Research map / Large language models / Language, safety and interpretability
Making Sense of LLM Decisions: A Prototype-based Framework for Explainable Classification
Bowen Wei, Mehrdad Fazli, Ziwei Zhu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add token attribution scores from target LLM (ablation) | ProtoSurE · Hotel/DBPedia/Consumer (Llama-3.1-8B) | Accuracy: 91.1 -> 91.9 | Improved |
| add token attribution scores from target LLM (ablation) | ProtoSurE · Qwen2.5-7B average | Accuracy: 92.7 -> 93.4 | Improved |
| make prototypes trainable after K-means init (ablation) | ProtoSurE · Llama-3.1-8B avg | Accuracy: 89.7 -> 90.9 | Improved |
| make prototypes trainable after K-means init (ablation) | ProtoSurE · Qwen2.5-7B avg | Accuracy: 92.6 -> 93.6 | Improved |
| swap sentence encoder (GTE vs SBERT/T5) (ablation) | ProtoSurE · Overall avg accuracy | Accuracy: 0.8952 vs 0.8881 | No clear effect |
| increase number of prototypes K (ablation) | ProtoSurE · DBPedia, Consumer (Llama-3.1-8B) | faithfulness metrics | Improved |
| increase training examples beyond 128 (ablation) | ProtoSurE · Hotel (Llama-3.1-8B) | Accuracy: 96.5 -> 98.4 | No clear effect |
| use ProtoSurE instead of SHAP/IG/Occlusion/DeepLIFT | ProtoSurE vs post-hoc baselines · 3 LLMs x 4 datasets | Comprehensiveness: .138 (IG) -> .192 | Improved |
| use ProtoSurE instead of baselines | ProtoSurE vs post-hoc baselines · 3 LLMs x 4 datasets | Deletion rank correlation: .092 (Occlusion) -> .100 | Mixed |
| use ProtoSurE instead of SELF-EXP/SHAP/Occlusion | ProtoSurE vs baselines · Hotel (GPT-4o-mini) | Sufficiency: 0.026 (SHAP) -> 0.001 | Improved |
Closest papers by meaning
- Sparsity-Guided Holistic Explanation for LLMs with Interpretable Inference-Time Intervention (AAAI 2024 · similarity 0.87)
- Faithful Explanations of Black-box NLP Models Using LLM-generated Counterfactuals (ICLR 2024 · similarity 0.87)
- Automated Assessment of Fidelity and Interpretability: An Evaluation Framework for Large Language Models’ Explanations (Student Abstract) (AAAI 2024 · similarity 0.87)
- Potemkin Understanding in Large Language Models (ICML 2025 · similarity 0.87)
- Walk the Talk? Measuring the Faithfulness of Large Language Model Explanations (ICLR 2025 · similarity 0.86)
- Large Language Models are Interpretable Learners (ICLR 2025 · similarity 0.86)
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.
