From Particles to Tokens: Structural Analogies and Open Gaps Between Sequential Monte Carlo and Sampling-Based Reasoning in Language Models

Recent work on test-time compute for large language models (LLMs) has converged on methods that generate, score, and aggregate multiple reasoning trajectories. Independently, the Sequential Monte Carlo (SMC) literature has developed rigorous theory for traversing populations of weighted particles through sequences of distributions, with proven guarantees on normalizing-constant estimation and handling of multimodal targets. This paper identifies precise structural analogies between these two bodies of work—and, equally importantly, documents where the analogy breaks down or remains unestablished. We show that autoregressive chain-of-thought generation admits a product-of-conditionals factorization (established in the DiffCoT preprint) that is superficially similar to sequential importance sampling recursions, but that no cited source formally establishes isomorphism between the two. We document that power-distribution sampling (proposed in the 'Reasoning with Sampling' preprint) shares motivational language with temperature-annealed SMC (analyzed in Salomone et al.), yet no formal derivation of one from the other exists in the available literature. We further note that DiffCoT's sliding diffusion window addresses error-accumulation in multi-step reasoning in ways that are evocative of SMC mutation kernels, but this interpretation is not advanced by the primary source. By mapping both the genuine analogies and the open theoretical gaps, we aim to provide a precise agenda for future work that could rigorously import SMC guarantees into LLM reasoning.

Authorship: Saluca Agentic AI Research Team (Saluca LLC). AI-drafted from arXiv preprint corpus on the date in the filename.

Cited arXiv preprints: 2605.30321v1, 2605.30327v1, 2605.30343v1