Rethinking Generalization in Reasoning SFT: A Conditional Analysis on Optimization, Data, and Model Capability

TL;DR

This paper reevaluates the generalization capabilities of reasoning supervised fine-tuning in large language models, emphasizing the roles of optimization, data quality, and model strength in cross-domain reasoning performance.

Contribution

It challenges the simplistic view that SFT memorizes and RL generalizes, showing that generalization depends on multiple factors including training dynamics, data, and model capability.

Findings

Cross-domain generalization is conditional and influenced by training dynamics.

Longer training can recover and improve initial performance dips.

Higher model capability leads to better transfer of reasoning skills.

Abstract

A prevailing narrative in LLM post-training holds that supervised finetuning (SFT) memorizes while reinforcement learning (RL) generalizes. We revisit this claim for reasoning SFT with long chain-of-thought (CoT) supervision and find that cross-domain generalization is not absent but conditional, jointly shaped by optimization dynamics, training data, and base-model capability. Some reported failures are under-optimization artifacts: cross-domain performance first degrades before recovering and improving with extended training (a dip-and-recovery pattern), so shorttraining checkpoints can underestimate generalization. Data quality and structure both matter: low-quality solutions broadly hurt generalization,while verified long-CoT traces yield consistent cross-domain gains. Model capability is essential: stronger models internalize transferable procedural patterns (e.g., backtracking) even from a toy arithmetic game, while weaker ones imitate surface verbosity. This generalization is asymmetric, however: reasoning improves while safety degrades, reframing the question from whether reasoning SFT generalizes to under what conditions and at what cost.