ZJU team proposes DTCA for token-level credit assignment in diffusion LM reinforcement learning
OmniAI-ZJU · hf · 2026-10-29
The OmniAI Group of ZJU ACES Lab, with NUS and SAP, released the DTCA (Denoising Token Credit Assignment) paper addressing credit assignment in RL for diffusion language models (DLMs).
- Problem: DLMs generate text via iterative denoising with bidirectional context and parallel token refinement, but the iterative process makes credit assignment hard—final outcome rewards reveal little about individual token contributions.
- Method: DTCA estimates each token's contribution by measuring the improvement in prediction confidence of remaining masked positions after the token commits, then groups these signals by part-of-speech categories to select informative denoising timesteps and assign token-level learning signals during RL training.
- Results: Boosts mathematical reasoning performance in RL-trained DLMs.
More from Research
- DeepMind's Jon Barron: every softmax and cross-entropy implementation has two subtle bugs — jon_barron · 2026-10-29
- H-Net update shows 3-5x efficiency gains in raw-byte language modeling — omarsar0 · 2026-10-29
- Math's existential crisis is about Lean crossing a threshold, not AI — idanbeck · 2026-10-29
- Nature Communications: diffusion models emerge from dense associative memory, with three computational phases — DimaKrotov · 2026-10-29
- Falcon: lightweight streaming audio-visual speaker extraction with RealSSA benchmark — thu-ml · 2026-10-29
- Dev post-trains Qwen3-4B into a tool-using agent: 66 MB llama.cpp LoRA runs on a Mac, val 0.51→0.83 — DesignerIsland5524 · 2026-10-29