Skip to content

Source. Causal Mediation Analysis for Interpreting Neural NLP: The Case of Gender Bias (Vig et al., 2020).

Description. Causal mediation analysis treats a gender edit to the input as a treatment and each neuron, layer, head or attention weight as a mediator, splitting the effect on a two-pronoun probability ratio into a natural direct and a natural indirect effect. Three properties of the effect distribution are reported — sparsity, with ten of 144 heads reproducing the effect of intervening on all of them; synergy among neurons; and decomposition into direct and indirect parts — against a randomly initialized GPT2-small carried through as a matched negative control. Evaluated on: five GPT-2 sizes on 17 templates over 169 professions plus WinoBias Type 1 and Winogender, with five further model families added in §5.5. Description mode: implementational-topographic.

Gender bias circuits. Readings of the claim, from the version the authors state to the version the field cites.

ReadingVerdictMissing
Primary. Ten of 144 attention heads carry the indirect effect of a gender edit on the pronoun probability ratio, and neurons combine synergistically to produce itCausally SuggestiveSpecificity, which has no second outcome to register on (I4); necessity, which the design’s substitution estimands cannot express (I1); correction for a greedy maximum over 144 candidates (M7); and dispersion on the head-level and neuron-level effects that carry the sparsity claim (M1)
These components mediate gender bias rather than gender processingUnderdeterminedA localized gender-competence mechanism to compute an overlap against; none is measured, and the definitionally gendered professions that would carry the test are excluded from the total-effect calculation (C4, V3, I5)
Training is what installs the sparse mediation structureUnderdeterminedA trajectory. The coupling is a single endpoint contrast against one untrained model whose total effects are small rather than absent — 0.07 on WinoBias against GPT2-small’s 0.25 (I12) — with no intermediate checkpoint anywhere (I11)
The effect grows with model size and the pattern holds across model familiesDisconfirmedNothing — tested and failed: the neuron-level pattern does not transfer to the masked language models and the authors state they have no theory for the difference (E4, M6), and the size trend holds on the Winograd sets but not on Professions, where model size is not well correlated with total effect (E4)

Gender bias circuits. Verdict: Causally Suggestive. The claim stops at the rung marked Blocked; rungs below it are Reached and rungs above it are blocked by that rung. Criteria in bold are absent or adverse at that rung; only those the rung requires can block it.

TierRequiresMissing
Proposedconstruct (C1–C2)Reached.
Causally Suggestivemeasurement (M2); internal (I1)Reached.
Mechanistically Supportedinternal (I2, I4); external (E1)Blocked. E1 intervention reach — One instrument throughout, so every agreement is the method with itself; I4 specificity — A two-element outcome set leaves off-target damage nowhere to register
Triangulatedconstruct (C3–C4); internal (I5–I7); external (E2, E4)Blocked at Mechanistically Supported. C4 discriminant validity — No second construct is ever localized, so nothing can separate from it; I6 double dissociation — One behavior by construction, so the second dissociation cannot be run
Validatedconstruct (C5–C6); measurement (M1–M6); internal (I3, I10–I12); external (E2–E6); interpretive (V1–V5)Blocked at Mechanistically Supported. E3 cross-task generalization — One two-candidate setup throughout; generality claimed, not shown; I10 rescue reversibility — Counterfactual activations exist for every example; no recovery is read; I11 onset coupling — Two states, before training and after; no sequence to locate an onset in

Gender bias circuits. Full 36-criterion audit. Status: C = Confirmed, PC = Partially confirmed, U = Untested, I = Inconclusive, D = Disconfirmed, N/A = Not applicable.

IDCriterionStatusEvidence
Construct Validity
C1FalsifiabilityCThe answer could have been 100 heads and was 10; one negative case fires
C2Structural plausibilityPCMediators taken from the architecture; no weight-level mechanism anywhere
C3Convergent validityCTwo mediator families, three datasets, four scales – all one instrument
C4Discriminant validityUNo second construct is ever localized, so nothing can separate from it
C5Nomological validityCPearl’s estimands under his definitions, with conditions stated (Pearl, 2001)
C6Complementation validityN/AMediators carry an index and an estimand, not a labeled role
Measurement Validity
M1ReliabilityPCDispersion reported where it is incidental and absent where it is load-bearing
M2Baseline separationCAn untrained model matched on everything but training, run at both levels
M3StabilityPCSelection rule and outcome scale both varied; the target word is not
M4CalibrationPCA fixed point at 1 and a sign convention; the effect scale has no referent
M5SensitivityPCEffects track published occupational statistics, but at the wrong level
M6InvarianceCSix axes tested, and where the invariance stops is reported too
M7Selection correctionUPool, rule and shortcut all disclosed; nothing corrected for multiplicity
Internal Validity
I1NecessityN/ANeither estimand can be put in the form necessity requires
I2SufficiencyPCTen heads of 144 reproduce the all-mediator effect; no capability measured
I3MinimalityPCGreedy admission by marginal gain; no ablation of the selected ten
I4SpecificityUA two-element outcome set leaves off-target damage nowhere to register
I5Rival mechanism exclusionPCThe untrained control excludes architecture; the stimulus rival stands
I6Double dissociationUOne behavior by construction, so the second dissociation cannot be run
I7Confound controlPCDefinitional gender dropped by crowdsourced rating; frequency untouched
I8Confounding sensitivityUUnattempted; nothing bounds an unmeasured confounder
I9Epistatic interactionCConcurrent against summed intervention, run at both levels, and it splits
I10Rescue reversibilityUCounterfactual activations exist for every example; no recovery is read
I11Onset couplingUTwo states, before training and after; no sequence to locate an onset in
I12Offset couplingPCBehavior and structure both absent untrained; nothing in between
External Validity
E1Intervention reachUOne instrument throughout, so every agreement is the method with itself
E2Prompt generalizationCThree prompt sets built differently, with the differences reported
E3Cross-task generalizationUOne two-candidate setup throughout; generality claimed, not shown
E4Cross-model recurrencePCTransfers within the autoregressive class and fails outside it
E5Graded responsePCDose is the number of mediators, not the strength of the intervention
E6Novel predictionPCOne prediction confirmed outside the model, one sharper one falsified
Interpretive Validity
V1Level declarationCStructural-behavioral declared twice, and no claim sits outside it
V2Level-evidence matchCEvery headline claim is a property of the measured effect distribution
V3Alternative levelPCThe architecture rival is defeated; the decisive one is never raised
V4Unlicensed labelingCComponents carry no names at all, only an index and an estimand
V5Scope declarationCDeclared where each restriction binds, not collected at the end
Total: 11 Confirmed, 14 Partially confirmed, 9 Untested, 2 Not applicable
Verdict: Causally Suggestive

Criterion judgments this audit record leaves contested: each carries an argument on both sides, and the status shipped is the one the record settled on.

CriterionIn tensionWhat would settle it
I6Untested or Not applicableWhether a design supplying one mechanism and one behavior leaves double dissociation unattempted or unaskable

An exploratory reading of this claim through the framework’s five lenses, written for this site and not part of the paper, is at Gender Bias Circuits — exploratory lens analysis.