💬 NLP

Schema for In-Context Learning

Inspired by cognitive schema theory, the paper introduces Schema-Activated In-Context Learning (SA-ICL), a framework that explicitly constructs abstracted reasoning templates from prior examples to significantly boost the performance and interpretability of large language models on complex scientific tasks while reducing reliance on numerous demonstration examples.

Original authors: Pan Chen, Shaohong Chen, Mark Wang, Shi Xuan Leong, Priscilla Fung, Varinia Bernales, Alan Aspuru-Guzik

Published 2026-03-31

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Pan Chen, Shaohong Chen, Mark Wang, Shi Xuan Leong, Priscilla Fung, Varinia Bernales, Alan Aspuru-Guzik

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Idea: Waking Up the Brain's "Filing System"

Imagine you are a brilliant student (the AI) sitting in a classroom. The teacher hands you a new, tricky math problem. To help you, the teacher gives you a similar example from last week's homework.

The Problem (Schema Dormancy):
Usually, you look at the example and try to copy the exact steps the teacher took. You might get the answer right, but often you get stuck because you're just mimicking the surface details (like the numbers used) rather than understanding the deep logic behind the problem. It's like trying to solve a puzzle by looking at the picture on the box, but the puzzle pieces you have are slightly different shapes. You're relying on "surface-level shortcuts" instead of deep reasoning.

The authors call this Schema Dormancy. It's like having a library of knowledge in your brain, but the lights are off. You have the books (the reasoning frameworks), but you can't find the right one to open because no one told you which aisle to go to.

The Solution (SA-ICL):
The paper introduces a new method called Schema-Activated In-Context Learning (SA-ICL).

Instead of just handing you the example, the teacher first says: "Before we look at the numbers, let's remember the concept of 'balancing scales'."

This is the "Schema Activation." The teacher explicitly wakes up the right mental framework (the "Schema") before showing you the specific problem. Once that framework is active, you can look at the new problem and instantly see how to fit the pieces together, regardless of whether the numbers or the story are slightly different.

The "Washing Machine" Analogy

To prove this works, the researchers ran a famous experiment (originally done with humans in 1972).

The Scenario: They gave an AI a confusing paragraph describing a process: "First, you arrange things into groups. If you don't have enough space, you go somewhere else. It's important not to do too many things at once..."
Without the Key: If you just read this, it sounds like a mess. Is it about cooking? Cleaning? Gardening? The AI gets confused and its attention is scattered.
With the Key: If you add the words "Washing Machine" at the very top, the paragraph suddenly makes perfect sense. The AI's brain "lights up." It knows exactly what "arranging things into groups" means (sorting clothes), and it understands the whole story instantly.

The paper shows that AI models suffer from the same confusion. They need that "Washing Machine" cue (the Schema) to organize the messy information before they can solve the problem.

How It Works: The REAP Process

The authors created a 4-step recipe to help the AI wake up its brain:

R (Representation): Look at the new problem and figure out what kind of puzzle it is.
R (Retrieval): Go to your mental library and find the best "rulebook" (Schema) for this type of puzzle.
E (Episodic Retrieval): Find a specific past example that fits this rulebook.
A (Activation): This is the magic step. Explicitly tell the AI: "Use the 'Physics Conservation of Energy' rulebook to solve this."

Once the AI is told which rulebook to use, it stops guessing and starts reasoning structurally.

Why This Matters: The Results

The researchers tested this on hard questions in Science, Statistics, and Finance.

The Old Way (One-Shot): Give the AI an example and say "Solve this."
- Result: The AI often fails or gives a weak answer because it's just copying the example without understanding the deep structure.
The New Way (SA-ICL): Tell the AI the rulebook first, then give the example, then ask it to solve.
- Result: The AI got significantly better at solving problems, especially when the examples were very similar to the question.

The "Aha!" Moment:
The most surprising finding was that when the AI did get the right "rulebook" (Schema), it didn't just get the answer right; it actually became more efficient. It used fewer words to explain its thinking because it didn't have to wander around guessing. It knew exactly where to go.

The Takeaway

Think of Large Language Models (LLMs) as incredibly smart but slightly absent-minded geniuses. They have all the knowledge in the world, but they sometimes forget how to organize it.

Schema-Activated ICL is like giving that genius a sticky note that says: "Remember, you are a Detective, not a Chef. Use your Detective logic."

By explicitly activating the right way of thinking before asking the AI to solve a problem, we stop it from relying on shallow tricks and help it use its true reasoning power. This makes AI smarter, more reliable, and better at handling complex real-world tasks.

1. Problem Statement: Schema Dormancy

The paper identifies a critical limitation in current Large Language Models (LLMs) utilizing In-Context Learning (ICL). While ICL allows models to perform tasks without parameter updates, models often rely on surface-level shortcuts (pattern matching) rather than abstract reasoning.

The authors attribute this to Schema Dormancy:

Definition: The phenomenon where an LLM fails to spontaneously activate a relevant abstract reasoning template (schema) during inference, despite possessing the latent capacity to use it when explicitly prompted.
Consequence: Models retrieve plausible analogs based on surface similarity but fail to align structural relationships, leading to errors in complex reasoning tasks (Science, Statistics, Finance).
Cognitive Basis: Drawing from Schema Theory in cognitive psychology (Rumelhart & Ortony, 1977), the authors argue that human learning involves organizing knowledge into mental frameworks (schemas) that guide interpretation. LLMs currently lack a mechanism to spontaneously "activate" these frameworks before processing a query.

2. Methodology: Schema-Activated ICL (SA-ICL)

The authors propose SA-ICL, a framework that explicitly decouples reasoning structure from task content to enforce structural mapping. The method mimics human cognitive processes through a five-step workflow (REAP + Activation):

Problem Latent Representation: The model generates an initial representation ( $S_x$ ) of the input problem.
Prior Schema Retrieval & Selection: The system retrieves the most relevant abstract schema ( $\hat{S}$ ) from a memory bank (e.g., Chain-of-Thought, Least-to-Most, Step-Back templates) that maximizes similarity to $S_x$ .
Episodic Examples Retrieval: Conditioned on the retrieved schema, the system selects a subset of past episodic examples ( $\hat{E}_\tau$ ) whose association weights exceed a decay threshold. This simulates "working memory" retrieval.
Schema Activation (Assimilation/Accommodation): The retrieved schema and examples are integrated with the current problem.
- Assimilation: If the schema fits, it is applied directly.
- Accommodation: If the fit is poor, the schema is restructured to fit the new context.
Schema-Guided Inference: The LLM solves the task by conditioning on the input and the activated/refined schema ( $S_{new}$ ), rather than just the raw examples.

Key Technical Mechanism:
The paper validates that schema activation is a tangible mechanical phenomenon. In preliminary experiments using the "Washing Machine" ambiguous text (Bransford & Johnson, 1972), injecting a schema keyword ("Washing Machine") caused attention entropy to collapse in specific induction heads. This shifted the model from diffuse, surface-level attention to focused, structural attention, effectively grounding abstract concepts.

3. Experimental Setup

Models: Evaluated across 7 diverse LLMs, including GPT-4o, GPT-4o Mini, DeepSeek-R1-Distill-Llama, Llama-3.1, Qwen-3, MistralSmall, and Gemini 2.0-Flash.
Datasets:
- GPQA-Diamond: Graduate-level Science (Chemistry, Physics, Biology).
- MMLUProStat: Statistics.
- MMLUProFinance: Finance.
- LabSafety: Multi-modal safety dataset.
Baselines: Compared against One-Shot and One-Shot + Chain-of-Thought (CoT).
Variables: Experiments were stratified by Knowledge Density (similarity between the in-context example and the target question):
- Essentially Same (High similarity)
- Similar (Medium similarity)
- Different (Low similarity/Out-of-domain)

4. Key Results

Performance Gains: SA-ICL consistently outperformed baselines. On average, it improved accuracy by +5.1% across all knowledge levels ( $p < 0.001$ ).
High-Knowledge Density: The most significant gains occurred when the example was "Essentially Same" to the target. In these high-similarity scenarios, Schema Dormancy was most pronounced as a bottleneck; SA-ICL resolved this by activating the latent schema, achieving a 6.5% improvement over baselines.
Robustness:
- SA-ICL showed superior robustness against retrieval noise (irrelevant examples).
- In "Schema Disruption" tests (using a Biology example for Chemistry/Physics questions), SA-ICL suffered less performance degradation than standard One-Shot + CoT.
Model Agnosticism: The method was effective for 5 out of 7 models tested. Notably, it provided significant gains for smaller models (e.g., Llama-3, DeepSeek) which often struggle with abstract reasoning.
Semantic Drift: The paper quantified "Schema Shift" (the semantic difference between latent and activated schemas). A greater semantic shift correlated with a higher probability of repairing incorrect reasoning paths (up to 7% repair rate).
Efficiency: In interpretability tests, SA-ICL generated fewer tokens in the final output compared to verbose CoT baselines while maintaining or improving accuracy, suggesting more efficient internal reasoning.

5. Key Contributions

Conceptualization of Schema Dormancy: Defined and empirically demonstrated that LLMs often fail to activate latent reasoning frameworks spontaneously, relying instead on surface patterns.
SA-ICL Framework: Introduced a unified, modular framework that explicitly activates abstract schemas before inference, bridging cognitive science theory with LLM mechanics.
Mechanistic Evidence: Provided attention-layer analysis showing that schema activation reduces entropy and restructures information flow from surface association to structural coherence.
Comprehensive Evaluation: Validated the approach across multiple domains (Science, Stats, Finance), model sizes, and knowledge densities, proving it is a generalizable improvement over standard ICL.

6. Significance and Future Directions

Paradigm Shift: SA-ICL moves ICL from "instance-driven retrieval" (matching examples) to "abstraction-driven cognition" (activating frameworks). This suggests that future ICL architectures should prioritize schema cultivation over simply increasing context window size or example quantity.
Interpretability & Safety: By making the reasoning process explicit through schema activation, the method offers a pathway to better AI interpretability and safety, as the model's "thought process" becomes more transparent and aligned with human cognitive structures.
Future Work: The authors suggest extending SA-ICL to agentic systems for dynamic schema generation and applying it to long-form scientific explanations and mathematical proofs.

In conclusion, the paper argues that the next frontier of In-Context Learning lies not in better examples, but in architectures that actively cultivate and activate latent cognitive schemas, thereby enabling LLMs to reason structurally rather than just statistically.