Swordsman: Entropy-Driven Adaptive Block Partition for Efficient Diffusion Language Models

Block-wise decoding effectively improves the inference speed and quality in diffusion language models (DLMs) by combining inter-block sequential denoising and intra-block parallel unmasking. However, existing block-wise decoding methods typically partition blocks in a rigid and fixed manner, which inevitably fragments complete semantic or syntactic constituents, leading to suboptimal performance. Inspired by the entropy reduction hypothesis (ERH), we recognize that constituent boundaries offer greater opportunities for uncertainty reduction, which motivates us to employ entropy analysis for identifying constituent boundaries. Therefore, we propose Swordsman, an entropy-driven adaptive block-wise decoding framework for DLMs. Swordsman adaptively partitions blocks by identifying entropy shifts between adjacent tokens to better align with semantic or syntactic constituent boundaries. In addition, Swordsman dynamically adjusts unmasking thresholds conditioned on the real-time unmasking status within a block, further improving both efficiency and stability. As a training-free framework, supported by KV Cache, Swordsman demonstrates state-of-the-art performance across extensive evaluations.

💡 Research Summary

The paper introduces Swordsman, a training‑free framework that improves both inference speed and generation quality of diffusion language models (DLMs) through entropy‑driven adaptive block partitioning.
Traditional block‑wise decoding methods accelerate DLMs by reusing KV caches and parallel unmasking within fixed‑size blocks, but the rigid boundaries often cut across semantic or syntactic constituents. This misalignment fragments strongly correlated tokens, raises uncertainty at block edges, and degrades the quality of the generated text.
Swordsman is built on the Entropy Reduction Hypothesis (ERH), which posits that the diffusion generation process progressively reduces uncertainty and that constituent boundaries correspond to sharp increases in predictive entropy. The authors formalize this by defining token‑wise predictive entropy (H_i) and the entropy shift (\Delta H_i = H_{i+1} - H_i). Within a constituent, the candidate vocabulary size changes smoothly, yielding small (|\Delta H_i|) (bounded by a constant (\delta)). At a constituent boundary, the candidate space expands dramatically (ratio (\rho \gg 1)), producing a large entropy jump (\Delta H_{\text{boundary}} \approx \log \rho). Consequently, detecting local maxima of (\Delta H_i) reliably identifies natural linguistic boundaries.
The adaptive partitioning algorithm works iteratively: after decoding a block (B_k), the KV cache is updated, the remaining masked positions are forward‑propagated again to obtain fresh entropy estimates, and the next block’s right boundary is set to the position with the maximal (\Delta H_i) that exceeds a minimal shift threshold (\tau_{\min}). This process naturally yields variable‑length blocks that align with semantic constituents and avoids over‑segmentation in the tail of the sequence where uncertainty has already converged.
Because different blocks exhibit varying unmasking difficulty, Swordsman replaces the static confidence threshold (\tau) with a dynamic threshold that adapts to the real‑time unmasking ratio inside the current block. When many tokens have already been unmasked, the threshold is raised to be conservative; when few have been unmasked, it is lowered to exploit parallelism. This “difficulty‑aware parallel unmasking” stabilizes decoding while preserving speed gains.
Extensive experiments demonstrate the effectiveness of the approach. On the GSM8K benchmark, Swordsman achieves an 8.79× speedup and improves accuracy from 77.40 % to 81.50 % compared with vanilla LLaDA. Against the LLaDA‑based Fast‑dLLM, it raises HumanEval accuracy from 35.59 % to 43.90 % while maintaining comparable latency. The method consistently outperforms prior block‑wise baselines that rely on fixed block sizes or punctuation‑based segmentation.
The paper’s contributions are fourfold: (1) a principled entropy‑based boundary detection mechanism that aligns blocks with linguistic constituents, (2) a dynamic, difficulty‑aware unmasking threshold within each block, (3) a training‑free design that can be plugged into existing DLM pipelines without additional model updates, and (4) comprehensive empirical validation showing state‑of‑the‑art speed‑quality trade‑offs.
In summary, Swordsman demonstrates that leveraging predictive entropy as a real‑time signal enables adaptive block partitioning, which in turn resolves the core inefficiency of fixed‑size block decoding in diffusion language models. This insight opens a path toward more responsive, high‑quality generative AI systems that can operate at scale in latency‑sensitive applications.