Prerequisite Circuits

Track what an intervention removes—and whether it stays removed as a model learns.

Code and evidence accompanying When Data Can Teach: Prerequisite Circuits Gate Learning, by Kunwar Kalra and Thanush Patlolla. Submitted to ICLR 2027; an arXiv link will be added when available.

Paper · Reproduction · Package guide · Evidence

The research studies how an existing computation changes what fixed future data can teach. Controlled removal/rescue experiments isolate this effect. Pretrained Pythia experiments test training-only head suppression and examine compensation over longer training. The demonstrated task family is induction and copying.

Start with the evidence

The package's reporting API uses only the Python standard library. From a clone:

python -m pip install .
prereq report results/reference/410m_*.json --html outputs/report.html --open

The interactive report works offline: compare runs, inspect exact observations, and export the embedded records. It reads saved evidence and does not train a model. For the paper's figures, raw-record verification, and experiment recipes, see Reproduction.

Run a small experiment

python -m pip install '.[train,plot]'
OMP_NUM_THREADS=1 prereq demo --output outputs/demo
prereq report outputs/demo/*.json --plot outputs/demo/learning.pdf

The CPU example compares available, suppressed, matched-control, correct-supply, and wrong-supply conditions under common evaluation. It is an illustrative fixed-feature model, not a replication or additional result from the paper. The paper's controlled-transformer replication is available separately.

Use the monitor in your experiment

from prerequisite_circuits import Criterion, Observation, Trajectory

trajectory = Trajectory(
    "head-suppression",
    criterion=Criterion(maximum_suppressed=0.20, minimum_drop=0.30, acquisition=0.50),
    metadata={"evaluation": "training intervention removed"},
)
trajectory.append(Observation(
    step=0,
    target_available=0.01,
    prerequisite_available=0.66,
    prerequisite_suppressed=0.06,
))
trajectory.save("outputs/trajectory.json")

These are illustrative values. Supply measurements from your own independent probe and choose the criterion before evaluating outcomes. The guide shows how to attach the monitor to a PyTorch loop without replacing its optimizer or trainer.

The optional PyTorch API provides scoped channel/head interventions, detached activation supply, exact batch replay checks, and a small matched-branch runner. Hooks are removed on exit, including exceptions. Evaluation preserves random states, module modes, and buffers. The runner initializes a fresh optimizer in each branch; it does not silently inherit optimizer history.

What version 0.1 establishes

The package reports intervention effectiveness on a declared probe at observed updates. It also measures target acquisition under the specified evaluation condition. It does not infer continuous suppression, discover a complete circuit, or turn temporal ordering into a causal mechanism.

A candidate can be evaluated with removal, controls, and rescue. Automatic candidate discovery and validated mechanisms beyond induction are future research directions. New task/mechanism adapters can be added without changing the record format. See Methods and scope.

Repository

Original execution sources and larger records are separate, checksum-verified release assets. They retain their historical identities. Refactored package code is a new implementation with explicit parity tests, not a relabeled historical execution. No training or multi-GB download occurs on import or installation.

Development

See Development and releases for the public/private research workflow and publication steps.

python -m pip install -e '.[dev,pythia]'
python -m pytest
ruff check src tests examples analysis experiments
python -m build

The scientific checks include channel/gradient scope, restoration cleanup, data-replay mismatches, observation neutrality, and independent reconstruction of GPT-NeoX head outputs. GPU experiments are opt-in. Original code is MIT licensed; third-party assets retain their own licenses. See CITATION.cff.