Inference-time sequential Monte Carlo (SMC) decoding helps masked diffusion language models reason more accurately by resampling a population of partial generations according to a reward. But the resampling step can quietly destroy diversity: multinomial resampling duplicates high-weight particles and drops lineages that might later recover, even when the effective sample size metric looks fine. The authors of a new arXiv paper argue that this loss is invisible to standard diagnostics and propose a fix called lineage-preserving SMC (LP-SMC).

LP-SMC replaces multinomial resampling with Chopthin resampling, which bounds the ratio between the largest and smallest weights rather than equalizing them, and adds an input weight floor so low-scoring lineages are not eliminated too early. The method works with either internal rewards (the model's own confidence) or external process reward models. The authors tested it on mathematical-reasoning and code-generation benchmarks using two base models, LLaDA-1.5 and Dream-7B.

Results show that LP-SMC improves oracle coverage—the fraction of problems with at least one correct particle—in every setting, by 1.8 to 25.6 percentage points. End-to-end accuracy under a common reward-model selector improved in eleven of twelve settings, with gains up to 7.9 points on LLaDA-1.5 and 18.3 points on Dream-7B. The coverage gain persisted even with an external reward model, suggesting that diversity-preserving resampling and reward guidance contribute independently to inference-time search. The paper is available on arXiv under identifier 2610.25932.