Saved in:
Bibliographic Details
Main Authors: Amani, Mohammad Hossein, Lotfi, Aryo, Baldwin, Nicolas Mario, Bengio, Samy, Farajtabar, Mehrdad, Abbe, Emmanuel, West, Robert
Format: Preprint
Published: 2025
Subjects:
Online Access:https://arxiv.org/abs/2506.18110
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866910037301002240
author Amani, Mohammad Hossein
Lotfi, Aryo
Baldwin, Nicolas Mario
Bengio, Samy
Farajtabar, Mehrdad
Abbe, Emmanuel
West, Robert
author_facet Amani, Mohammad Hossein
Lotfi, Aryo
Baldwin, Nicolas Mario
Bengio, Samy
Farajtabar, Mehrdad
Abbe, Emmanuel
West, Robert
contents Learning in the combinatorially large output space of sequence generation problems is challenging as providing expert demonstrations scales poorly with sequence length, and RL struggles with sparse rewards. Between dense demonstrations in supervised training and no demonstrations in reinforcement learning lies an underexplored regime: partial supervision. We ask whether some classes of sequence learning problems become efficiently learnable by exploiting this gap. We address this by introducing adaptive backtracking (AdaBack), a per-sample curriculum learning algorithm that reveals a partial prefix of the target output. The supervision length is adjusted dynamically for each sample based on the model's past reward signal, allowing it to incrementally learn to complete reasoning chains by conditioning on correct partial solutions. We investigate this intermediate regime between SFT and RL and argue that per-sample curriculum learning is more than a trade-off between efficiency and generality--it can succeed in tasks with long sequences of latent dependencies where SFT and RL both fail to generalize. Using a synthetic task with latent parity constraints, we show that AdaBack reliably solves problems that are otherwise intractable. On three mathematical reasoning benchmarks, DeepScaleR, MATH, and GSM8k, we find that AdaBack enables models to solve problems that RL alone cannot, acquiring new reasoning capabilities through incremental exposure to partial solutions.
format Preprint
id arxiv_https___arxiv_org_abs_2506_18110
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle RL for Reasoning by Adaptively Revealing Rationales
Amani, Mohammad Hossein
Lotfi, Aryo
Baldwin, Nicolas Mario
Bengio, Samy
Farajtabar, Mehrdad
Abbe, Emmanuel
West, Robert
Machine Learning
Artificial Intelligence
Learning in the combinatorially large output space of sequence generation problems is challenging as providing expert demonstrations scales poorly with sequence length, and RL struggles with sparse rewards. Between dense demonstrations in supervised training and no demonstrations in reinforcement learning lies an underexplored regime: partial supervision. We ask whether some classes of sequence learning problems become efficiently learnable by exploiting this gap. We address this by introducing adaptive backtracking (AdaBack), a per-sample curriculum learning algorithm that reveals a partial prefix of the target output. The supervision length is adjusted dynamically for each sample based on the model's past reward signal, allowing it to incrementally learn to complete reasoning chains by conditioning on correct partial solutions. We investigate this intermediate regime between SFT and RL and argue that per-sample curriculum learning is more than a trade-off between efficiency and generality--it can succeed in tasks with long sequences of latent dependencies where SFT and RL both fail to generalize. Using a synthetic task with latent parity constraints, we show that AdaBack reliably solves problems that are otherwise intractable. On three mathematical reasoning benchmarks, DeepScaleR, MATH, and GSM8k, we find that AdaBack enables models to solve problems that RL alone cannot, acquiring new reasoning capabilities through incremental exposure to partial solutions.
title RL for Reasoning by Adaptively Revealing Rationales
topic Machine Learning
Artificial Intelligence
url https://arxiv.org/abs/2506.18110