IC-1053Instruction tuning suppresses in-context learning in LLaMA, Vicuna, and OPT-IML, with the suppression being largest for English prompts and partially recoverable via translation to other languages

Suhas Kotha, Jacob Mitchell Springer, Aditi Raghunathan

SourceUnderstanding Catastrophic Forgetting in Language Models via Implicit Inference

The paper constructs a controlled prompt where in-context exemplars define a latent task (repeat or capitalize) but the final query is phrased as an instruction (math or fill-in-the-blank). Pretrained models (LLaMA-7B, OPT-1.3B) perform the ICL task, while their instruction-tuned counterparts (Alpaca, Vicuna-7B, OPT-IML-1.3B) follow the instruction instead. The drop in ICL accuracy is largest in English (56.75% for Alpaca, 40.00% for Vicuna, 21.00% for OPT-IML) and much smaller in French, Spanish, Dutch, Hungarian, leetspeak, or Pig Latin. For example, Alpaca's ICL drop under leetspeak is only 1.0% versus 56.75% in English.

Evidence
correlational
Key metric
Alpaca ICL drop: 56.75% (English), 29.00% (French), 47.75% (Spanish), 51.00% (Dutch), 45.75% (Hungarian), 1.50% (leetspeak), 13.50% (Pig Latin); Vicuna ICL drop: 33.00% (English), 19.50% (French), 11.00% (Spanish), 39.00% (Dutch), 36.50% (Hungarian), 1.00% (leetspeak), 25.00% (Pig Latin); OPT-IML ICL drop: 21.00% (English), 9.25% (French), 5.25% (Spanish), 5.75% (Dutch), 4.25% (Hungarian), 4.00% (leetspeak), 10.00% (Pig Latin)
Caveat
Translation is implemented via Google Translate and may introduce errors; the paper notes that translation can violate invertibility for tasks requiring contextual knowledge that varies across languages. Evaluation is limited to first-word completion accuracy over 400 samples and 4 ICL vs IF task combinations.
Model
LLaMA, Alpaca, Vicuna, OPT OPT-IML
Concepts
Failure mode
Related findings
IC-1054, IC-1055
Extraction
automatic-extraction