Label only six of ten foot types, and full correction still leaves a CBM ~2 points behind a DNN

Measuring What Matters: Synthetic Benchmarks for Concept Bottleneck Models

Julian Skirzynski, Harry Cheon, Shreyas Kadekodi, Meredith Stewart, Berk Ustun

NeurIPS 2026 (Evaluations and D

cs.LG, cs.AI

2026-06-03

On synthetic robots, full intervention lifts a CBM 4.0 points past a DNN when concepts match the rule, and fails or hurts when a foot type is missing or concepts are auto-labeled.

What problem this solves

A concept bottleneck model (CBM) predicts concepts a person can check, then predicts the label from those concepts alone. In dermoscopy that means calling streaks before anyone trusts a melanoma probability. Annotation is the blocker. With m concepts, the labeling bill is about m times a normal dataset. Across 123 CBM papers since 2020, experiments keep returning to a few sets such as CUB-200, CelebA, and CheXpert. Those sets were not built for concept supervision. Noise and missing labels are hard to measure, and correcting a concept often has no decision at stake, as when a smile is used to predict attractiveness. A failed run does not say whether the architecture is weak or the concept set is.

The generators assign labels with a known concept-to-label rule. Modality, granularity, concept source, noise, and missingness can be changed one at a time. The benchmarks attribute failures. They do not rank models.

Method

Decision support separates Glorp robots from Drent robots. Nine binary body parts produce 7,680 bodies. A person can check the head or the knees and still not know the type. Feet have 10 subtypes. Labels come from a fixed score spread across several parts. A model that sees every part tops out near 99% accuracy. Training shows six foot types; the test set has all ten, and 40% of test feet never appeared in training. Images are 32 by 32.

Automation checks whether a 9 by 9 Sudoku board is valid. Twenty-seven concepts say whether each row, column, and block contains digits 1 through 9 once. The label is their logical AND, so one wrong concept flips a valid board. Valid boards come from a mixed-integer program; invalid boards are lightly corrupted copies. Renderings include printed givens, handwriting, and candidate marks. Half of 1,000 boards are valid. Handwriting is readable at 50 pixels per cell and hard to read at 10. The rule stays fixed.

Gain is accuracy after intervention minus an end-to-end DNN. Automation locks selective accuracy at 0.99 and abstains when unsure. WorkSaved is coverage minus the share of concepts a person checks. Interventions edit the concepts most likely to flip the prediction, with a budget of 1, 3, or all. The same protocol runs a standard CBM, CEM (concept embeddings), ProbCBM (Gaussian concept latents), ECBM (joint concept and label training), and the DNN, over 10 seeds.

Results

SettingComparisonResult
Robots, no interventionDNN88%
Oracle concepts, none to all correctedCBM84.5%±0.3 to 92.0%±0.5 (Gain 4.0±0.5 points)
Discovered concepts, all correctedCBM85.7%±0.4, about 2 points below the DNN
Sudoku, 50 px, no interventionCBM / DNNWorkSaved 79.8%±5.9 / 2.6%±0.8
Sudoku, 10 px, no interventionCBMWorkSaved 58.8%±3.8

The oracle set has 7 concepts. With no intervention, HasKnees is detected near chance and the standard CBM trails the DNN at 88%. After every concept is set to its true value, Gain is 4.0±0.5 points.

A missing foot type is not repaired by a different bottleneck. Discovered concepts keep the six foot subtypes seen in training, 12 concepts in all. The standard CBM is significantly worse at every budget, and 85.7%±0.4 after a perfect correction, still about 2 points behind. Accuracy is 77% on unseen feet and 92% on seen feet, against 91% and 93% with oracle concepts. CEM, ProbCBM, and ECBM also stay at or below the DNN under perfect interventions. A probe reads unseen foot shape from CEM embeddings at 99% accuracy; a probe trained only on seen feet manages 30% to 90%. ProbCBM replaces the image embedding with a learned present or absent vector, and those two vectors are nearly identical when the detector cannot see the concept. Removing the concept term from the ECBM loss adds 8 points of accuracy.

Automated concepts make intervention harmful. On CLIP labels, LLM-written concepts, and CLIP words, writing in the true values makes almost every model significantly worse. CEM on the machine-labeled set is the exception. Averaged over budgets of 1, 3, and all concepts, the worst mean change for the standard CBM is -26.6 points, from LLM concepts plus an LLM intervener. Perfect intervention on the machine-labeled set is -14.7 points. Intervening with the model's own concepts thresholded at 0.5 changes no value and still drops accuracy. The extra signal lives in the detector's continuous output, and an intervention strips it. On hand-specified concepts, true values beat that empty edit.

Matching concept accuracy does not match usefulness. A simulated expert and an LLM both get about 80% of concepts right, yet they recover 42% and 11% of the gain from perfect interventions. The LLM fails on HasKnees, the concept that most often needs a fix, where it is near chance. Swap only those answers and the gains trade places.

An alignment constraint can spend the value of later edits. The true weight on HasKnees is negative. Forcing it to be non-negative sets it to zero. Training accuracy moves from 93.0% to 91.5%, so an accuracy check would keep the constraint. Before any intervention the constrained model is more accurate, because it no longer trusts a near-random concept. Correcting every concept then adds only 1 to 2 points, against about 8 without the constraint. In Fig. 7 the oracle changes are +7.5%±0.5 versus +0.9%±0.5, and the discovered changes are +8.1%±0.8 versus +2.2%±0.5.

On Sudoku, how the rule is split matters more than which bottleneck is used. At 50 pixels, with selective accuracy fixed at 0.99 and no interventions, WorkSaved is 79.8%±5.9 for the CBM, 72.5%±9.5 for CEM, 79.5%±2.6 for ProbCBM, 53.1%±11.8 for ECBM, and 2.6%±0.8 for the DNN. A plausible reason is that concept models learn each constraint and the AND separately, while the DNN has to learn the whole validity rule and is rarely confident at 0.99. Checking up to three concepts on each abstained board raises WorkSaved, significantly for ProbCBM and ECBM. After every concept is checked, 38% to 49% of the boards the standard CBM, ProbCBM, and ECBM still abstain on stay abstained. Only CEM saves significantly more work at that point. At 10 pixels each detector loses only a few points, but boards with at least one wrong concept rise from about a quarter to more than half. WorkSaved falls to 58.8%±3.8, 36.4%±8.6, 39.6%±6.9, and 34.5%±11.1 for CBM, CEM, ProbCBM, and ECBM, with the DNN at 4.3%±1.3. No architecture gets all of that loss back by intervention.

Why it matters

Before paying for concept labels, three factors can be turned separately: whether the set misses something that flips the label, whether the person editing concepts fails on the ones that matter, and whether every condition must hold at once. Common bird-part and face-attribute sets tangle the three in one noisy annotation.

In decision support the CBM passes the DNN through intervention, and only when the concepts are the ones in the rule. Building concepts with CLIP or an LLM and then asking a person to edit them lowers accuracy in these setups. Clause-by-clause document checks look more like Sudoku. After the rule is split into concepts, the share of work cleared at a 0.99 selective-accuracy bar is far above the black box. A few points of detector error then stack across concepts and become a large block of abstentions.

Limitations

The paper says the tasks are simplified, so the reported accuracies and labor savings may overstate real data, and that only some mechanisms were tested. A high score is not a license to deploy.

The robots are designed 32 by 32 drawings, far from subjective, correlated concepts in skin or chest images. Sudoku's AND has no exceptions. The paper expects document checking to inherit the result because it shares that mechanism, and it does not measure how much of the effect survives the move. Standard errors on automation are large: ECBM's WorkSaved is 53.1%±11.8 with no intervention, so who saves more work is unstable. The main results charge the same cost for every concept. The drop after intervening on automated concepts also appeared on two real datasets, without numbers in the main text, so it should not be read straight onto CheXpert.

Terms

Source

What people are saying

Related papers

All paper explainers