Mechanistic story
Awaiting repeated prefixStep forward until the current token repeats an earlier prefix.
See a pattern once. Copy what follows.
Induction heads are a proposed attention circuit for in-context learning. One head carries the previous token; another finds an earlier matching token and copies what came next. Ablate, patch, and distract the circuit to test the mechanism.
[A][B] … [A] → predict [B]previous-token head + induction head
Token stream
Repeated prefixes create candidate copy paths.
Previous-token signal
Match attention
Copy logit
Attention circuit
Arcs show the previous-token relay and the earlier matching prefix.
Token 0
Causal test
Circuit intactAblation should reduce copying; patching should restore the missing signal.
Copy confidence0%
Match position—
Distractor mass0%
Circuit effect0%
A convincing circuit must survive intervention.
Attention patterns alone are correlational. A mechanistic claim gets stronger when ablation removes behavior, activation patching restores it, and the effect generalizes across controlled token patterns.
Two heads compose across layers.
The earlier head writes previous-token information into the residual stream. A later head uses that signal to match an earlier prefix and increase the logit of its continuation.
In-context copying appears suddenly.
Training curves can show a phase change when induction heads form.
AB…A→B
Pattern is not proof.
A diagonal attention stripe can coexist with redundant or compensating mechanisms.
Ablation tests necessity.
Remove the head or path and measure the behavior’s drop against matched controls.
Patching tests sufficiency.
Restore a clean activation into a corrupted run and inspect whether copying returns.
Intervene on the path.
Freeze prompts, tokenization, positions, model checkpoint, head identities, ablation rule, patch source, and behavioral metric before claiming a circuit.
Counterfactual tokens
Change the repeated prefix while preserving length and frequency. The proposed match should move predictably.
Matched ablations
Compare the target head with random heads, same-layer heads, and path-specific interventions.
Activation patching
Patch clean signals into corrupted runs and localize where the correct continuation reappears.
Three observations, only one causal chain.
Hover or focus each view to expose what it can and cannot establish.
Primary sources behind the circuit.
The values are illustrative. These papers introduce transformer circuits, induction heads, causal tracing, and activation patching methodology.
CIRCUITS FRAMEWORK
A Mathematical Framework for Transformer Circuits
Develops composition, QK/OV circuits, and interpretable attention paths.
INDUCTION HEADSIn-context Learning and Induction Heads
Connects induction heads with a phase change in in-context learning.
CAUSAL TRACINGLocating and Editing Factual Associations in GPT
Uses causal tracing to localize model computations.
PATCHING METHODSTowards Automated Circuit Discovery
Formalizes edge-level activation patching and circuit discovery.
Export a circuit audit.
Preserve the tokens, intervention, effect size, and controls. Then see how Superpowers builds inspectable agent tools and caches repeatable computer-use evaluations.
0 of 4 ready