acceptodds
Withdrawn from ICLR 2027

Faithfulness Above One: How a Circuit Benchmark Rewards Leaving Out Suppression

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

MIB, the Mechanistic Interpretability Benchmark, ranks circuit-discovery methods by two areas of a faithfulness curve. On its grid, the higher-is-better CPR is a constant plus the overshoot of the curve above the full model minus its undershoot below it, and the lower-is-better CMD is their sum, so every published pair can be checked, and CPR rewards circuits for exceeding the model they are meant to explain. MIB computes CPR on circuits of the highest-scoring edges. For additive edge effects and exact scores this rule maximizes CPR, and its most faithful circuits contain no suppressive component. Among MIB's own entries, the leading CPR in 9 of its 11 model and task cells can only come from a curve above the model, and on GPT-2 small the CPR spread of five public submissions is overshoot. Removing the logit edges of the two negative name mover heads, part of the known IOI circuit, from the MIB example's magnitude-selected circuits makes them less complete and raises their CPR from 0.97 to 1.58, above the 1.24 of the circuits whose CPR MIB reports. Rescored under the ABC counterfactual, EAP, the MIB example's method, gives value-selected circuits that exceed the model at the sizes that carry most of the weight when evaluated under the answer flip that MIB runs and fall below it under ABC. The largest overshoot, MAttr's, is not explained by the two edges and arises in patched states far outside the model's range. We recommend computing both scores on one declared circuit family, with a per-example distance from the model as the headline behavioral score.

Primary area:
interpretability and explainable AI
Keywords:
mechanistic interpretability, circuit discovery, faithfulness, benchmark validity, construct validity, evaluation methodology, negative name movers, copy suppression

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_HLU3OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper audits how MIB evaluates discovered circuits. It decomposes CPR and CMD into overshoot and undershoot, examines published results, and re-evaluates public submissions. Its central experiment shows that removing two known suppressive connections from an IOI circuit increases CPR while making the circuit less complete. Additional experiments examine the effects of counterfactual choice and distinguish suppression removal from more extreme overshoot.

Critical strengths

The paper connects a concrete evaluation problem to a direct intervention: removing known components of the IOI mechanism improves the benchmark score. The deletion controls and restoration experiments make this more convincing than an observation that faithfulness sometimes exceeds one.

The analysis also leads to useful recommendations about evaluating the same circuit family and distinguishing behavioral agreement from task performance. The authors acknowledge prior observations and report limitations and unsuccessful predictions.

Critical weaknesses
  1. The presentation could be clearer. The abstract includes many results and numerical details before the scoring setup is established. A simpler introduction to the selection rules and metrics would help. Figure 3b also feels cramped: the labels and legend take up substantial space relative to the plotting area, and several markers overlap. The layout should be revised, and the caption should explain the horizontal bars.

  2. The experimental coverage limits the broader conclusions. The detailed intervention study centers on GPT-2 Small and IOI. The Qwen experiments provide useful extensions but use only 32 IOI and 50 MCQA examples, and only EAP is rescored under both counterfactuals. The published-table audit establishes that overshoot is widespread, but it does not establish that the same mechanism explains it across those settings.

  3. The novelty is mainly in the empirical audit. As Section 1 acknowledges, prior work had already identified over-optimization of task metrics and the omission of negatively contributing components. The score decomposition is useful but mathematically straightforward. The contribution lies in quantifying the consequences for MIB and demonstrating them through interventions. The abstract and discussion could emphasize this more clearly and expand on what the findings imply for MIB’s two explicitly different evaluation goals.

Questions for authors
  1. Could the scoring/evaluation counterfactual comparison be repeated for another method? This would help determine whether the result that the evaluation counterfactual sets the side of one extends beyond EAP.

  2. How stable are the Qwen IOI and MCQA intervention effects with more evaluation examples? Additional examples and uncertainty estimates would strengthen the evidence beyond the main GPT-2 IOI setting.

  3. Section 9 notes that the recommended per-example distance does not reliably distinguish four of the submissions at the current sample size. Could you discuss whether larger samples are likely to resolve this, or whether these methods are simply behaviorally similar under magnitude selection? This matters for using the proposed distance as a headline ranking metric.

  4. Could you expand the discussion of CPR’s remaining role? Since MIB explicitly assigns it a task-performance objective, would retaining CPR under that interpretation, alongside a separate behavioral-faithfulness ranking, address the main concern?

  5. Figure presentation: In Figure 3b, the text and labels take up too much space relative to the plotting area, making the panel feel cramped. Could you adjust the font sizes, enlarge the plotting area, and tidy up the layout to make the comparisons easier to follow? Could you adjust the layout and font sizes, and check the other figures for similar readability issues?

Reviewer_aXjpOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The authors discuss widespread circuit faithfulness metrics proposed in MIB (Mueller et al., 2025), namely CPR and CMD. They make some recommendations on how to correctly report results when using the two metrics.

Critical strengths

None

Critical weaknesses

A first list of weaknesses concerns the presentation of the paper:

  • The paper reads like a very dense analysis of results. This makes the paper hard to follow and its actual contributions unclear. It lacks an abstract, critical discussion of results, making it of limited use for practitioners.
  • The paper does not convincingly describe its motivation, the gap it addressed, nor provides a compelling narrative for why the existing MIB metrics are insufficient.
  • The authors make some rather odd choices of exposition. To cite a few: Section 2 contains references to specific sections of prior literature, mentions of GitHub issues, or explicit mentions of the arXiv link of a cited work; the appendix on reproducibility states every single source codebase used in the project and the exact changes; Table 7 refers to certain values in an edge removal experiment as “doses”. The paper is laden with several other stylistic (and unscientific) oddities.
  • Captions do not provide any interpretation at all of the associated figure.

A second list concerns the actual content:

  • Two out of three alleged contributions are just restatements and trivial analyses of results from prior works. The “decomposition of both scores into overshoot and undershoot” reported as a key contribution in the paper is a simple function of MIB’s own metrics. The “model-free check” of existing results is again just a reformulation of prior results under the new metrics.
  • It’s unclear what the “causal test” of Section 7 adds to the discussion. Even if it did, its scope is limited to a single circuit and model. The authors acknowledge this yet provide no motivation.
  • The paper contains several proofs. While they may be correct, they add little over the already confusing exposition of this work.
  • While the description and analysis of results presented in the paper may be correct, they are of limited utility to a reader.

Contrary to the stated AI Use Statement, I believe very little of this paper was reviewed or produced by a human.

Questions for authors

None

Reviewer_4o8TOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper analyzes the evaluation metrics (CPR and CMD) in MIB. The authors show that MIB computes CPR and CMD on two different circuit families (value-selected vs. magnitude-selected) and prove that CPR=W+O−U\text{CPR} = W + O - U while CMD=O+U\text{CMD} = O + U. Consequently, CPR rewards circuits for overshooting the full model (f(k)>1f(k) > 1), which happens when suppressive components are excluded or when patching produces out-of-range states. They validate this across published MIB leaderboards and experiments on GPT-2 small and Qwen-2.5 0.5B, and recommend a per-example distance on a single circuit family.

Critical strengths

The paper is clearly written with a solid, easy-to-follow logical flow.

It accurately pinpoints the fundamental flaws in MIB's evaluation setup.

Critical weaknesses

While I really appreciate the authors calling out MIB's flaws, I don't think this alone is enough for a standalone conference paper. First, MIB is not yet a universally accepted benchmark, and most current circuit discovery works do not rely on it. Second, the underlying flaw in MIB is actually quite obvious: the nonlinear distortion of logit differences. Specifically, f(k)f(k) is linearly tied to logit differences, whereas actual circuit capability does not scale linearly with logits (e.g., doubling circuit faithfulness does not simply double the logit difference). Pointing out such an obvious metric formulation issue does not warrant a full paper on its own. Moreoevr, the narrow scope is also reflected in the proposed "per-example distance". It only fixes the minor issue of positive and negative causal effects canceling out across examples, while still suffering from the exact same nonlinear logit distortion. Moreover, as the authors admit, this metric lacks discriminative power to rank most methods under magnitude selection.

The empirical evaluation feels incomplete and statistically underpowered. Sample sizes are very small (e.g., 96 validation/100 test examples on GPT-2 small, and only 32 IOI/50 MCQA examples on Qwen-2.5 0.5B compared to 1,000 in MIB). The evaluated methods are also heavily homogenized—three of the five submissions (Warm-start EP, Submission A, and Submission B) come from the same team, with Submission B being a simple linear combination of the other two. Furthermore, several grid evaluations only cover 3 out of 10 points due to time constraints.

Questions for authors
  1. Can the authors scale up the experiments during the rebuttal (e.g., full 10-point grids, larger sample sizes, and more diverse, independent circuit discovery methods)?

  2. I don't expect the remaining fundamental issues to be resolved within a short rebuttal window, but I am listing them here in the hope that the authors will substantially expand the work for a future resubmission:

  1. Most circuit faithfulness evaluations today rely on metrics like KL divergence rather than just MIB's setup. A comprehensive analysis across all mainstream faithfulness metrics—rather than focusing solely on MIB—would make a much stronger paper.

  2. If proposing a new metric is a core contribution, it should move beyond relying on a linear relationship with logits to reflect faithfulness, while maintaining enough resolution to distinguish between different methods. I don't have a silver-bullet metric in mind either, but exploring and validating such a metric is what would make this a genuinely valuable contribution.

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

Related papers

Discussion (0)

Sign in to comment.