Source
Min, Lyu, Holtzman, Artetxe, Lewis, Hajishirzi & Zettlemoyer (2022), LIT-814 (EMNLP 2022), read in full in NOTE-607.
What was actually shown
The decomposition. A demonstration carries four things: the input–label mapping, the distribution of the inputs, the label space and the format (inputs paired with labels). Each is ablated alone, on classification and multiple-choice tasks with small answer sets: 26 datasets, six language models from 774M to 175B parameters, each with direct (P(y|x)) and channel (P(x|y)) inference.
Labels (§4.1–4.2). With k = 16, replacing gold labels by labels drawn uniformly from the label set lowers accuracy by 0–5 points absolute: 2.6 on classification and 1.7 on multiple choice on average, and 0.1–0.9 for the meta-trained MetaICL. All-incorrect labels keep 92%, 100% and 97% of the gain in three of four settings. The gold–random gap stays at 0.8–1.6 points for k from 4 to 32.
The rest (§5, four model–method pairs, nine datasets). Out-of-distribution sentences in place of the inputs cost 3–16 points for most models. Random English words in place of the labels cost 5–16 points for direct models and 0–2 for channel models. Removing the pairing, inputs only or labels only, is about as bad as no demonstrations; keeping the format with only one side real retains up to 95% (direct MetaICL) and 75–87% (channel) of the gain.
What could have come out otherwise. Random labels could have cost as much as removing the demonstrations. They cost a few points.
What this does not say
- That labels do not matter. The random labels always come from the correct label set, so the label space is always given. What matters little is which label goes with which input.
- That the models learn nothing at test time. That is the authors’ interpretation (§6), drawn from the result and not tested on tasks whose correspondence cannot have been learned in pretraining. The authors predict the gap would open there.
- That it holds for every dataset. Macro-averages hide gaps up to about 14 points (financial_phrasebank with GPT-J), and Kim et al. (2022) find negated labels hurt substantially.
- That it holds for generation. Only classification and multiple choice were tested.
- That one context does one thing. Which parts act depends on the read-out: the label space matters for direct inference and hardly for channel inference, on the same model.
Connections
- Lu et al. (LIT-833, THEORY-683) vary a different axis: the order of a fixed set of demonstrations, which moves accuracy far more. Both say that what a context does to a fixed model is carried by its form more than by the correctness of its examples.
- Xie et al. (LIT-797) give the latent-concept account the authors cite. On it, inputs, label space and format are evidence about the task, and the mapping is evidence the model can do without.
- Akyürek et al. (LIT-796, THEORY-439) cite this paper for in-context learning not resembling standard learning, and show that updating the weights on the same demonstrations does more than conditioning on them.
- CLAIM-029 reads this result as support for treating a context as a frame.