Post-Training as Reweighting: A Stochastic View of Reasoning Trajectories in Language Models

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Bu, Dake, Huang, Wei, Han, Andi, Nitanda, Atsushi, Xue, Bo, Zhang, Qingfu, Wong, Hau-San, Suzuki, Taiji
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866908771958128640
author Bu, Dake
Huang, Wei
Han, Andi
Nitanda, Atsushi
Xue, Bo
Zhang, Qingfu
Wong, Hau-San
Suzuki, Taiji
author_facet Bu, Dake
Huang, Wei
Han, Andi
Nitanda, Atsushi
Xue, Bo
Zhang, Qingfu
Wong, Hau-San
Suzuki, Taiji
contents Foundation models encode rich structural knowledge but often rely on post-training procedures to adapt their reasoning behavior to specific tasks. Popular approaches such as reinforcement learning with verifiable rewards (RLVR) and inference-time reward aggregation are typically analyzed from a performance perspective, leaving their effects on the underlying reasoning distribution less understood. In this work, we study post-training reasoning from a stochastic trajectory viewpoint. Following Kim et al. (2025), we model reasoning steps of varying difficulty as Markov transitions with different probabilities, and formalize reasoning processes using tree-structured Markov chains. Within this framework, pretraining corresponds to discovering the reasoning structure, while post-training primarily reweights existing chains of thought. We show that both RLVR and inference-time reward aggregation concentrate probability mass on a small number of high-probability trajectories, leading to the suppression of rare but essential reasoning paths. As a consequence, solving hard instances often depends on low-probability trajectories already present in the base model. We further prove that exploration-oriented mechanisms, such as rejecting easy instances and applying KL regularization, help preserve these rare trajectories. Empirical simulations support our theoretical analysis.
format Preprint
id arxiv_https___arxiv_org_abs_2511_07368
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Post-Training as Reweighting: A Stochastic View of Reasoning Trajectories in Language Models
Bu, Dake
Huang, Wei
Han, Andi
Nitanda, Atsushi
Xue, Bo
Zhang, Qingfu
Wong, Hau-San
Suzuki, Taiji
Machine Learning
Artificial Intelligence
Foundation models encode rich structural knowledge but often rely on post-training procedures to adapt their reasoning behavior to specific tasks. Popular approaches such as reinforcement learning with verifiable rewards (RLVR) and inference-time reward aggregation are typically analyzed from a performance perspective, leaving their effects on the underlying reasoning distribution less understood. In this work, we study post-training reasoning from a stochastic trajectory viewpoint. Following Kim et al. (2025), we model reasoning steps of varying difficulty as Markov transitions with different probabilities, and formalize reasoning processes using tree-structured Markov chains. Within this framework, pretraining corresponds to discovering the reasoning structure, while post-training primarily reweights existing chains of thought. We show that both RLVR and inference-time reward aggregation concentrate probability mass on a small number of high-probability trajectories, leading to the suppression of rare but essential reasoning paths. As a consequence, solving hard instances often depends on low-probability trajectories already present in the base model. We further prove that exploration-oriented mechanisms, such as rejecting easy instances and applying KL regularization, help preserve these rare trajectories. Empirical simulations support our theoretical analysis.
title Post-Training as Reweighting: A Stochastic View of Reasoning Trajectories in Language Models
topic Machine Learning
Artificial Intelligence
url https://arxiv.org/abs/2511.07368