Why Fine-Tuned Memories Still Struggle to Generalize
burny_tech · x · 2026-07-13
This paper examines a common phenomenon: LLMs can quickly memorize new facts during fine-tuning, but generalize poorly when multi-hop reasoning is required.
The author suggests the issue isn't "failure to learn knowledge," but a misalignment between knowledge and reasoning circuits. While facts exist inside the model, they don't reside in the layers actively involved in reasoning. By using self-patching to move appropriate hidden representations to the correct layers, the paper restored a significant portion of generalization ability; a simple heuristic recovered 58–75% of the oracle headroom.
Related event: Study Reveals Knowing-Using Gap in LLM Fine-Tuning(2 posts)→
More from Models
- Benchmark scores drop from 89% to 19% on new evals — how benchmaxxing breaks leaderboard trust — airesearch12 · 2026-09-11
- Developer Building a Unified Leaderboard of All Model Benchmark Scores — airesearch12 · 2026-09-11
- Rumor claims Kimi faked performance by serving Claude; DeepSeek new model surprises in evals — realsohamparekh · 2026-09-11
- GPT-5.6 writes well but is instantly forgettable, user complains — BasedRaddka · 2026-09-11
- Opus Refuses Protein Research Codebase Over 'Safety' Concerns, Dev Considers Rolling His Own — josephdviviano · 2026-09-11
- User Hails Unconfirmed 'DeepSeek 4.1 Flash' as an Inflection Point in LLMs — himanshustwts · 2026-09-11