Research map / Large language models / Language, safety and interpretability
Concept Bottleneck Large Language 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.
Introduces Concept Bottleneck LLMs (CB-LLMs) that add a concept bottleneck layer to pretrained LLMs, using automatic LLM-free concept labeling for classification and an adversarially disentangled hybrid CBM for interpretable text generation.
Problem
Black-box LLMs are opaque, and existing interpretable language models (concept bottleneck models for text) are limited to small text classification datasets and cannot handle generation or require costly GPT-4/ChatGPT concept labeling.
Approach
Transform any pretrained LLM into an inherently interpretable model by inserting a ReLU concept bottleneck layer trained against automatically scored concept labels (sentence-embedding similarity plus concept correction using class mapping); for generation, an unsupervised side layer is kept disentangled from concepts via an adversarial entropy-based loss so that concept neurons steer generation.
Key results
- CB-LLM classification with ACC reaches 0.9407 (SST2), 0.9806 (YelpP), 0.9453 (AGnews), 0.9928 (DBpedia), matching or exceeding the fine-tuned RoBERTa-base black-box (e.g., 0.9778 on YelpP).
- Human activation-faithfulness rating averages 4.12 for CB-LLM w/ ACC vs 2.75 for TBM & C3M on a 1-5 scale.
- CB-LLM generation (Llama3-8B) reaches steerability 0.95 on YelpP and 0.82 on SST2 versus 0.21-0.69 without adversarial training, with accuracy close to the black-box Llama3-8B.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Automatic Concept Correction (ACC) (ablation) | CB-LLMs classification with RoBERTa-base · SST2, YelpP, AGnews, DBpedia | accuracy: 0.9012 -> 0.9407; 0.9312 -> 0.9806; 0.9009 -> 0.9453; 0.9831 -> 0.9928 | Helped |
| add adversarial training to disentangle the unsupervised layer from concept information (ablation) | CB-LLMs generation · SST2, YelpP, AGnews, DBpedia | steerability: 0.57 -> 0.82; 0.69 -> 0.95; 0.52 -> 0.85; 0.21 -> 0.76 | Helped |
| add adversarial training (ablation) | CB-LLMs generation · SST2 | perplexity: 59.19 -> 116.22 | Hurt |
| add adversarial training (ablation) | CB-LLMs generation · YelpP | perplexity: 12.39 -> 13.03 | Hurt |
| add adversarial training (ablation) | CB-LLMs generation · AGnews | perplexity: 17.93 -> 18.25 | Hurt |
| add adversarial training (ablation) | CB-LLMs generation · DBpedia | perplexity: 35.13 -> 37.59 | Helped |
| add adversarial training (ablation) | CB-LLMs generation · SST2, YelpP, AGnews, DBpedia | accuracy: 0.9676 -> 0.9638; 0.9830 -> 0.9855; 0.9418 -> 0.9439; 0.9934 -> 0.9924 | Mixed |
Surprising findings
- Without adversarial training, the hybrid CB-LLM generation model achieves steerability close to random generation on some datasets (e.g., 0.21 on DBpedia).
Limitations
- Generation perplexity on SST2 is notably higher (116.22) than the black-box model (84.70), and steerability is not perfect.
- Concept quality depends on the sentence embedding model and the ChatGPT-generated concept sets, so concept scores can be inaccurate without correction.
- Generation experiments use concept labels from dataset labels and reduced 100k-sample subsets.
Experimental setup
Datasets: SST2 Yelp Polarity AGnews DBpedia ToxicDPOqa toxic-chat
Benchmarks: SST2 Yelp Polarity AGnews DBpedia
Metrics: test accuracy concept accuracy steerability score perplexity human faithfulness rating toxicity detection accuracy
Baselines: TBM & C3M (Llama3-8B labeling) Fine-tuned RoBERTa-base (black-box) Fine-tuned GPT2 (black-box) Fine-tuned Llama3-8B (black-box) CB-LLM w/o ADV training
Builds on: Concept Bottleneck Models Text Bottleneck Models C3M Label-free CBM hybrid CBM
Closest papers by meaning
- Bayesian Concept Bottleneck Models with LLM Priors (NeurIPS 2025 · similarity 0.89)
- Sparsity-Guided Holistic Explanation for LLMs with Interpretable Inference-Time Intervention (AAAI 2024 · similarity 0.89)
- Extracting Interpretable Task-Specific Circuits from Large Language Models for Faster Inference (AAAI 2025 · similarity 0.88)
- Hybrid Concept Bottleneck Models (CVPR 2025 · similarity 0.88)
- Large Language Models are Interpretable Learners (ICLR 2025 · similarity 0.88)
- VLG-CBM: Training Concept Bottleneck Models with Vision-Language Guidance (NeurIPS 2024 · similarity 0.88)
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.