Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Task Schema and Binding: A Double Dissociation Study of In-Context Learning

Chaeha Kim

TL;DR

This paper argues that in-context learning is not a single mechanism but a composition of two separable processes: Task Schema, an abstract representation of the task type encoded in late MLP layers, and Binding, the input–output associations encoded in the residual stream. Through activation patching across diverse Transformer families and the Mamba model, the authors demonstrate a robust 100% transfer for Task Schema and a 62% transfer for Binding, revealing a principled double dissociation. They quantify a Prior‑Schema trade‑off showing that higher prior knowledge reduces schema reliance, while recency biases explain most binding failures rather than direct prior competition. The findings generalize across architectures, provide a compositional model of ICL, and offer practical guidelines for prompt design and deployment, including attention‑level interventions and prior‑aware routing to improve reliability in production systems.

Abstract

We provide causal mechanistic validation that in-context learning (ICL) decomposes into two separable mechanisms: Task Schema (abstract task type recognition) and Binding (specific input-output associations). Through activation patching experiments across 9 models from 7 Transformer families plus Mamba (370M-13B parameters), we establish three key findings: 1. Double dissociation: Task Schema transfers at 100% via late MLP patching; Binding transfers at 62% via residual stream patching -- proving separable mechanisms 2. Prior-Schema trade-off: Schema reliance inversely correlates with prior knowledge (Spearman rho = -0.596, p < 0.001, N=28 task-model pairs) 3. Architecture generality: The mechanism operates across all tested architectures including the non-Transformer Mamba These findings offer a mechanistic account of the ICL puzzle that contrasts with prior views treating ICL as a monolithic mechanism (whether retrieval-based, gradient descent-like, or purely Bayesian). By establishing that Schema and Binding are neurally dissociable -- not merely behavioral modes -- we provide causal evidence for dual-process theories of ICL. Models rely on Task Schema when prior knowledge is absent, but prior knowledge interferes through attentional mis-routing (72.7% recency bias) rather than direct output competition (0%). This explains why arbitrary mappings succeed (zero prior leads to full Schema reliance) while factual overrides fail -- and reveals that the true bottleneck is attentional, not output-level. Practical implications: Understanding these dual mechanisms enables more efficient prompt engineering -- reliable schema transfer reduces required demonstrations for novel tasks, while prior-aware design can mitigate the 38% binding failure rate in high-prior scenarios, improving ICL system reliability in production deployments.

Task Schema and Binding: A Double Dissociation Study of In-Context Learning

TL;DR

This paper argues that in-context learning is not a single mechanism but a composition of two separable processes: Task Schema, an abstract representation of the task type encoded in late MLP layers, and Binding, the input–output associations encoded in the residual stream. Through activation patching across diverse Transformer families and the Mamba model, the authors demonstrate a robust 100% transfer for Task Schema and a 62% transfer for Binding, revealing a principled double dissociation. They quantify a Prior‑Schema trade‑off showing that higher prior knowledge reduces schema reliance, while recency biases explain most binding failures rather than direct prior competition. The findings generalize across architectures, provide a compositional model of ICL, and offer practical guidelines for prompt design and deployment, including attention‑level interventions and prior‑aware routing to improve reliability in production systems.

Abstract

We provide causal mechanistic validation that in-context learning (ICL) decomposes into two separable mechanisms: Task Schema (abstract task type recognition) and Binding (specific input-output associations). Through activation patching experiments across 9 models from 7 Transformer families plus Mamba (370M-13B parameters), we establish three key findings: 1. Double dissociation: Task Schema transfers at 100% via late MLP patching; Binding transfers at 62% via residual stream patching -- proving separable mechanisms 2. Prior-Schema trade-off: Schema reliance inversely correlates with prior knowledge (Spearman rho = -0.596, p < 0.001, N=28 task-model pairs) 3. Architecture generality: The mechanism operates across all tested architectures including the non-Transformer Mamba These findings offer a mechanistic account of the ICL puzzle that contrasts with prior views treating ICL as a monolithic mechanism (whether retrieval-based, gradient descent-like, or purely Bayesian). By establishing that Schema and Binding are neurally dissociable -- not merely behavioral modes -- we provide causal evidence for dual-process theories of ICL. Models rely on Task Schema when prior knowledge is absent, but prior knowledge interferes through attentional mis-routing (72.7% recency bias) rather than direct output competition (0%). This explains why arbitrary mappings succeed (zero prior leads to full Schema reliance) while factual overrides fail -- and reveals that the true bottleneck is attentional, not output-level. Practical implications: Understanding these dual mechanisms enables more efficient prompt engineering -- reliable schema transfer reduces required demonstrations for novel tasks, while prior-aware design can mitigate the 38% binding failure rate in high-prior scenarios, improving ICL system reliability in production deployments.

Paper Structure

This paper contains 65 sections, 8 equations, 3 figures, 13 tables.

Table of Contents

  1. Introduction
  2. Explaining the ICL Asymmetry.
  3. Contributions.
  4. Related Work
  5. Dual Mechanisms in ICL.
  6. Mechanistic Interpretability of ICL.
  7. ICL as Implicit Inference.
  8. Role of MLP Layers.
  9. Prior Knowledge and ICL.
  10. ICL Beyond Transformers.
  11. Methods
  12. Models
  13. Tasks
  14. Activation Patching Protocol
  15. Schema Vector Extraction.
  16. ...and 50 more sections

Figures (3)

  • Figure 1: Layer-wise Task Schema Gradient. Schema encoding concentrates in late MLP layers (75--95% depth), with minimal signal in early/middle layers.
  • Figure 2: Prior-Schema Trade-off. Task Schema Gradient inversely correlates with prior knowledge across 28 task-model pairs (Spearman $\rho = -0.596$, $p < 0.001$). The 8 task types from Table \ref{['tab:tasks']} are evaluated across representative models; additional task variants (e.g., number_next, fruit_number) are included to increase statistical power. Low-prior tasks (left) show stronger Schema reliance; high-prior tasks (right) show reduced Schema reliance.
  • Figure 3: Ordering Robustness. (a) Absolute performance across demo orderings (CV=1.8%). (b) Schema selectivity remains stable across all 24 orderings (CV=2.0%, mean=0.967). Both metrics show low variance, confirming Task Schema encoding is robust to ordering variations.