DTCA Lifts LLaDA-8B RL Math Scores +9 on MATH500 via POS Credit

A new method called Denoising Token Credit Assignment (DTCA) raises LLaDA-8B-Instruct's math reasoning by up to 9.0 points on MATH500 and 5.3 on GSM8K over the base model, according to a preprint posted to arXiv by Hongze Chen and co-authors.

The problem the authors target is specific to diffusion language models, which the paper describes as systems that "generate text through iterative denoising, enabling bidirectional context modeling and parallel token refinement." Unlike autoregressive setups, "the iterative denoising process makes credit assignment more challenging, as final outcome rewards provide limited information about how different tokens contribute to the final output." DTCA's answer is to estimate each token's contribution "by measuring the improvement in prediction confidence of remaining masked positions after token commitment," then group those signals by part-of-speech category. Those weights are used to pick informative denoising timesteps and assign token-level learning signals during RL, with a further calibration against token-level entropy changes induced by policy updates.

The abstract reports results on two math benchmarks and one base model. It does not break out per-POS numbers, show ablations, or compare DTCA against RL recipes designed for autoregressive LMs at the same scale.