pith. sign in

arxiv: 2606.06828 · v1 · pith:RPSMOH7Jnew · submitted 2026-06-05 · 💻 cs.CV · cs.LG

AdaGRPO: A Capability-Aware Adaptive Enhancement for Flow-based GRPO

Jiazi Bu , Pengyang Ling , Yujie Zhou , Yibin Wang , Yuhang Zang , Tianyi Wei , Xiaohang Zhan , Jiaqi Wang
show 3 more authors
Tong Wu Xingang Pan Dahua Lin
This is my paper

Pith reviewed 2026-06-27 22:52 UTC · model grok-4.3

classification 💻 cs.CV cs.LG
keywords AdaGRPOGRPOflow-based modelstext-to-imagereinforcement learningcurriculum filteringadvantage fusionpolicy optimization
0
0 comments X

The pith

AdaGRPO corrects capability-blind spots in GRPO for flow-based text-to-image models through adaptive prompt selection and advantage fusion.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Current GRPO methods for flow models randomly sample prompts and rely only on intra-group statistics, ignoring how well the model is performing at any moment. AdaGRPO introduces capability tracking to choose prompts at the right difficulty level and fuses local group advantages with global ones for better signals. A sympathetic reader cares because this decoupling can waste training effort and cause unstable learning in preference alignment. If the approach works, flow models could reach higher alignment quality with less variance in the RL process.

Core claim

AdaGRPO is a capability-aware RL algorithm for flow models that uses an Online Curriculum Filtering Strategy to dynamically select prompts matching the model's proficiency and a Cross-Level Advantage Fusion to combine intra-group and global advantages, serving as a lightweight plug-in that boosts performance and stabilizes training when added to existing flow-based GRPO frameworks.

What carries the argument

The combination of Online Curriculum Filtering Strategy, which tracks model proficiency to select boundary-matching prompts, and Cross-Level Advantage Fusion, which integrates fine-grained intra-group with macro-level global advantages for unbiased evaluation.

If this is right

  • Consistent performance improvements in text-to-image flow model alignment.
  • Significantly more stable GRPO training dynamics.
  • Easy integration into frameworks like Flow-GRPO, DanceGRPO, and Flow-CPS.
  • Addresses both prompt selection and advantage estimation issues in flow model RL.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Similar capability-aware adaptations could benefit RL alignment in other generative architectures.
  • The global advantage component may help identify when local statistics mislead about true progress.
  • Dynamic filtering might reduce overall training compute by avoiding mismatched data.

Load-bearing premise

Dynamically tracking the model's proficiency and combining intra-group with global advantages produces unbiased improvements without adding new biases or instability to the reinforcement learning process.

What would settle it

Comparative training runs on standard text-to-image preference benchmarks showing that AdaGRPO versions do not outperform or stabilize better than standard GRPO baselines.

Figures

Figures reproduced from arXiv: 2606.06828 by Dahua Lin, Jiaqi Wang, Jiazi Bu, Pengyang Ling, Tianyi Wei, Tong Wu, Xiaohang Zhan, Xingang Pan, Yibin Wang, Yuhang Zang, Yujie Zhou.

Figure 1
Figure 1. Figure 1: Gallery of AdaGRPO. By integrating the proposed AdaGRPO, flow models (Flux.1-dev in this figure) experience a substantial leap in the generation performance, yielding remarkable improvements in intricate textures and visual fidelity. All prompts are listed in the appendix. ABSTRACT Group Relative Policy Optimization (GRPO) has demonstrated remarkable success in aligning text-to-image (T2I) flow models with… view at source ↗
Figure 2
Figure 2. Figure 2: Observations. (a) Random sampling (current GRPO methods) and extreme prompts (“Easiest”/“Hardest”) yield suboptimal alignment efficacy. While selecting locally moderate prompts (“Medium”) offers improvements, it remains biased by the current batch. In contrast, our Online Curriculum Filtering Strategy maximizes performance by dynamically identifying moderate tasks through the model’s global capability. (b)… view at source ↗
Figure 3
Figure 3. Figure 3: Pipeline of AdaGRPO. (a) First, Online Curriculum Filtering Strategy evaluates candidate prompts via deterministic ODE sampling and adaptively selects the one that best matches the model’s current capability anchor (µema). The selected prompt is utilized for stochastic SDE rollout. (b) Then, Cross-Level Advantage Fusion integrates the intra-group local advantage with the history-calibrated global advantage… view at source ↗
Figure 4
Figure 4. Figure 4: Reward Curves during Training. The proposed AdaGRPO facilitates significantly smoother training dynamics and higher performance ceilings across diverse training configurations. In the standard case (both sets non-empty), this operation scales the sum of positive terms to 1 and negative terms to −1, guaranteeing P i A¯i global = 0. In the rare event of a unilateral batch (i.e., P = ∅ or N = ∅), we dynamical… view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative Comparisons with Baselines on HPS-v2. Best viewed zoomed in [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative Comparisons with Baselines on HPS-v3. Best viewed zoomed in [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Additional Comparison Results on HPS-v2. (1/2) 16 [PITH_FULL_IMAGE:figures/full_fig_p016_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Additional Comparison Results on HPS-v2. (2/2) 17 [PITH_FULL_IMAGE:figures/full_fig_p017_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Additional Comparison Results on HPS-v3. (1/2) 18 [PITH_FULL_IMAGE:figures/full_fig_p018_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Additional Comparison Results on HPS-v3. (2/2) 19 [PITH_FULL_IMAGE:figures/full_fig_p019_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Additional Visual Samples of AdaGRPO. (1/4) 20 [PITH_FULL_IMAGE:figures/full_fig_p020_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Additional Visual Samples of AdaGRPO. (2/4) 21 [PITH_FULL_IMAGE:figures/full_fig_p021_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Additional Visual Samples of AdaGRPO. (3/4) 22 [PITH_FULL_IMAGE:figures/full_fig_p022_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Additional Visual Samples of AdaGRPO. (4/4) 23 [PITH_FULL_IMAGE:figures/full_fig_p023_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Results using same prompts and different seeds. (HPS-v2) Seed = 3407 Seed = 3408 Seed = 3409 “In the evening, under the Eiffel Tower in Paris, two red retro sports cars parked side by side on the wet stone road with their lights on.” “National Geographic-style photography shows a lazy orange cat yawns on a sun￾filled windowsill with a blurred indoor potted plant in the background.” “A crystal clear crysta… view at source ↗
Figure 16
Figure 16. Figure 16: Results using same prompts and different seeds. (HPS-v3) 24 [PITH_FULL_IMAGE:figures/full_fig_p024_16.png] view at source ↗
read the original abstract

Group Relative Policy Optimization (GRPO) has demonstrated remarkable success in aligning text-to-image (T2I) flow models with human preferences. However, we have identified that the learning loop of current flow-based GRPO is fundamentally decoupled from the learner's current capability, suffering from critical blind spots at both prompt selection and advantage estimation: (i) Existing methods sample prompts randomly, overlooking the substantial impact of data selection on reinforcement learning (RL) efficacy--a factor proven crucial in GRPO for large language models; (ii) They evaluate sample quality solely relying on intra-group statistics, lacking a global perspective to accurately measure true policy improvement. To address these issues, we propose Adaptive GRPO (AdaGRPO), a novel capability-aware RL algorithm tailored for flow models. Specifically, AdaGRPO consists of two principal components: (i) Online Curriculum Filtering Strategy: Dynamically tracks the model's proficiency and adaptively selects prompts that best match its current learning boundary; (ii) Cross-Level Advantage Fusion: Synergistically integrates fine-grained intra-group advantages with macro-level global advantages, providing a comprehensive and unbiased policy evaluation. As a lightweight, plug-and-play module, AdaGRPO can be seamlessly integrated with existing frameworks such as Flow-GRPO, DanceGRPO, and Flow-CPS. Extensive experiments demonstrate that AdaGRPO consistently drives performance gains while significantly stabilizes GRPO training for flow models.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

0 major / 1 minor

Summary. The paper proposes AdaGRPO as a capability-aware enhancement to Group Relative Policy Optimization (GRPO) for aligning flow-based text-to-image models. It identifies two blind spots in existing flow-based GRPO: random prompt sampling that ignores learner proficiency and advantage estimation based solely on intra-group statistics without a global view. AdaGRPO introduces an Online Curriculum Filtering Strategy to dynamically select boundary-matching prompts and Cross-Level Advantage Fusion to combine intra-group and global advantages. The method is presented as a lightweight, plug-and-play module compatible with frameworks such as Flow-GRPO, DanceGRPO, and Flow-CPS, with experiments claimed to demonstrate consistent performance gains and improved training stability.

Significance. If the claimed results hold, the work could improve sample efficiency and stability in RL fine-tuning of flow models by making the training loop adaptive to current model capability. The two proposed mechanisms directly target known issues in prompt selection and advantage estimation that affect RL efficacy, and the plug-and-play design would allow easy adoption in existing pipelines for preference alignment in generative models.

minor comments (1)
  1. [Abstract] Abstract: the claim of 'extensive experiments demonstrate that AdaGRPO consistently drives performance gains while significantly stabilizes GRPO training' is stated without any quantitative metrics, baselines, ablation results, or protocol details, which weakens the ability to evaluate the central empirical claim from the summary alone.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for their positive assessment of our work on AdaGRPO and for recommending minor revision. The referee's summary correctly identifies the core issues addressed and the plug-and-play nature of the proposed method. No major comments were raised in the report.

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The provided abstract and description contain no equations, derivations, fitted parameters, or mathematical claims. The method is introduced descriptively as two components (online curriculum filtering and cross-level advantage fusion) without any reduction of predictions to inputs by construction, self-citation chains, or ansatz smuggling. Performance and stability claims are presented as experimental outcomes rather than derived results that collapse to the inputs. This is a standard non-finding for a methods paper lacking visible derivation structure.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review supplies no explicit free parameters, axioms, or invented entities; the method description implies standard RL assumptions but does not enumerate them.

pith-pipeline@v0.9.1-grok · 5820 in / 1042 out tokens · 16601 ms · 2026-06-27T22:52:47.236925+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

31 extracted references · 30 canonical work pages · 20 internal anchors

  1. [1]

    Training Diffusion Models with Reinforcement Learning

    Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training diffusion models with reinforcement learning.arXiv preprint arXiv:2305.13301,

    work page internal anchor Pith review Pith/arXiv arXiv

  2. [2]

    Hiflow: Training-free high-resolution image generation with flow-aligned guidance.arXiv preprint arXiv:2504.06232,

    Jiazi Bu, Pengyang Ling, Yujie Zhou, Pan Zhang, Tong Wu, Xiaoyi Dong, Yuhang Zang, Yuhang Cao, Dahua Lin, and Jiaqi Wang. Hiflow: Training-free high-resolution image generation with flow-aligned guidance.arXiv preprint arXiv:2504.06232,

    work page arXiv

  3. [3]

    From sparse to dense: Multi-view grpo for flow models via augmented condition space.arXiv preprint arXiv:2603.12648,

    Jiazi Bu, Pengyang Ling, Yujie Zhou, Yibin Wang, Yuhang Zang, Tianyi Wei, Xiaohang Zhan, Jiaqi Wang, Tong Wu, Xingang Pan, et al. From sparse to dense: Multi-view grpo for flow models via augmented condition space.arXiv preprint arXiv:2603.12648,

    work page arXiv

  4. [4]

    Process Reinforcement through Implicit Rewards

    Ganqu Cui, Lifan Yuan, Zefan Wang, Hanbin Wang, Yuchen Zhang, Jiacheng Chen, Wendi Li, Bingxiang He, Yuchen Fan, Tianyu Yu, et al. Process reinforcement through implicit rewards. arXiv preprint arXiv:2502.01456,

    work page internal anchor Pith review Pith/arXiv arXiv

  5. [5]

    Dynamic-TreeRPO: Breaking the Independent Trajectory Bottleneck with Structured Sampling

    Xiaolong Fu, Lichen Ma, Zipeng Guo, Gaojing Zhou, Chongxiao Wang, ShiPing Dong, Shizhe Zhou, Ximan Liu, Jingling Fu, Tan Lit Sin, et al. Dynamic-treerpo: Breaking the independent trajectory bottleneck with structured sampling.arXiv preprint arXiv:2509.23352,

    work page internal anchor Pith review Pith/arXiv arXiv

  6. [6]

    AnimateDiff: Animate Your Personalized Text-to-Image Diffusion Models without Specific Tuning

    Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning.arXiv preprint arXiv:2307.04725,

    work page internal anchor Pith review Pith/arXiv arXiv

  7. [7]

    TempFlow-GRPO: When Timing Matters for GRPO in Flow Models

    Xiaoxuan He, Siming Fu, Yuke Zhao, Wanli Li, Jian Yang, Dacheng Yin, Fengyun Rao, and Bo Zhang. Tempflow-grpo: When timing matters for grpo in flow models.arXiv preprint arXiv:2508.04324,

    work page internal anchor Pith review Pith/arXiv arXiv

  8. [8]

    HunyuanVideo: A Systematic Framework For Large Video Generative Models

    Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, et al. Hunyuanvideo: A systematic framework for large video generative models.arXiv preprint arXiv:2412.03603,

    work page internal anchor Pith review Pith/arXiv arXiv

  9. [9]

    MixGRPO: Unlocking Flow-based GRPO Efficiency with Mixed ODE-SDE

    Junzhe Li, Yutao Cui, Tao Huang, Yinping Ma, Chun Fan, Miles Yang, and Zhao Zhong. Mixgrpo: Unlocking flow-based grpo efficiency with mixed ode-sde.arXiv preprint arXiv:2507.21802, 2025a. Yuming Li, Yikai Wang, Yuying Zhu, Zhongyu Zhao, Ming Lu, Qi She, and Shanghang Zhang. Branchgrpo: Stable and efficient grpo with structured branching in diffusion model...

    work page internal anchor Pith review Pith/arXiv arXiv

  10. [10]

    Flow-GRPO: Training Flow Matching Models via Online RL

    Jie Liu, Gongye Liu, Jiajun Liang, Yangguang Li, Jiaheng Liu, Xintao Wang, Pengfei Wan, Di Zhang, and Wanli Ouyang. Flow-grpo: Training flow matching models via online rl.arXiv preprint arXiv:2505.05470,

    work page internal anchor Pith review Pith/arXiv arXiv

  11. [11]

    Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow

    Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow.arXiv preprint arXiv:2209.03003,

    work page internal anchor Pith review Pith/arXiv arXiv

  12. [12]

    Dynamics-predictive sampling for active rl finetuning of large reasoning models.arXiv preprint arXiv:2603.10887,

    Yixiu Mao, Yun Qu, Qi Wang, Heming Zou, and Xiangyang Ji. Dynamics-predictive sampling for active rl finetuning of large reasoning models.arXiv preprint arXiv:2603.10887,

    work page arXiv

  13. [13]

    OpenVid-1M: A Large-Scale High-Quality Dataset for Text-to-video Generation

    Kepan Nan, Rui Xie, Penghao Zhou, Tiehan Fan, Zhenheng Yang, Zhijie Chen, Xiang Li, Jian Yang, and Ying Tai. Openvid-1m: A large-scale high-quality dataset for text-to-video generation.arXiv preprint arXiv:2407.02371,

    work page internal anchor Pith review Pith/arXiv arXiv

  14. [14]

    Sudo: Enhancing text-to- image diffusion models with self-supervised direct preference optimization.arXiv preprint arXiv:2504.14534,

    Liang Peng, Boxi Wu, Haoran Cheng, Yibo Zhao, and Xiaofei He. Sudo: Enhancing text-to- image diffusion models with self-supervised direct preference optimization.arXiv preprint arXiv:2504.14534,

    work page arXiv

  15. [15]

    SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis

    Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis.arXiv preprint arXiv:2307.01952,

    work page internal anchor Pith review Pith/arXiv arXiv

  16. [16]

    Proximal Policy Optimization Algorithms

    John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms.arXiv preprint arXiv:1707.06347,

    work page internal anchor Pith review Pith/arXiv arXiv

  17. [17]

    DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, et al. Deepseekmath: Pushing the limits of mathemat- ical reasoning in open language models.arXiv preprint arXiv:2402.03300,

    work page internal anchor Pith review Pith/arXiv arXiv

  18. [18]

    Bots: A unified framework for bayesian online task selection in llm reinforcement finetuning

    Qianli Shen, Daoyuan Chen, Yilun Huang, Zhenqing Ling, Yaliang Li, Bolin Ding, and Jingren Zhou. Bots: A unified framework for bayesian online task selection in llm reinforcement finetuning. arXiv preprint arXiv:2510.26374,

    work page arXiv

  19. [19]

    Denoising Diffusion Implicit Models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models.arXiv preprint arXiv:2010.02502, 2020a. Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations.arXiv preprint arXiv:2011.13456, 2020b. Petru Soviany, Radu T...

    work page internal anchor Pith review Pith/arXiv arXiv 2010

  20. [20]

    Longcat-video technical report.arXiv preprint arXiv:2510.22200,

    Meituan LongCat Team, Xunliang Cai, Qilong Huang, Zhuoliang Kang, Hongyu Li, Shijun Liang, Liya Ma, Siyu Ren, Xiaoming Wei, Rixu Xie, et al. Longcat-video technical report.arXiv preprint arXiv:2510.22200,

    work page arXiv

  21. [21]

    HunyuanVideo 1.5 Technical Report

    URL https: //arxiv.org/abs/2511.18870. Georgios Tzannetos, Bárbara Gomes Ribeiro, Parameswaran Kamalaruban, and Adish Singla. Proxi- mal curriculum for reinforcement learning agents.arXiv preprint arXiv:2304.12877,

    work page internal anchor Pith review Pith/arXiv arXiv

  22. [22]

    Wan: Open and Advanced Large-Scale Video Generative Models

    Team Wan, Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianxiao Yang, et al. Wan: Open and advanced large-scale video generative models. arXiv preprint arXiv:2503.20314,

    work page internal anchor Pith review Pith/arXiv arXiv

  23. [23]

    Coefficients-preserving sampling for reinforcement learning with flow matching.arXiv preprint arXiv:2509.05952,

    Feng Wang and Zihao Yu. Coefficients-preserving sampling for reinforcement learning with flow matching.arXiv preprint arXiv:2509.05952,

    work page arXiv

  24. [24]

    Unigenbench++: A unified semantic evaluation benchmark for text-to-image generation.arXiv preprint arXiv:2510.18701, 2025a

    Yibin Wang, Zhimin Li, Yuhang Zang, Jiazi Bu, Yujie Zhou, Yi Xin, Junjun He, Chunyu Wang, Qinglin Lu, Cheng Jin, et al. Unigenbench++: A unified semantic evaluation benchmark for text-to-image generation.arXiv preprint arXiv:2510.18701, 2025a. Yibin Wang, Zhimin Li, Yuhang Zang, Yujie Zhou, Jiazi Bu, Chunyu Wang, Qinglin Lu, Cheng Jin, and Jiaqi Wang. Pre...

    work page arXiv

  25. [25]

    Human Preference Score v2: A Solid Benchmark for Evaluating Human Preferences of Text-to-Image Synthesis

    Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis.arXiv preprint arXiv:2306.09341,

    work page internal anchor Pith review Pith/arXiv arXiv

  26. [26]

    DanceGRPO: Unleashing GRPO on Visual Generation

    Zeyue Xue, Jie Wu, Yu Gao, Fangyuan Kong, Lingting Zhu, Mengzhao Chen, Zhiheng Liu, Wei Liu, Qiushan Guo, Weilin Huang, et al. Dancegrpo: Unleashing grpo on visual generation.arXiv preprint arXiv:2505.07818,

    work page internal anchor Pith review Pith/arXiv arXiv

  27. [27]

    CogVideoX: Text-to-Video Diffusion Models with An Expert Transformer

    Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer.arXiv preprint arXiv:2408.06072,

    work page internal anchor Pith review Pith/arXiv arXiv

  28. [28]

    DAPO: An Open-Source LLM Reinforcement Learning System at Scale

    Qiying Yu, Zheng Zhang, Ruofei Zhu, Yufeng Yuan, Xiaochen Zuo, Yu Yue, Weinan Dai, Tiantian Fan, Gaohong Liu, Lingjun Liu, et al. Dapo: An open-source llm reinforcement learning system at scale.arXiv preprint arXiv:2503.14476,

    work page internal anchor Pith review Pith/arXiv arXiv

  29. [29]

    Srpo: A cross-domain implementation of large-scale reinforcement learning on llm.arXiv preprint arXiv:2504.14286,

    Xiaojiang Zhang, Jinghui Wang, Zifei Cheng, Wenhao Zhuang, Zheng Lin, Minglei Zhang, Shaojie Wang, Yinghan Cui, Chao Wang, Junyi Peng, et al. Srpo: A cross-domain implementation of large-scale reinforcement learning on llm.arXiv preprint arXiv:2504.14286,

    work page arXiv

  30. [30]

    Act only when it pays: Efficient reinforcement learning for llm reasoning via selective rollouts.arXiv preprint arXiv:2506.02177, 2025a

    Haizhong Zheng, Yang Zhou, Brian R Bartoldson, Bhavya Kailkhura, Fan Lai, Jiawei Zhao, and Beidi Chen. Act only when it pays: Efficient reinforcement learning for llm reasoning via selective rollouts.arXiv preprint arXiv:2506.02177, 2025a. Kaiwen Zheng, Huayu Chen, Haotian Ye, Haoxiang Wang, Qinsheng Zhang, Kai Jiang, Hang Su, Stefano Ermon, Jun Zhu, and ...

    work page arXiv

  31. [31]

    Generate a game concept setting diagram: a huge turtle carries a small castle on its back, which serves as a mobile base for players and travels through the fantasy forest

    F LIMITATION ANDDISCUSSION While AdaGRPO demonstrates superior performance and training stability, it faces certain constraints. Similar to dynamic data sampling strategies in LLM alignment (Yu et al., 2025; Bae et al., 2026), our online prompt filtering mechanism inevitably introduces some computational overhead. However, given the relatively modest VRAM...

    2025