Early Decisions Matter: Proximity Bias and Initial Trajectory Shaping in Non-Autoregressive Diffusion Language Models
For researchers working on non-autoregressive decoding in diffusion language models, this work addresses a critical failure mode and provides a practical solution to improve performance on reasoning and planning tasks.
The paper identifies a proximity bias in non-autoregressive diffusion language models that causes spatial error propagation, and proposes a minimal-intervention approach with a lightweight planner and temperature annealing to guide early token selection, achieving substantial improvements over heuristic baselines on reasoning and planning tasks without significant computational overhead.
Diffusion-based language models (dLLMs) have emerged as a promising alternative to autoregressive language models, offering the potential for parallel token generation and bidirectional context modeling. However, harnessing this flexibility for fully non-autoregressive decoding remains an open question, particularly for reasoning and planning tasks. In this work, we investigate non-autoregressive decoding in dLLMs by systematically analyzing its inference dynamics along the temporal axis. Specifically, we uncover an inherent failure mode in confidence-based non-autoregressive generation stemming from a strong proximity bias-the tendency for the denoising order to concentrate on spatially adjacent tokens. This local dependency leads to spatial error propagation, rendering the entire trajectory critically contingent on the initial unmasking position. Leveraging this insight, we present a minimal-intervention approach that guides early token selection, employing a lightweight planner and end-of-sequence temperature annealing. We thoroughly evaluate our method on various reasoning and planning tasks and observe substantial overall improvement over existing heuristic baselines without significant computational overhead.