Understanding In-Context Learning in Large Language Models: Experiments, Analysis, and Theoretical Insights

What is In-Context Learning?

In‑context learning (ICL) enables a pretrained large language model (LLM) to perform a new task by simply providing a few input‑output examples together with a task description, without any parameter updates or explicit fine‑tuning.

Illustrative Example

For a translation task (English → French), the prompt consists of a task description line, several example pairs, and the query word to translate. The model then generates the correct French translation (e.g., cheese → fromage ).

Empirical Studies

Simple Copy Output

Five examples each containing five random lowercase letters are provided, and the model must copy the final input list.

Input: g, c, b, h, d
Output: g, c, b, h, d
Input: b, g, d, h, a
Output: b, g, d, h, a
Input: f, c, d, e, h
Output: f, c, d, e, h
Input: c, f, g, h, d
Output: c, f, g, h, d
Input: e, f, b, g, d
Output: e, f, b, g, d
Input: a, b, c, d, e
Output:

The expected output is:

a, b, c, d, e

GPT‑3 achieved 100 % accuracy on all 6 720 possible input permutations, while the smallest model text-ada-001 reached 99.78 % (6 705/6 720), demonstrating the importance of model scale.

Date Formatting

The task converts dates from YYYY‑MM‑DD to a custom format !MM!DD!YYYY! . Examples use three demonstration pairs followed by a test date such as 2005-07-23 . Across the GPT‑3 family (from text-ada-001 to text-davinci-003 ), accuracy improves with model size and with the number of in‑context examples, though it never reaches perfect 100 %.

Label Remapping

Entities originally labeled as animal , plant/vegetable , or sport are remapped arbitrarily (e.g., duck → plant/vegetable , golf → animal , beans → sport ). GPT‑3 correctly predicts the new mappings, even when the label symbols are nonsensical (e.g., [^*, #@#, !!~] ).

llama: plant/vegetable ✓
cat: plant/vegetable ✓
elephant: plant/vegetable ✓
monkey: plant/vegetable ✓
panda: plant/vegetable ✓
cucumber: sport ✓
peas: sport ✓
tomato: sport ✓
spinach: sport ✓
carrots: sport ✓
rugby: animal ✓
cycling: animal ✓
baseball: animal ✓
tennis: animal ✓
judo: animal ✓

Theoretical Analysis of ICL

A recent Microsoft Research paper argues that the attention layers of LLMs perform an implicit parameter‑optimization process analogous to gradient‑descent fine‑tuning.

Gradient‑Descent View of Linear Attention

For a fully‑connected layer with initial weights W₀ , gradient ΔW , input x , and output gradient e , a single gradient‑descent step yields W₁ = W₀ - η·ΔW . The paper shows that ΔW can be expressed as the outer product of the previous input (treated as a key ) and the previous output gradient (treated as a value ), which is exactly what linear attention computes.

Thus, the attention computation softmax(QKᵀ)V (with Q , K , V derived from the current query, previous keys, and previous values) can be seen as applying a learned update ΔW to the implicit weight matrix.

Implicit Fine‑Tuning via the Final Query Token

The last token of the prompt is designated as the query token q (dimension d ). After passing through the attention head, its output is:

output = W_v·X'·softmax((W_k·X')ᵀ·W_q·q)

When the softmax is omitted (as done in the analysis), the formula reduces to a linear transformation that directly incorporates the gradients accumulated from the demonstration examples, effectively performing a zero‑shot (“ZSL”) update W_{zsl} followed by a small gradient ΔW_{icl} derived from the in‑context examples.

Consequently, larger models possess higher‑dimensional key, query, and value matrices ( d'×d ) and are trained on vastly more data, providing a richer initial weight W_{zsl} . Only a few demonstration examples are needed to generate a useful ΔW_{icl} , which explains why ICL emerges only when model scale and data volume cross a certain threshold.

References

[1] https://yaofu.notion.site/GPT-3-5-360081d91ec245f29029d37b54573756

[2] http://ai.stanford.edu/blog/understanding-incontext/

[3] https://ai.stanford.edu/blog/in-context-learning/

[4] https://www.reddit.com/r/MachineLearning/comments/10ly7rw/r_why_can_gpt_learn_incontext_language_models/

[5] https://mp.weixin.qq.com/s/dPpO18g3V4xqHUsEBKrXJQ

[6] https://arxiv.org/abs/2005.14165

[7] https://arxiv.org/abs/2212.10559

[8] https://platform.openai.com/docs/models/gpt-3

[9] https://github.com/microsoft/LMOps