arXiv AI By Jaemoo Choi, Yuchen Zhu, Wei Guo, Petr Molodyk, Bo Yuan, Jinbin Bai, Yi Xin, Molei Tao, Yongxin Chen

Rethinking the Design Space of Reinforcement Learning for Diffusion Models: On the Importance of Likelihood Estimation Beyond Loss Design

Read the original on arXiv AI →

The paper investigates how reinforcement learning can be effectively applied to diffusion models for visual tasks, focusing on the role of likelihood estimation. By systematically separating policy‑gradient objectives, likelihood estimators, and rollout sampling schemes, the authors find that using an evidence lower bound (ELBO) based likelihood estimator computed from the final generated sample is the key factor for stable and efficient RL optimization, outweighing the choice of loss function. Experiments on SD 3.5 Medium across multiple reward benchmarks confirm that this approach improves GenEval scores from 0.24 to 0.95 in 90 GPU hours, outperforming existing methods such as FlowGRPO and the current state‑of‑the‑art without reward hacking.

Machine-generated by The Flow from the publisher's headline and feed description — not written or checked by a human. The full article lives at arXiv AI.

arXiv AI
Oct 1

CAST: Causal Advantage-Structured Training with Spatially Grounded Compositional Rewards for Diffusion Models

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
arXiv AI
Oct 27

VaLoP: Value Gradient Lookahead and Pullback for Reward-Guided Diffusion

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
Hugging Face Trending Papers
Oct 15

Routing Policy Optimization for Online Diffusion Reinforcement Learning

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.

arXiv Machine Learning
Aug 27

A Comedy of Estimators: On KL Regularization in RL Training of LLMs

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