跳转至

文章背景与核心概要

在大语言模型（LLM）的监督微调（SFT）过程中，模型在目标领域之外的事实性表现往往会退化，这一现象通常被称为“灾难性遗忘”。然而，本研究表明，开放式的事实生成失败并不一定意味着底层事实已经被彻底抹去。

作者在此项研究中识别出一个更细微的问题，将其定义为事实访问失效（factual access failure）：在经历特定领域SFT之后，模型在受限评估（如多项选择题）下通常仍然能够识别或正确排序正确答案，却在闭卷生成过程中无法将其输出。为了解决这一问题，论文引入了召回锚定蒸馏（Recall-Anchored Distillation, RAD）。这是一种基于基座锚定的自蒸馏目标，能够在不需要标准外部答案、外部评判器或标注事实数据的情况下，保持模型在分布外（OOD）的生成行为。

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

Authors: Haodong Chen, Yadong Wang, Shengtao Wen, Dong Liang, Xiang Chen
Subjects: Artificial Intelligence (cs.AI)
ArXiv ID: arXiv:2608.20794 | DOI: 10.48550/arXiv.2608.20794
Submitted: 21 August 2026

Executive Summary

监督微调（SFT）大语言模型（LLM）常常会导致目标领域外的事实行为退化，传统上这一现象被称为灾难性遗忘。然而，本研究表明，开放式的事实性失效并不一定意味着底层的知识事实已被擦除。

作者指出了一种更为细致的问题，即事实访问失效：在领域SFT之后，模型在受限评估（如多项选择）下往往仍能识别或正确排序正确答案，但在闭卷生成时却无法产出这些答案。为了应对这一挑战，本文引入了召回锚定蒸馏（Recall-Anchored Distillation, RAD），这是一种基于基座锚定的自蒸馏目标，它能够在无需标准OOD答案、外部评判员或标记事实数据的前提下，保持分布外（OOD）的生成行为。

Supervised fine-tuning (SFT) of Large Language Models (LLMs) often degrades factual behavior outside the target domain, a phenomenon traditionally labeled as catastrophic forgetting. However, this research demonstrates that open-ended factual failures do not necessarily mean underlying facts have been erased.

The authors identify a more nuanced issue termed factual access failure: post-domain SFT, models can frequently still recognize or correctly rank the right answers under constrained evaluations (e.g., multiple-choice), yet fail to produce them during closed-book generation. To combat this, the paper introduces Recall-Anchored Distillation (RAD), a base-anchored self-distillation objective that preserves out-of-distribution (OOD) generation behavior without requiring gold OOD answers, external judges, or labeled factual data.

Abstract

监督微调（SFT）可能会导致目标领域之外的事实行为退化。这种退化通常被描述为灾难性遗忘，但开放式的事实性失效并不一定暗示底层事实已被擦除。在这项工作中，我们识别出一个更具体的现象——事实访问失效：在进行领域SFT后，模型在受限评估下仍然能够识别或排序正确答案，却在闭卷生成中未能将其产出。

通过基准层面的对比、同事实多项选择与生成探针以及失效模式分析，我们表明SFT引起的事实退化既反映了真正的错误答案生成，也反映了表达层面的失效，例如冗长、格式不匹配以及精确匹配伪影。为了解决这个问题，我们引入了召回锚定蒸馏（RAD），这是一种基座锚定的自蒸馏目标，通过在未标记的OOD文本上将适配后的模型与原始基座模型的软续写分布进行对齐，从而保持分布外的生成行为。RAD不需要标准的OOD答案、外部评判员或标记的事实数据。在基于MedMCQA微调的三种主干模型上，RAD在保持目标领域适应性的同时，恢复了相当一部分丢失的OOD召回率。与在相同OOD文本上进行回放相比，RAD表明关键的保持信号是基座模型的软分布，而不仅仅是额外的文本曝光。

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.

Key Takeaways & Contributions

  1. 事实访问失效的识别： 证明了模型在SFT后通常知道事实（通过多项选择识别证明），但由于表达层面的障碍（如格式问题、冗长），在闭卷生成中未能准确表述它们。
  2. 召回锚定蒸馏（RAD）： 一种新颖、高效的自蒸馏目标，通过将适应后模型的软续写分布与原始基座模型进行对齐，来保持分布外的生成行为。
  3. 零外部标签依赖： RAD完全基于未标记的OOD文本运行，无需金标准答案、外部评判员或精心策划的事实数据集。
  4. 实证验证： 在基于MedMCQA数据集微调的三种不同主干模型上进行了测试，证明RAD在维持强大的领域适应性能的同时，成功恢复了丢失的OOD召回率。
  1. Identification of Factual Access Failure: Proves that models often know facts post-SFT (demonstrated via multiple-choice recognition) but fail to state them accurately in closed-book generation due to expression-level barriers (e.g., formatting issues, verbosity).
  2. Recall-Anchored Distillation (RAD): A novel, efficient self-distillation objective that preserves out-of-distribution generation behaviors by aligning the adapted model's soft continuation distribution with its original base model.
  3. Zero Dependence on External Labels: RAD operates entirely on unlabeled OOD text, bypassing the need for gold standard answers, external judges, or meticulously curated factual datasets.
  4. Empirical Validation: Tested across three different backbones fine-tuned on the MedMCQA dataset, proving that RAD successfully recovers lost OOD recall while maintaining strong domain adaptation performance.