IC-1221LLM performance on in-context boolean function learning is scale-dependent, with GPT-2 failing and Llama-2 models improving gradually with size

Satwik Bhattamishra, Arkil Patel, Phil Blunsom, Varun Kanade

SourceUnderstanding In-Context Learning in Transformers and LLMs by Learning to Learn Discrete Functions

In appendix H.3, the authors evaluate GPT-2, Llama-2-7B, Llama-2-13B, and Llama-2-70B directly on conjunction and majority tasks at 5, 7, and 15 dimensions. GPT-2 fails on both tasks even at 5 dimensions, performing near the null classifier. Llama-2-7B achieves non-trivial performance, and accuracy increases gradually through 13B to 70B. This indicates that the ability to implement learning algorithms in-context emerges and strengthens with model scale.

Evidence
correlational
Key metric
GPT-2 fails on both tasks even for 5 dimensions; gradual increase in performance as Llama-2 size increases from 7B to 13B to 70B
Caveat
Only three Llama-2 sizes are tested; the relationship between scale and performance is observed, not controlled. GPT-2 is a different architecture family, so the comparison is not purely within one family.
Model
GPT-2, Llama 2 / Llama 2 base
Concepts
Scale-dependent behaviour
Methods
k-nearest neighbours classifier / Nearest-neighbor baseline [compared-to]
Related findings
IC-1219, IC-1220
Extraction
automatic-extraction