Algorithm Overview¶

CircuitKIT ships 13 discovery algorithms across 4 backends. This page explains how to choose the right one. Note that "ships" is not "validated". Six are Stable (eap, eap-ig, eap-gp, acdc, ibcircuit, cdt) and have been tested across the GPT-2, Llama, Gemma, and Qwen families. The remaining seven are Research: implemented and validated on GPT-2/IOI but not yet exercised at scale or across architectures. Three of the Stable algorithms carry documented caveats: acdc is node-only by construction and slow, ibcircuit has a memory ceiling on multi-billion-parameter models at aggressive settings, and cdt scores on RoPE models are approximate.

EAP Family Stable

Gradient-based edge attribution patching. Default eap-ig is validated across GPT-2 through Llama-3B and Gemma-4B.

EAP variants
ACDC Stable

Greedy edge-pruning. Produces minimal circuits. Node-only and slow.

ACDC details
IBCircuit Stable

Information-bottleneck noise model. No paired corruption data needed.

IBCircuit details
CD-T Stable

Contextual decomposition through transformers. Clean inputs only. Scores on RoPE models are approximate.

CD-T details

Quick selection guide¶

Goal	Algorithm	Why
New to CircuitKIT, any model	`eap-ig`	Stable, fast, validated across model families
Speed over precision	`eap`	~30% faster; slightly noisier
Minimal circuit	`acdc`	Greedy edge-pruning. Node-only and slow
Information-flow analysis	`ibcircuit`	No paired data needed
Large model (3B+)	`eap-ig`	Validated at this scale. `ibcircuit` has a memory ceiling here and `acdc` is slow

All 13 algorithms¶

EAP family¶

Algorithm	Tier	Description
`eap-ig`	Stable	EAP + Integrated Gradients — default
`eap`	Stable	Vanilla EAP — fast baseline
`eap-gp`	Stable	EAP-GP / GradPath
`eap-ig-activations`	Research	IG over node activations
`eap-clean-corrupted`	Research	EAP with both clean/corrupted passes
`eap-exact`	Research	Exact EAP (quadratic cost)
`atp-gd`	Research	Attribution Patching with GradDrop (AtP+GD)
`relp`	Research	Relevance Patching (LRP-style)
`peap`	Research	Position-aware EAP (PEAP)
`eap-ifr`	Research	Information Flow Routes (IFR)

When to use: Start with eap-ig. Use eap if speed is the bottleneck. Use Research variants only for algorithm comparison studies.

ACDC¶

Algorithm	Tier	Description
`acdc`	Stable	Greedy edge-pruning; node-only by construction (edge search); slow

When you want a minimal circuit and can afford a slow search.

IBCircuit¶

Algorithm	Tier	Description
`ibcircuit`	Stable	Noise-model approach; clean-only data

When paired (clean, corrupted) examples are hard to construct.

Memory ceiling on multi-billion-parameter models

IBCircuit trains a noise model end-to-end, doubling memory. At aggressive settings it can OOM on models above ~3B parameters.

CD-T¶

Algorithm	Tier	Description
`cdt`	Stable	Clean inputs only; frozen-RoPE attention approximation

CD-T uses a frozen-RoPE attention approximation (Q/K are not decomposed) and a 50/50 gated-MLP cross-term split, so its scores on RoPE models are approximate.

Model compatibility¶

Algorithm	GPT-2	Llama 3.x	Gemma 2/3	Qwen 2.5
`eap-ig`	✅	✅	✅	✅
`eap`	✅	✅	✅	✅
`eap-gp`	✅	✅	✅	✅
`acdc`	✅	⚠️	⚠️	⚠️
`ibcircuit`	✅	⚠️	⚠️	⚠️
`cdt`	✅	⚠️	⚠️	⚠️
`eap-ig-activations`	✅	❌	❌	❌
`eap-clean-corrupted`	✅	❌	❌	❌
Research tier	✅	❌	❌	❌

✅ Tested ⚠️ Tested, with a documented caveat ❌ Not validated

Caveats: acdc is slow above GPT-2 scale, ibcircuit has a memory ceiling on multi-billion-parameter models at aggressive settings, and cdt scores on RoPE models (Llama, Gemma, Qwen) are approximate.

Algorithm-specific config keys¶

Algorithm	Key	Default	Description
`eap-ig`	`ig_steps`	`5`	Integration steps for IG
`acdc`	`tao_bases`	`[1, 3, 5, 7, 9]`	Bases for the tao threshold sweep
`acdc`	`tao_exps`	`[-5, -4, -3, -2]`	Exponents for the tao threshold sweep
`acdc`	`faithfulness_target`	`kl_div`	Metric optimized during pruning (`kl_div` or `mse`)
`ibcircuit`	`num_epochs`	`1000`	Training epochs for noise model
`ibcircuit`	`beta`	`0.001`	IB regularization weight