IC-324Fine-tuning Gemini Nano 1 on 8 memorization examples causes it to override in-context predictions with in-weight predictions in 2 of 8 cases, while the base model always follows in-context predictions
Bryan Chan, Xinyi Chen, András György, Dale Schuurmans
The authors fine-tuned the released Gemini Nano 1 (1.8B) on eight name-city pairs (four real, four invented) to force in-weight memorization. When tested with a contradictory in-context prompt (e.g., 'kaitlyn lives in kjheergg. Where does kaitlyn live?'), the base model always answers with the in-context city (8/8 ICL). After fine-tuning, the model reverts to the memorized in-weight answer in 2 of 8 cases (trrrrqe, kaitlyn) and produces an unrelated answer in 2 more, while still following ICL in 4 cases. On the pure IWL test (no contradictory context), the fine-tuned model answers 6/8 correctly (4/4 real, 2/4 invented), whereas the base model answers 0/8. This demonstrates that in-weight memorization can partially override in-context learning in a released LLM.
Evidence
interventional
Key metric
Base model ICL: 8/8 answers = in-context city (kjheergg). Fine-tuned model ICL: 4/8 ICL, 2/8 IWL override (trrrrqe answers qtbnaaa log-prob 0.0000 vs kjheergg -21.6897; kaitlyn answers kingston log-prob -0.1156 vs kjheergg -2.4694), 2/8 other. Fine-tuned IWL: 6/8 correct (4/4 real, 2/4 invented) vs base 0/8.
Caveat
Only 8 questions were used for fine-tuning and evaluation; the effect is demonstrated on a single small set of name-city pairs. The paper notes that with longer or more informative prompts the ICL behaviour is partially restored, and that the results are specific to this small-scale memorization setup.