This repository collects all relevant resources about interpretability in LLMs

MICCAI 2022 (Oral): Interpretable Graph Neural Networks for Connectome-Based Brain Disorder Analysis

[KDD'22] Source codes of "Graph Rationalization with Environment-based Augmentations"

(ICML 2023) Discover and Cure: Concept-aware Mitigation of Spurious Correlation

Official code for the CVPR 2022 (oral) paper "OrphicX: A Causality-Inspired Latent Variable Model for Interpreting Graph Neural Networks."

[ICCV 2023] Learning Support and Trivial Prototypes for Interpretable Image Classification

Codebase the paper "The Remarkable Robustness of LLMs: Stages of Inference?"

[TPAMI 2025] Mixture of Gaussian-distributed Prototypes with Generative Modelling for Interpretable and Trustworthy Image Recognition

[CVPR 2025] Concept Bottleneck Autoencoder (CB-AE) -- efficiently transform any pretrained (black-box) image generative model into an interpretable generative concept bottleneck model (CBM) with minimal concept supervision, while preserving image quality

TraceFL is a novel mechanism for Federated Learning that achieves interpretability by tracking neuron provenance. It identifies clients responsible for global model predictions, achieving 99% accuracy across diverse datasets (e.g., medical imaging) and neural networks (e.g., GPT).

hopwise: A Python Library for Explainable Recommendation based on Path Reasoning over Knowledge Graphs, ACM CIKM '25

Explainable AI: From Simple Rules to Complex Generative Models

Explainable Boosting Machines

Semi-supervised Concept Bottleneck Models (SSCBM)

This repository contains the official code of the paper: "Comprehensive Information Bottleneck for Unveiling Universal Attribution to Interpret Vision Transformers", which is published in CVPR 2025.

Explainable Speaker Recognition

Layer-wise Semantic Dynamics (LSD) is a model-agnostic framework for hallucination detection in Large Language Models (LLMs). It analyzes the geometric evolution of hidden-state semantics across transformer layers, using contrastive alignment between model activations and ground-truth embeddings to detect factual drift and semantic inconsistency.

Build a Neural net from scratch without keras or pytorch just by using numpy for calculus, pandas for data loading.

Visualization methods to interpret CNNs and Vision Transformers, trained in a supervised or self-supervised way. The methods are based on CAM or on the attention mechanism of Transformers. The results are evaluated qualitatively and quantitatively.

Implementation of the gradient-based t-SNE sttribution method described in our GLBIO oral presentation: 'Towards Computing Attributions for Dimensionality Reduction Techniques'