Generative RL 进展

An estimated 1 billion people worldwide live with vision impairment, yet current vision-language models (VLMs) produce descriptions too vague for safe navigation by blind and low-vision (BLV) users. Large VLMs can generate high-quality audio-description-compliant narrations but cannot run on mobile devices; small VLMs offer competitive latency but lack spatial detail, directional cues, and hazard awareness for navigational assistance. We present Smol-VL-BLV, a compact VLM for blind and low-vision users that closes this gap using a 500M decoder transformer model and two post-training mechanisms: (1) teacher-student distillation and (2) Group Relative Policy Optimization (GRPO) with a composite BLV reward targeting directional language, metric distances, and hazard detection. Because multi-stage post-training can induce catastrophic forgetting, we add a lightweight finetuning stage after the last stage GRPO finetuning to recover general descriptive quality while preserving BLV-specific spatial grounding. Our best model substantially outperforms the baseline across various benchmarks, including tasks: VQA, BLV captioning, OCR, and latency. Compared with the baseline for relative improvement, it improves the Spatial score gain of 19.3%, and the Social score gain of 14.8%. It also increases OCR-Bench by 101.5%, and raises TextVQA accuracy by 44.2%. These results show that BLV-focused post-training improves both accessibility-specific spatial grounding and general visual-text reasoning. Deployed on a mid-range Android smartphone via Mixed-Precision Quantization, the model remains approx. 450 MB and runs entirely on-device, offline and without network dependency, generating descriptions with latency dependent on host hardware capabilities. Our model, dataset, and code is publicly released at https://smol-vl-blv.github.io/Smol-VL-BLV-website/

Test-time scaling with parallel branches is widely adopted to improve performance on challenging reasoning tasks. The predominant approach, repeated sampling, draws branches independently from a single policy, which can produce redundant attempts and thereby limit the gains from additional inference compute. To address this limitation, we propose Planned Test-Time Scaling (PTTS), which replaces independent sampling with a coordinated joint policy: a planner generates a solution outline for each branch, steering the branches toward distinct reasoning paths, and an executor produces a full solution conditioned on each outline. Formally, we show that PTTS strictly generalizes repeated sampling and, in a stylized setting, provably promotes coverage of complementary reasoning modes and yields better pass@k scaling. We instantiate PTTS on top of strong reasoning models, keeping them fixed as executors while replacing repeated sampling with PTTS inference to further enhance test-time scaling. Concretely, we develop two variants: PTTS-ZS prompts a model to jointly generate outlines for all branches in a single autoregressive pass, while PTTS-RL directly optimizes the planner against the pass@k reward using truncated execution rollouts for efficient training and a sharper reward signal. Across five mathematical reasoning benchmarks with Qwen3-1.7B and 4B, PTTS-ZS improves pass@64 over repeated sampling by up to 6.7 points, while PTTS-RL further increases the gain to up to 13.4 points. Further analysis indicates that broader coverage of distinct reasoning paths contributes to these gains. Overall, PTTS provides a general framework for improving test-time scaling by coordinating reasoning branches, with zero-shot and trainable instantiations that yield substantial performance gains.

Diffusion Large Language Models (dLLMs) have emerged as an efficient alternative to autoregressive models, yet aligning them via Reinforcement Learning (RL) requires likelihood surrogates estimated from masked reconstruction subproblems under a small Monte Carlo budget per rollout. Existing methods construct these subproblems by uniform random masking, leaving open the question of which subproblems to prioritize. We identify a systematic upstream/downstream structure in dLLM rollouts. Some tokens, when revealed, trigger large confidence changes in nearby undecoded positions; we call them upstream. Others induce only small local changes and are therefore downstream. We find masking downstream tokens yields substantially better-posed subproblems than masking upstream tokens, a phenomenon we term subproblem difficulty asymmetry. Based on the observation, we propose Informed Masking (IM), which derives a per-token priority score from the denoising trajectory at zero extra inference cost and biases mask sampling toward downstream tokens. IM is plug-and-play: when plugged into three state-of-the-art dLLM RL methods on LLaDA-8B-Instruct, it delivers up to 2.01%, 8.68%, and 5.77% relative average gains on math and planning benchmarks with improved training stability.

In LLM-based speech translation, transcription-based chain-of-thought (CoT) suffers from a mismatch between reference transcripts used in supervised fine-tuning (SFT) and model-generated transcripts at inference. To address this, we propose joint recognition and translation fine-tuning via group relative policy optimization (GRPO). We score both transcripts and translations, with translation conditioned on model-generated transcripts, and compare three token advantage strategies. Using Qwen2.5-Omni-3B across four languages, we evaluate CoT against direct speech translation (Direct ST) under SFT and GRPO, training on CoVoST 2 and testing on CoVoST 2 and FLEURS. CoT GRPO outperforms Direct ST GRPO by 1.77 and 0.83 average BLEU points on CoVoST 2 and FLEURS. Compared to CoT SFT, GRPO boosts BLEU by 0.82 and 0.67 points and reduces word error rate (WER) by 8.8% and 7.2% relatively. These results highlight reinforcement fine-tuning as an effective method to mitigate the training-inference mismatch, jointly improving recognition and translation.

Hindsight relabeling which retroactively replacing a transition's goal with the outcome the agent actually achieved is an effective tool for improving sample-efficiency in Reinforcement Learning (RL). A natural extension to preference-conditioned multi-objective RL (MORL) relabels transitions with the preference direction the agent achieved rather than the one asked for. We show that this extension is frequently harmful: across four preference-conditioned off-policy algorithms spanning two critic backbones and two preference-sampling schemes on the continuous-control MO-Gymnasium suite, it degrades 19 of 36 algorithm-environment settings by as much as four standard deviations, improves only one, and leaves the rest unaffected. The harm is not a symptom of noisy relabels; denoising the target recovers almost nothing, and neither prioritized sampling nor any buffer-structural choice reproduces it. Instead, repeated relabeling collapses the critic's coverage onto whatever narrow region of the preference space the agent happened to visit. We name this failure mode Preference Coverage Collapse, and quantify it with abandoned preference mass (APM), a value-aware statistic that tracks the harm ($ρ= -0.73$) where a purely structural coverage count does not. We then introduce her_mix, a single-parameter convex combination pulling the achieved direction back towards the requested preference. At one fixed value across every algorithm and environment, it returns 16 of the 19 harmed settings to baseline, preserves and even improves the one setting in which relabeling helps, and cuts abandoned preference mass from $69%$ to $6%$. Protecting coverage over the preference simplex, not filtering noisy relabels, is what makes hindsight relabeling safe for MORL.

Reinforcement learning (RL)-based post-training has become an effective approach for eliciting reasoning capabilities in large language models (LLMs). However, adapting post-RL models to new knowledge domains or behaviors through subsequent supervised fine-tuning (SFT) can severely overwrite these capabilities. Existing approaches mitigate such forgetting through experience replay, specialized initialization, or constrained optimization using gradient projection, but either provide limited preservation or incur substantial training overhead. Our analysis shows that reasoning activations concentrate in low-dimensional subspaces, leaving substantial null-space capacity for adaptation, and that the corresponding approximate null spaces can be reliably estimated from a modest number of examples. Motivated by these observations, we propose Null-Basis Low-Rank Adaptation (NB-LoRA), a parameter-efficient method for adapting post-RL LLMs while preserving their acquired reasoning ability. We formulate reasoning retention as a layer-wise hidden-state preservation constraint and construct a fixed approximate null basis from reasoning activations. LoRA updates are then reparameterized through this basis, enforcing the preservation constraint throughout fine-tuning. Extensive experiments across multiple RL-trained LLMs and diverse downstream tasks show that NB-LoRA matches standard LoRA in adaptation performance, maintains reasoning accuracy near pre-fine-tuning levels, and generalizes this preservation to held-out reasoning benchmarks.

Group Relative Policy Optimization (GRPO) and related policy-gradient methods for training language model agents collapse an entire multi-turn rollout into a single scalar trajectory reward before it enters the policy update. When the task composes distinct skills, especially under sparse and delayed environmental feedback, this collapsing is lossy: the optimizer must implicitly infer which competency drove the outcome and how that should change behavior. We argue the right primitive is not a better scalar but a decomposition: trajectory reward should be split along subtasks before it enters the policy update. We introduce Reinforcement Learning with Decomposed Subtasks (RLDS), whose core is Subtask-Decomposed Advantage Estimation (SDAE): a replacement for the scalar GRPO advantage that splits trajectory reward into per-subtask shares on a fixed taxonomy, computes a group-relative advantage per subtask, and distributes per-token credit by weighting each subtask's advantage by its importance, concentrating it around the step where a reflection marks that subtask's execution as consequential. We evaluate on four agentic benchmarks: FrozenLake (sparse grid navigation), HotpotQA (multi-hop QA, one retrieval tool), ScienceWorld (long-horizon embodied science), and DeepResearch (long-form research, four tools, composite rubric reward). Heterogeneity diagnostics emitted during training show where decomposition pays off - gains scale with subtask heterogeneity, largest on the high-heterogeneity tasks ScienceWorld (+11.5 points, paired-bootstrap 95% CI [+9.8, +13.3]) and FrozenLake (+9.8 points, [+7.0, +12.8]), and within noise on HotpotQA and DeepResearch, where the diagnostics predicted little to recover. ScienceWorld is also more compute-efficient under RLDS than scalar GRPO (-10.9% wall-clock per step), as long rollouts amortize the fixed reflect-and-grade overhead.

Reward fine-tuning aims to update a pre-trained flow-based generative model to improve the downstream reward of its generated samples. Existing methods typically formulate this problem as sampling from a reward-tilted distribution, the solution to a KL-regularized reward-maximization problem. Here, we introduce an optimal transport regularizer built directly from the pre-trained drift. Unlike KL reward tilting, the resulting objective transports individual samples toward higher reward rather than reweighting the base distribution. We show that the resulting problem is equivalent to a deterministic optimal control problem on the flow. Given a pre-trained flow map, this equivalence yields a simulation-free reinforcement learning algorithm for fine-tuning generative flows. We call the resulting framework Wasserstein-Tilted Flow Maps (WTF), the first end-to-end fine-tuning recipe native to flow maps. The output is a fine-tuned flow map that retains strong reward-aligned performance at few-step inference budgets without post-hoc distillation. Experiments on ImageNet-256 and text-to-image show that WTF achieves higher reward with comparable or higher diversity than baselines, while requiring up to $280\times$ less training compute. More broadly, we argue that accelerated samplers such as flow maps are essential infrastructure for efficient post-training, and that the dominant KL-regularized formulation is only one of many choices worth revisiting.

While Large Language Models (LLMs) have achieved remarkable results across various benchmarks, their alignment with normative values often results in homogenized responses that fail to address diverse user preferences. Existing training-free methods often occupy valuable context windows through prompt engineering, while training-based methods typically remain static post-training, failing to support the continual optimization required in real-world settings. To address these challenges, we propose COPE (Continual Optimization with Personalized embedding and self-Evaluation), a novel optimization framework tailored for real-world-motivated interaction settings with sparse user feedback. Our framework assigns learnable personalized embeddings to each user and synergistically integrates preference capture, self-evaluation calibration, and personalized response optimization within a single update step. A key innovation of our method is the use of self-evaluation to generate proxy rewards, enabling continuous model updates even when explicit user feedback is unavailable. Experiments show that COPE consistently outperforms strong training-free and training-based baselines under sparse feedback, and remains complementary to Retrieval-Augmented Prompting (RAP). Further analyses confirm COPE's reliable self-evaluation, meaningful preference patterns, stable general capabilities, and robustness under shifting preferences and alternative evaluators.

Large language models increasingly tackle hard reasoning problems by spending more test-time compute, yet the dominant strategy remains naive repeated sampling: draw many independent solutions and hope one is correct. Because such sampling explores only through local decoding noise, it tends to produce many near duplicate attempts rather than genuinely different ideas. We ask whether exploration can instead be steered at a semantic level, by first sampling problem specific concepts, hints, or strategies and then conditioning answer generation on them. We refine this into a simple, more exploratory procedure that emits many diverse concepts in a single trajectory, and evaluate it on hard problems where repeated sampling struggles. We then go a step further and make concept generation trainable: a small concept generator is optimized with reinforcement learning so that its concepts maximize the downstream success of a larger, frozen answer generator. On hard mathematical reasoning problems, the trained concept generator substantially improves the answer generator's pass@k over naive repeated sampling at the same answer generation allocation, surpasses concepts drawn from much larger untuned models, and transfers to answer generators it was never trained against, including a model from a different family. A small model can thus be trained into an effective, reusable search policy for a much larger one.

HopChain has shown on still images that multi-hop data synthesis improves vision-language reasoning, because long chain-of-thought reasoning exposes errors that compound across steps, while most data used for reinforcement learning with verifiable rewards (RLVR) rarely demands a chain of visual evidence, so these weaknesses are likely to stay unexposed. We observe the same problem in video, where this framework has not yet been explored. We therefore build Video-HopChain, a dataset of 22,550 multi-hop video questions over 13,378 videos, together with a held-out benchmark of 1,000 questions. Each question chains three to six yes/no questions about moments in one video, and each yields one of two integers depending on its answer. The final answer is the sum of these integers, so an exact match on that sum gives the verifiable reward that RLVR needs. We first train Qwen3-VL-8B with GRPO on a standard video dataset, and a second stage on Video-HopChain then raises the mean over eight video understanding and reasoning benchmarks from 55.4 to 57.9 and improves every one of them. Training on such a dataset, however, exposes a known limitation of GRPO: its learning signal comes from the reward variance within a group, so hard questions whose rollouts are all incorrect and easy questions whose rollouts are all correct both leave the group with no gradient. To recover these groups at the same compute budget, we introduce Confidence-Gated Exploration (CGE). With 8 rollouts per question, CGE samples the first 4 as usual. If these 4 are either all correct or all incorrect, it samples the last 4 with the policy's most confident token masked inside the reasoning span, and removes the masked positions from the loss while all 8 rollouts enter the advantage. With CGE, the mean rises further to 59.3. We release the dataset, the checkpoint, and the data generation and training code.

Offline reinforcement learning and off-policy evaluation evaluates dynamic treatment rules based on retrospectively collected data prior to deployment. In recent AI applications, state and reward information is recorded as complex text or image, which recent AI advancements such as LLM-as-a-judge can label with unknown bias. Expert annotation may be available but at a higher cost. For example, safety classification via cheap but imperfect classifiers vs. expensive expert review. We show how a limited budget for ground-truth data-annotation can be used via doubly-robust OPE with missing rewards, and we optimize variance-optimal annotation probabilities for sequential off-policy evaluation, where the target policy value is estimated from annotated data. We characterize the optimal annotation probabilities for sequential forward-monotone annotation protocols, and provide a feasible batch-adaptive implementation. Our work is motivated by a collaboration with a homelessness services nonprofit that writes casenotes for individuals over time. Our method can be used to unlock trustworthy inference from casenote data and answer new inferential questions such as: how does expanding outreach effort over time affect progress towards a housing application and improvement in housing placement? In simulations and on two real datasets - casenotes from the nonprofit and human-preference votes from LMArena - we see reductions in RMSE of 34-65% for housing placement and 17-68% for progress towards a housing application at budgets of 40% of full annotation and above, and by 55-62% at every budget on LMArena.

Applying large language models to Traditional Chinese Medicine (TCM) prescription generation reveals three clinically critical gaps: models produce end-to-end mappings without auditable reasoning following the li-fa-fang-yao paradigm (SR Gap), treat each encounter in isolation without follow-up adjustment via sui zheng jia jian (LA Gap), and fail to enforce absolute contraindication rules such as Shi Ba Fan (SC Gap). We propose a progressive four-stage framework (SFT $\to$ PG-CoT $\to$ Dynamic $\to$ K-RL) that addresses each gap: PG-CoT constrains CoT distillation under the li-fa-fang-yao paradigm to produce auditable diagnostic chains, Dynamic SFT models patient trajectories with explicit transition reasoning, and K-RL encodes deterministic pharmacological rules as rule-based DPO preference signals. Across 12 fine-tuned models and 6 zero-shot baselines, our framework substantially improves prescription quality over zero-shot baselines---with a 7B model (Mistral-7B) surpassing zero-shot GPT-5 on all three TCM evaluation metrics.

Omni-modal models have expanded multimodal interaction across vision, audio, speech, and language. However, their training is predominantly organized around semantic descriptions and general-purpose objectives, leaving physical attributes, interaction states, and causal mechanisms only partially specified. This gap is not simply a matter of modality coverage: adding more modalities does not by itself provide the supervision needed to connect observations with the physical structure of the world. We present OmniFysics-Nano-V2, a compact omni-modal model for physical-world perception and understanding. The model supports image, video, audio, speech, and text inputs within a shared reasoning framework, together with text and speech generation. To address the lack of explicit physical supervision, we construct a dual-branch physics-aware data pipeline that grounds salient objects in structured physical attributes and aligns visual changes with acoustic events, intermediate responses, and interaction outcomes. To address homogeneous training objectives, we curate reinforcement-learning prompts by reward diversity and adopt a two-stage Group Relative Policy Optimization curriculum that progresses from general task correctness to fine-grained physical perceptual reasoning. Experiments across multimodal, audio-visual, and physical reasoning benchmarks show that the proposed data and training strategy improves physical-world understanding while preserving broad omni-modal competence. The proposed model achieves leading result on 17 of 21 benchmarks against SOTA omni-modal models. By equipping AI systems with both omni-modal and physical-world perception capabilities, OmniFysics-Nano-V2 is poised to become a cornerstone of next-generation Physical AI.

Collaboration topology shapes both the performance and execution cost of LLM-based multi-agent systems. Because tasks differ in complexity and required capabilities, recent approaches generate task-specific collaboration graphs that specify agent participation and information flow. However, representative topology generators use either individual agents or predefined groups throughout an organization, overlooking differing collaboration needs across subtasks. Our key insight is to select granularity locally for each functional role, combining fine-grained control with reusable collaboration patterns within one organization. Learning such organizations requires exploring a combinatorial construction space with limited intermediate feedback from final-answer rewards. Therefore, we propose MAGIC, a dense-reward reinforcement learning framework for mixed-granularity graph generation. Specifically, MAGIC constructs a mixed-granularity agent graph by sequentially selecting a functional role, instantiating it as a single agent or reusable group, and connecting it to existing units. We directly optimize the construction policy using returns from trajectories sampled under the current policy and use potential-based reward shaping to provide intermediate feedback from probe-based utility and structural signals while preserving the cumulative task reward. MAGIC outperforms state-of-the-art baselines across eight benchmarks and demonstrates strong inference efficiency in our efficiency study.

Large language models (LLMs) offer great potential to automate a broad range of telecom engineering tasks by reasoning over standards, network configurations, mathematical models, source code, and operational logs. However, existing telecom LLMs struggle to reliably reason across these diverse tasks and data types. General-purpose LLMs often lack reliable grounding in telecom-specific knowledge, while telecom-specialized models are typically developed for narrower task families and exhibit limited multi-task performance. To fill this gap, we introduce TelecomGPT-R1, a family of open source unified telecom reasoning models structured around four complementary axes: protocol, knowledge, modeling, and fault. We first develop an axis-aware data generation framework that refines coarse public telecom artifacts into verified question-answer pairs and high quality chain-of-thought (CoT) reasoning trajectories, yielding a training corpus containing 104,880 examples. Building on this corpus, supervised fine-tuning (SFT) instills telecom knowledge and evidence-grounded reasoning patterns to overcome the cold start barrier for reinforcement learning (RL). We then apply dynamic sampling policy optimization (DAPO) with task-routed rubric rewards to keep RL updates informative and stable across heterogeneous telecom reasoning tasks. These rewards decompose axis-specific CoT traces into verifiable reasoning units and combine grounded dense process credit with outcome correctness, allowing RL to learn generalizable problem solving behaviors from verifiable telecom evidence. We release the TelecomGPT-R1 models and a reproducible training recipe to support further community development. Evaluations on seven benchmarks of the GSMA Open Telco Leaderboard show that the open-source TelecomGPT-R1-27B achieves an 89.64% mean score, outperforming leading proprietary models, including GPT-5, Claude, and Gemini.

Reasoning-oriented reinforcement learning enables large language models to solve mathematical, coding, and other multi-step tasks, but shifts a substantial portion of the training cost to rollout, where trajectories are generated for policy updates. Efficient rollout mechanisms are therefore essential to reduce this cost while maintaining the freshness, consistency, and statistical validity of training data. This survey provides a systematic taxonomy of recent research on rollout efficiency for reasoning-oriented reinforcement learning, classifying existing approaches from both mechanism and bottleneck perspectives. Based on this taxonomy, we analyze how different technique families address distinct sources of rollout inefficiency, examine opportunities and potential conflicts for combining them, identify gaps in the evaluation and reporting of efficiency gains, and discuss open challenges and future research directions.

Large Audio Language Models (LALMs) can follow diverse instructions to synthesize speech in specified styles. However, complex instructions that require simultaneous control over pitch dynamics, speaking rate, and emotional tone often exceed what a single-pass generation can faithfully realize. While recent reasoning models have shown that intermediate "thinking" tokens improve output quality, this paradigm has been confined to the text modality. In this work, we extend reasoning to the audio token space by training a LALM with reinforcement learning to reason over its own speech output. The model first generates a draft speech as a form of audio-token reasoning, critiques its own generation by reflecting on the acoustic realization in text, and then produces a refined version conditioned on both the first-pass speech and the critique, all within a single model. After RL training, the refined two-hop outputs achieve a relative improvement of 7.15% on the InstructTTSEval benchmark, demonstrating the model's reflective ability.

RLVR has substantially improved the reasoning capabilities of LLMs. However, existing methods typically parameterize temporal progression in the Markov Decision Process by token-by-token generation, despite the highly non-uniform information flow along autoregressive trajectories. In this paper, we propose InfoPPO, which reparameterizes temporal progression using information density rather than raw token count. This reparameterization induces a common state-dependent structure for both temporal credit propagation and policy updates. InfoPPO restores the effectiveness of non-trivial discounting in long-horizon reasoning, retaining effective-horizon contraction while avoiding excessive attenuation of terminal supervision over long token sequences. Moreover, the information-time policy-improvement analysis naturally leads to a state-dependent update constraint, which we implement through adaptive clipping. By adapting the clipping threshold at each token position to the information density of its corresponding state, this mechanism enables more targeted policy updates while preserving proximal control. Theoretically, we extend performance-difference and policy-improvement analyses to the information-time MDP, deriving a policy-improvement lower bound when policy changes are regulated by information density. We further connect the general information-time analysis to practical LLM policy optimization by relating state-wise information density to local policy movement, while also providing theoretical grounding for the adaptive update mechanism. Experiments on Qwen3 models demonstrate consistent gains over competitive baselines across five challenging competition-style mathematical reasoning benchmarks. InfoPPO also maintains stable accuracy and response length across non-trivial discount settings under which token-time PPO deteriorates.

Generating Scalable Vector Graphics (SVG) code from natural-language instructions is an open-ended task without absolute visual ground truth, leaving both evaluation and policy optimization without a faithful signal. Scalar metrics (CLIP, Aesthetic) calibrated on natural images transfer poorly to stylized vector content, and reusing them as RL rewards triggers reward hacking. We address both limitations with rubric-based scoring. We first establish empirically that prompting a vision-language judge with a multi-axis rubric correlates with human judgments far better than scalar metrics, both across samples and within instructions. Building on this finding, we introduce RULER (Instance-aware Rubric Rewards for Reinforcement Learning), which converts each instruction into an instance-aware rubric of six items spanning semantic, visual, and stylistic axes; a judge VLM scores rendered rollouts item-by-item, and the weighted satisfactions form a fine-grained reward optimized via Group Relative Policy Optimization. Because the rubric is derived from text alone, RULER requires neither paired SVG ground truth nor human preference labels. On MMSVG-Illustration and MMSVG-Icon, RULER lifts the rubric score from 0.432/0.395 to 0.693/0.683, surpassing dedicated SVG specialists and matching the substantially larger DeepSeek-V3, with ablations identifying rubric design as the active lever for RL on open-ended SVG generation. The project page is available at https://hangyuran.github.io/RULER/.

Recent advances in video generative models have enabled high-fidelity, temporally coherent video generation. However, these models often struggle to satisfy prompts requiring specialized knowledge, specific identities, physical consistency, or ordered events. In this paper, we present VideoGen-Agent, a multimodal agent trained through multitask agentic reinforcement learning to use external tools for video generation. The agent coordinates augmentation, generation, and verification tools through multi-turn interactions, using the prompt and intermediate observations to guide its decisions. We train a shared policy on a category-balanced dataset spanning six tasks. Supervised fine-tuning on teacher-generated trajectories establishes tool-use behavior, which is then refined through reinforcement learning. A category-aware hybrid reward evaluates tool-call validity, task-appropriate tool use, and generated video quality. We further introduce VABench, a held-out benchmark of 600 prompts covering procedural knowledge, single- and multi-entity identity preservation, physical consistency, scene composition, and multi-shot temporal structure. On VABench, VideoGen-Agent improves over its base text-to-video generator by 19.1 points, from 56.5 to 75.6. Upgrading the generation tools further raises the score to 86.1 without additional agent training. Human raters prefer the upgraded configuration over the strongest standalone baseline in 84.3% of comparisons. These results support learning tool use across video-generation tasks and show that the trained agent can benefit from subsequent advances in generation tools. Project page: https://andyca111.github.io/VideoGen_Agent/

Building strong chart-to-code systems increasingly relies on reinforcement learning, whose effectiveness depends critically on the quality of the reward signal. Large Multimodal Models (LMMs) play a natural critical role in jointly assessing chart visual appearance and task requirements. They are therefore increasingly used as visual critics and reward models, yet their reliability as judges remains largely unexplored. To this end, we introduce ChartJudgeBench, a diagnostic vision-language benchmark for assessing LMM judges in chart-to-code workflows. It includes 1,003 Chart Perception Alignment (CPA) instances for pairwise chart comparison and 650 Chart Reasoning Judgment (CRJ) instances for binary Accept/Reject verification in Chart Reproduction and Chart Editing. Together, these tasks emulate the core judging decisions required in agentic refinement and RL-based chart optimization. Our evaluation of strong LMMs reveals four systematic limitations: (i) positional bias in pairwise comparison, (ii) a strong tendency to overpredict Accept, (iii) difficulty in matching visual styles and aesthetics, and (iv) an unexpected leniency bias in RL-trained models. These findings show that current LMM judges require explicit reliability validation before being used as critics or reward models in chart-to-code optimization. The code and data are available on ChartJudgeBench.

Reinforcement learning (RL) improves reasoning in vision-language models (VLMs) but can induce chain-of-thought (CoT) obfuscation: an operational, non-intentional outcome where task reward or accuracy rises while traces become less grounded and monitorable. Prior work largely documents this decay behaviorally, leaving its representation-level correlates and actionable controls unclear. We find that template- and ground-associated activations become less separable during RL; matched interventions support the contribution of selected features to monitorability degradation. Guided by this evidence, we propose Targeted Anti-obfuscation with Mechanistic Enforcement (TAME), which uses Sparse Autoencoders (SAEs) to combine behavioral feedback with targeted suppression of template-associated activations during RL. Its asymmetric constraint penalizes template activations only above their pre-RL baseline, anchoring the localized features while behavioral feedback promotes grounded refinements. Across VIRL-39k, SPA-VL, and two model families, TAME improves CoT monitorability by up to 30.9 and 16.7 percentage points over Group Relative Policy Optimization (GRPO), respectively. Blinded human evaluation finds higher human monitorability on both datasets, and two held-out monitor families reproduce the monitorability gains. Task accuracy changes are small and mixed, and general-capability benchmarks show task-specific trade-offs. These results provide a path from behavioral monitoring to representation-level oversight for more auditable RL-trained multimodal systems.

We present MimicAgent, a prompt-to-trajectory generation framework for learning dynamic quadruped skills. Although reward shaping is extensively used when training quadruped policies, navigating the resulting reward landscape is notoriously difficult, requiring hours of "graduate student descent". Eureka attempts to automate reward design with LLMs, but we find that it struggles to generalize across diverse skills and morphologies. Our key observation is that it is far easier for a human - and by association, an LLM - to generate reference motions than to shape reward functions. Our hypothesis is motivated by the success of example-guided RL for humanoids, which exploits large-scale motion capture datasets as references for training locomotion policies. Unlike humanoids, quadrupeds lack such reference motion data. Towards this end, we propose MimicAgent, an agentic harness that, given a skill prompt, generates quadruped reference trajectories with coding agents. These coarse reference trajectories are then used to train example-guided RL policies that are deployable in simulation and in the real-world. Notably, we find that when prompting Claude Fable 5.1 within our agentic harness, 87% of prompts yield semantically aligned reference trajectories.

Speculative decoding accelerates large language model (LLM) inference by using a lightweight draft model to generate multiple candidate tokens that are verified by the target model in a single forward pass. Its speedup is largely determined by the acceptance length, yet existing draft-model training methods mainly optimize cross-entropy or Kullback-Leibler (KL) divergence as proxies. These objectives encourage distribution matching but do not directly optimize acceptance length, and the acceptance mechanism also differs between greedy and sampling-based decoding. In this work, we propose acceptance-length-aware training losses that directly optimize the expected number of accepted tokens within a speculative window. For greedy verification, we derive an expected accepted length (EAL) loss that explicitly maximizes expected acceptance length. For sampling-based decoding, we introduce a window total variation (WTV) loss that optimizes the overlap between temperature-scaled draft and target distributions while accounting for sequential acceptance dependencies. Both objectives can be further combined with a group-relative reinforcement learning stage (GRPO) using simulated acceptance length as the reward. Experiments across different target and draft models, tasks, and decoding settings show that our losses consistently improve acceptance length over KL-based training. WTV provides particularly strong gains under sampling-based decoding, while EAL better matches greedy verification. These results show that directly optimizing the acceptance objective, with losses tailored to the decoding mode, is more effective than conventional distribution-matching objectives.

Reinforcement Learning with Verifiable Rewards (RLVR) is expanding from tasks with well-defined correctness signals, such as mathematics and code, toward multifaceted quality requirements specified by multi-dimensional rubrics. Since policy optimization consumes one scalar per rollout, rubric-based pipelines must map multiple criterion scores into a scalar reward. This aggregation is often treated as score scaling, but it implicitly determines how quality dimensions trade off during training. The prevailing practice, normalizing each criterion and taking a linear combination, assumes that cardinal score differences are comparable across criteria and that gains on one criterion compensate for failures on another; both assumptions are unreliable when criteria are semantically heterogeneous. We propose Reinforcement Learning with Verifiable Rubric-based Ranking (RLVR$^2$), a verifiable ranking paradigm for rubric-based RLVR. For each criterion, RLVR$^2$ converts rubric scores into criterion-specific within-group ordinal outcomes, recovers a latent utility from the resulting comparison matrix, and merges these utilities into one training signal. By retaining only within-group ordering and discarding raw score magnitudes, RLVR$^2$ avoids calibrating heterogeneous rubric scales. It further supports objective-preserving attribute adjustment: auxiliary attributes that correlate with observed rankings but are not training objectives can enter the estimation without expanding the rubric or rewarding them directly. Across three model scales and 16 benchmarks, RLVR$^2$ consistently outperforms representative rubric-based baselines, achieving the best overall performance on most benchmarks at every scale. Analysis shows it controls systematic effects tied to reasoning efficiency and response formatting while preserving the quality objective.

Constraint-following music generation asks a score to satisfy several user-specified properties at once, each checkable programmatically (key, meter, length, range, final note, rhythm, motion and form), yet no existing benchmark isolates this capability. We construct MusicConstraintBench, 2,180 items over eight constraint families, on which current models fail once a few constraints are combined. The natural remedy is reinforcement learning with these verifiers as reward, yet we observe that a reward paid only when every property holds leaves most training groups without a learning signal: over the first 50 updates, 0.550 of rollout groups score identically and receive no gradient, even though a failing score typically misses only one requested property. Under the joint criterion, rollouts for a prompt tend to fail together, so a binary reward cannot separate a nearly correct score from a malformed one. We therefore introduce MusicRLVR, which pays graded per-property credit behind a hard validation gate that rejects malformed outputs, plus a joint-satisfaction bonus, requiring no human annotation, learned reward model, or music-domain fine-tuning. On MusicConstraintBench, MusicRLVR lifts Qwen3-4B-Instruct from 0.160 to 0.807 on mixed constraints and leads every zero-shot baseline including Llama-3.1-70B at 0.380. It also generalises to property combinations unseen in training and to out-of-range parameter values, showing that verifiable rewards need not presuppose a target output.

Large audio-language models (LALMs) are increasingly used for a broader range of audio reasoning tasks. These models typically incorporate audio representations into a large language model (LLM) backbone to enable multimodal reasoning. Recent test-time reinforcement learning (TTRL) methods further improve LLM reasoning capability by leveraging unlabelled test data after pre-training. However, the importance of the perceptual capability of LALMs remains underexplored, particularly how much acoustic evidence is integrated and relied upon during reasoning, and how this contributes to final task performance. This gap limits the development of effective post-training methods like TTRL for audio reasoning. In this work, we first analyse how audio information is integrated and utilised during reasoning process. We quantify layer-wise perceptual reliance and show that stronger acoustic reliance is associated with higher accuracy and a larger performance gain attributable to the audio input. Building on this, we propose Perception-Grounded TTRL (PG-TTRL), which aligns label-free test-time optimisation with perceptually grounded reasoning, encouraging the model to structure its reasoning more strongly on the audio input. Experiments across LALMs and benchmarks show that PG-TTRL consistently improves reasoning performance over both the base models and standard TTRL, showing the value of perceptual-grounding optimisation for test-time audio reasoning.

While test-time scaling enhances Large Language Model (LLM) agents in long-horizon software engineering (SWE), sparse binary rewards (Pass/Fail) create a severe credit assignment crisis and waste failed exploratory trajectories. Current trajectory optimization and scaling methods are costly and structurally limited, relying on heuristic state reuse without causal diagnosis or delayed scalar scoring without actionable online guidance. We propose FLARE (Full-Lifecycle Alignment and Reward Engine), a novel dense supervision paradigm driven by a lightweight Generative Reward Model (GRM). First, RADAR, an offline causal-aware diagnostic framework, extracts high-fidelity, hindsight-free supervision through causal-chain backtracking to distill a GRM providing real-time, step-level risk feedback. Second, FLARE uses this GRM to continuously optimize the agent across its entire lifecycle. During inference, FLARE acts as an Active Scaffold, autonomously intercepting high-risk generation steps for localized breakpoint re-execution, drastically reducing compute overhead. During post-training, the GRM's structured signals serve as process-supervised reranking scores for Supervised Fine-Tuning (SFT) and step-level dense rewards for Reinforcement Learning (RL), mitigating policy collapse in sparse environments. Extensive evaluations show that FLARE establishes a new Pareto frontier across the agent lifecycle: FLARE (N=1) outperforms Global Rollout (N=5) with a 5x reduction in token consumption. Extending FLARE to training overcomes the sparse reward problem in long-horizon interactive tasks, delivering relative performance gains of 19.13% in SFT through process-aware data curation and a consistent 9.19% improvement in RL.

Training a language model on data with a narrow flaw can sometimes make the model broadly misaligned. Inspecting the data at face value often does not settle whether it will emerge, and today it is caught only after training, by auditing the resulting model. To complement post-hoc audits, we introduce Alignment Forecasting: the task of predicting alignment failures before training. Given a target model, a fine-tuning dataset, and a failure mode such as deception or sycophancy, a forecaster outputs the probability that fine-tuning would meaningfully increase that failure mode. To measure progress on alignment forecasting, we introduce ALIGNMENTFORECASTBENCH, a benchmark of over 5,000 forecasting questions spanning 17 target models, 32 datasets, and 16 failure modes. Frontier models prompted directly perform poorly on ALIGNMENTFORECASTBENCH. We therefore propose a forecasting scaffold in which an LLM reads the dataset and rates how strongly and broadly it pushes the model toward misbehavior, and a simple learned model combines that rating with the failure mode's base rate and the target model's prior tendency. This forecasts well above chance, and beats a model fine-tuned on the task and a simple forecaster allowed to see how weaker models behaved after fine-tuning on the same data. Its signals also flag problematic training examples that a frontier-model classifier misses. Filtering those examples out from real post-training data such as UltraChat results in more aligned models on our multiple-choice evaluation in most cases, though the benefit in open-ended conversations is unclear. More progress is needed before forecasts can reliably guide training data curation in practice, but our results suggest that forecasting many alignment failures before training can be tractable in the SFT setting.