Why LLMs Remember but Fail to Apply After Fine-Tuning

HKUST · hf · 2026-07-13

This paper investigates a common issue in LLM fine-tuning: models rapidly memorize new knowledge but fail to apply it during downstream reasoning tasks. The authors formalize this phenomenon as the **Knowing–Using Gap**, characterized by two aspects: - The accuracy discrepancy between memorization and generalization. - The time lag between a fact being memorized and becoming practically usable. To explain the underlying mechanics, the paper introduces a **self-patching** intervention method. This tracks the spatial diffusion dynamics of knowledge within the model and identifies locations where migrating representations to the correct layers significantly resolves failed cases. The findings support a "knowledge circuit mismatch" hypothesis: although the knowledge exists internally, it is not routed to the layers engaged in computation. Finally, the authors propose a simple heuristic strategy that recovers about 58%–75% of the oracle margin in generalization failures.

Related event: Study Reveals Knowing-Using Gap in LLM Fine-Tuning(2 posts)→

Original post →

More from Research

Research channel →