Stabilizing LLM Supervised Fine-Tuning via Explicit Distributional Control

Fuente: arXiv
Salvato in:
Dettagli Bibliografici
Autori principali: Wang, Xinyu, Sun, Changzhi, Wu, Yuanbin, Wang, Xiaoling
Natura: Preprint
Pubblicazione: 2026
Soggetti:
Accesso online:
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866913093913673728
author Wang, Xinyu
Sun, Changzhi
Wu, Yuanbin
Wang, Xiaoling
author_facet Wang, Xinyu
Sun, Changzhi
Wu, Yuanbin
Wang, Xiaoling
contents Post-training large language models (LLMs) often suffers from catastrophic forgetting, where improvements on a target objective degrade previously acquired capabilities. Recent evidence suggests that this phenomenon is primarily driven by excessive distributional drift during optimization. Motivated by this perspective, we propose Anchored Learning, a simple framework that explicitly controls distributional updates during offline fine-tuning via a dynamically evolving moving anchor. Instead of matching a fixed reference distribution, the anchor interpolates between the current model and a frozen reference to construct an intermediate target that the model distills toward, transforming global fine-tuning into a sequence of local trust-region updates in distribution space. Theoretically, we prove this anchor-based update admits a linear KL-divergence upper bound per iteration, ensuring a stable transition between model distributions. Extensive experiments on iGSM, MedCalc, and IFEval show that Anchored Learning consistently lies on the Pareto frontier of gain-stability trade-offs, achieving near-optimal performance improvements while substantially reducing degradation compared to strong baselines. For example, while standard SFT suffers from over 53% performance degradation on iGSM and MedCalc, Anchored Learning slashes this drop to under 5% while maintaining near-optimal gains (e.g., 75.2% on iGSM).
format Preprint
id arxiv_https___arxiv_org_abs_2605_04468
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Stabilizing LLM Supervised Fine-Tuning via Explicit Distributional Control
Wang, Xinyu
Sun, Changzhi
Wu, Yuanbin
Wang, Xiaoling
Machine Learning
Artificial Intelligence
Computation and Language
Post-training large language models (LLMs) often suffers from catastrophic forgetting, where improvements on a target objective degrade previously acquired capabilities. Recent evidence suggests that this phenomenon is primarily driven by excessive distributional drift during optimization. Motivated by this perspective, we propose Anchored Learning, a simple framework that explicitly controls distributional updates during offline fine-tuning via a dynamically evolving moving anchor. Instead of matching a fixed reference distribution, the anchor interpolates between the current model and a frozen reference to construct an intermediate target that the model distills toward, transforming global fine-tuning into a sequence of local trust-region updates in distribution space. Theoretically, we prove this anchor-based update admits a linear KL-divergence upper bound per iteration, ensuring a stable transition between model distributions. Extensive experiments on iGSM, MedCalc, and IFEval show that Anchored Learning consistently lies on the Pareto frontier of gain-stability trade-offs, achieving near-optimal performance improvements while substantially reducing degradation compared to strong baselines. For example, while standard SFT suffers from over 53% performance degradation on iGSM and MedCalc, Anchored Learning slashes this drop to under 5% while maintaining near-optimal gains (e.g., 75.2% on iGSM).
title Stabilizing LLM Supervised Fine-Tuning via Explicit Distributional Control
topic Machine Learning
Artificial Intelligence
Computation and Language
url https://arxiv.org/abs/2605.04468