We present Jina-OCR-v1, an end-to-end document parsing model built to serve on low-budget GPUs. It combines the compressed-vision encoder and the 3B mixture-of-experts decoder of DeepSeek-OCR, which activates about 570M parameters per token, with a FastMTP speculative decoding head that shares a single draft block recursively across K=3 prediction steps. Greedy verification makes decoding lossless. Post-training combines instruction alignment, robustness fine-tuning on difficult documents, and GRPO under dense verifiable rewards: deterministic formula, table, and structural checks that award partial credit. The training data mixes cleaned public corpora with targeted synthetic pages. At the default dynamic-resolution setting, Jina-OCR-v1 scores 91.14 on OmniDocBench v1.6 and 83.4 on olmOCR-Bench, and reaches the highest page throughput in our comparison at 2.57 pages per second. On a low-budget GPU such as the NVIDIA L4, FastMTP doubles decoding speed over greedy autoregressive decoding. The model is publicly available at https://huggingface.co/jinaai/jina-ocr-v1.
On-policy distillation (OPD) accelerates post-training by providing dense token-level supervision from a frozen teacher on the student's own rollouts. Vanilla OPD applies this supervision uniformly across prompts, without checking whether the teacher is reliable for each prompt. Because reverse KL is mode-seeking, a confidently wrong teacher can induce a strong yet misleading update. Distributional proxies, such as entropy or teacher-student likelihood agreement, measure uncertainty or agreement but do not directly verify outcome correctness. We introduce Teacher-Gated On-Policy Distillation (TGOPD), built on the principle that teacher reliability should be verified at the prompt level before dense supervision is admitted. TGOPD estimates reliability from a small set of verifier-scored teacher probes and routes each prompt exclusively to dense OPD when the reliability check passes or to verifier-grounded GRPO otherwise. Across 4B and 35B students in mathematics, code, and instruction following, TGOPD outperforms Vanilla OPD in all six single-domain settings and achieves higher seven-benchmark averages at both scales under multi-domain training. By using otherwise-idle teacher capacity for reliability estimation, TGOPD also reduces teacher-side compute waste in asynchronous OPD, increasing teacher-node GPU utilization from 9.8% to 78.9% in the measured 4B single-domain run.
Post-training large language models usually applies a single training recipe to all samples, even though the model's own rollouts reveal different sample-level learning states. We propose Self-Routing, a behavior-conditioned post-training framework that uses rollout correctness and confidence to decide how each sample should be optimized. Depending on its behavior state, a sample is routed to GRPO, on-policy self-distillation, regularization, or skipping, allowing training to adapt without external teachers, extra annotations, or additional sampling. Experiments on mathematical reasoning across Qwen3 and Qwen3.5 backbones show that Self-Routing consistently improves over uniform GRPO, uniform OPSD, fixed mixtures, and simpler routing baselines. Further analyses show that the routing distribution changes over training and reduces unnecessary updates on low-signal or already stable samples.
Group relative policy optimization for reinforcement learning with verifiable rewards (RLVR) typically uses a fixed importance-sampling (IS) ratio clipping boundary across all rollouts. We identify a key limitation: rare correct rollouts on harder problems and abundant correct rollouts on easier problems are clipped at comparable rates, despite contributing very different learning signals. Rollouts with low group success exhibit larger IS ratios and carry stronger gradient signal for exploration and solving new problems, yet are disproportionately suppressed by fixed clipping. To address this, we propose Group Adaptive Clipping Policy Optimization (GAPO), a plug-in modification to GRPO methods that adapts the clipping boundary to the rollout advantage. GAPO is motivated by a reverse-KL trust-region perspective, which suggests that rollouts with larger learning signal should receive proportionally greater update headroom. GAPO requires no reward shaping and preserves the standard PPO/GSPO surrogate while adapting only the clipping threshold. Across Qwen and Llama models, GAPO consistently improves both Pass@1 and Pass@k over fixed clipping and advantage-shaping baselines on math reasoning and coding benchmarks where the pass rates by the base model are relatively low.
Reinforcement learning fine-tuning of large language models increasingly adopts multiple reward dimensions, including verifiable rules, task-specific evaluators, and learned reward models, to provide richer supervision across diverse capabilities. These dimensions are commonly scalarized with fixed aggregation weights. We identify a failure mode in which aggregation itself induces reward hacking: static projection aliases qualitatively different reward profiles into a single scalar, steering optimization toward whichever dimensions are easiest, densest, or systematically favored by the reward signal. Over training, this traps the policy in suboptimal profiles and prevents convergence to better-balanced ones that would yield higher task performance. To address this, we propose Adaptive Multi-Reward Projection (AMRP), a lightweight online method that reallocates aggregation weights using three signals, relative shortfall, reward volatility, and recent progress, increasing pressure on lagging, unstable, or stagnant dimensions while relieving saturated ones. Across structured reasoning, citation-grounded generation, and open-ended alignment under GRPO, AMRP consistently improves reward-profile balance and downstream performance over fixed and dynamic weighting baselines; it also remains effective with GDPO and PPO, supporting compatibility across RL algorithms. Our code is available at https://github.com/yyhappier/AMRP.git.
Outcome-based reinforcement learning provides verified feedback for language-model agents, but assigns trajectory-level advantage uniformly to all decisions, yielding coarse credit over long-horizon interactions. On-policy self-distillation offers finer supervision by re-evaluating sampled behavior with privileged information (PI) available only during training. However, fine-grained supervision is not necessarily fine-grained credit: PI-induced likelihood changes describe how additional information alters policy preference, but do not directly determine how an executable action should inherit the verified task outcome. This creates a supervision-credit gap. Privileged signals may be irrelevant to the current interaction state, operate at a token granularity misaligned with executable decisions, and lack the outcome semantics required for reinforcement. We introduce TASPO, which converts privileged supervision into outcome-grounded action credit. TASPO constructs decision-applicable PI from verified successful experience, aggregates PI-induced likelihood shifts at the executable-action level, and converts relative action support into positive, bounded, mean-preserving weights on the original trajectory advantage. Thus, the verified outcome determines the update direction and average scale, while PI only redistributes credit across actions. Across three agentic benchmarks, TASPO improves over GRPO by 10.6\% and generalizes better to unseen tasks. Further analysis indicates that TASPO reduces supervision mismatch and that action-level assignment stabilizes the policy optimization process. These findings offer the community another interesting perspective.
Shangqing Tu, Daniel Zhang-Li, Yucheng Wang +21cs.CL cs.AI
We present CogEvol, a family of models trained specifically for Learning Environment Generation: turning a course brief into a finished learning artifact (structured-JSON slides or self-contained interactive HTML pages) in a single pass. Across 220k production requests, CogEvol completes a slide in a median of 17 seconds and an interactive page in 59, replacing minutes-long multi-turn agent scaffolding. Reliability is enforced rather than hoped for: a production-grounded data pipeline turns real failures into 53,687 verified SFT samples, and a hybrid rule-plus-VLM reward drives GRPO-based RL, hardened after we caught and fixed a reward-hacking episode that produced visually convincing but unplayable games. CogEvol-27B scores 83.7 on slide quality and 63.7 on a 500-case interactive-HTML benchmark with 26.9x fewer parameters than flagship coding models, and, in collaboration with the OpenMAIC team, serves their live production traffic. CogEvol-4B is released openly under the Apache 2.0 license at https://github.com/CogEvol/CogEvol-4B; external flagships are measured on the same suites under the identical harness. Scaffold editing cuts interactive-page generation cost by a further ~76%, and the full stack runs on domestic Ascend accelerators at application-level parity with A800 GPUs, lowering the unit cost of AI-native education at scale.
Modern video generators excel at synthesizing individual clips, but complete video production requires coordinating a long sequence of interdependent creative steps, including scripting, storyboarding, generation, and editing. It further demands persistent asset management and dynamic task orchestration as intermediate outputs, dependencies, and execution states evolve over time. Existing automated systems typically rely on rigid pipelines that are difficult to adapt to diverse inputs and changing workflows, while general-purpose large language models (LLMs) remain unreliable for long-horizon orchestration and multimodal asset routing. We introduce FRAMEWORKERS, a task-centric and workspace-grounded multi-agent framework for open-ended video production. A central Director formulates video creation as dynamic task management, continuously editing a Task Stack to determine which subtask to execute next and which sub-agent to invoke. An Assistant serves as the execution layer, grounding each selected task in a shared Workspace, retrieving the required assets and context, invoking the assigned sub-agent, and persisting the resulting artifacts. Execution capabilities are exposed through modular sub-agents with registered descriptors, allowing new sub-agents to be integrated without redesigning the orchestration workflow. To improve orchestration reliability, we fine-tune the Director via supervised fine-tuning (SFT) followed by Group Relative Policy Optimization (GRPO) for descriptor-conditioned task routing. Experiments show that FRAMEWORKERS outperforms strong LLM planners in routing accuracy, recovers reliably from runtime failures, generalizes to unseen sub-agents without retraining, and achieves higher end-to-end video quality and broader task coverage than fixed pipelines, single-agent systems, and prior multi-agent approaches.
Industrial anomaly detection is a critical component of modern manufacturing. Most traditional unsupervised methods rely on modelling normal feature distributions, inherently limiting generalization to unknown categories. To improve generalizability, some recent methods incorporate vision-language models (VLMs) for zero-shot detection via text prompts. However, we observe that reasoning-oriented post-training can cause anomaly discrimination to collapse, with some fine-tuned models performing worse than their base VLMs. Existing methods also provide only textual decisions or coarse boxes, without pixel-level segmentation. A more explicit detection principle comes from human inspection: anomalies are identified by comparing a query image with a defect-free reference. Inspired by this, we propose InspectorGPT, a VLM framework centered on comparative reasoning. Given a normal reference and a query image, InspectorGPT compares them to identify discrepancies and perform multiple inspection tasks with detailed reasoning. We internalize this capability through Chain-of-Thought (CoT) fine-tuning and Group Relative Policy Optimization (GRPO) with tailored, verifiable rewards. We further introduce InspectorGPT-Seg for pixel-level anomaly masks. Segmentation supervision improves anomaly discrimination but weakens semantic reasoning, while joint training fails to balance them. We therefore train the two branches separately and combine them through task-vector fusion. Extensive experiments demonstrate superior multi-dimensional performance and generalization to unseen benchmarks, validating comparative reasoning for comprehensive industrial inspection.
Across audit applications, judgments must be supported by reasonable evidence. However, standard financial language models prioritize fluency over evidence. They are built for general financial reasoning and may produce plausible but ambiguous answers, creating a grounding gap that makes them unsuitable for audit work. We address this gap with VERA-8B, a new end-to-end audit reasoning system that identifies audit risks before enforcement actions occur. Constructing such a model raises several challenges, as no prior machine learning work targets pre-enforcement audit prediction. To our knowledge, we are the first to unify SFT and GRPO for evidence-grounded audit reasoning under one evidence standard, achieving performance that surpasses all evaluated baselines. Because auditing cannot tolerate unsupported claims, we introduce abstention and uncertainty qualification to defer uncertain or evidence-incomplete cases. Finally, we design an AuditBridge to ground model reasoning for practical audit work. It transforms raw filings into verified records and then into reviewer-ready reports, bridging finance and computation with broad generality. Together, these components produce auditable, review-ready outputs suitable for practical audit work.
Evolution Strategies (ES) have recently emerged as a memory-efficient post-training paradigm for LLM reasoning. However, the optimization behavior of ES remains understudied, making it hard to define its advantage scope compared to mainstream post-training paradigms (e.g., Group Relative Policy Optimization (GRPO)). By systematically investigating ES dynamics and mechanisms, this paper first identifies a performance advantage of ES over GRPO, theoretically and empirically showing that ES can lead to broader reasoning coverage, thereby better exploiting the reasoning capabilities of pretrained LLMs. Theoretically, we show that verifier-projected Jensen-Shannon diversity across the ES population is helpful to higher Pass@K performances. Empirically, unlike GRPO, which exhibits entropy collapse, ES improves Pass@1 while attaining higher Pass@K than GRPO. We further develop a sequential GRPO-ES training strategy that combines GRPO's strength in Pass@1 with ES's gains in Pass@K. Second, we find that despite substantial whole-model parameter drift, the task-performance gains of ES are only contributed to a sparse subset of larger-magnitude updates. This functional sparsity suggests that large parameter movement need not imply widespread functional change, and held-out evaluations further show that it does not necessarily lead to catastrophic forgetting. Finally, we study how hyperparameter design affects the effectiveness of ES, demonstrating that ES requires a smaller population size in a larger LLM. These findings position ES as a distinct reasoning post-training paradigm rather than a less effective, memory-efficient alternative to GRPO.
Reward models play an essential role in aligning visual generative models, yet most existing visual reward models use a single scalar score or rely on fixed criteria that cannot adapt to different instructions. This limits both interpretability and task sensitivity, especially for text-to-image generation and instruction-based image editing, where different inputs require different evaluation dimensions. We propose RubricRM, a pairwise generative reward modeling framework that first produces an input-specific rubric with evaluation dimensions, weights, and scoring criteria, and then applies the rubric to score candidate images. We train dedicated RubricRM models for text-to-image generation and image editing using a two-stage training pipeline: supervised fine-tuning teaches the model the rubric-based scoring paradigm, while GRPO further improves scoring through fine-grained dimension-level rewards. Experiments on multiple generation and editing benchmarks show that RubricRM outperforms existing specialized reward models and remains competitive with strong proprietary MLLM judges despite using smaller backbones. Our models, data, and code are available at https://github.com/zijiankan/RubricRM.
Reinforcement learning post-training drives reasoning and agentic capabilities in modern AI systems, yet a growing body of work shows that it is most effective when used to fine-tune an already capable base model. We question whether existing pipelines yield models that are most suitable for reinforcement learning. Building on prior work highlighting the role of coverage and pass@K as predictors of post-RL performance, we design a simple modification to supervised fine-tuning, TailSFT, which filters out already fit sequences during training, thereby focusing learning on under-modeled regions, or the tail, of the data distribution. We justify and validate the design choices in TailSFT, particularly the specific filtering criteria, through a combination of controlled experiments and theoretical analysis. On OLMo-3 7B, TailSFT often improves pass@16 performance on math and coding evaluations, with gains up to 17% absolute, while incurring minimal computational overhead. These higher-coverage checkpoints consistently translate to up to 4% absolute pass@1 gains in subsequent GRPO runs, demonstrating that TailSFT checkpoints serve as better initializations for RL. We further introduce a lightweight diagnostic for identifying settings where TailSFT is most likely to help. More broadly, our results motivate a principled, stage-aware approach to model development, in which intermediate checkpoints are judged by how effectively they support subsequent training.
Longitudinal radiology report generation (LRRG) requires identifying both current findings and their changes relative to a prior study. Existing methods jointly model diagnosis, attribute estimation, temporal comparison, and language generation within implicit representations, which can cause task interference, obscure the evidence underlying each decision, and limit error traceability. They also model progression states as independent labels, ignoring their ordered structure and thus treating missed changes and direction reversals equally. We present STRIVE, Multi-Agent Structured Temporal Reasoning with Integrated Verification for LRRG, which decomposes clinical reasoning into specialized Diagnosis, Attribute, and Temporal Change Agents that produce explicit intermediate evidence. In particular, the Temporal Change Agent is further post-trained using Progression-Aware GRPO, a verifiable, shaped reward that assigns partial credit to direction-preserving errors while scoring direction reversals lowest. STRIVE performs verification at two stages: a deterministic Consistency Gate reconciles the agent outputs before report generation, and a Validation Agent checks whether the generated report is supported by the aggregated clinical evidence. On Longitudinal-MIMIC, STRIVE attains the best clinical efficacy among recent methods and more than doubles Longitudinal Change Concordance (LCC), a measure of temporal agreement with the reference report, over the strongest baseline.
Penghui Qi, Xiangxin Zhou, Wee Sun Leecs.LG cs.AI cs.CL
Group-based reinforcement learning methods such as GRPO for large language models avoid training a critic by sampling multiple responses for each prompt. A reliable critic could instead estimate token-level advantages from one response, but standard critic-based training recipes are often unstable. We study this instability and develop \textbf{Best-Practice Critic Optimization (BPCO)}, a recipe that combines DPPO, value predictions bounded to the reward range, Monte Carlo value targets, unnormalized policy advantages, and length-adaptive generalized advantage estimation. Because the critic is used only during training, BPCO can also condition it on reward-defining information, such as a reference answer or grading rubric, that is hidden from the policy. Controlled experiments isolate the effect of each design choice. Across mathematical reasoning tasks with models ranging from 1.5B parameters to 30B-A3B mixtures of experts, BPCO improves a strong critic-based baseline consistently, and matches or exceeds a group-based baseline while sampling one response per prompt. The same recipe also improves learning with rubric-based rewards. These results show that a carefully designed critic provides a reliable alternative to group-relative advantage estimation. Code is available at https://github.com/QPHutu/golden_critic
Recent work has proposed that reasoning and memorization in language models can be characterized by a single representation direction, including methods that keep this direction fixed during reinforcement learning. We test two assumptions behind this view. First, are reasoning-oriented and factual-recall task groups approximately single-direction separable? Second, does the resulting geometry remain stable after GRPO? Using Qwen3-0.6B and a controlled 400-example dataset, we find that a one-dimensional projection can match a full 1024-dimensional linear probe with AUROC = 1.00 on the studied task groups. However, after GRPO, the corresponding direction is substantially reorganized: mean-direction cosine averages 0.453, probe-direction cosine 0.445, while direct representation drift reaches 0.511 at the final layer. Probe AUROC nevertheless remains 1.00. The evidence therefore supports single-direction decodability for the studied task groups but challenges fixed-direction stability: the information persists while its geometric realization changes.
Credit assignment in large-language-model reinforcement learning (LLM RL) can be separated into three objects: evidence about success, a transport operator that converts this evidence into token-level advantages, and an update geometry that turns advantages into policy changes. Recent work has greatly improved evidence, sampling, and update geometry, but the transport operator is usually architecture-agnostic. Fixed-discount GAE applies a stationary geometric kernel along token time; group-relative methods broadcast an outcome statistic across an entire response. Neither operator represents the trajectory-specific computation used by the Transformer policy itself. We introduce computation-conditioned credit transport (CCT), a general framework in which a detached statistic of the behavior policy's internal computation parameterizes the causal kernel that transports downstream value through a rollout. Our concrete algorithm, CompPO, maps native attention concentration to a bounded per-token retention gate, uses the gate in both the one-step bootstrap and a path-dependent generalized-advantage trace (Comp-GAE), and co-designs a transport-aligned critic (TAC) that reuses the actor's hidden states and routing information without a second same-scale Transformer. The task reward and clipped PPO policy objective remain unchanged; a constant gate recovers fixed-coefficient GAE. Across five Qwen3-4B seeds, CompPO reaches 61.4% final held-out accuracy (95% CI [60.8,62.0]) versus 53.8% [52.9,54.7] for tuned GRPO. Neither Comp-GAE with a standard critic (55.2%) nor TAC with a fixed gate (56.4%) matches the full model (interaction +2.4 [1.9,2.9]). Shuffle and position controls confirm trajectory-specific alignment; CompPO is stable in 10/12 PPO-grid runs versus 3/12. Frozen evaluation improves over GRPO by 4.3 and 3.9 greedy pass@1 macro points on Qwen3-4B and Llama-3.1-8B-Instruct.
Lars Benedikt Kaesberg, Tianyu Yang, Florian Valentin Wunderlich +4cs.CV cs.AI
Vision-language models (VLMs) have advanced rapidly in multimodal reasoning, yet recent work shows that their failures often reflect an interaction between visual grounding and downstream reasoning. What remains less clear is how the visual presentation of a task shapes model performance and failure modes when the underlying reasoning problem is unchanged. We study this question in SPaRC, a benchmark for grid-based visual spatial planning, by introducing lightweight input-side scaffolds that preserve the visual modality while making spatial structure more accessible. Across multiple VLMs, these scaffolds improve task accuracy over the original visual setting by up to 34.0 percentage points and further complement GRPO-based training, yielding up to 4.6 additional accuracy points compared with near-zero gains on the original visual input. Analyses on both end-to-end task solving and object detection show that these gains are closely tied to reductions in grounding-related errors, while rule reasoning remains comparatively challenging. We find that visual presentation is a central factor that determines whether VLM benchmarks measure grounded perception, downstream reasoning, or a mixture of both.
Chenghua Zhu, Zhaolu Kang, Qifan Shi +8cs.CV cs.CL cs.LG
Video multimodal large language models have advanced significantly, yet fine-grained motion-temporal understanding remains fragile. The core bottleneck is not only sparse frame sampling, but also the lack of a complete temporal modeling pipeline for explicitly representing frame-to-frame change, enabling appearance-motion interaction, and optimizing temporal direction sensitivity. We propose COMET, a temporally grounded framework that systematically strengthens video MLLMs through explicit temporal representation, appearance-motion fusion, and direction-aware optimization. Architecturally, COMET introduces a temporal motion branch built on Taylor frame differences and injects its motion evidence into the appearance stream via temporal attention bias-enhanced cross-attention. For optimization, COMET combines temporal prior distillation with a forward-reverse TC-GRPO stage that turns temporal order into a direct learning signal and strengthens the model's use of directional motion patterns encoded by the temporal motion branch. The method achieves consistent overall improvements with a pronounced motion-temporal bias: on Qwen3-VL-8B, action-centric tasks (STAR, SSv2) improve by 4.9% on average, temporal reasoning tasks (NExT-QA, CLEVRER, LLaVA-178K) by 2.1% over BL-GRPO, while static perception tasks (PerceptionTest) remain on par. The same gain pattern also transfers to InternVL2.5-8B, indicating that COMET generalizes across model families.
Large multimodal models (LMMs) have demonstrated strong OCR recognition capabilities, yet remain vulnerable to adversarial visual text that is readable to humans but challenging for models to localize and recognize. Existing OCR benchmarks mainly focus on natural or document-style text, while adversarial OCR evaluations remain limited in scale, task coverage, or region-aware evaluation. In this paper, we formulate adversarial OCR as a \textbf{grounded OCR perception} task and introduce \textbf{AdvSpot}, the first benchmark for grounded adversarial OCR evaluation. AdvSpot comprises 390 images with region-level annotations, spanning 5 primary categories and 13 fine-grained adversarial OCR types. To address this challenge, we propose \textbf{ArmorOCR}, a two-stage training framework for robust adversarial OCR perception. ArmorOCR first acquires missing adversarial OCR perception from privileged transformed observations through On-Policy Self-Distillation (OPSD), and then refines grounded OCR perception through Group Relative Policy Optimization (GRPO) with task-conditioned rewards for localization, recognition, full spotting, and visual question answering (VQA). Experiments on our AdvSpot, other adversarial OCR benchmarks, and general OCR benchmarks demonstrate that ArmorOCR consistently improves adversarial OCR perception while preserving competitive general OCR capability.
Group Relative Policy Optimization (GRPO) has become a widely used approach for post-training Large Language Models (LLMs) for reasoning. In GRPO, the group gradients induced by different queries within the same mini-batch are directly averaged to form the policy update. However, these group gradients can point in conflicting directions. Our empirical analysis suggests that group-gradient conflicts tend to be associated with less effective policy updates, motivating the need for a reliable aggregated update direction under such conflicts. Standard GRPO aggregation treats the realized group gradients as deterministic contributions and does not account for differences in their reliability during aggregation. To address this issue, we propose Gradient Uncertainty-Aware Policy Optimization (GUPO), which models each group gradient as a random variable under a Bayesian formulation and estimates its probability distribution. GUPO then derives gradient uncertainty using a Dirichlet-based formulation and uses it to calibrate the contribution of each group gradient during aggregation. Extensive experiments on multiple benchmarks demonstrate the effectiveness of GUPO.
What a language model internalizes from fine-tuning is usually diagnosed after the fact. We make it an experimental variable. OraclePhys is a systematic fine-tuning framework with three components: OraclePhys-Bench, an exactly-graded structural-mechanics benchmark whose finite-element oracle scores every answer and counterfactual edit -- no human labels, no LLM judging; OraclePhys-30K, a supervision dataset of seven answer forms over byte-identical structure descriptions; and a controlled training study across the seven forms and three verifier roles. The study yields two findings. First, the label's answer form -- not its bit count -- causally determines what fine-tuning teaches: a ranking objective installs an out-of-distribution forward model where the untrained base sits at the guessing prior, a scalar objective at best a partial one, a boolean nothing detectable; the vector-scalar gulf survives a second physics domain, a second model family, and a paraphrased evaluation surface. Second, written or score-filtered answers install this capability, while advantage-weighted scores (GRPO) raise reward yet leave the model statistically equivalent to its start on held-out physics -- within the recipes and budgets tested -- sufficing only for routing. The trained 8B -- the first LLM on spatial structural response -- reaches the task's data-precision frontier: above a frontier LLM at zero- and 32-shot, at a specialist's level. What the label spells out about the target computation is what fine-tuning teaches; what you train on is what you route.
Tony Alex, Wish Suharitdamrong, Sara Atito +5cs.SD cs.AI cs.CL eess.AS
Large Audio Language Models (LALMs) have made rapid progress on standardized benchmarks, yet their deployment in practical media workflows, curation, archival indexing, and content distribution remains largely unrealized. We identify automated audio chapterization, the task of segmenting continuous audio streams into thematically coherent chapters, as a demanding and commercially consequential setting that exposes this gap. Chapterization is challenging because boundaries are defined less by objective acoustic events than by subjective editorial judgment, requiring models to reason sequentially over long acoustic contexts and approximate creator-authored boundary decisions. We present AudioChaps, a post-training framework for aligning end-to-end LALMs for this task via Group Relative Policy Optimization (GRPO) guided by Chain-of-Thought (CoT) reasoning. To support training and evaluation, we curate three datasets: AudioChaps-Alignment, derived from creator-annotated chapter boundaries on YouTube; AudioChaps-CoT, which provides structured supervision for well-formatted, high-quality, and evidence-grounded boundary reasoning; and AudioChaps-Eval, a held-out benchmark for audio chapterization. Applying GRPO directly without a Supervised Fine-Tuning (SFT) cold start, AudioChaps-R1-Zero already improves average F1 by 33 points over the state-of-the-art LALM Audio-Flamingo-3-Think. The AudioChaps framework produces our final aligned LALM, AudioChaps-R1, which improves average F1 by 49 points. These results demonstrate that GRPO-trained LALMs can reliably transform unstructured auditory streams into navigable, structured media. Our code, models, and dataset resources will be released upon acceptance at https://github.com/ta012/AudioChaps.
Victor Ye Dong, Reid Pryzant, Yi Liu +1cs.CL cs.AI
Large language models (LLMs) are increasingly deployed as agents that rely on system prompts to use tools and complete tasks. Such deployments impose distinct operational requirements, including appropriate tool use, concise prompts and solution paths, and compliance with safety and formatting policies. For many practitioners, however, assembling domain-specific supervised data to post-train models to meet these requirements is infeasible. We introduce CAPO (Constraint-Aware Prompt Optimization), a primal-dual method that combines pool-based rewrites with adaptive constraint weighting to optimize system prompts under explicit operational constraints. Across agentic benchmarks, CAPO more reliably reaches empirically feasible operating points while improving task performance. CAPO also generalizes beyond agentic settings, achieving strong results on assistant-style evaluations with output-format and safety/privacy constraints. We further introduce DCAPO (Dynamically Trained CAPO), which trains a feedback- and dual-conditioned rewriter with pool-based GRPO while keeping the task agent frozen. Across task agents of different sizes, DCAPO produces a feasible prompt in every evaluated domain and matches or improves the task accuracy achieved by the evaluated baselines. A surrogate analysis characterizes how finite-pool and discrete-rewrite errors enter the inexact primal-dual procedure.
On-policy self-distillation (OPSD) gives language agents dense token-level supervision from a privileged self-teacher on the policy's own trajectories. Existing methods allocate this supervision mainly by teacher trust, but trust does not reveal whether emphasizing a token supports the current policy objective. We call this the trust-utility mismatch and introduce Influence Calibration for Self-Distillation (ICSD). For each supervised token, ICSD measures the first-order response of its importance-weighted RL surrogate contribution to a teacher-directed output perturbation. Batch-adaptive calibration converts this non-stationary signal into a bounded allocation weight while preserving the original auxiliary-loss mass within each action turn. These detached weights affect only the distillation loss and require no additional model pass. Across ALFWorld, WebShop, and Search-QA, ICSD improves all matched aggregate metrics over trust-only allocation under Group Relative Policy Optimization (GRPO) and Group-in-Group Policy Optimization (GiGPO), across two model families spanning 1.5B to 7B. At 7B, it reaches 96.1% ALFWorld success and a WebShop score of 93.1. Frozen-batch analyses show that ICSD reduces teacher-supported mass assigned to objective-opposed tokens from 60.1% to 37.8% and raises cosine compatibility with the RL gradient by 0.192. A companion repository is avail- able at https://github.com/lanqz7766/Influence-Calibration-for-On-Policy-Self-Distillation-in-Agentic-RL.
Yi-Chung Chen, Philip Jacobson, Tom Lampo +6cs.CV cs.LG
Efficiently retrieving relevant clips from large-scale driving logs is essential for data curation, model development, and safety analysis. Structured and rule-based retrieval systems can explicitly target driving events, but typically require expert-defined rules, auxiliary data, and multi-stage perception pipelines. Multimodal embedding models offer a simpler and more efficient alternative by representing each video with a single searchable vector. However, general-purpose models often rely on shortcuts from static scene context and struggle to distinguish motion-centric events, such as turning left versus right or accelerating versus decelerating. In this work, we study how to adapt a general-purpose multimodal embedding model to driving-video retrieval. We first fine-tune Qwen3-VL-Embedding on paired clips and reasoning traces from nuReasoning using an InfoNCE objective. While this stage substantially improves overall retrieval, caption supervision alone remains insufficient for fine-grained motion understanding. We therefore introduce TraVEL (Trajectory-Guided Video Embedding Learning), a motion-aware fine-tuning framework that uses ego-trajectory similarity as a reward within Group Relative Policy Optimization. Trajectories serve only as privileged training supervision; retrieval still operates on single-vector video embeddings without ego poses, expert rules, or auxiliary perception outputs. We further construct a driving-video retrieval benchmark from nuReasoning. Experiments show that TraVEL improves motion-centric retrieval across model scales: relative to SFT, it raises longitudinal and lateral mAP by 9.8 and 4.7 points at 2B, with corresponding gains of 7.2 and 1.5 points at 8B. TraVEL thus combines physically grounded supervision with efficient embedding-based search.
Group Relative Policy Optimization (GRPO) learns from reward differences within a rollout group, but receives no useful relative signal when every sampled response is incorrect. Privileged self-distillation can fill this gap with dense token supervision, yet applying it throughout training creates a different failure mode: the teacher is a biased, low-variance surrogate for the reward objective, so persistent imitation can oppose reward-improving updates after the policy becomes capable of producing successful trajectories. We introduce I-SDPO (Instance-Level Adaptive Self-Distillation Policy Optimization), which treats teacher reliance as capability-dependent. I-SDPO makes one routing decision per input instance and shares it across that instance's rollout group: all-incorrect groups use a privileged self-distillation objective, whereas any-success groups remain intact for GRPO. This design uses imitation only where group-relative rewards are uninformative. A local analysis characterizes when teacher and reward directions align and shows that a non-vanishing biased distillation weight induces an optimization bias floor. The routing rule automatically reduces the expected distillation rate as success probability rises, withdrawing teacher influence without a hand-designed schedule. On SciKnowEval, I-SDPO obtains the best result in all four scientific domains and improves average mean@16 accuracy from 56.67% with GRPO to 70.31%, with a maximum domain gain of 18.24 points.
Deep search agents operate over trajectories spanning dozens of steps, yet standard reinforcement learning provides only a single outcome reward per trajectory, which is far too sparse for effective credit assignment. On-policy self-distillation (OPSD) addresses this by using the model's own logits as dense token-level teachers, but extending it to search agents introduces a fundamental tension: the teacher, having access to privileged information such as the correct answer, produces a distribution that differs systematically from the student's exploration-based reasoning, and naive distillation causes the student to inherit this information asymmetry rather than learn better search strategies. We resolve this tension through two contributions. First, we construct Evidence Anchors, which are concise, step-level evidence snippets extracted from the web, as privileged information that captures key reasoning steps without revealing the entire answer path. Second, we propose Step-Level Self-Distilled Policy Optimization (SSPO), which converts teacher-student disagreement into step-level advantage weights within GRPO, applied exclusively to incorrect trajectories. This design decouples what to update from how much to update: the outcome reward determines the direction of policy change, while the teacher modulates its magnitude at each step. Correct trajectories are left untouched, preserving their diversity. On Qwen3-8B, SSPO consistently outperforms GRPO across BrowseComp, GAIA, and FRAMES, surpassing or matching GRPO trained with twice as many gradient steps while adding only about 5 percent overhead per step from a single additional forward pass.
Kangning Zhang, Haotian Fang, Xukun Luo +6cs.IR cs.AI
Semantic-ID generative recommenders represent each item as a short sequence of discrete semantic tokens and predict the next item by autoregressively generating this token sequence. This paradigm enables a unified generation interface for item IDs, histories, and item text, but it also creates a structured optimization bottleneck during reward-based post-training: when an early semantic token enters the wrong branch of the item-token space, finite rollout groups rarely reach the ground-truth item, so group-relative optimization receives identical zero rewards and produces no useful advantage. We propose Hint-Conditioned Generative Recommendation (HCGRec), a semantic-ID generative recommendation framework that recovers learning signal for such hard training instances. HCGRec diagnoses each instance with checkpoint rollouts and supplies a minimal target-prefix hint only when the current generator cannot reach the correct item. The model then generates the unhinted suffix under the hinted semantic branch, turning zero-reward groups into informative comparisons over item-token completions. Hinting also changes token identity: hinted prefix tokens are oracle-provided item context, while unhinted suffix tokens are sampled generation actions. We therefore introduce hint-aware credit decomposition, using supervised learning to preserve item-semantic and prefix-structure alignment for hinted tokens and GRPO to optimize the sampled suffix. Experiments on sequential recommendation benchmarks show that HCGRec substantially improves over supervised fine-tuning and vanilla reward-based post-training, while reducing zero-advantage training samples from over 70% to below 20%. The code is accessible at https://github.com/WncFht/GRec.
Jiabao Zhuang, Changhao Jiang, Hanchen Wang +11cs.SD cs.CL
Long-form song generation models continue to improve in duration, structural integrity, and acoustic complexity, making reliable aesthetic rewards increasingly important for aligning these models with human preferences. However, reward models for complete songs remain limited, and existing evaluators typically predict scores in a single forward pass without providing readable explanations. We introduce MUSECRITIC, a semi-scalar reward model that generates a natural-language critique covering five aesthetic dimensions and uses it as an intermediate representation to predict continuous reward scores. MUSECRITIC follows a two-stage training pipeline: a teacher model first provides high-quality critiques for supervised fine-tuning, after which the fine-tuned model generates its own critiques for reward learning, mitigating distribution shift between training and inference. On an in-domain test set of 200 SongEval songs, MUSECRITIC reduces macro-averaged mean squared error from 0.2875 to 0.2316 and improves macro-averaged LCC, SRCC, and Kendall's tau to 0.9068, 0.8838, and 0.7178, respectively. On the out-of-domain Music Arena benchmark with 733 preference pairs, it achieves the highest accuracy of 71.35%. Moreover, using MUSECRITIC with GRPO improves Muse-0.6B on all nine aesthetic metrics from SongEval and Audiobox Aesthetics. These results demonstrate that critique-conditioned reward modeling reduces scoring error and provides an effective optimization signal for song generation. The project repository is available at https://github.com/WuqnEl/MuseCritic.