HeyHomees Lab Notes · No. 12 · ZJU-OmniAI · Published October 29, 2026
By HeyHomees Newsroom ·
In 30 seconds
- DTCA gives token-level credit signals for RL in diffusion language models.
- It improves MATH500 scores by up to 9.0 points and GSM8K by up to 5.3 points.
- Code and models are released under Apache License 2.0 and run on Linux with Python 3.10/3.11 and CUDA.
What is it?
DTCA is a framework for token-level credit assignment in reinforcement learning for diffusion language models. It measures token contributions by the improvement in prediction confidence of remaining masked positions after token commitment and groups signals by part-of-speech categories.
How does it work?
DTCA selects informative denoising timesteps using contribution weights, assigns token-level learning signals during RL training, and calibrates these signals according to token-level entropy changes caused by policy updates. The implementation builds on RLDF/Drift and incorporates ideas from dLLM-RL, Fast-dLLM, and OPEFO.
How well does it do?
Experiments on LLaDA-8B-Instruct show DTCA consistently improves mathematical reasoning, achieving gains of up to 9.0 points on MATH500 and 5.3 points on GSM8K over the base model.
What does it need to run?
The code runs on Linux with Python 3.10 or 3.11, requires a CUDA-compatible PyTorch installation, and uses Accelerate and DeepSpeed ZeRO-3 for training.
By the numbers
- MATH500 improvement: 9.0 points
- GSM8K improvement: 5.3 points
- License: Apache License 2.0
The honest read
The reported gains of up to 9.0 points on MATH500 and 5.3 points on GSM8K are self-reported by the authors and reflect improvements over their base LLaDA-8B-Instruct model under the specific evaluation settings described. The figures do not indicate absolute performance levels, nor do they compare DTCA to other RL or diffusion methods beyond the base model. No external benchmarks or third-party evaluations are provided. The code and experiments are limited to the LLaDA-8B-Instruct model and the listed math datasets; applicability to other models or tasks is not demonstrated.
What we don't know yet
- Exact model parameter count or download size is not specified.
- Performance on non-math tasks or with different generation budgets beyond those reported is not provided.
Sources
- Paper (arXiv abstract) (primary)
- Code, GitHub
Researched and drafted with AI from the sources above; every figure is checked against them by code before it is published, and a line that fails is dropped.