Stanford AI Lab: QuasiMoTTo Cuts LLM Samples 47%

Stanford AI Lab introduced QuasiMoTTo to fix wasteful independent sampling during inference compute scaling. The method generates correlated samples that maintain marginal exactness from the LLM while boosting coverage and cutting redundancy. Results show the same benchmark performance with 25-47% fewer samples in test-time scaling and 50% fewer training steps in RL, according to co-authors Michael Y. Li and team. This approach directly targets the core inefficiency in parallel inference attempts that repeatedly rediscover identical solutions.