The paper tests whether ICL predictions are independent of the conditional label distribution of in-context examples (NH1). By randomizing in-context labels, they show that log likelihoods and entropies degrade significantly across all 10 model variants and 9 tasks, rejecting NH1. They further demonstrate that ICL can learn a completely novel author-identification task (private Slack messages between two authors) that is guaranteed absent from pre-training data, with all models except Falcon-7B-Instruct achieving better-than-random accuracy. This confirms ICL genuinely uses and learns from in-context label information rather than merely recognizing task format.
Evidence
correlational
Key metric
Δ log likelihood (default minus random) at max context size for llama-2 70b: 0.51 (SST-2), 0.53 (Subj), 0.57 (FP), 0.34 (HS), 0.80 (AGN), 0.29 (MQP), 0.04 (MRPC), 0.22 (RTE), 0.18 (WNLI); for falcon 40b: 0.34, 0.35, 0.31, 0.18, 0.90, 0.06, 0.01, 0.01, 0.02
Caveat
On entailment tasks (MQP, MRPC, RTE, WNLI) smaller models struggle to reach better-than-random performance on default labels, limiting the scope of the randomization comparison for those tasks. The novel author-identification dataset is not publicly released.