acceptodds
Under review as a conference paper at ICLR 2027

The Floor That Barely Falls: Untrained Encoders as the Operative Null for Concept-Discovery Metrics

16 views

Abstract

Concept-discovery methods are scored with NMI, MIG, DCI and relatives, and often against an implicit floor of zero: of sampled papers reporting one of these metrics, none reports a chance level. That a floor exists is known: Carbonneau et al. (2022) checked, and found most of these metrics do score near zero on a random representation — all but DCI under a lasso regressor. We measure a strictly larger floor, for a different reason. Their random baseline is noise independent of the factors; ours is an untrained encoder applied to the real inputs, which carry information about the factors that generated them, and that difference is the whole paper. On our cell the untrained-encoder floor is NMI at held-out samples, in the decomposition run of the same cell whose canonical estimate is . Adjusted mutual information (Vinh et al., 2010) removes of it. Under the equal-frequency binning these metrics use the correction is affine, so it cannot reorder methods — but it rescales every difference by about , which matters against a fixed equivalence margin. What it leaves, , is the operative quantity, and it barely shrinks: at samples, a change across a fivefold rise in evaluation data while the removable part falls . Matching inflates it only at small samples: in a Gaussian reference that reproduces this floor, 90% of it is population alignment at samples and 99% at , and cross-fitting the matching agrees. Its size is a property of the configuration, and not in the direction a reader would guess: coarsening the discretisation from bins to shrinks the total floor while raising this component from 30% of it to 85%. It has to be measured per cell rather than looked up. Doing so changes conclusions, including ours: on a tabular benchmark we build, one cell places no task-supervised method above its null at the sample size we ship it at, and the gain a task label plus its architecture buys over that architecture's own untrained state differs by an order of magnitude across architectures (a post-hoc largest-to-smallest ratio of whose interval is informative only at its lower end, ) — and the smallest gain clears its floor only narrowly (under Holm, not under plain Bonferroni). It also disqualifies an equivalence margin this literature would call tight: NMI is – the largest excess over chance among the methods it is used to separate, and – their between-method range. Read against measured chance, concept labels still beat every architecture we tested, in every cell that resolves.

Primary area:
datasets and benchmarks
Keywords:
disentanglement metrics, concept bottleneck models, chance correction, normalized mutual information, evaluation protocols, benchmarks, concept discovery, interpretability, negative results

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_DpzJOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper argues we should interpret concept discovery measurements compared to an untrained encoder, not on their own. The score of an untrained encoder can be measured by using randomly initialized parameters. The paper decomposes the score of the random encoder into estimator bias and concept information. The quantity of interest is concept information, and the estimator bias can be largely removed by AMI. In terms of the margin over an untrained encoder, level 1 methods that don't use concept labels show marginal gain. However, CBM, which uses concept labels, shows a larger gain. The paper proposes ConceptDiscoveryBench, which is a tabular benchmark with the score of an untrained encoder.

Critical strengths
  • The argument of comparing the score with the untrained encoder is valid. The decomposition of the current measurement into components supports this argument by showing that part of the untrained encoder's score cannot be removed by existing corrections such as AMI.
  • The paper provides ConceptDiscoveryBench, which contains the score of the untrained encoder that the paper recommends. This enables future work to compare its scores to the untrained encoder without evaluating it.
Critical weaknesses
  • The paper lacks novelty. The main argument of comparing the performance against an untrained encoder is valid. However, it provides limited new insights beyond previous work. As the paper already mentioned, previous work also compared to the untrained encoder, which takes away from the paper's contribution.
  • The paper is really hard to read, so it should be reorganized to be published. The main content of the paper contains too much technical detail, including the file names, scores, and settings, which can be sufficiently explained in the appendices. The main content repeatedly explains what the paper changed from the previous version.
Questions for authors

Please see the weaknesses.

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

In concept discovery, the authors argue that, too often, papers use an implicit zero baseline. Instead, they argue that a randomly initialized model, using the same architecture and inputs as the tested case, should be used. Although they acknowledge that this suggestion is not new, they propose a to use AMI rather than NMI in order to compute the correlation between discovered concepts and annotated ones, and argue that the difference between NMI and AMI captures bias that is sometimes reported as a real signal if the baseline is not correctly handled. This evaluation is then applied to several families of methods, with an increasing level of supervision (unsupervised, class-supervised, concept-supervised).

Critical strengths

The general aim of the paper (proposing a fair evaluation of concept discovery) corresponds to a real need of the concept discovery community. Using AMI rather than NMI and subtracting the score of a random model sounds like good advice.

Critical weaknesses
  • The paper seems to be mostly LLM-written, and I struggled understanding what the message of the paper is. Results are included starting in page 2, the contribution are many and fuzzy, the wording is too LLM-like for me to be able to focus on reading it. There’s no conclusion section and the appendix goes up to number Z, including what seems to be clarification about previous versions of the manuscript. A heavy LLM usage is acknowledged by the authors.
  • After reading it, I’m not 100% sure what the contribution is, whether the authors manage to answer some question or propose a general evaluation framework. I agree with the main message, as stated in the strengths, but this message is hard to distil from the paper and be relatively trivial, even if totally reasonable.
Questions for authors

Not really a question, but I would appreciate if the authors managed to distil a human-crafted manuscript out of the results of this, apparently, agentic research experiment. Maybe a worthy contribution is there, but I couldn't disentangle it from the heavy LLMish language.

Reviewer_R3QoOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper studies how concept-discovery methods should be scored when using alignment metrics such as NMI, MIG, and DCI, proposing the use of an untrained encoder applied to the inputs under the same evaluation setup as a proper default baseline. The authors contrast this with treating a score of 0 as the baseline, which they claim is the implicit default in prior work, and report that the choice of baseline significantly changes the interpretation of results. They also introduce ConceptDiscoveryBench, a synthetic tabular benchmark that reports the untrained encoder floor for each setting, and demonstrate that many task-supervised methods perform just above the untrained encoder. Experiments find that an equivalence margin of δ\delta = 0.05 NMI is too loose to distinguish between the methods that they compared, and that concept supervision (CBM) outperforms every task-label-only architecture that they test.

Critical strengths
  1. The proposed baseline is simple, cheap, and broadly applicable within concept discovery.

  2. The decomposition of the untrained-encoder floor is insightful, separating the estimator bias that AMI removes from a remainder that does not shrink with more test data.

Critical weaknesses
  1. The writing is unclear, making the contribution difficult to understand despite its simplicity. For example, the main text never defines the benchmark task: how inputs, concepts, and targets are generated, and what methods are trained to predict. The writing seems to assume familiarity with previous versions of the paper or project, referring to "ConceptNAM, the architecture we previously proposed," and "the corrected fixed-W protocol," without prior introductions in the main text.

  2. There is limited demonstration of the impact of the paper's contributions. The changed conclusions as a result of changing the comparison baseline only comes from the authors' own trained models, mostly on their own synthetic benchmark. The one test on externally trained models (Appendix M.2) does not result in any reordering.

Questions for authors
  1. Can the authors identify a published comparison between methods with different encoder architectures where applying the architecture-matched floor changes a ranking or removes a claimed improvement?

  2. When comparing methods that share an architecture, does the correction affect anything beyond absolute scores and ratios?

  3. Can the authors more clearly define the benchmark task in the main text, and revise the text so that it does not rely on familiarity with earlier versions of the work?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.