arXiv:2608. 14430v1 Announce Type: new Abstract: Reinforcement learning (RL) post-training provides a direct way to align diffusion models with human preferences and task-specific rewards.
By Yixian Xu, Yuanrui Zhang, Shengjie Luo, Liwei Wang, Di He
CAST introduces a reinforcement‑learning fine‑tuning framework for diffusion models that addresses three key limitations: it automatically selects the denoising window based on each model’s trajectory, decomposes prompts into verifiable semantic atoms via Causal Scene Graphs, and applies atom‑level rewards spatially weighted in the policy objective. The method is applied to FLUX.2‑dev and Qwen‑Image‑2512, yielding up to 3.07× improvement on the hardest GenEval 2 prompts compared with Flow‑GRPO while also enhancing overall generation quality.
By Shu Yu, Chaochao Lu
VaLoP introduces a new method for reward-guided diffusion called Value Gradient Lookahead and Pullback, which evaluates the value function at a less noisy state and pulls its gradient back to the current state to reduce posterior covariance errors. The approach unifies two common gradient estimators and adds a correction term computed with a single vector-Jacobian product, while VaLoP‑Fast amortizes this correction with a lightweight head trained on a single reward. Experiments across multiple models, rewards, and datasets show that both variants consistently outperform existing value‑gradient guidance, and VaLoP‑Fast achieves the highest success rates on robot policy benchmarks such as LIBERO and RoboCasa.
By Kwanyoung Kim
Routing Policy Optimization (RPO) introduces stochastic exploration within a diffusion visual generator by treating attention routing as a stochastic action. Instead of perturbing the entire reverse denoising trajectory, RPO samples key attention blocks at selected denoising levels, perturbs the transition velocity, and evaluates the perturbation with a score‑function objective. A two‑rollout estimator provides an advantage signal, and after training the model restores dense attention, yielding higher generation quality and nearly four times the training efficiency compared to Flow‑GRPO‑style methods.
The paper investigates how different estimators of the reverse Kullback–Leibler (KL) divergence used as a regularization term in reinforcement learning (RL) training of large language models (LLMs) affect training stability and downstream performance. By analyzing gradient bias across various estimator configurations, the authors demonstrate that biased gradients can cause training instabilities, while unbiased configurations improve performance on both in‑domain and out‑of‑domain tasks. Experiments on Qwen2.5‑7B, Llama‑3.1‑8B‑Instruct, and Qwen3‑4B‑Instruct‑2507 confirm these findings and show that KL regularization also stabilizes off‑policy RL training in asynchronous setups.
By Vedant Shah, Johan Obando-Ceron, Vineet Jain, Brian Bartoldson, Bhavya Kailkhura, Sarthak Mittal, Glen Berseth, Pablo Samuel Castro, Yoshua Bengio, Esmeralda S. Whitammer, Moksh Jain, Siddarth Venkatraman, Aaron Courville
arXiv:2608.23664v1 Announce Type: cross
Abstract: Reward fine-tuning is becoming an important tool for adapting diffusion models to human preferences and task-specific objectives, but existing method...
By Jaemoo Choi, Wei Guo, Yuchen Zhu, Arash Vahdat, Molei Tao, Julius Berner, Yongxin Chen
The paper introduces DRILL, a bilevel training framework that optimizes world models for policy improvement rather than pixel accuracy. By reweighting prediction errors with policy visitation and confidence (DRILL‑VWI) or using a full meta‑gradient approach (DRILL‑IMG), the authors demonstrate significant gains in downstream policy success rates across multiple manipulation simulators. The work shows that world models can be more effectively trained using the policy gradients they generate, rather than relying solely on pixel‑level prediction metrics.
By Zhaolu Kang, Tailong Luo, Lei Wei, Guangyuan Dong, Sihan Liu, Yingjie He, Siheng Wang, Shuaibo Li, Rongchao Zhang, Zhen Tian, Richeng Xuan, Zhichao Hu
Routing Policy Optimization (RPO) introduces stochastic exploration directly within a visual generator’s attention mechanism, avoiding costly rollout-based optimization. By sampling key attention blocks at selected denoising levels and perturbing the transition velocity, RPO uses a Gaussian likelihood and a two‑rollout estimator to provide an advantage signal for learning. Experiments on text‑to‑image and text‑to‑video tasks demonstrate improved generation quality and nearly four times higher training efficiency compared to Flow‑GRPO‑style methods.
By Qiuhong Shen, Yiding Yang, Yuxin Zhang, Xinwei Huang, Minxuan Lin, Bo Liu, Hao Kang, Haotian Yang, Yizhi Wang, Chongyang Ma, Xinchao Wang
arXiv:2601. 00898v3 Announce Type: replace Abstract: Diffusion-based policies have gained growing popularity in solving a wide range of decision-making tasks due to their superior expressiveness and controllable generation during inference.
By Ruiming Liang, Yinan Zheng, Kexin Zheng, Tianyi Tan, Jianxiong Li, Liyuan Mao, Zhihao Wang, Guang Chen, Hangjun Ye, Jingjing Liu, Jinqiao Wang, Xianyuan Zhan
arXiv:2607. 07693v1 Announce Type: cross Abstract: Reinforcement learning from human feedback (RLHF) has emerged as a powerful paradigm for aligning generative models with human preferences.
By Eric Zhu, Abhinav Shrivastava, Soumik Mukhopadhyay
The paper introduces DRILL, a bilevel training framework that optimizes world models for policy improvement rather than pixel accuracy. DRILL-VWI reweights prediction errors using policy visitation and model confidence, while DRILL-IMG employs meta‑gradients to directly maximize downstream policy returns. Experiments on manipulation simulators show DRILL-IMG boosts success rates by 13.1 points over WMPO and 7.9 points over RLVR-World, demonstrating that world models can be more effectively trained using policy gradients.
arXiv:2606. 02884v1 Announce Type: cross Abstract: Reward guidance algorithms steer a learned generative process toward the reward-tilted measure at inference time.
By Sanjit Dandapanthula, Nicholas M. Boffi