Is In-Context Learning Learning?

Setup at scale. Four LLMs, nine formal-task families (regular and context-free), multiple prompting styles, and 0–100 exemplars yielded 1.89M predictions per model. Results are reported as accuracies with OOD stress tests at growing distribution distances.

More shots help, and models converge. Accuracy rises steadily with the number of exemplars, with the steepest gains for plain few-shot modus ponens. Gaps between models and prompts shrink as shots grow, which suggests ICL’s effect is tied to the autoregressive mechanism rather than model choice. Peak performance typically arrives at 50–100 shots, not a few.

Robustness is the weak spot, especially for CoT and APO. Changing the test distribution depresses accuracy across the board; chain-of-thought has the largest OOD drop (average slope about −1.4), while plain few-shot is the least affected.

Language in the prompt matters less in the limit. Replacing natural language descriptions with word salad eventually matches non-randomized prompts once enough exemplars are shown, implying the models key on statistical regularities in the exemplars more than lexical semantics. By contrast, “salad-of-thought,” which randomizes CoT exemplars, performs poorly.

Mixed task competence and classical baselines. Some regular-language tasks are essentially solved (Pattern Matching ≈94% ±1), while others remain hard (Reversal, Maze Solve, Vending-Machine Sum). Closely related tasks can differ by up to 31% in peak accuracy. Simple ML baselines (DT, kNN, MLP) beat ICL’s average on roughly half the tasks.