acceptodds
Under review as a conference paper at ICLR 2027

CLEAR: Calibrated Linear Erasure Auditing of Representations

Abstract

What does a performance drop after representation erasure reveal about a model's use of the targeted information? Erasure aims to remove target information from internal activations, but can also disrupt other task-relevant features. The resulting performance drop can therefore reflect both effects. Interpreting this drop requires checking how much target information remains, how sensitively residual tests can detect it, and how intervention choices affect behaviour. We introduce CLEAR, a framework for calibrated linear erasure auditing of representations. CLEAR interprets behavioural change together with the remaining linear target association, the sensitivity of the residual test, and the response to different intervention strengths and replacement values. Calibration makes the resolution of each residual verdict explicit, connecting the validity of an internal edit to the behavioural evidence it provides. Experiments demonstrate how this joint assessment reveals ambiguities hidden by effect magnitude alone. Synthetic controls show that error from fitting the eraser can produce behavioural effects even at zero target dependence. In the SSR planner, full ego state erasure raises 3 s average displacement error to 7.3 times its baseline value while failing residual checks at every audited layer. Error also grows beyond the intervention strength that minimizes target association and varies with replacement values. CLEAR makes explicit how residual information, detection sensitivity, and intervention design constrain the interpretation of behavioural changes caused by erasure changes.

Primary area:
interpretability and explainable AI
Keywords:
Representation Erasure, Mechanistic Interpretability, Representation Auditing, Causal Intervention, Autonomous Driving

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_F1hBOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces CLEAR, a framework for auditing behavioural changes induced by linear representation erasure. It combines held-out residual association tests, calibration through injected leakage, auxiliary readout checks, and sensitivity analyses over intervention strength and replacement values. Synthetic experiments demonstrate that eraser estimation error can produce behavioural changes even when the target is unused. In SSR, full ego-state erasure increases planning ADE to approximately 7.3 times the baseline while failing residual checks at every audited layer. A complementary LAW example illustrates that strong decodability need not imply contribution to the current output.

Critical strengths
  • The paper targets on an important and well-motivated problem and clearly distinguishes target decodability, intervention validity, and behavioural impact. This distinction is valuable for interpreting representation interventions.

  • Scene-disjoint data roles, held-out residual evaluation, and fixed-projection dose sweeps improve the interpretability of the measurements. The SSR results provide a concrete example where a large performance drop coexists with incomplete linear erasure.

  • The authors report high failure floors, limited detection power, failed preregistered predictions, and readout-control failures. The discussion appropriately acknowledges that preservation of non-target information is an additional requirement for attribution.

Critical weaknesses
  • Section 3.3 introduces matched semisynthetic activations and an exact-projection oracle, but the manuscript does not provide an explicit generator or sufficient details to reconstruct the calibration. It remains unclear which aspects of the real activation distribution and scene dependence are preserved, how the exact projection is obtained, and what is resampled or refitted across repetitions.
  • The paper states that behavioural attribution additionally requires preservation of other task-relevant information. However, G3 and G4 in Table 1 evaluate clean activations. They establish properties of the original representations and readouts, rather than preservation after erasure.
Questions for authors
  • Can authors provide the explicit semisynthetic generator, oracle construction, repetition protocol, and definition of the fidelity check? What does β quantify in the fitted and oracle pipelines?
  • Can existing synthetic experiments be extended to demonstrate cases where CLEAR correctly supports interpretation, and quantify its improvement over an uncalibrated audit at comparable abstention rates?
  • What evidence assesses preservation of non-target information after erasure? Also the precise L0–L3 intervention locations and whether the behavioural effects in Table 3 correspond to individual-layer or sequential interventions seems unclear.
Reviewer_Rhj5OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the reliability of behavioral conclusions drawn from linear representation erasure. The authors argue that a performance change after erasure is not, by itself, sufficient evidence that the erased variable causally contributed to the model output, because the intervention may leave residual target information, distort other task-relevant features, or induce effects that depend strongly on the intervention implementation. To address this issue, the paper proposes CLEAR, a calibrated auditing framework for linear representation erasure. CLEAR combines held-out residual linear-association tests with a matched null distribution, injected-leakage calibration, a fitted-pipeline failure floor p(0)p(0), and a target- and layer-specific detection limit β∗\beta^*. The authors further interpret behavioral effects alongside auxiliary readout checks, fixed-projection dose sweeps, and alternative replacement interventions. The authors evaluate the framework using synthetic controls, a LAW computational-path diagnostic, and an SSR autonomous-driving planner.

Critical strengths
  1. The paper addresses an important weakness in representation-erasure-based causal interpretation.
  2. The synthetic controls are informative and directly support the paper's motivating concern.
Critical weaknesses
  1. The incremental contribution over the closest prior work is not sufficiently isolated, and the paper provides limited evidence for reliable positive attribution. Prior work has already studied removal failure, completeness/selectivity, and intervention reliability, so the main novelty of CLEAR lies in calibrating the residual-validity judgment itself. However, the experiments do not directly compare CLEAR against conventional effect-only, completeness, or completeness/selectivity-style baselines in a setting with known ground truth. Moreover, the strongest SSR results mainly show when attribution should be withheld, since full ego state and velocity fail the residual criterion at every audited layer. It therefore remains unclear how much CLEAR reduces erroneous causal attribution relative to existing methods and under what realistic conditions it enables a trustworthy positive conclusion about target use.

  2. The residual test has a substantial failure floor, which limits the practical utility of the audit. The fitted-pipeline p(0)p(0) values are often high, indicating that the auditing procedure can reject an intervention even when it introduces no additional leakage. CLEAR usefully quantifies this unreliability, but does not directly improve the underlying measurement procedure. The paper should therefore distinguish more clearly between calibrating an imperfect test and making the test itself more reliable.

Questions for authors
  1. Could the authors further quantify the incremental benefit of CLEAR relative to existing intervention-validity approaches?
  2. Under what conditions should a CLEAR result be interpreted as a positive causal attribution rather than only as an absence of evidence against the intervention?
Reviewer_zCCjOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper asks what a drop in model performance after linear representation erasure actually tells us about whether the model uses the erased variable. It proposes CLEAR, an auditing framework rather than a new erasure method. CLEAR fits a LEACE eraser on held-out scenes and, on separate audit scenes, tests whether residual linear association with the target remains, using a scene-level permutation null. It then calibrates this residual test by injecting known amounts of leakage, reporting the failure rate at zero injected leakage and the leakage level detected with 80% power, and compares against an exact-projection oracle. Behavioural effects are classified with an equivalence-style rule, and sensitivity to the intervention is examined by scaling a fixed projection and changing the replacement values.

The framework is evaluated on (1) a synthetic system with known target dependence, where a poorly fitted eraser produces sizable behavioural effects even when the target is unused; (2) the LAW planner, where speed is highly decodable from a representation computed after the trajectory is produced, showing decodability without use; and (3) the SSR planner on nuScenes, where erasing the full ego state raises 3 s ADE to 7.3× its baseline but fails the residual check at every audited layer, and planning error keeps increasing beyond the erasure strength that minimises residual association. The authors conclude that effect magnitude, residual validity and intervention sensitivity capture distinct aspects of an erasure and must be interpreted together

Critical strengths
  1. The paper treats the residual check as a measurement with its own resolution: it calibrates the test with injected leakage and an exact-projection oracle, showing that a “pass” can reflect low sensitivity and that eraser fitting error matters.
  2. Simple, convincing controls (the zero-dependence synthetic case and the fixed-projection dose sweep) show that effect size alone is misleading. The use of disjoint data, preregistration, and reporting of all conditions, including failed predictions, is careful research practice.
Critical weaknesses
  1. The paper asks whether a model actually uses a target variable. However, both the eraser (LEACE) and the residual test only deal with linear information, while the rest of the model is nonlinear. Figure 7 shows that a small MLP can still recover the target from the erased activations (R² of 0.56–0.99). So even when an erasure passes every check, the model may still be able to access the target through nonlinear patterns. A passing audit therefore cannot show that the target has been removed from the model’s point of view. The authors should either explain why linear erasure is enough for the conclusions they want to draw, or discuss how their calibration approach could be extended to nonlinear erasers and nonlinear tests.
  2. The paper does not directly measure whether other task-relevant information survives the edit. The decodability check (G3) uses clean activations only. The random-subspace placebo and the geometry diagnostics (Figures 4, 6) give indirect signs, but they do not show which information is damaged or how much of the behavioural effect it explains. Since the paper itself states that preservation is required for attribution, a post-erasure readout of other task variables (navigation command, surrounding agents) would be a natural and inexpensive addition.
  3. The full CLEAR pipeline is applied to only one setting: a single SSR checkpoint from one training seed, evaluated with open-loop 3 s ADE on nuScenes, using LEACE as the eraser. LAW is used only for readouts, without erasure, and the eraser comparison covers a single layer. This setting is also unusual in ways that favour the paper’s conclusions: speed is strongly aligned with the first principal component (correlation −0.90 to −0.97), so linear erasure inevitably distorts the representation, and open-loop ADE on nuScenes is known to depend heavily on ego status (Li et al., 2024b). It is therefore hard to tell whether the findings reflect general properties of erasure-based attribution or specifics of this model and metric. Since the paper presents CLEAR as a general interpretability tool, at least one additional model (ideally a language model, where most erasure work is done), multiple training seeds, and a closed-loop or NAVSIM-style metric would substantially strengthen the claims.
  4. The methodological novelty is limited. CLEAR combines standard statistical tools (permutation nulls, power analysis via injected signal, and TOST-style equivalence testing) and applies them to linear erasure. The main empirical messages, that erasure can leave residual information, can damage other features, and that effect size alone is not interpretable, have already been raised by Kumar et al. (2022), Canby et al. (2025) and Dobrzeniecka et al. (2025). The genuinely new elements are the leakage-injection calibration, the oracle comparison and the scene-count scaling of the null, which are useful but incremental. Moreover, the paper stops at diagnosing ambiguity: it does not propose a corrected test or show any real case where CLEAR yields a reliable, resolved attribution. The paper would need to show more clearly what CLEAR enables that prior evaluations could not.
Questions for authors

See Weaknesses.

2 comments

Thank you for the detailed rebuttal and the additional experiments. However, my main concerns are not fully addressed. These are the linear-only scope of the audit, the specificity of the behavioural effects relative to non-target perturbations, and the limited contribution beyond diagnosing ambiguity. I therefore keep my original score.

AuthorsOpenReview ↗

Follow-up on audit scope, behavioural specificity, and practical contribution

Thank you for reading our rebuttal and additional experiments. We appreciate your clear summary of the remaining concerns. We would like to highlight how the revised manuscript bears on each point.

Scope of the audit. CLEAR uses a calibrated linear test to assess residual target information after erasure. Detectable linear residuals establish that target information remains, regardless of any additional nonlinear encoding. A pass is interpreted relative to the fitted failure floor and detection limit. Table R1.1 and Fig. 7 of the manuscript complement this audit by quantifying nonlinear target information that remains recoverable after erasure.

Specificity of the behavioural effects. Table 11 (Appendix E.3) compares target edits with random perturbations matched in rank and displacement on the fitting scenes. Full ego and velocity edits increase ADE more than the random controls, with paired differences of 1.400 m and 1.897 m. Acceleration and rotation edits, which pass the residual check at all four layers, produce smaller effects than their controls, with differences of -0.416 m and -0.487 m. Random controls cause greater activation displacement on the audit scenes (RMS ratio 1.05 to 1.44), yet smaller behavioural changes for full ego and velocity.

Revised Table 3 shows that a nuisance placebo leaves the plan unchanged at both dependency weights, while removing the target together with its context changes the plan in both cases. After full ego state erasure, navigation command accuracy remains 99.56 to 99.98%, and the refitted object readouts in Table 5 retain positive scores. Object information remains recoverable after refitting.

Practical contribution. We have also added Table 10 (Appendix E.2) to the revised PDF to test whether the audit helps identify behavioural changes that support attribution. Known computational dependence provides ground truth for evaluating these decisions. Effect magnitude alone gives 99.14% coverage with 35.21% error. At the calibrated 1% quota, its error is 50%, while the matched null rule with preservation yields zero observed errors among 64 decisions. At its overall coverage of 5.20%, the latter achieves 11.71% error. CLEAR therefore turns the diagnosis of ambiguity into an explicit decision procedure whose reliability and coverage can be measured.

We hope these results clarify what the audit yields on the cases it resolves, and we are happy to provide further clarifications should there be any remaining questions.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.