Probing Classifiers

Source. Probing Classifiers: Promises, Shortcomings, and Advances (Belinkov, 2022).

Description. A classifier $g$ is trained on a model’s intermediate output $f_l(x)$ to predict a property $z$, and its accuracy is reported as evidence that the model has learned information relevant to $z$. The audited object is the method rather than one experiment: the anchor is a review that runs no experiments of its own, so the evidence sits in the studies it formalizes, and its Figure 1 separates the nine components a probing result depends on from the ten controls the literature has added to them. Evaluated on: every architecture family the method has been applied to, from static word embeddings through recurrent and recursive networks to transformers, and outward to speech recognition and computer vision. Description mode: representational.

Readings of the claim

Probing classifiers. Readings of the claim, from the version the authors state to the version the field cites.

Reading	Verdict	Missing
Primary. A probing result licenses that property $z$ is extractable from the representation by a classifier of the stated capacity	Proposed	Necessity: the method performs no removal of its own, and where removal was performed the studies the anchor reports disagree four ways (I1, E1)
A good probe score shows that the model represents $z$	Disconfirmed	Nothing — tested and failed: control tasks attribute most of a nonlinear probe’s accuracy to the probe itself, and even random features decode the property (C4)
A good probe score shows that the model uses $z$ in producing its output	Disconfirmed	Nothing — tested and failed: a control dataset holds $z$ non-discriminative for the original task and the probe recovers it anyway, so decodability does not localize to the task (I4)
Interventions on a probe-identified direction show which features the model uses	Underdetermined	A result that decides the four-way disagreement among the intervention studies of §4.3 (I1, E1). This is the broad scope, under which I2 and E5 are unattempted; under probing as §2 defines it, a read-out with no path writing back in, they are not applicable

Verdict

Probing classifiers. Verdict: Proposed. 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.

Tier	Requires	Missing
Proposed	construct (C1–C2)	Reached.
Causally Suggestive	measurement (M2); internal (I1)	Blocked. I1 necessity — Standard probing performs no removal; removal studies disagree four ways
Mechanistically Supported	internal (I2, I4); external (E1)	Blocked at Causally Suggestive. E1 intervention reach — Probing has no intervention of its own; those applied to it disagree; I4 specificity — A control dataset holds the property constant and probing fails on it
Triangulated	construct (C3–C4); internal (I5–I7); external (E2, E4)	Blocked at Causally Suggestive. C4 discriminant validity — Two neighbors the instrument must separate: probe memorization, random features; I6 double dissociation — Needs two properties and two behaviors; the framework supplies one of each
Validated	construct (C5–C6); measurement (M1–M6); internal (I3, I10–I12); external (E2–E6); interpretive (V1–V5)	Blocked at Causally Suggestive. I3 minimality — The unit is a whole intermediate output; no operation removes part of one; I10 rescue reversibility — Every catalogued intervention runs one way, projecting or training out; I11 onset coupling — The original model is taken as given and trained once; I12 offset coupling — The converse needs the training sequence onset coupling also lacks; M1 reliability — A review reports no variance; the anchor states what it requires of studies

36-criterion audit

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

ID	Criterion	Status	Evidence
Construct Validity
C1	Falsifiability	C	The prediction can fail and did (Giulianelli et al., 2018; Elazar et al., 2021)
C2	Structural plausibility	PC	Every symbol a probing result depends on is named and typed
C3	Convergent validity	PC	Probe families agree on accuracy, reverse under selectivity (Hewitt & Liang, 2019)
C4	Discriminant validity	D	Two neighbors the instrument must separate: probe memorization, random features
C5	Nomological validity	PC	Tied to mutual information, description length and information gain
C6	Complementation validity	N/A	The audited object is a method, not a construct with named parts
Measurement Validity
M1	Reliability	U	A review reports no variance; the anchor states what it requires of studies
M2	Baseline separation	C	The criterion the literature solved; three control families formalized
M3	Stability	I	Ordering holds under accuracy and inverts under selectivity (Hewitt & Liang, 2019)
M4	Calibration	D	The shortcomings section opens by denying the measurement has a scale
M5	Sensitivity	PC	Skylines and floors exist; no planted property is ever recovered
M6	Invariance	PC	Invariance is what probing measures, so the instrument is the question
M7	Selection correction	U	Three control families catalogued; the omission from the catalogue is the gap
Internal Validity
I1	Necessity	I	Standard probing performs no removal; removal studies disagree four ways
I2	Sufficiency	N/A	Probing is a map out of the representation, with no path writing back in
I3	Minimality	U	The unit is a whole intermediate output; no operation removes part of one
I4	Specificity	D	A control dataset holds the property constant and probing fails on it
I5	Rival mechanism exclusion	PC	Control tasks, solvable only by memorization, separate cleanly (Hewitt & Liang, 2019)
I6	Double dissociation	U	Needs two properties and two behaviors; the framework supplies one of each
I7	Confound control	PC	Probe capacity, information in a random baseline and task difficulty
I8	Confounding sensitivity	U	The anchor answers confounding with designed controls, never with a bound
I9	Epistatic interaction	U	Non-additivity needs two manipulable components; the object is one output
I10	Rescue reversibility	U	Every catalogued intervention runs one way, projecting or training out
I11	Onset coupling	U	The original model is taken as given and trained once
I12	Offset coupling	U	The converse needs the training sequence onset coupling also lacks
External Validity
E1	Intervention reach	I	Probing has no intervention of its own; those applied to it disagree
E2	Prompt generalization	PC	One line of work reports the same property across several datasets
E3	Cross-task generalization	PC	Transfers to any property with an annotated dataset, which is the appeal
E4	Cross-model recurrence	C	Portability is established without qualification, from static embeddings on
E5	Graded response	N/A	No intervention, so no strength to dial and no response to grade
E6	Novel prediction	PC	Four design changes followed probing results and then worked
Interpretive Validity
V1	Level declaration	PC	The formalism fixes which objects a probing result is about
V2	Level-evidence match	D	The evidence supports extractability; the claim made is representation
V3	Alternative level	C	Every rival reading of a good score is given a measure of its own
V4	Unlicensed labeling	PC	“The model knows z” projects an epistemic relation onto decodability
V5	Scope declaration	C	A scope declaration end to end, conceding limits before any content
Total: 5 Confirmed, 12 Partially confirmed, 3 Inconclusive, 9 Untested, 4 Disconfirmed, 3 Not applicable
Verdict: Proposed

Sensitivity

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.

Criterion	In tension	What would settle it
I6	Untested or Not applicable	Whether a design supplying one mechanism and one behavior leaves double dissociation unattempted or unaskable

Exploratory Lens Analysis

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