Research map / Vision-language and multimodal / Human and embodied modeling
Explanation Bottleneck Models
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 explanation bottleneck models (XBMs), which generate free-form text explanations from images with a pre-trained BLIP/LLaVA encoder-decoder and classify from those explanations, using explanation distillation from a frozen decoder to prevent collapse.
Problem
Concept bottleneck models are restricted to a fixed, small pre-defined concept set, limiting the interpretability of their explanations; the goal is an interpretable model that explains predictions in natural language without pre-defined concepts.
Approach
Replace pre-defined concepts with a text explanation decoded from a pre-trained vision-language encoder-decoder, train end-to-end with the classification loss via Gumbel-softmax, and regularize the decoder with sequence-level knowledge distillation from a frozen copy of the pre-trained decoder (explanation distillation) to avoid collapse of the explanations.
Key results
- XBM-BLIP reaches 74.09% test accuracy on Aircraft and 89.47% on Car, versus 44.47% and 77.91% for Label-free CBM (GPT-3) and 77.86% and 90.08% for fine-tuned BLIP-ViT.
- Without explanation distillation, XBM collapses to degenerate text (CLIP-Score 0.4730 vs 0.7151 on Aircraft; GPT-2 perplexity 322.6 vs 129.8).
- XBM-LLaVA-v1.6-Vicuna-7B reaches 74.42% multi-modal test accuracy on ImageNet, up from 56.04% for frozen BLIP with a multi-modal classifier.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add explanation distillation (ablation) | XBM-BLIP · Aircraft | CLIP-Score: 0.4730 -> 0.7151 | Helped |
| add explanation distillation (ablation) | XBM-BLIP · Bird | GPT-2 Perplexity: 431.0 -> 166.8 | Helped |
| add explanation distillation (ablation) | XBM-BLIP · Car | Test accuracy: 86.59 -> 89.47 | Helped |
| add explanation distillation (ablation) | XBM-BLIP · ImageNet | Test accuracy: 66.58 -> 67.83 | Helped |
| fine-tune the explanation bottleneck rather than use frozen BLIP captions | XBM-BLIP · Aircraft | Test accuracy: 45.23$\pm$.32 -> 74.09$\pm$.07 | Helped |
| fine-tune the explanation bottleneck rather than use frozen BLIP captions | XBM-BLIP · Bird | CLIP-Score: 0.7535 -> 0.7942 | Helped |
| fine-tune the explanation bottleneck rather than use frozen BLIP captions | XBM-BLIP · Car | GPT-2 Perplexity: 168.8 -> 131.8 | Helped |
| use XBM-BLIP instead of the strongest reported black-box baseline | XBM-BLIP · Aircraft | Test accuracy: 77.86$\pm$.30 -> 74.09$\pm$.07 | Hurt |
| use XBM-BLIP instead of the strongest reported black-box baseline | XBM-BLIP · Bird | Test accuracy: 83.48$\pm$.15 -> 80.99$\pm$.18 | Hurt |
| use XBM-BLIP instead of the strongest reported black-box baseline | XBM-BLIP · Car | Test accuracy: 90.08$\pm$.35 -> 89.47$\pm$.10 | Hurt |
| use XBM-BLIP instead of the strongest reported black-box baseline | XBM-BLIP · ImageNet | Test accuracy: 65.21$\pm$.14 -> 67.83$\pm$.33 | Helped |
| apply exponential temperature annealing (initial temperature 10) (ablation) | XBM-BLIP · Car | Test accuracy: 86.71 -> 88.65 | Helped |
Surprising findings
- Without explanation distillation the text decoder collapses into degenerate repetitive output (e.g., repeated 'when' tokens), even though classification accuracy stays high.
- Randomly replacing explanations drops XBM-BLIP accuracy to 44.42% on Bird, while ground-truth explanation intervention raises it to 82.21%.
Limitations
- XBMs cannot generate explanations based on user-defined or expert-specified concepts, which CBMs can.
- Explanations inherit biases from the backbone language model.
- Evaluation of explanation quality relies on CLIP-Score and GPT-2 perplexity rather than human judgments.
Experimental setup
Datasets: FGVC Aircraft CUB-200-2011 Stanford Cars ImageNet ImageNet Segmentation
Benchmarks: FGVC Aircraft CUB-200-2011 Stanford Cars ImageNet ImageNet Segmentation
Metrics: test accuracy CLIP-Score GPT-2 perplexity pixel accuracy mIoU mAP
Baselines: Fine-tuned BLIP-ViT Label-free CBM Frozen BLIP XBM w/o R_int L2SP
Builds on: concept bottleneck models BLIP label-free CBM sequence-level knowledge distillation LoRA
Closest papers by meaning
- Show and Tell: Visually Explainable Deep Neural Nets via Spatially-Aware Concept Bottleneck Models (CVPR 2025 · similarity 0.87)
- Relational Concept Bottleneck Models (NeurIPS 2024 · similarity 0.87)
- Concept Bottleneck Large Language Models (ICLR 2025 · similarity 0.87)
- Predictive Concept Decoders: Training Scalable End-to-End Interpretability Assistants (NeurIPS 2026 · similarity 0.87)
- Interpretable Concept-Based Memory Reasoning (NeurIPS 2024 · similarity 0.86)
- Hyperbolic Concept Bottleneck Models (NeurIPS 2026 · similarity 0.85)
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.