Sampling as Reasoning, Reasoning as Sampling: Structural Bridges and Their Limits Across Generative Inference
Description
Version 2 — revised in response to an external structural review and an automated critique pass. See "Response to Review" appendix in the PDF for the change log.
Several recent lines of work—antithetic noise in diffusion models, learned proposal distributions for sequential Monte Carlo, power-distribution sampling for reasoning elicitation, autoregressive factorization of chain-of-thought traces, and faithfulness failures in formal verification—have each been developed within their own technical communities. This paper offers a **heuristic reading** of structural analogies connecting these lines, while carefully marking where the analogies are grounded in primary sources and where they remain conjectural. We show that antithetic noise in diffusion models is a documented instance of classical antithetic-variate variance reduction, generalized by the same authors to quasi-Monte Carlo and randomized QMC estimators. We observe that learned inference networks for SMC instantiate amortized proposal design in directed graphical models, sharing a high-level motivation with other amortized inference approaches without establishing a direct formal bridge to chain-of-thought reasoning. We note that the autoregressive factorization of chain-of-thought traces structurally resembles sequential importance-sampling decompositions, while stopping short of claiming a formal identity. Finally, we document that formal verification guarantees proof validity but not formalization faithfulness, and that specification gaming in reasoning models constitutes an empirically documented phenomenon whose decision-theoretic interpretation remains conjectural. Throughout, we distinguish grounded structural parallels from speculative analogies, providing a map of what is and is not yet established. All cited sources are preprints and have not undergone formal peer review. ---
Authorship: Saluca Agentic AI Research Team (Saluca LLC). AI-drafted from arXiv preprint corpus on the date in the filename.
Cited arXiv preprints: 2605.30327v1
AI disclosure. This work was produced with an agentic AI research apparatus operated by Saluca Labs. The apparatus drafted, searched and analysed under direction. Cristian Ruvalcaba is the human author and is accountable for the content. No AI system is listed as an author or contributor, because authorship entails accountability that a model cannot hold; this disclosure is the credit, and it is deliberately the whole of it.