Amin Karimi Monsefi

Amin Karimi Monsefi

Ph.D. candidate at The Ohio State University. Efficient and controllable generative models — diffusion and flow matching across vision and language.

DACA-GRPO: Denoising-Aware Credit Assignment for Reinforcement Learning in Diffusion Language Models

Amin Karimi Monsefi, Dominic Culver, Nikhil Bhendawade, Lokesh Boominathan, Manuel R. Ciosici, Yizhe Zhang, Irina Belousova

arXiv, 2026

Figure from DACA-GRPO: Denoising-Aware Credit Assignment for Reinforcement Learning in Diffusion Language Models

Existing RL methods for diffusion language models treat every denoising step as equally important and rely on biased, high-variance likelihood estimates. DACA-GRPO adds Denoising Progress Scores, which extract per-token importance weights from intermediate predictions at no extra forward cost, and Stratified Masking Likelihood, which reduces mean-field bias. Applied on top of three GRPO base methods, it improves results across seven benchmarks spanning mathematical reasoning, code generation, constraint satisfaction and constrained generation.

Paper