IC-581Induction heads for ICL operate on task-specific attention subspaces, with partial overlap across tasks, and the geometry of these subspaces explains demonstration saturation

Hakaze Cho, Mariko Kato, Yoshihiro Sakai, Naoya Inoue

SourceRevisiting In-context Learning Inference Circuit in Large Language Models

The paper identifies induction heads (attention heads with high scores from label tokens to the query forerunner) in Llama 3 70B and finds that more than half are not 'correct' (do not preferentially attend to the correct label). Correct induction heads show significant but incomplete overlap across the 6 evaluation datasets (e.g., 0.83 overlap between SST-2 and MR, 0.55 between SST-2 and AG News), indicating a mix of task-inherent and task-specific subspaces. PCA visualization of label representations mapped through the best induction head (correct rate 0.95) shows that attention assignment morphology changes significantly from k=1 to k=2 demonstrations but stabilizes from k=15 to k=16, explaining the submodular improvement and saturation of ICL performance with more demonstrations.

Evidence
correlational
Key metric
correct induction head overlap rates (SST-2 vs. others): 0.83 (MR), 0.67 (FP), 0.76 (SST-5), 0.59 (TREC), 0.55 (AG News); best induction head correct rate 0.95 (layer 31, head 32) vs. worst 0.00 (head 9)
Caveat
The overlap analysis is on 6 classification datasets; the degree of task-specificity may differ for other task types. The PCA visualization is on a single sample from SST-2, though appendix H.3 shows 4 additional samples with consistent results.
Model
Llama 3
Concepts
Linear representation
Datasets
SST-2 [eval], MR [eval], Financial Phrasebank [eval], SST-5 [eval], TREC [eval], AG-News [eval]
Methods
Principal component analysis [supporting]
Related work
induction circuit [builds-on]
Related findings
IC-578, IC-579, IC-580
Extraction
automatic-extraction