Saved in:
Bibliographic Details
Main Authors: Ran-Milo, Yuval, Alexander, Yotam, Mendel, Shahar, Cohen, Nadav
Format: Preprint
Published: 2026
Subjects:
Online Access:https://arxiv.org/abs/2601.15158
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866915768275304448
author Ran-Milo, Yuval
Alexander, Yotam
Mendel, Shahar
Cohen, Nadav
author_facet Ran-Milo, Yuval
Alexander, Yotam
Mendel, Shahar
Cohen, Nadav
contents Transformers trained via Reinforcement Learning (RL) with outcome-based supervision can spontaneously develop the ability to generate intermediate reasoning steps (Chain-of-Thought). Yet the mechanism by which sparse rewards drive policy gradient to discover such systematic reasoning remains poorly understood. We address this by analyzing the policy gradient dynamics of single-layer Transformers on a synthetic graph traversal task that cannot be solved without Chain-of-Thought but admits a simple iterative solution. We prove that despite training solely on final-answer correctness, policy gradient drives the Transformer to converge to a structured, interpretable algorithm that iteratively traverses the graph vertex-by-vertex. We characterize the distributional properties required for this emergence, identifying the critical role of "simple examples": instances requiring fewer reasoning steps. When the training distribution places sufficient mass on these simpler examples, the Transformer learns a generalizable traversal strategy that extrapolates to longer chains; when this mass vanishes, policy gradient learning becomes infeasible. We corroborate our theoretical results through experiments on synthetic data and with real-world language models on mathematical reasoning tasks, validating that our theoretical findings carry over to practical settings.
format Preprint
id arxiv_https___arxiv_org_abs_2601_15158
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Outcome-Based RL Provably Leads Transformers to Reason, but Only With the Right Data
Ran-Milo, Yuval
Alexander, Yotam
Mendel, Shahar
Cohen, Nadav
Machine Learning
Artificial Intelligence
Transformers trained via Reinforcement Learning (RL) with outcome-based supervision can spontaneously develop the ability to generate intermediate reasoning steps (Chain-of-Thought). Yet the mechanism by which sparse rewards drive policy gradient to discover such systematic reasoning remains poorly understood. We address this by analyzing the policy gradient dynamics of single-layer Transformers on a synthetic graph traversal task that cannot be solved without Chain-of-Thought but admits a simple iterative solution. We prove that despite training solely on final-answer correctness, policy gradient drives the Transformer to converge to a structured, interpretable algorithm that iteratively traverses the graph vertex-by-vertex. We characterize the distributional properties required for this emergence, identifying the critical role of "simple examples": instances requiring fewer reasoning steps. When the training distribution places sufficient mass on these simpler examples, the Transformer learns a generalizable traversal strategy that extrapolates to longer chains; when this mass vanishes, policy gradient learning becomes infeasible. We corroborate our theoretical results through experiments on synthetic data and with real-world language models on mathematical reasoning tasks, validating that our theoretical findings carry over to practical settings.
title Outcome-Based RL Provably Leads Transformers to Reason, but Only With the Right Data
topic Machine Learning
Artificial Intelligence
url https://arxiv.org/abs/2601.15158