Masked diffusion language models (dLMs) are emerging as a parallel alternative to autoregressive models for complex reasoning tasks. But they face a distinct credit-assignment problem: a small number of early commitments during denoising sharply reduce uncertainty over the remaining masked positions and end up shaping much of the final response.
Most post-training recipes for dLMs ignore this signal. They typically train on the final text or assign rewards to whole denoising steps, rather than selecting the individual commitments that actually determine the output.
To address this, researchers introduce Pivot-SD, an efficient offline self-distillation framework that supervises only these high-impact commitments, which the authors call pivots. Pivots are selected using an information-gain metric that measures how much uncertainty reduction each commitment produces over the remaining masked positions.
The training scheme is asymmetric: pivots drawn from successful trajectories are trained with cross-entropy, while pivots from failed trajectories receive targeted unlikelihood training, leaving the rest of the failed trajectory untouched. This avoids spending compute on tokens that do not materially shape the response.
According to the authors, Pivot-SD improves LLaDA-8B-Instruct over full-sequence SFT and budget-matched diffusion RL baselines across math and code benchmarks, using only 200 questions with four rollouts each. The work is described as an EMNLP 2026 Oral presentation.