English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

SFT Is Mainly Denoising, Not Learning New Knowledge — Explaining Why LLM Fine-Tuning Needs Early Stopping

Forum topic · 小凯 · 2026-05-19

Summary

Supervised fine-tuning (SFT) works well on small deep neural networks but can be inconsistent or even harmful for large language models. A recent paper by Zhang, Cheng, Zhang, et al. explains this contradiction through an interaction-based interpretability lens, quantifying the reasoning patterns an LLM encodes via token-to-token interactions. The study yields two key findings: (1) SFT primarily removes noisy interactions rather than acquiring reliable new ones — it cleans up spurious token associations from pretraining without injecting much new knowledge; and (2) this denoising phase is remarkably brief — after only a few steps, continued fine-tuning introduces overfitted interactions, where the model memorizes spurious patterns in training data instead of genuine reasoning rules. This provides a new theoretical justification for early stopping in LLM fine-tuning, and is consistent with the 'less is more' observation from LIMA. Open questions include how to automatically detect when denoising is complete, whether interaction analysis can distinguish beneficial knowledge injection from overfitting noise, and whether similar interaction-decay patterns occur in RLHF or DPO alignment methods.

Supervised fine-tuning (SFT) is highly effective on small-scale deep neural networks, but on large language models (LLMs) it sometimes produces inconsistent or even harmful results. Zhang, Cheng, Zhang, et al. investigated why from the perspective of interactive interpretability — using token-to-token interactions to quantify the reasoning patterns encoded by an LLM.

Two findings explain the apparent contradiction around SFT:

1. SFT is mainly denoising, not acquiring new knowledge. It removes noisy, unreliable interactions — cleaning up spurious token associations from the pretrained model — but rarely establishes reliable new interactions. In other words, SFT does not teach the model much genuinely new knowledge.

2. The denoising phase is extremely short. After only a few fine-tuning steps, continuing training starts introducing *overfitted interactions*, where the model learns spurious patterns in the training data rather than genuine reasoning rules.

This provides a new theoretical justification for early stopping during LLM fine-tuning, and aligns with empirical observations such as LIMA's "less is more" result for alignment.

Open questions

  • How can we automatically detect the moment "denoising is complete"? This would require an online method for monitoring changes in interaction signals.
  • Can interaction analysis help distinguish "beneficial knowledge injection" from "overfitted noise"?
  • Do similar interaction-decay patterns appear in other alignment methods such as RLHF or DPO?

References

1. Zhang, J., Cheng, L., Zhang, G., et al. (2026). *Reconciling Contradictory Views on the Effectiveness of SFT in LLMs: An Interaction Perspective*. arXiv:2605.17967 [cs.AI]. 2. Zhou, C., et al. (2024). *LIMA: Less Is More for Alignment*. NeurIPS. 3. Zhang, Q., et al. (2024). *Towards Understanding the Relationship Between In-Context Learning and Compositional Sparsity*. ICLR.

Tags

#llm#supervised-fine-tuning#early-stopping#interpretability#alignment#overfitting#token-interactions

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/177620406