mlsanityData to 14 Oct 2026

Reconciling Contradictory Views on the Effectiveness of SFT in LLMs: An Interaction Perspective

arXiv 2605.17967 ↗First posted 18 May 2026cs.AIcited by 0

Not on the map. Why?

Abstract (arXiv)

This paper explores a scientific question in supervised fine-tuning (SFT): why SFT is broadly effective for small-scale deep neural networks, yet can produce inconsistent or even detrimental effects when applied to large language models (LLMs). Recent advances in interaction-based explanations suggest that interactions between words/tokens provide a faithful metric for quantifying the inference patterns encoded by LLMs. We find that the evolution of interactions during SFT can effectively explain the inconsistent effectiveness of SFT for LLMs. Specifically, we find that (1) SFT primarily removes noise-like interactions, while rarely acquiring reliable new interactions. (2) This denoising stage is extremely brief, after which continued fine-tuning tends to introduce overfitted interactions. We validate these findings across multiple LLMs and datasets. Our findings provide new insights into early stopping and offer practical guidance for LLM training.

Key facts. No citing papers in the map yet (data through 14 Oct 2026). Builds on 30 of its 38 references that are in the map. First posted on arXiv 18 May 2026.

Builds on

30 of 38 references.

References · 30

  1. 2018BERT: Pre-training of Deep Bidirectional Transformers for Language Understandingcited by 40,399
  2. 2022Training language models to follow instructions with human feedbackcited by 15,971
  3. 2024The Llama 3 Herd of Modelscited by 14,715
  4. 2019Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformercited by 14,599
  5. 2023Llama 2: Open Foundation and Fine-Tuned Chat Modelscited by 13,028
  6. 2021Finetuned Language Models Are Zero-Shot Learnerscited by 3,780
  7. 2023Qwen Technical Reportcited by 3,344
  8. 2022Scaling Instruction-Finetuned Language Modelscited by 3,024
  9. 2024LlamaFactory: Unified Efficient Fine-Tuning of 100+ Language Modelscited by 1,464
  10. 2023Safe RLHF: Safe Reinforcement Learning from Human Feedbackcited by 656
  11. 2023An Empirical Study of Catastrophic Forgetting in Large Language Models During Continual Fine-tuningcited by 639
  12. 2020GoEmotions: A Dataset of Fine-Grained Emotionscited by 486
18 more
  1. 2023The False Promise of Imitating Proprietary LLMscited by 263
  2. 2020A Unified Approach to Interpreting and Boosting Adversarial Transferabilitycited by 71
  3. 2021Discovering and Explaining the Representation Bottleneck of DNNscited by 42
  4. 2022Two-stage LLM Fine-tuning with Less Specialization and More Generalizationcited by 41
  5. 2024Instruction Tuning With Loss Over Instructionscited by 40
  6. 2021Defining and Quantifying the Emergence of Sparse Concepts in DNNscited by 36
  7. 2023Does a Neural Network Really Encode Symbolic Concepts?cited by 27
  8. 2023Explaining Generalization Power of a DNN Using Interactive Conceptscited by 25
  9. 2025ProxySPEX: Inference-Efficient Interpretability via Sparse Feature Interactions in LLMscited by 16
  10. 2024MA-RLHF: Reinforcement Learning from Human Feedback with Macro Actionscited by 15
  11. 2024Defining and Extracting generalizable interaction primitives from DNNscited by 14
  12. 2024Learning to Understand: Identifying Interactions via the M"obius Transformcited by 13
  13. 2024Towards the Dynamics of a DNN Learning Symbolic Interactionscited by 13
  14. 2025Robust Reinforcement Learning from Human Feedback for Large Language Models Fine-Tuningcited by 13
  15. 2023Technical Note: Defining and Quantifying AND-OR Interactions for Faithful and Concise Explanation of DNNscited by 12
  16. 2025SPEX: Scaling Feature Interaction Explanations for LLMscited by 12
  17. 2025Unilaw-R1: A Large Language Model for Legal Reasoning with Reinforcement Learning and Iterative Inferencecited by 12
  18. 2023Bayesian Neural Networks Avoid Encoding Complex and Perturbation-Sensitive Conceptscited by 9

Built on by

Not cited yet.