# Rethinking Generalization in Reasoning SFT

> Plain SFT can generalize across domains in reasoning — once you fix optimization, data, and base-model strength.

- Org: Shanghai AI Lab / SJTU / Rice
- Date: 2026-04-08
- Category: paper, algorithm
- Source: https://arxiv.org/abs/2604.06628
- Canonical page: https://ai-tldr.dev/releases/rethink-sft-generalization/

## In short

Top-trending HF paper challenges the 'SFT memorizes, RL generalizes' narrative. Shows cross-domain reasoning gains from SFT are conditional on optimization, data quality, and base-model capability; releases code and datasets.

## What it is

A new Shanghai AI Lab / SJTU / Rice paper (arXiv 2604.06628) that pushes back on the popular claim 'SFT memorizes, RL generalizes' for reasoning models. It is currently the #1 trending paper on Hugging Face with 306 upvotes. The authors release code, 33 models, and four datasets (Math-CoT-44k with token-level logprobs, Math-CoT-20k, DeepSeek-R1-20k, and Countdown-CoT-20k) so the results are fully reproducible.

## How it works

The paper runs a conditional analysis over three axes — optimization dynamics, training data, and base-model capability — and finds that cross-domain generalization from long-CoT SFT follows a 'dip-and-recovery' pattern: during training the cross-domain score first drops, then recovers and improves, so short training runs underestimate how well SFT transfers. Verified long-CoT traces consistently help; low-quality solutions hurt. Stronger base models internalize transferable procedural patterns like backtracking even from a toy arithmetic game (Countdown), while weaker models imitate surface verbosity. They also document an asymmetric side effect: reasoning improves while safety behaviour degrades.

## Why it matters

The 'RL is the only path to generalization' framing has been shaping post-training pipelines for the last year. This paper argues the premise is partly an optimization artifact: with better data and longer training, plain SFT can match or beat the generalization RL papers claim as their own. That reopens a much cheaper, more stable recipe for building reasoning models — and the open datasets + logprobs make it easy for others to test the claim directly.

## Who it's for

Researchers and practitioners training reasoning models who were about to commit to an RL pipeline.

## Try it

github.com/Nebularaid2000/rethink_sft_generalization

## Links

- [Paper](https://arxiv.org/abs/2604.06628)
- [HF paper page](https://huggingface.co/papers/2604.06628)
- [Repo](https://github.com/Nebularaid2000/rethink_sft_generalization)
- [Collection](https://huggingface.co/collections/jasonrqh/rethink-sft-generalization)