Keep the Lineages Alive: Diversity-Preserving SMC for Diffusion LLM Decoding

Key points

• The effective sample size (ESS) does not register this loss: multinomial resampling resets it to its maximum even as distinct lineages disappear.1
• To retain such lineages, we introduce lineage-preserving SMC (LP-SMC), which combines Chopthin resampling with an input weight floor.1
• Across mathematical-reasoning and code-generation benchmarks on two base models, LP-SMC raises oracle coverage, the fraction of problems with at least one correct particle, in every setting, by 1.8 to 25.6 points, and raises end-to-end accuracy under a common reward-model selector in eleven of twelve settings, by up to 7.9 points on LLaDA-1.5 and 18.3 on Dream-7B.1
• The coverage gain persists under the external reward model, which suggests that diversity-preserving resampling and the guidance signal contribute separately to inference-time search.1

Sources (1)

• [1]Keep the Lineages Alive: Diversity-Preserving SMC for Diffusion LLM Decoding
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 19, 07:30 AM

  Inference-time sequential Monte Carlo (SMC) decoding can improve the reasoning accuracy of masked diffusion language models (diffusion LLMs) without post-training by resampling a population of partial generations according to a reward.

  The effective sample size (ESS) does not register this loss: multinomial resampling resets it to its maximum even as distinct lineages disappear.

Extractive summary: sentences quoted from the sources.