Training LLMs for Divide-and-Conquer Reasoning Elevates Test-Time Scalability

Fuente: arXiv
Enregistré dans:
Détails bibliographiques
Auteurs principaux: Liang, Xiao, Li, Zhong-Zhi, Lin, Zhenghao, Jiang, Eric Hancheng, Zhang, Hengyuan, Shen, Yelong, Chang, Kai-Wei, Wu, Ying Nian, Gong, Yeyun, Chen, Weizhu
Format: Preprint
Publié: 2026
Sujets:
Accès en ligne:
Tags: Ajouter un tag
Pas de tags, Soyez le premier à ajouter un tag!
_version_ 1866917242517585920
author Liang, Xiao
Li, Zhong-Zhi
Lin, Zhenghao
Jiang, Eric Hancheng
Zhang, Hengyuan
Shen, Yelong
Chang, Kai-Wei
Wu, Ying Nian
Gong, Yeyun
Chen, Weizhu
author_facet Liang, Xiao
Li, Zhong-Zhi
Lin, Zhenghao
Jiang, Eric Hancheng
Zhang, Hengyuan
Shen, Yelong
Chang, Kai-Wei
Wu, Ying Nian
Gong, Yeyun
Chen, Weizhu
contents Large language models (LLMs) have demonstrated strong reasoning capabilities through step-by-step chain-of-thought (CoT) reasoning. Nevertheless, at the limits of model capability, CoT often proves insufficient, and its strictly sequential nature constrains test-time scalability. A potential alternative is divide-and-conquer (DAC) reasoning, which decomposes a complex problem into subproblems to facilitate more effective exploration of the solution. Although promising, our analysis reveals a fundamental misalignment between general-purpose post-training and DAC-style inference, which limits the model's capacity to fully leverage this potential. To bridge this gap and fully unlock LLMs' reasoning capabilities on the most challenging tasks, we propose an end-to-end reinforcement learning (RL) framework to enhance their DAC-style reasoning capacity. At each step, the policy decomposes a problem into a group of subproblems, solves them sequentially, and addresses the original one conditioned on the subproblem solutions, with both decomposition and solution integrated into RL training. Under comparable training, our DAC-style framework endows the model with a higher performance ceiling and stronger test-time scalability, surpassing CoT by 8.6% in Pass@1 and 6.3% in Pass@32 on competition-level benchmarks.
format Preprint
id arxiv_https___arxiv_org_abs_2602_02477
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Training LLMs for Divide-and-Conquer Reasoning Elevates Test-Time Scalability
Liang, Xiao
Li, Zhong-Zhi
Lin, Zhenghao
Jiang, Eric Hancheng
Zhang, Hengyuan
Shen, Yelong
Chang, Kai-Wei
Wu, Ying Nian
Gong, Yeyun
Chen, Weizhu
Computation and Language
Large language models (LLMs) have demonstrated strong reasoning capabilities through step-by-step chain-of-thought (CoT) reasoning. Nevertheless, at the limits of model capability, CoT often proves insufficient, and its strictly sequential nature constrains test-time scalability. A potential alternative is divide-and-conquer (DAC) reasoning, which decomposes a complex problem into subproblems to facilitate more effective exploration of the solution. Although promising, our analysis reveals a fundamental misalignment between general-purpose post-training and DAC-style inference, which limits the model's capacity to fully leverage this potential. To bridge this gap and fully unlock LLMs' reasoning capabilities on the most challenging tasks, we propose an end-to-end reinforcement learning (RL) framework to enhance their DAC-style reasoning capacity. At each step, the policy decomposes a problem into a group of subproblems, solves them sequentially, and addresses the original one conditioned on the subproblem solutions, with both decomposition and solution integrated into RL training. Under comparable training, our DAC-style framework endows the model with a higher performance ceiling and stronger test-time scalability, surpassing CoT by 8.6% in Pass@1 and 6.3% in Pass@32 on competition-level benchmarks.
title Training LLMs for Divide-and-Conquer Reasoning Elevates Test-Time Scalability
topic Computation and Language
url https://arxiv.org/abs/2602.02477