Search papers, labs, and topics across Lattice.
The paper introduces dVoting, a novel test-time technique for diffusion large language models (dLLMs) that leverages their parallel token generation capability to improve reasoning performance. dVoting iteratively refines dLLM outputs by identifying inconsistent token predictions across multiple samples for the same prompt and regenerating these uncertain tokens through a voting mechanism. Experiments on GSM8K, MATH500, ARC-C, and MMLU demonstrate consistent performance improvements, highlighting the effectiveness of dVoting in enhancing the reasoning capabilities of dLLMs without additional training.
dVoting unlocks significant reasoning gains for diffusion LMs at test time by iteratively refining only the most uncertain tokens, sidestepping the computational bottleneck of full re-sampling.
Diffusion Large Language Models (dLLMs) represent a new paradigm beyond autoregressive modeling, offering competitive performance while naturally enabling a flexible decoding process. Specifically, dLLMs can generate tokens at arbitrary positions in parallel, endowing them with significant potential for parallel test-time scaling, which was previously constrained by severe inefficiency in autoregressive modeling. In this work, we introduce dVoting, a fast voting technique that boosts reasoning capability without training, with only an acceptable extra computational overhead. dVoting is motivated by the observation that, across multiple samples for the same prompt, token predictions remain largely consistent, whereas performance is determined by a small subset of tokens exhibiting cross-sample variability. Leveraging the arbitrary-position generation capability of dLLMs, dVoting performs iterative refinement by sampling, identifying uncertain tokens via consistency analysis, regenerating them through voting, and repeating this process until convergence. Extensive evaluations demonstrate that dVoting consistently improves performance across various benchmarks. It achieves gains of 6.22%-7.66% on GSM8K, 4.40%-7.20% on MATH500, 3.16%-14.84% on ARC-C, and 4.83%-5.74% on MMLU. Our code is available at https://github.com/fscdc/dVoting
NUS