cs.LGNov 2, 2026

Do Visual RL Gains Scale at Test Time? A Quantile View of Post-Training

Authors: Austin Wang, Lexing Ying

Organizations: ByteDance Seed

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Reinforcement learning (RL) post-training improves the single-sample quality of visual generative models, but deployment often provides additional inference compute through Best-of-N sampling. Whether RL gains transfer under such test-time scaling, particularly on difficult held-out tasks, remains poorly understood in the visual domain. Our central contribution is to systematically identify and characterize a failure mode of scaling transfer in visual RL: across multiple RL methods, model backbones, and evaluations spanning compositional generation, text rendering, and preference alignment, higher single-sample reward does not necessarily imply a better quality-compute frontier. RL gains may attenuate, saturate, or even reverse at large sampling budgets, with the base model eventually outperforming its post-trained counterpart. We explain these regimes through a quantile-transfer framework, showing that Best-of-N places increasing weight on the upper tail of the reward distribution as N grows. Consequently, RL can improve typical samples while compressing or degrading the rare high-quality outcomes that test-time scaling exploits. As a simple intervention motivated by this analysis, we introduce Quantile Utility Policy Optimization (QUPO), which optimizes a controllable mixture of quantile utilities corresponding to different inference budgets. Experiments on SD3.5-M and BAGEL show that QUPO better preserves upper-tail quality and substantially improves scaling on challenging held-out prompts relative to standard RL post-training. Our results suggest that visual generative models should be evaluated and optimized not only for single-sample quality, but for how post-training reshapes the full reward distribution that inference-time scaling exploits.

More in this topic

Reinforcement Learning

Reinforcement learning studies how agents learn sequential decisions from rewards and interaction, balancing exploration, performance, safety, and adaptation.

See all 1,774 papers

Explore similar work

CardsList
Jul 2, 2026

Optimizing Visual Generative Models via Distribution-wise Rewards

Conventional reinforcement learning strategies for visual generation typically employ sample-wise reward functions, yet this practice frequently results in reward hacking that degrades image diversity and introduces visual anomalies. To address these limitations, we present a novel framework that finetunes generative models using distribution-wise rewards, ensuring better alignment with real-world data distributions. Unlike rewards that evaluate samples individually, distribution-wise reward accounts for the data distribution of the samples, mitigating the mode collapse problem that occurs when all samples optimize towards the same direction independently. To overcome the prohibitive computational cost of estimating these rewards, we introduce a subset-replace strategy that efficiently provides reward signals by updating only a small subset of a generated reference set. Additionally, we apply RL to optimize post-hoc model merging coefficients, potentially mitigating the train-inference inconsistency caused by introducing stochastic differential equation (SDE) in regular RL practices. Extensive experiments show our approach significantly improves FID-50K across various base models, from 8.30 to 5.77 for SiT and from 3.74 to 3.52 for EDM2. Qualitative evaluation also confirms that our method enhances perceptual quality while preserving sample diversity.

Ruihang Li, Mengde Xu, Shuyang Gu +4

University of Science and Technology of China · Shanghai Innovation Institute · Hunyuan Frontier Lab, Tencent +1

Diffusion Model Fine-TuningRL for Image GenerationRL for Generative ModelsMode Collapse
AskPDF
Oct 23, 2026

Of Chains and Gains: What Reinforcement Learning changes on Reasoning-based Image Quality Assessment

Reinforcement Learning with Verifiable Rewards (RLVR) has improved the performance of large Vision-Language Models (VLMs) on image quality assessment (IQA). Existing approaches often attribute these gains to improved reasoning before scoring, supported by increasingly plausible Chain-of-Thought (CoT). However, whether these gains exceed sampling effects and how they depend on explicit reasoning remain unclear. We rigorously investigate these questions through sampling-aware evaluation, controlled reasoning interventions, and shortcut experiments. We find that RL-trained models outperform base models under large-budget max⁡g@k\operatorname{max}_g@k evaluation, while improvements beyond averaging are concentrated in-distribution. RLVR produces chains that raise base-model performance when transplanted and improves scoring with the chain held fixed. Yet scoring becomes less dependent on the chain during inference, more clearly at 7B than at 32B. Steering the probed quality direction in thinking tokens also has little effect on scoring. Finally, RL can improve scoring accuracy by exploiting textual cues correlated with mean opinion scores (MOSs), which could inflate benchmark performance. Together, these findings show that RLVR raises quality scoring and produces more useful chains, but that better scoring does not imply a stronger reliance on explicit reasoning, while score-based rewards can also encourage reliance on quality-irrelevant shortcuts. We hope this preliminary study can enlighten more and more concrete reasoning schemes in IQA.

Yixuan Li, Yuxuan Jiang, Hanwei Zhu +4

University of Bristol, UK · Jiangxi University of Finance and Economics, China · China University of Mining and Technology, China +1

Image Quality AssessmentReinforcement Learning with Verifiable RewardsRL for VLMsReasoning for VLMsVLM EvaluationFaithfulness in VLMs
AskPDF
Jun 1, 2026

TRON: Targeted Rule-Verifiable Online Environments for Visual Reasoning RL

Reinforcement learning (RL) for visual reasoning needs scalable, verifiable, and controllable training signals. Existing visual RL post-training trains on static curated datasets, with fixed image-question-answer samples bounded by their collection budget. In this work, we introduce TRON (Targeted, Rule-verifiable Online eNvironments), an online environment substrate: a training rollout is generated on demand by a controllable generator-verifier program that samples a fresh latent visual state, renders an image, asks a question, and exactly verifies the answer. A single run can therefore draw an unbounded stream of fresh instances at the difficulty level required by the current curriculum. The current TRON suite contains 520 environments organized into five ability buckets (spatial, mathematical, diagram, pattern/logic, and counting); the same substrate supports both a single full model trained on all buckets and per-bucket ability-specialist models, with no additional data collection. We also introduce a substrate analysis covering generation reliability, instance and level diversity, cross-environment near-duplicates, and base-model pass rate by difficulty level. RL post-training with METHOD consistently improves performance on ten external multimodal reasoning benchmarks across Qwen3-VL-4B, Qwen2.5-VL-7B, and MiMo-VL-7B-SFT.

Tianze Yang, Yucheng Shi, Ruitong Sun +3

University of Georgia

Reinforcement LearningVLM ReasoningReinforcement Learning with Verifiable RewardsVisually Grounded ReasoningRL Post-TrainingVisual RL
AskPDF