Skip to content
Mechanistic Validity

MI Artifact Quality Metrics

Section titled “MI Artifact Quality Metrics”

This page documents metrics that evaluate the quality, interpretability, and validity of learned decomposition artifacts — SAE features, transcoder circuits, and crosscoder representations. These metrics test whether the artifacts produced by a decomposition method are faithful, interpretable, and free from known pathologies.

SAE and Feature-Level Quality

Section titled “SAE and Feature-Level Quality”

These metrics evaluate whether individual features from sparse autoencoders or similar decompositions are interpretable, stable, and semantically valid.

AQ01 — Automated Interpretability

Section titled “AQ01 — Automated Interpretability”

ID. AQ01.autointerp | File. EX3_autointerp.py

What it computes. Tests whether artifact features have human-interpretable descriptions that predict activation patterns, via detection (can a judge identify activating examples?), fuzzing (do synthetic inputs activate the feature?), and intervention (does ablation produce predicted changes). Falls back to activation-statistics proxies (monosemanticity, sparsity) without an LLM API.

Evidence family. Measurement (M4 Sensitivity)

Pass threshold. Proxy-based; varies by sub-metric.

Reference. Bills et al. (2023); Templeton et al. (2024).

AQ02 — Feature Absorption

Section titled “AQ02 — Feature Absorption”

ID. AQ02.feature_absorption | File. EX4_feature_absorption.py

What it computes. Detects the absorption pathology where a parent feature fails to fire because a more specific child feature absorbs its activation (e.g., an “A” feature fails on “Attention” because an “Attention”-concept feature captures the signal instead).

Evidence family. Measurement (M4 Sensitivity)

Pass threshold. absorption_rate < 0.15 (lower is better)

Reference. Klindt, Bloom et al. (2025), NeurIPS 2025 Oral.

AQ03 — NLA Reconstruction Fidelity

Section titled “AQ03 — NLA Reconstruction Fidelity”

ID. AQ03.nla_reconstruction | File. EX5_nla_reconstruction.py

What it computes. Tests whether a feature’s activation pattern is coherent enough to survive a compress-decompress cycle by identifying top-k activating positions, reconstructing a concept direction via PCA, and comparing against the original feature direction.

Evidence family. Measurement (M4 Sensitivity)

Pass threshold. mean_roundtrip_fidelity > 0.3

Reference. Anthropic (2026), transformer-circuits.pub/2026/nla/.

AQ04 — TopK SAE Scaling

Section titled “AQ04 — TopK SAE Scaling”

ID. AQ04.topk_scaling | File. EX7_topk_scaling.py

What it computes. Three complementary SAE quality sub-metrics: (a) planted feature recovery (whether SAE directions recover planted random unit directions), (b) downstream effect sparsity (L0 of per-feature logit effects through W_U), and (c) activation explainability (kurtosis and sparsity of activation patterns).

Evidence family. Measurement (M4 Sensitivity)

Pass threshold. recovery > 0.3; sparsity > 0.8; explainability > 0.5

Reference. Gao, Dupre la Tour et al. (2025), ICLR 2025 Oral.

AQ05 — RoPE Massive Value Filter

Section titled “AQ05 — RoPE Massive Value Filter”

ID. AQ05.rope_massive_value | File. EX8_rope_massive_value.py

What it computes. Detects SAE features contaminated by RoPE positional encoding artifacts (massive values in Q/K projections) rather than representing genuine semantic features, via alignment with RoPE frequency bases, position-dependence, and magnitude checks.

Evidence family. Measurement (M4 Sensitivity)

Pass threshold. clean_fraction > 0.9

Reference. Jin et al. (2025), ICML 2025.

AQ06 — Latent Self-Consistency

Section titled “AQ06 — Latent Self-Consistency”

ID. AQ06.latent_self_consistency | File. EX11_self_consistency.py

What it computes. Tests SAE encoder-decoder stability via iterated encode-decode roundtrips, measuring whether latent representations converge (stable fixed point) or drift (saddle point instability) across multiple cycles.

Evidence family. Measurement (M4 Sensitivity)

Pass threshold. single_roundtrip_consistency > 0.85

Reference. Buchanan et al. (2026), ICLR 2026.

AQ07 — GradSAE Causal Influence Score

Section titled “AQ07 — GradSAE Causal Influence Score”

ID. AQ07.gradsae_causal_influence | File. 113_gradsae_causal.py

What it computes. For each SAE latent, computes activation times gradient (of output probability w.r.t. latent activation) to produce per-feature causal influence scores. The causal dissociation rate measures how many top-by-activation features are NOT in the top-by-causal-influence ranking.

Evidence family. Internal (I1 Necessity, I2 Sufficiency)

Pass threshold. causal_dissociation_rate < 0.3

Reference. Shu et al. (2025), EMNLP 2025.

AQ08 — Output-Centric Description Score

Section titled “AQ08 — Output-Centric Description Score”

ID. AQ08.output_centric_description | File. 116_output_centric.py

What it computes. Compares input-centric descriptions (top activating tokens) against output-centric descriptions (decoder direction projected through the unembedding matrix W_U) to detect correlational features that lack causal execution validity. Classifies features into a 2x2 encoding/execution grid (HH, HL, LH, LL).

Evidence family. Internal (I2 Sufficiency), Construct (C5 Convergent Validity)

Pass threshold. encoding_execution_agreement > 0.2

Reference. Gur-Arieh et al. (2025), ACL 2025.

AQ09 — Neuronpedia Agreement

Section titled “AQ09 — Neuronpedia Agreement”

ID. AQ09.neuronpedia_agreement | File. 126_neuronpedia_agreement.py

What it computes. Cross-references weight-based top-k token analysis (decoder direction projected through W_U) against Neuronpedia’s automated feature descriptions and top activating tokens, measuring external convergent validity via Jaccard overlap.

Evidence family. Construct (C5 Convergent Validity)

Pass threshold. neuronpedia_weight_agreement > 0.2

Reference. Neuronpedia community infrastructure (neuronpedia.org).

Transcoder Quality

Section titled “Transcoder Quality”

These metrics evaluate validity properties specific to transcoders — models that decompose MLP computation into input-output feature circuits rather than internal representations.

AQ10 — Transcoder Circuit Composability

Section titled “AQ10 — Transcoder Circuit Composability”

ID. AQ10.transcoder_composability | File. 139_transcoder_composability.py

What it computes. Tests whether transcoder decoder directions compose correctly with downstream weight matrices by comparing weight-predicted effects (decoder direction projected through next layer’s input weights) against patching-measured effects (observed change when zeroing a feature’s contribution).

Evidence family. Internal (C2 Structural Plausibility, I2 Sufficiency)

Pass threshold. composability_score > 0.5

Reference. Dunefsky, Chanin, Nanda (2024), arXiv:2406.11944.

AQ11 — Transcoder vs. SAE Feature Agreement

Section titled “AQ11 — Transcoder vs. SAE Feature Agreement”

ID. AQ11.transcoder_sae_agreement | File. 140_transcoder_sae_agreement.py

What it computes. Compares transcoder features (MLP input-to-output decomposition) against SAE features (MLP output state decomposition) at the same layer, measuring direction similarity and top-activating-token overlap of matched feature pairs.

Evidence family. Construct (C5 Convergent Validity)

Pass threshold. transcoder_sae_agreement > 0.3

Reference. Dunefsky, Chanin, Nanda (2024), arXiv:2406.11944.

AQ12 — Transcoder Decomposition Fraction

Section titled “AQ12 — Transcoder Decomposition Fraction”

ID. AQ12.transcoder_decomposition | File. 141_transcoder_decomposition.py

What it computes. Measures what fraction of MLP output variance is captured by the input-dependent sparse features versus the input-invariant bias/mean term, testing whether sparse features do real computational work or the MLP is bias-dominated.

Evidence family. Construct (M6 Construct Coverage), Measurement (C2 Structural Plausibility)

Pass threshold. input_dependent_fraction > 0.6

Reference. Dunefsky, Chanin, Nanda (2024), arXiv:2406.11944.

Crosscoder Quality

Section titled “Crosscoder Quality”

These metrics evaluate validity properties specific to crosscoders — models that jointly decompose representations across multiple layers or multiple models.

AQ13 — Crosscoder Cross-Layer Persistence

Section titled “AQ13 — Crosscoder Cross-Layer Persistence”

ID. AQ13.crosscoder_persistence | File. 142_crosscoder_persistence.py

What it computes. Tests whether crosscoder features claimed to span multiple layers actually activate coherently at all those layers by measuring Pearson correlation of per-position projection magnitudes between adjacent layers.

Evidence family. Construct (C2 Structural Plausibility), Measurement (M2 Invariance)

Pass threshold. cross_layer_coherence > 0.5

Reference. Lindsey et al. (2024), transformer-circuits.pub.

AQ14 — Crosscoder Model Diffing Validity

Section titled “AQ14 — Crosscoder Model Diffing Validity”

ID. AQ14.crosscoder_model_diff | File. 143_crosscoder_model_diff.py

What it computes. Validates the shared/exclusive feature classification by checking activation behavior: exclusive features should have near-zero activation in the excluded model, shared features should activate similarly in both models.

Evidence family. Internal (C3 Task Specificity), Measurement (M3 Baseline Separation)

Pass threshold. diff_classification_accuracy > 0.7

Reference. Lindsey et al. (2024), transformer-circuits.pub.

AQ15 — Crosscoder L1 Artifact Detection

Section titled “AQ15 — Crosscoder L1 Artifact Detection”

ID. AQ15.crosscoder_artifact_detection | File. 144_crosscoder_artifact_detection.py

What it computes. Detects the Muto et al. sparsity artifact where L1-penalized crosscoder training inflates apparent feature exclusivity, reclassifying genuinely shared features as exclusive. Compares decoder-norm classification against activation-based classification.

Evidence family. Measurement (M6 Construct Coverage), Internal (I5 Confound Control)

Pass threshold. artifact_rate < 0.15

Reference. Muto et al. (2025), NeurIPS 2025.

Summary Table

Section titled “Summary Table”
IDNameFileEvidence FamilyThreshold
AQ01Automated InterpretabilityEX3_autointerp.pyMeasurementproxy-based
AQ02Feature AbsorptionEX4_feature_absorption.pyMeasurement< 0.15
AQ03NLA Reconstruction FidelityEX5_nla_reconstruction.pyMeasurement> 0.3
AQ04TopK SAE ScalingEX7_topk_scaling.pyMeasurementvaries
AQ05RoPE Massive Value FilterEX8_rope_massive_value.pyMeasurement> 0.9
AQ06Latent Self-ConsistencyEX11_self_consistency.pyMeasurement> 0.85
AQ07GradSAE Causal Influence113_gradsae_causal.pyInternal< 0.3
AQ08Output-Centric Description116_output_centric.pyInternal / Construct> 0.2
AQ09Neuronpedia Agreement126_neuronpedia_agreement.pyConstruct> 0.2
AQ10Transcoder Composability139_transcoder_composability.pyInternal> 0.5
AQ11Transcoder-SAE Agreement140_transcoder_sae_agreement.pyConstruct> 0.3
AQ12Transcoder Decomposition141_transcoder_decomposition.pyConstruct> 0.6
AQ13Crosscoder Persistence142_crosscoder_persistence.pyConstruct> 0.5
AQ14Crosscoder Model Diff143_crosscoder_model_diff.pyInternal / Measurement> 0.7
AQ15Crosscoder Artifact Detection144_crosscoder_artifact_detection.pyMeasurement< 0.15

Relationship to Other Pages

Section titled “Relationship to Other Pages”

For circuit faithfulness metrics (CLT attribution, cross-prompt consistency, minimality), see MI Faithfulness Metrics. For safety-relevant construct validation, see MI Safety Metrics. For the overall MI lens framework, see Mechanistic Interpretability.