{"id":"arxiv/2603.07887","url":"https://arxiv.org/abs/2603.07887","tags":["cs.LG","cs.AI","cs.CL","math.ST","stat.ML"],"score":0,"title":"Reject, Resample, Repeat: Understanding Parallel Reasoning in Language Model Inference","domain":"cs.LG","source":"arxiv","authors":["Noah Golowich","Fan Chen","Dhruv Rohatgi","Raghav Singhal","Carles Domingo-Enrich","Dylan J. Foster","Akshay Krishnamurthy"],"reviews":0,"subject":"cs.LG","summary":"Inference-time methods that aggregate and prune multiple samples have emerged as a powerful paradigm for steering large language models, yet we lack any principled understanding of their accuracy-cost tradeoffs. In this paper, we introduce a route to rigorously study such approaches using the lens of *particle filtering* algorithms such as Sequential Monte Carlo (SMC). Given a base language model and a *process reward model* estimating expected terminal rewards, we ask: *how accurately can we sample from a target distribution given some number of process reward evaluations?* Theoretically,…","category":"research","comments":0,"disputed":false,"division":"cs","mentions":0,"discussed":0,"published":"2026-03-09","reactions":{}}