Research map / Image generation / Autoregressive and latent generation
PRIME: Prioritizing Interpretability in Failure Mode Extraction
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes PRIME, which extracts failure modes of image classifiers by exhaustively searching combinations of RAM-generated tags for minimal, significantly lower-accuracy subpopulations, yielding more similar, coherent, and specific text descriptions than DOMINO.
Problem
Failure modes of image classifiers found via latent-space clustering often have descriptions that do not match the images, because latent proximity does not reliably reflect shared semantic attributes.
Approach
Reverse the usual pipeline: tag images with a tagging model first, then search over combinations of human-readable tags for minimal tag sets whose presence yields a significant accuracy drop, so the tag set itself is the failure-mode description.
Key results
- On Living17, a failure mode with three tags (black ape hanging from tree branch) drops accuracy from 86.23% to 41.88%, while single tags keep accuracy above 69%.
- PRIME improves over DOMINO in average similarity, standard deviation, and AUROC of description-image similarity across datasets, using similar numbers of failure modes.
- Failure modes with 3 tags show average accuracy drops on unseen data of 26.82% on Living17 versus 8.18% for single tags.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Require all three tags hang, branch, and black together rather than using only one or two (ablation) | PRIME on Living17 ape images · Living17, ape (unseen images) | accuracy: 70.09\%, 69.23\%, 73.21\% -> 41.18\% | Hurt |
| Use three tags rather than groups matching at least two or at least one tag (ablation) | PRIME · Entity13, unseen images | accuracy drop: $25.29\%$ -> $34.75\%; $14.86\%$ -> $34.75\% | Helped |
| Use three tags rather than groups matching at least two or at least one tag (ablation) | PRIME · Living17, unseen images | accuracy drop: $17.13\%$ -> $26.82\%; $8.18\%$ -> $26.82\% | Helped |
| Use three tags rather than groups matching at least two or at least one tag (ablation) | PRIME · Waterbirds, unseen images | accuracy drop: $14.43\%$ -> $23.35\%; $7.19\%$ -> $23.35\% | Helped |
| Use three tags rather than groups matching at least two or at least one tag (ablation) | PRIME · CelebA, unseen images | accuracy drop: $16.84\%$ -> $23.25\%; $9.02\%$ -> $23.25\% | Helped |
| Increase the minimum group size s (ablation) | PRIME · Living17 and CelebA, unseen data | generalization | Helped |
| Increase the minimum accuracy drop a (ablation) | PRIME · Failure mode detection | difficulty of detected subpopulations | Helped |
| Use PRIME descriptions instead of DOMINO descriptions | Failure mode description extraction · Multiple datasets | description quality | Helped |
Surprising findings
- Points sharing many semantic tags in CelebA can still be far apart in CLIP latent space; with at least 5 shared tags, the distance exceeds that of random pairs with probability 0.34.
- Average number of shared tags in semantic space is much larger than in CLIP representation space among nearest neighbors (e.g., 7.55 vs. 4.17 on CelebA).
Limitations
- Depends on the accuracy of the auxiliary tagging model RAM; tags may be misaligned with the main object (e.g., 'white').
- Exhaustive search complexity grows exponentially with the number of tags considered; greedy search misses some failure modes.
- Descriptions are bag-of-tags rather than fluent sentences, since language models may associate tags with class labels incorrectly.
Experimental setup
Datasets: ImageNet Living17 Entity13 NonLiving26 Waterbirds CelebA CUB-200
Benchmarks: Living17 Entity13 Waterbirds CelebA NonLiving26
Metrics: average similarity standard deviation of similarity AUROC accuracy drop on unseen data accuracy on generated images
Baselines: DOMINO Distilling Failure Directions (Jain et al.) Bias2Text
Builds on: DOMINO Distilling Failure Directions Recognize Anything Model
Closest papers by meaning
- Interpretable Failure Detection with Human-Level Concepts (AAAI 2025 · similarity 0.88)
- A Multimodal Automated Interpretability Agent (ICML 2024 · similarity 0.83)
- Enhancing Performance of Explainable AI Models with Constrained Concept Refinement (ICML 2025 · similarity 0.83)
- Failures Are Fated, But Can Be Faded: Characterizing and Mitigating Unwanted Behaviors in Large-Scale Vision and Language Models (ICML 2024 · similarity 0.83)
- FineGRAIN: Evaluating Failure Modes of Text-to-Image Models with Vision Language Model Judges (NeurIPS 2025 · similarity 0.83)
- Towards Fine-Grained Interpretability: Counterfactual Explanations for Misclassification with Saliency Partition (CVPR 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.