, chart tables, or protected region masks, and re-renders them to produce label-valid training samples whose difficulty is calibrated to the actor's current ability through a pass-rate-based reward. This loop continuously realigns task difficulty with actor capability without any additional human annotation. Across nine multimodal benchmarks spanning mathematical reasoning and visually grounded understanding, VICO-8B improves over its base model by up to +5.0% on out-of-domain tasks, surpasses the strongest self-evolution and text-editing co-evolution baselines by +4.3% and +8.4% respectively, and stays comparable to chart-specialized RLVR methods using 16-160 times fewer labeled samples. By shifting from human-labeled supervision to image-editing co-evolution, VICO offers a scalable path beyond static-corpus RLVR for visual reasoning.

Text-to-image diffusion models are increasingly distilled into few-step variants and being deployed to enable fast inference. However, their ability to generate harmful or undesired content poses significant safety risks. Data-driven unlearning methods suppress targeted generations by fine-tuning model weights using specialized unlearning objectives. Crucially, these objectives implicitly rely on multi-step denoising dynamics, an assumption that breaks down for few-step distilled (FSD) models, resulting in ineffective forgetting. Furthermore, performing unlearning on the non-distilled base model and subsequently re-distilling it to obtain an unlearned FSD model incurs substantial computational and time overhead, making it impractical in many settings. Hence, we address this limitation with a preference-driven unlearning framework that revisits Direct Preference Optimization (DPO) for diffusion models. We show that standard DPO and its unlearning derivatives, formulated around noise-prediction error, transfer poorly to FSD models due to their altered generation dynamics. To overcome this, we introduce a modified preference optimization formulation explicitly aligned with the few-step generation properties, enabling direct concept removal in FSD models while preserving few-step efficiency and maintaining strong retention of desirable (non-targeted) capabilities. We evaluate our framework primarily on identity and NSFW (nudity) removal tasks and also extend our method to object-level unlearning. Extensive experiments demonstrate consistent and effective forgetting, and strong retention performance, establishing our method as a practical and principled solution for unlearning in FSD models.

We study whether small LLM agents can operate effectively under explicit wall-clock time budgets by both respecting the allocated runtime and using available time productively. We evaluate Qwen3.6-27B on five competitions from MLE-Bench Lite and Qwen3-4B on Zork I (Jericho), two agentic benchmarks where additional computational time can meaningfully improve performance. In the simplest setting, where the budget is stated only in the prompt, agents fail to translate the stated budget into controlled use of time. These failures arise from gaps in time awareness, since the harness provides no timing feedback, but also because they cannot reliably anticipate the duration of actions, and do not have a learned mapping from available time to an appropriate strategy. We investigate two complementary classes of interventions: harness-based mechanisms that expose timing information and enforce deadlines, and reinforcement learning with budget-aware rewards. Injecting timing information through the harness substantially improves budget adherence for Qwen3.6-27B without measurable loss in performance, while enforcement hooks tighten adherence further. RL with GRPO achieves near-perfect budget adherence on Zork I and generalizes to held-out budgets not seen during training, but does not improve task performance over the untrained harness on MLE-Bench. Once agents are made to respect the budget, they still fail to use additional time to improve task performance. RL-trained policies learn when to stop but often fill extra time with repeated actions, and GRPO training on multiple budgets tends to collapse toward the strategy learned for the shortest budget. Our results reveal a gap between time adherence and productive time allocation, which remains a central challenge for budget-conditioned agents.

For better or worse, LLMs are by now used routinely for scientific writing.\footnote{This paper is no exception; we did use AI to assist with writing some of the sections (see Acknowledgments).} Many have noticed that recent models fill papers with unnecessary antithesis, stating over and over what the work does not do, in ways that do not contribute to its precision or quality of expression and annoy reviewers rather than impressing them. We study the construction rather than in ACL papers from 2019, ACL-style arXiv papers from 2026, and papers written by GPT models from the same titles and abstracts. Its rate in 2026 is seven times the 2019 rate, and higher still in the GPT papers. Two annotators, blind to the source, find almost no 2019 use annoying and about one in ten 2026 uses; they seldom agree on which, yet about half of 2026 papers contain a use that annoys each of them. Annoying uses present the rejected alternative less favorably than legitimate uses. Raters of preference data and open reward models favor the construction, and an instruction to be honest promotes it. We conjecture that it is a side effect of post-training on pairwise preferences, which credit a disavowal in a single response and cannot register its cost across a text.

Reconstructing 3D Computed Tomography (CT) images from a few X-ray projections is a highly ill-posed inverse problem due to the loss of volumetric information. We propose PhyDiCT, a training-free framework that integrates a differentiable Physics-based forward model, grounded in the Beer-Lambert law, with a text-conditioned Diffusion as a strong prior to reconstruct 3D lung CT images. We refer to our approach as training-free since the prior model is used without fine-tuning, and our goal is to steer the denoising procedure to generate samples consistent with X-ray observations. We guide the diffusion generation using Split Gibbs sampling to jointly optimize for projection fidelity (reward) and consistency with prior knowledge. Also, we introduce a test-time refinement step that enhances image realism and anatomical coherence. We extensively evaluate our method on publicly available 3D CT datasets using both perceptual and semantic metrics, demonstrating that it surpasses existing plug-and-play diffusion and fully trained reconstruction approaches. Our findings highlight that combining a strong generative prior with the underlying physics of image formation substantially improves reconstruction quality, e.g., 7.5% improvement on SSIM compared to full training methods. Code will be released at https://github.com/batmanlab/PhyDiCT.

Reward modeling often requires jointly representing and reasoning over multiple evaluation criteria, yet verbalizing this process token by token can incur substantial inference cost. Recent work on latent reasoning suggests that continuous states may support this computation more compactly. We introduce LatentGRM, a latent evaluation framework built on semantic chunking, compression, and reconstruction. By using the structure of rubric-guided evaluations to guide compression, LatentGRM learns compact continuous trajectories that support autonomous pairwise judgments without generating textual assessments. A separate interpreter reconstructs evaluation text from these trajectories, providing an offline view of the information retained under compression. Under matched training data and backbones, LatentGRM achieves competitive aggregate preference accuracy relative to explicit Supervised Fine-Tuning (SFT) judges at both 4B and 8B scales. Across four benchmark domains, LatentGRM-8B compresses evaluation trajectories by 8.9--9.2x and reduces total judge inference time by 6.1--7.0x at vote@5. Controlled rubric interventions show that criterion-dependent preference information is carried through the latent sequence. Together, these results demonstrate that continuous latent evaluation can substantially reduce inference cost while preserving competitive judgment quality.

Long-horizon language agents often receive supervision only from terminal task outcomes, leaving little signal for distinguishing productive intermediate behavior from stagnation or even regression. Rather than learning a separate value function or process reward model for every task, we ask whether pretrained models can recognize task progress from their existing world knowledge. We formalize this capability with a World Potential Model (WPM), a goal-conditioned evaluator of task-relative realized progress in agent contexts. In ALFWorld and ScienceWorld, off-the-shelf pretrained models substantially outperform chance at recovering realized-progress structure without task-specific evaluator fine-tuning. We further anchor these progress judgments to task-specific milestones to obtain scalar world potentials, whose temporal differences provide process-sensitive step-level credit for policy optimization. Under matched comparisons, WPM-guided optimization improves success over outcome-only GRPO across all evaluated configurations. Together, these results provide initial evidence that pretrained world knowledge can support reusable realized-progress evaluation and provide useful supervision for long-horizon agents.

Adaptive rollout methods for group-relative reinforcement learning typically allocate a fixed per-update budget across prompts. We instead study how to coordinate a finite rollout budget over the entire training horizon. We formulate this problem using a concave surrogate utility of cumulative prompt exposure and introduce CERO, an online primal dual scheduler for prompt admission and budget pacing. In our experiments, each admitted prompt receives a fixed-size response group. CERO instead adapts which prompts are selected, how often they are revisited across rounds, and how many groups are generated in each round. A compact Fenchel representation linearizes the dependence on cumulative exposure, while projected online gradient descent updates prompt-specific supporting slopes and a shared budget price using reward-variation feedback and budget deviations. We establish pathwise guarantees for the surrogate allocation objective against fixed-rate and same-path time-varying benchmarks, with explicit terms for proxy discrepancy and rate variation. Under matched training-response budgets, CERO attains the highest avg@16 macro-average on each of three backbones across five mathematical reasoning benchmarks. Mechanistic analyses link CERO's prompt choices to within-group reward contrast, while multi-seed ablations show gains from adaptive pacing over both uniform and preset spending schedules.

When reinforcement learning teaches a language model a new behavior, can we find the training rollouts that taught it? And when an attribution method says it can, how do we know the answer is real? We study both questions on online RL fine-tuning with GRPO, using a planted behavior with a known cause. We release BehaviorTrace, an open evaluation harness that combines full-gradient sketching, the planted-behavior setup, and controls for gradient magnitude, fluency, headroom, and variation across seeds and generation draws. Across three seeds on Qwen2.5-1.5B, much of the apparent attribution signal comes from confounds. A control that ranks training steps by gradient size alone, with no behavior target, reaches 4.2 to 4.5 times chance and matches or beats the best targeted estimator on two of three seeds. At saturated checkpoints, model fluency predicts the behavior label at least as well as every gradient method we compared it with. Once fluency is controlled, the per-rollout results change from seed to seed and from one generation draw to the next, so a single run cannot settle the question. One signal does hold on all three seeds. The gradient of the trigger tokens aligns with a target built where the behavior actually occurs. We turn these findings into a checklist for evaluating attribution in RL. We test existing estimators, including GAS (renormalized TracInCP) and a TRAK-style estimator, and do not propose a new one.

Audio-visual multi-segment grounding (AV-MSG) in untrimmed videos, reasoning over audio-visual evidence and predicting multiple segments for a query, is a fundamental problem but remains challenging. Visual-only models overlook complementary acoustic cues, while audio-visual models often fail to calibrate the number of events - a phenomenon we refer to as count miscalibration. We present TiTok, an audio-visual large language model (AV-LLM) that localizes an arbitrary number of temporal event segments for each query. For precise boundary prediction, we introduce the Time Token Interleaving (TTI) method, which explicitly injects special time tokens into the audio-visual stream to align input-side temporal perception with output-side temporal prediction. We further propose decoupled, multi-segment-oriented rewards for reinforcement learning, consisting of global, local, count, precision, and format rewards, optimized with Group reward-Decoupled Normalization Policy Optimization (GDPO). To assess the performance on AV-MSG, we establish a new UnAV-100-based evaluation protocol, and propose the CountF1 metric for quantifying count miscalibration that overlap metrics fail to capture. TiTok reaches 65.7 mIoU and 0.58 CountF1, achieving state-of-the-art performance. Our code is available at this link.

Large language models (LLMs) can be tuned to influence human attitudes, yet the respective contributions of successive post-training stages remain un-clear. This study examines how three successive training stages affect LLM persuasiveness: (1) misalignment through supervised fine-tuning (SFT) on conspiracy data, (2) additional persuasive SFT on argumentative data, and (3) Identity Preference Optimization (IPO), a preference-optimization method. A total of 835 participants recruited on Prolific were randomly assigned to five between-subject conditions (neutral text, conspiracy-trained model, persuasion-trained model, preference-optimized model, and GPT-4) and were exposed to texts on 10 divisive political issues, personalized from their individual profiles in all model conditions. Attitude change was measured as the difference between pre- and post-exposure positions on continuous Likert scales and analyzed with an analysis of covariance (ANCOVA). A significant condition x baseline-attitude interaction, F (4, 825) = 5.33, p < .001, indicated that training effects depended on participants' initial attitudes. Persuasive SFT produced greater attitude change than conspiracy training alone, d = 0.30, whereas IPO provided no additional benefit, d = 0.03, and GPT-4 did not differ from neutral text, d = --0.01. These results show that targeted supervised training on persuasive data increases LLM persuasiveness, whereas preference optimization yields no significant gains beyond it.

Best-of-N (BoN) is a common inference-time alignment method that selects the highest-scoring response among N samples from a reference model. Evaluating BoN policies from logged data is challenging under sample-only access because standard off-policy estimators require density ratios that depend on unavailable response likelihoods. In this paper, we propose a sample-only framework for evaluating and selecting BoN policies without access to these likelihoods. We show that the order-statistic structure of BoN allows the required density ratios to be expressed through score-rank probabilities that are estimable from samples alone. We then develop a doubly robust estimator of the BoN policy value (BoN-DR) that efficiently reuses a shared auxiliary sample pool across candidate budgets. We establish valid asymptotic inference even under reward estimator misspecification and prove the efficiency of our BoN-DR estimator. Since larger budgets can amplify errors in the score function and lead to reward overoptimization, we derive two selection rules: (i) maximizing the estimated policy value and (ii) maximizing a lower confidence bound on the improvement over the reference policy, which accounts for estimation uncertainty and provides a no-harm guarantee. Across synthetic experiments and GSM8K with multiple reference and reward models, our framework accurately estimates BoN policy values and selects effective sampling budgets.

On-policy post-training can improve large language models by learning from their own trajectories, but requires costly rollout generation. We ask whether fixed demonstrations can support competitive off-policy learning through better supervision. Our premise is that their usefulness depends not only on the training trajectories, but also on whether supervision provides informative preferences among continuations and connects this guidance to the reasoning decision being learned. We introduce Step-Aligned Privileged Distillation (SAPD), a rollout-free self-distillation method that turns demonstrations into step-aligned distributional supervision. Its key insight is to use the known progression of a reference solution to associate each reasoning transition with targeted privileged guidance, rather than treating the solution as undifferentiated context. On mathematical reasoning benchmarks, SAPD outperforms supervised fine-tuning and label smoothing on average while remaining competitive with on-policy reinforcement learning and self-distillation. Analyses support both the value of context-dependent distributional guidance and the benefit of aligning privileged information with the current step. SAPD also largely preserves out-of-domain coding performance and achieves approximately 2x training-loop speedups over the on-policy baselines. These findings suggest that carefully constructed supervision can make fully off-policy post-training a competitive and computationally efficient alternative. Our code is available at https://github.com/Miaow-Lab/SAPD.