Inference-time power sampling with Sequential Monte Carlo (SMC) can boost LLM reasoning without fine-tuning, but standard equal-weight resampling tends to prune low-weight trajectories, discarding potentially correct reasoning paths. A new method, Chopthin-Consensus Power Sampling (CCPS), replaces equal-weight resampling with the Chopthin resampler, which enforces an upper bound on the ratio between the largest and smallest weights and carries unequal weights forward. This preserves a richer set of distinct reasoning paths while keeping the weighted approximation unchanged in conditional expectation and guaranteeing a lower bound on post-resampling effective sample size.
To exploit the enriched population, CCPS adds a semantic-majority selection step: it merges token-identical final trajectories, clusters semantically equivalent answers, and returns the answer supported by the largest number of distinct trajectories. The authors evaluated CCPS across three open-weight models and five reasoning benchmarks. Chopthin alone increased oracle coverage in 13 of 15 settings, and when combined with semantic-majority selection, CCPS matched or exceeded the Power-SMC baseline's final-answer accuracy in 14 of 15 settings, with absolute gains of up to 10.6 percentage points.
The results indicate that diversity-preserving resampling and diversity-aware selection are complementary mechanisms for training-free LLM reasoning. The paper, accepted at the COLM 2026 Workshop on Efficient Reasoning, is available on arXiv with accompanying code.