Rethinking Easy-to-Hard: Limits of Curriculum Learning in Post-Training for Deductive Reasoning

arXiv cs.CL / 3/31/2026

💬 OpinionIdeas & Deep AnalysisModels & Research

共有:

Key Points

The paper systematically tests curriculum learning (CL) for post-training of LLMs on synthetic arithmetic and logical benchmarks where difficulty is defined by reasoning complexity rather than surface proxies.
Contrary to the intuition that ordering from easy to hard should improve generalization in compositional/deductive reasoning, the study finds no robust accuracy or response-length gains from difficulty-based example sequencing versus random sampling.
The negative effect holds across multiple model families and curriculum schedules, indicating the result is not dependent on a specific architecture or curriculum design.
Findings persist across both supervised fine-tuning (SFT) and reinforcement learning (RL) post-training, suggesting limited practical value of CL ordering for compositional generalization in this setting.
The authors conclude that, for deductive reasoning post-training, the specific order of training examples appears to play a negligible role in achieving compositional generalization, challenging common CL practices.

Abstract

Curriculum learning (CL), motivated by the intuition that learning in increasing order of difficulty should ease generalization, is commonly adopted both in pre-training and post-training of large language models (LLMs). The intuition of CL is particularly compelling for compositional reasoning, where complex problems are built from elementary inference rules; however, the actual impact of CL on such tasks remains largely underexplored. We present a systematic empirical study of CL for post-training of LLMs, using synthetic arithmetic and logical benchmarks where difficulty is characterized by reasoning complexity rather than surface-level proxies. Surprisingly, across multiple model families and curriculum schedules, we find no robust advantage in difficulty-based sequencing over standard random sampling in either accuracy or response length. These findings persist across both supervised fine-tuning (SFT) and reinforcement learning (RL) methods. Our study suggests that, in the context of deductive reasoning, the specific ordering of training examples plays a negligible role in achieving compositional generalization, challenging the practical utility of curriculum-based post-training.