On the Step Length Confounding in LLM Reasoning Data Selection

TL;DR

This paper identifies a bias in data selection for LLM reasoning datasets, where longer reasoning steps are favored, and proposes methods to mitigate this confounding effect, improving dataset quality.

Contribution

The paper reveals the step length confounding issue in naturalness-based data selection and introduces two novel methods to address it, enhancing reasoning dataset construction.

Findings

The confounding bias favors longer reasoning steps in data selection.

Proposed methods effectively reduce the bias across multiple models and benchmarks.

Mitigating step length confounding improves the quality of reasoning datasets.

Abstract

Large reasoning models have recently demonstrated strong performance on complex tasks that require long chain-of-thought reasoning, through supervised fine-tuning on large-scale and high-quality datasets. To construct such datasets, existing pipelines generate long reasoning data from more capable Large Language Models (LLMs) and apply manually heuristic or naturalness-based selection methods to filter high-quality samples. Despite the proven effectiveness of naturalness-based data selection, which ranks data by the average log probability assigned by LLMs, our analysis shows that, when applied to LLM reasoning datasets, it systematically prefers samples with longer reasoning steps (i.e., more tokens per step) rather than higher-quality ones, a phenomenon we term step length confounding. Through quantitative analysis, we attribute this phenomenon to low-probability first tokens in reasoning steps; longer steps dilute their influence, thereby inflating the average log probabilities. To address this issue, we propose two variant methods: ASLEC-DROP, which drops first-token probabilities when computing average log probability, and ASLEC-CASL, which applies a causal debiasing regression to remove the first tokens' confounding effect. Experiments across four LLMs and five evaluation benchmarks demonstrate the effectiveness of our approach in mitigating the step length confounding problem.