← News

ResearchSources: arXiv · 18 Nov 2026

Ambient SFT uses the frozen base model as a noisy teacher to curb fine-tuning overfitting

Researchers behind a new arXiv paper propose Ambient SFT, a fine-tuning method that constrains a post-trained model to match its pre-trained version at a coarse level, reporting gains over standard SFT.

In plain words

When a language model is fine-tuned on a small dataset, it can overfit and lose abilities it learned during pre-training. Ambient SFT adds a constraint: the fine-tuned model must roughly match the original pre-trained model's predictions, using a temperature setting to control how much the original model is trusted. The authors report better results than standard fine-tuning across several task types.

What was announced

The authors frame the problem as a data scarcity issue: high-quality post-training datasets are hard to collect, so Supervised Fine-Tuning usually runs on far smaller datasets than pre-training, causing overfitting and catastrophic forgetting of the tails of the data distribution. Their framework, Ambient SFT, addresses this by using a frozen copy of the pre-trained model as a noisy teacher that the fine-tuned model must match at a coarse level.

Technically, the method requires that high-temperature versions of the fine-tuned model's conditional distributions match those of the pre-trained model. The temperature hyperparameter sets the level of trust in the pre-trained model: as it approaches infinity the method reverts to regular SFT, while at a temperature of one it becomes knowledge distillation. The authors report that the approach outperforms standard SFT on domain adaptation, instruction-tuning, and reasoning, with up to 21% average relative improvement. These claims come from the paper's abstract and have not been independently verified.

What it does not show

  • The evidence is limited to the paper's abstract; experimental setup, baselines, and evaluation details are not available in the supplied passages.
  • The reported improvements, including the 21% figure, are the authors' own claims and have not been independently confirmed.

Sources

Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 19 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.