SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2502.01384 · 2025

Fine-Tuning Discrete Diffusion Models with Policy Gradient Methods

arXiv · PDF · Open in the Atlas

Code that ran

We lifted 8 functions out of this paper's own repositories and ran 1 of them in a sandbox. "Ran" means the function executed on a synthesized input and returned a value. It is not a reproduction of the paper's results.

RepositoryRoleRan
ozekri/SEPO canonical 1 of 8
FunctionStatusWhere it lives
fsspec_exists Ran ozekri/SEPO/GRPO_MDLM_DNA/utils.py
code served (permissive licence) · get_code("dec1de00023a923b")
batch_dna_detokenize Not yet run ozekri/SEPO/GRPO_MDLM_DNA/dataloader_gosai.py
code served (permissive licence) · get_code("4ec57ee85c940684")
binary_discretization Not yet run ozekri/SEPO/GRPO_MDLM_DNA/noise_schedule.py
code served (permissive licence) · get_code("9b9814cca50339d7")
dna_detokenize Not yet run ozekri/SEPO/GRPO_MDLM_DNA/dataloader_gosai.py
code served (permissive licence) · get_code("b988ba95b058558a")
dna_tokenize Not yet run ozekri/SEPO/GRPO_MDLM_DNA/dataloader_gosai.py
code served (permissive licence) · get_code("733f359f0f876fb7")
fsspec_listdir Not yet run ozekri/SEPO/GRPO_MDLM_DNA/utils.py
code served (permissive licence) · get_code("db0136eb80a50383")
get_logger Not yet run ozekri/SEPO/GRPO_MDLM_DNA/utils.py
code served (permissive licence) · get_code("60dbd956810fafb6")
get_noise Not yet run ozekri/SEPO/GRPO_MDLM_DNA/noise_schedule.py
code served (permissive licence) · get_code("fe6314378d2def00")

Repositories linked to this paper

Some links come from the archived Papers with Code dataset (CC BY-SA 4.0): attribution and licence.

Abstract

Discrete diffusion models have recently gained significant attention due to their ability to process complex discrete structures for language modeling. However, fine-tuning these models with policy gradient methods, as is commonly done in Reinforcement Learning from Human Feedback (RLHF), remains a challenging task. We propose an efficient, broadly applicable, and theoretically justified policy gradient algorithm, called Score Entropy Policy Optimization (\SEPO), for fine-tuning discrete diffusion models over non-differentiable rewards. Our numerical experiments across several discrete generative tasks demonstrate the scalability and efficiency of our method. Our code is available at https://github.com/ozekri/SEPO.

For agents

The same record, over MCP at https://syntology.ai/mcp:

get_harvested_code_for_paper("2502.01384")
get_code_for_paper("2502.01384")
have("2502.01384")

Connect an agent — have() is free.