Share
Embed
In-Context Meta Learning Induces Multi-Phase Circuit Emergence
In ICLR 2025 Workshop on Building Trust in Language Models and Applications, 2025, oral presentation
Gouki Minegishi
March 01, 2026
Transcript
11 slides
-
©MATSUO-IWASAWA LAB, THE UNIVERSITY OF TOKYO In-Context Meta Learning Induces
Multi-Phase Circuit Emergence Gouki Minegishi paper -
©MATSUO-IWASAWA LAB, THE UNIVERSITY OF TOKYO *O$POUFYU-FBSOJOH *$-
• In-context learning means inferring the task from the prompt examples and predicting the answer accordingly. • Why large language models can perform in-context learning so effectively remains an open research question. • Understanding the mechanism of in-context learning is a step toward predictable & verifiable LLM behaviour. Brown et al., “Language Models are Few-Shot Learners” -
©MATSUO-IWASAWA LAB, THE UNIVERSITY OF TOKYO .FDIBOJTUJD*OUFSQSFUBCJMJUZ • Reverse-engineering
methods that causally uncover a model’s internal computation, beyond surface I/O or attribution analyses • Circuits: functional units inside the network that implements a specific capability • Induction Head – the canonical in-context-learning circuit • Operation: locate a matching token earlier in the prompt and copy the next token, letting the model predict the correct continuation … Harry Potter … Harry … ??? Match Copy Conmy et al., “Towards Automated Circuit Discovery for Mechanistic Interpretability” Induction Head -
©MATSUO-IWASAWA LAB, THE UNIVERSITY OF TOKYO 3FTFBSDI2VFTUJPOT Few-shots prompts
Transformer fromage Prediction What kind of circuit works? How does the circuit grow? Induction heads is the simple match-and- circuit. We still donʼt know what kind of circuits emerge in real few-shot prompts. -
©MATSUO-IWASAWA LAB, THE UNIVERSITY OF TOKYO 5PZ&YQFSJNFOU4FUVQ • Problem
Setup • We designed a task involving 64 classes and 32 labels, with context-specific class–label pairings. • To answer accurately, the model must infer the underlying task from the context. • Network Structure • 2-layer attention only Transformer + 1-layer MLP (classification) … Task1 Task2 Task ! !! !" !# !$ Copy Task (Reddy, 2023) In-Context Meta Learning (Ours) " Classes … … … … … … … … # Labels !! ℓ! !# !$ ℓ$ ? !" ℓ" copy match ! Classes # Labels infer predict ℓ! % ℓ" % ℓ$ % ? … … Problem setup Network structure examples query -
©MATSUO-IWASAWA LAB, THE UNIVERSITY OF TOKYO 1IBTF5SBOTJUJPOBOE$JSDVJU&NFSHFODF • The
model arrives at 100% undergoing three accuracy plateaus • 3-Circuits at each plateau • Non-Context Circuit; NCC (Phase1): The model Ignore the context and relying solely on the modelʼs weights. • Semi-Context Circuit; SCC (Phase2): The model not only leverages weights memory but also attends to label tokens (i.e., half of the context) • Full-Context Circuit; FCC (Phase3): The model use the entire context. 1-layer Bigram Bigram Phase 1 Phase 2 Bigram Label Attention 1-layer 2-layer Chunk Example Label Attention 2-layer 1-layer 2-layer Phase 3 -
©MATSUO-IWASAWA LAB, THE UNIVERSITY OF TOKYO $JSDVJU.FUSJDT 1. Bigram
Metric: Attention from a query token to itself. 2. Label Attention Metric: Total attention from the query to all label tokens in context. 3. Chunk Example Metric: Attention from each example token 𝑥 to its paired label ℓ. Correlation with Accuracy: The timing of these metric transitions closely matches the modelʼs discrete accuracy jumps, validating that our metrics quantitatively capture the internal circuit reconfigurations across all three learning phases attention weights 𝑝!,# $,% (where 𝜇is the layer and ℎ the head), for a context window of length 2𝑁 + 1 -
©MATSUO-IWASAWA LAB, THE UNIVERSITY OF TOKYO $JSDVJU4FQBSBUJPOJO.VMUJ)FBE • Multi-Head
vs. Single-Head: Two-head models show smoother accuracy curves, lacking the sharp phase jumps seen with one head. • Head Specialization: One head implements an NCC-like circuit, the other an FCC-like circuit. • Parallel Learning: Running circuits concurrently yields more efficient and gradual accuracy gains. • LLM Implication: Confirms that multi-head attention in LLMs enables distinct functional roles per head, smoothing learning dynamics. NCC-like circuit FCC-like circuit -
©MATSUO-IWASAWA LAB, THE UNIVERSITY OF TOKYO $JSDVJUTJO--. • Setup
• Test whether our identified circuits appear in LLMs by evaluating a pretrained GPT2-XL (48 layers) on SST2 (872 samples). • 2-shot prompt: two labeled examples (Review: {text}, Sentiment: {label}) and a third, unlabeled query • Results • Chunk Example scores peak in earlier layers while Label Attention scores are higher in middle or later layers, consistent with the final circuit (FCC) behavior in our 2-layer attention-only model -
©MATSUO-IWASAWA LAB, THE UNIVERSITY OF TOKYO $PODMVTJPOBOE'VUVSF%JSFDUJPO • Conclusion
• We watched what happens inside a model in a practical few-shot setting. • The model passed through three distinct circuits̶NCC, SCC, and FCC̶before it reached perfect accuracy. • Those same circuits also show up inside a pretrained LLM (GPT-2 XL), so the toy findings really scale. Youʼll find extra details in the paper̶multi-head results and how the circuits depend on data distribution property. • Future Direction • We plan to test these circuit metrics in application-level scenarios where transparency matters: bias detection, hallucination control, and other safety- critical cases. paper