Knowing but Not Saying: Preventing Factual Access Failures in LLM SFT via Recall-Anchored Distillation

2026-08-21

AI Summary

本文针对领域微调导致的事实访问失败问题，通过引入召回锚定蒸馏（RAD）方法，在无需额外标注数据的情况下，恢复模型在非目标域上的事实生成能力。

Abstract

Supervised fine-tuning (SFT) can degrade factual behavior outside the target domain. This degradation is often described as catastrophic forgetting, yet open-ended factual failures do not necessarily imply that the underlying facts have been erased. In this work, we identify a more specific phenomenon, factual access failure: after domain SFT, models can still recognize or rank the correct answer under constrained evaluation, while failing to produce it in closed-book generation. Through benchmark-level comparisons, same-fact multiple-choice and generation probes, and failure-mode analysis, we show that SFT-induced factual degradation reflects both genuine wrong-answer generations and expression-level failures such as verbosity, formatting mismatch, and exact-match artifacts. To address this problem, we introduce Recall-Anchored Distillation (RAD), a base-anchored self-distillation objective that preserves out-of-distribution generation behavior by aligning the adapted model with the original base model's soft continuation distribution on unlabeled OOD text. RAD requires no gold OOD answers, external judges, or labeled factual data. Across three backbones fine-tuned on MedMCQA, RAD recovers a consistent portion of the lost OOD recall while preserving target-domain adaptation. Compared with replay on the same OOD text, RAD shows that the key preservation signal is the base model's soft distribution rather than additional text exposure alone.

Problem

• Supervised Fine-Tuning (SFT)
• Factual Access Failure
• Closed-book Generation

Innovation

• Recall-Anchored Distillation
• factual access failure
• supervised fine-tuning
• out-of-distribution
• soft continuation distribution