Skip to content

This page maps MI techniques to the metrics that use them. Methods are not an evidence category — they are implementation techniques that produce evidence in various families (causal, structural, representational, etc.). Each metric’s canonical page is in its evidence-type category; this page is a cross-reference for readers who want to find “all metrics that use ACDC” or “all metrics that use DAS.”

MethodMetricEvidence familyCanonical page
ACDC (Conmy et al. 2023)C10 ACDCCausalMI Causal
EAP (Syed et al. 2023)C14 Position-Aware EAPCausalMI Causal
Information BottleneckC13 IB Circuit DiscoveryCausalMI Causal
CircuitLensC20 CircuitLensCausalMI Causal
MethodMetricEvidence familyCanonical page
Sparse Feature CircuitsC08 SFCCausalMI Causal
Relevance Patching / LRPC11, C19 RelPCausalMI Causal
Contextual DecompositionC12 CDCausalMI Causal
VPDC18 VPDCausalMI Causal
Activation PatchingC2CausalCore A01 Pearl SCM
Path PatchingC33CausalMI Causal
MethodMetricEvidence familyCanonical page
CAA (Panickssery et al. 2024)C09 CAASteeringMI Steering
LEACE (Belrose et al. 2023)C15 Concept ErasureSteeringMI Steering
RepE (Zou et al. 2023)C16 RepESteeringMI Steering
Steering-BenchB21SteeringMI Steering
MethodMetricEvidence familyCanonical page
DAS / IIA (Geiger et al. 2024)C1 DAS-IIACausalMI Causal
Causal Scrubbing (Chan et al. 2022)C4CausalMI Causal
MethodMetricEvidence familyCanonical page
NOTEARS (Zheng et al. 2018)C9CausalMI Causal
oCSE (Sun et al. 2023)C7CausalMI Causal
PC AlgorithmC42CausalMI Causal
Granger CausalityC56Information-theoreticMI Information
MethodMetricEvidence familyCanonical page
SAEAQ01-AQ09Artifact qualityMI Artifact Quality
TranscoderAQ10-AQ12Artifact qualityMI Artifact Quality
CrosscoderAQ13-AQ15Artifact qualityMI Artifact Quality
CLT / Circuit TracingFH01-FH05FaithfulnessMI Faithfulness