How K-Composition and Induction Circuits Drive In-Context Learning in AI

The binding constraint of any deployed large language model is that its memory is frozen. Once training finishes, the billions of neural weights that define the model cannot change, meaning it cannot learn new facts by updating its parameters. The architecture is entirely static during inference.

Yet, when a user types a prompt with a completely made-up word or a novel formatting pattern, the model learns it instantly. It adapts to the context without a single parameter update, a phenomenon known as in-context learning. This capability allows models to perform tasks they were never explicitly trained to do.

For years, how a static matrix of numbers could dynamically learn on the fly remained a black box to developers. Now, mechanistic interpretability researchers have isolated the exact mathematical engine responsible: a two-layer mechanism called an induction circuit. This discovery provides a physical explanation for an emergent behavior.[1][3]

By tracing the flow of data through these circuits, researchers found that models use a process called K-composition to temporarily rewire their own attention. It allows the model to detect and copy token sequences directly within the prompt, effectively building a temporary memory bank.[2]

The Anatomy of the Residual Stream

To understand how induction circuits work, one must first look at the transformer's central highway, known as the residual stream. This is a high-dimensional vector space where every token in the prompt maintains its own running state, constantly updated as it passes through the network.[2]

As information flows forward through the model's layers, attention heads read from this stream and write new information back into it. The stream acts as a shared memory buffer, accumulating context and enriching the mathematical representation of each word as it moves deeper into the architecture.

"All components of a transformer communicate with each other by reading and writing to different subspaces of the residual stream," explains the foundational 2021 paper by Elhage et al. on transformer circuits. This additive writing is the foundation that makes complex multi-layer behaviors possible.[2]

Crucially, attention heads operate in two parallel tracks to manage this flow. The Query-Key (QK) circuit determines where the head looks by calculating attention scores, while the Output-Value (OV) circuit determines what specific information it copies from the target back to the current token's residual stream.[2]

Step One: The Previous Token Head

The induction circuit requires a minimum of two attention layers to function, making it impossible for single-layer models to learn in-context. The process begins in the first layer with a specialized component known as a "previous token head," which performs a highly specific copying task.[1][2]

When the previous token head looks at a sequence, its QK circuit is hardwired to attend exactly one position backward. If the prompt contains the sequence "cat sat," the head at the "sat" position looks exclusively at the preceding "cat" token, ignoring all other context.[1]

Once it focuses on the preceding token, the head's OV circuit copies the identity of that token into the current token's residual stream. After this layer executes, the token "sat" carries a mathematical label advertising that its predecessor was "cat," fundamentally altering its representation.[1][2]

This operation enriches the data passing through the network. The model is no longer looking at isolated words; each position now holds a composite representation of both itself and the token that came immediately before it, setting the stage for complex pattern matching.[4]

Step Two: K-Composition in Action

The magic of in-context learning happens in the second layer, where the actual "induction head" resides. This head uses the enriched data from the first layer to perform a sophisticated pattern-matching operation called K-composition, linking the two layers into a single functional circuit.[1][3]

K-composition occurs when a layer-two head uses the output written by a layer-one head to construct its Key matrix. Instead of just looking at the raw token, the induction head's keys now include the predecessor information, changing what each position advertises about itself.[2]

Imagine the prompt later repeats the word "cat" and asks for the next word. The induction head's Query at this new "cat" position searches backward through the sequence, looking for any token whose Key explicitly advertises the label "predecessor = cat."[1]

Because the previous token head already wrote "predecessor = cat" into the residual stream of the original "sat" token, the induction head finds a perfect mathematical match. The QK circuit locks onto "sat" as the target, ignoring the rest of the sequence.[1][4]

Copying the Future

With the attention pattern locked onto the correct historical token, the induction head executes its final move. Its OV circuit copies the identity of that target token—"sat"—and writes it into the current position's residual stream, preparing the model to generate its prediction.[1]

The model then translates this updated residual stream into its final output, predicting "sat" as the next word. It has successfully recognized a pattern, retrieved the historical continuation, and copied it, all without altering a single static weight in its billions of parameters.[1][3]

"Induction heads search the context for previous instances of the present token, attend to the token which would come next if the pattern repeated, and increase its probability," notes the 2022 Anthropic study by Olsson et al., which first formalized the mechanism.[1]

This mechanism explains why large language models are so adept at few-shot prompting. When a user provides three examples of translating English to French, the induction circuits simply match the formatting prefix and copy the corresponding structural output, executing the task seamlessly.[1][4]

The Phase Change in Training

The discovery of induction circuits also solved a lingering mystery about how language models learn during their initial training phase. Researchers observed that models do not acquire in-context learning capabilities gradually over millions of training steps, as one might expect from a neural network.[1]

Instead, models undergo a sudden "phase change" early in their training run. Over a very narrow window of optimization steps, the model's ability to learn from its context window spikes dramatically, visible as a sharp drop in training loss when processing repeated sequences.[1][3]

By tracking the internal weights of the network, researchers proved that this phase change occurs exactly when the previous token heads and induction heads align to form K-composition circuits. The capability emerges the moment the two layers connect and begin sharing information.[1][2]

In models with only a single attention layer, this phase change never happens. Because K-composition requires one head to read the output of another, a one-layer architecture physically cannot form an induction circuit, permanently capping its reasoning abilities regardless of how long it trains.[1][2]

Measuring the Circuit's Strength

To prove that K-composition is the primary driver of this behavior, researchers do not just observe the model; they actively intervene. By mathematically ablating, or deleting, specific induction heads during inference, they can measure the exact drop in the model's performance.[1]

When the induction heads are removed from a small attention-only model, its ability to perform in-context learning vanishes almost entirely. The model reverts to acting like a basic statistical predictor, unable to follow novel prompt instructions or recognize repeated patterns in the text.[1][3]

Furthermore, researchers can quantify the exact strength of the K-composition between any two heads. They calculate the Frobenius norm of the matrix product between the first head's OV circuit and the second head's QK circuit, providing a concrete metric for how much information flows between them.[2]

High composition scores reliably identify which heads are working together across layers. This mathematical precision allows engineers to map the computational graph of a transformer just as an electrician would trace the wiring in a physical circuit board, moving AI from alchemy to engineering.[2][4]

Bypassing the Weight Bottleneck

Understanding K-composition fundamentally shifts how researchers view artificial intelligence. The neural weights are not merely a database of memorized facts; they are a collection of algorithms designed to manipulate the residual stream in real time, creating dynamic programs on the fly.[4]

By offloading the pattern-matching to the context window, the model bypasses the bottleneck of its frozen parameters. The induction circuit acts like a dynamic RAM module, storing and retrieving temporary variables without needing to write them to the permanent hard drive of its weights.[1][4]

This mechanistic transparency is crucial for AI safety and alignment. If engineers know exactly which attention heads are responsible for copying text, they can monitor those specific circuits to prevent the model from repeating toxic or classified information, intervening before the output is generated.[3][4]

As models scale to trillions of parameters, these basic induction circuits combine into increasingly complex structures that handle abstract reasoning. Yet, the foundational math remains the same: one layer writes the history, and the next layer uses K-composition to copy the future.[1][2]