Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Talking to Yourself: Defying Forgetting in Large Language Models

Yutao Sun, Mingshuai Chen, Tiancheng Zhao, Phillip Miao, Zilun Zhang, Haozhan Shen, Ruizhe Zhu, Jianwei Yin

TL;DR

SA-SFT, a lightweight self-augmentation routine in which an LLM generates self-dialogues prior to fine-tuning, and the resulting self-authored data are mixed with task data without modifying optimization or training schedules, consistently mitigates catastrophic forgetting while improving in-domain performance.

Abstract

Catastrophic forgetting remains a major challenge when fine-tuning large language models (LLMs) on narrow, task-specific data, often degrading their general knowledge and reasoning abilities. We propose SA-SFT, a lightweight self-augmentation routine in which an LLM generates self-dialogues prior to fine-tuning, and the resulting self-authored data are mixed with task data without modifying optimization or training schedules. Despite requiring no external data or additional tuning, SA-SFT consistently mitigates catastrophic forgetting while improving in-domain performance. Across 50 evaluation scenarios, it maintains performance comparable to the original model and achieves the best results in 40 cases, outperforming common baselines such as layer freezing and external data mixing. Guided by these empirical findings, we further present a theoretical analysis suggesting that forgetting can partly stem from style-induced parameter drift, and that self-alignment through self-generated data provides an effective means to counteract this effect. Overall, our results indicate that self-augmentation offers a simple and effective mechanism for robust LLM adaptation without incurring catastrophic forgetting.

Talking to Yourself: Defying Forgetting in Large Language Models

TL;DR

SA-SFT, a lightweight self-augmentation routine in which an LLM generates self-dialogues prior to fine-tuning, and the resulting self-authored data are mixed with task data without modifying optimization or training schedules, consistently mitigates catastrophic forgetting while improving in-domain performance.

Abstract

Catastrophic forgetting remains a major challenge when fine-tuning large language models (LLMs) on narrow, task-specific data, often degrading their general knowledge and reasoning abilities. We propose SA-SFT, a lightweight self-augmentation routine in which an LLM generates self-dialogues prior to fine-tuning, and the resulting self-authored data are mixed with task data without modifying optimization or training schedules. Despite requiring no external data or additional tuning, SA-SFT consistently mitigates catastrophic forgetting while improving in-domain performance. Across 50 evaluation scenarios, it maintains performance comparable to the original model and achieves the best results in 40 cases, outperforming common baselines such as layer freezing and external data mixing. Guided by these empirical findings, we further present a theoretical analysis suggesting that forgetting can partly stem from style-induced parameter drift, and that self-alignment through self-generated data provides an effective means to counteract this effect. Overall, our results indicate that self-augmentation offers a simple and effective mechanism for robust LLM adaptation without incurring catastrophic forgetting.
Paper Structure (50 sections, 29 equations, 3 figures, 4 tables)

This paper contains 50 sections, 29 equations, 3 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Synthetic Data.
  4. Mitigation Methods of Catastrophic Forgetting.
  5. Methodology
  6. Notation.
  7. Step 1: Self-augmentation.
  8. Generation hyper-parameters.
  9. Step 2: Mix-in.
  10. Fine-tuning objective.
  11. Implementation notes.
  12. Main Experiments
  13. Experimental Setup
  14. Models.
  15. Fine-tuning tasks.
  16. ...and 35 more sections

Figures (3)

  • Figure 1: External rehearsal (top) fine-tunes on unfamiliar data and risks forgetting, whereas Sa-sft (bottom) lets the model “talk to itself,” mixing a processing-free, self-generated corpus with task data to retain prior skills.
  • Figure 2: Results of Sa-sft and Task+UC under varying mix ratios. The dashed line indicates the Task-only baseline.
  • Figure 3: GSM8K accuracy over training steps.