A complete end-to-end pipeline for LLM interpretability with sparse autoencoders (SAEs) using Llama 3.2, written in pure PyTorch and fully reproducible.

Real-time 3D visualisation of SAE feature activations inside GPT-2, token by token

Fast XAI with interactions at large scale. SPEX can help you understand the output of your LLM, even if you have a long context!

A toolbox for exploring the informational nature of LLMs and language

[Under Review] Not All Tokens Are Equally Useful for Steering: Robust Directions and Prefix Steering

Automates attribution-graph analysis via probe prompting: circuit-trace a prompt, auto-generate concept probes, profile feature activations, cluster supernodes.

Spectral Attention Divergence — dynamical systems probe for LLM inference via dual-path attention comparison and delay-coordinate attractor reconstruction

[ICLR 2026] AbsTopK: Rethinking Sparse Autoencoders For Bidirectional Features

A minimal mech-interp project for steering LLMs

Replication of 'From Reasoning to Answer' (EMNLP 2025) — Reasoning-Focus Heads + Activation Patching on DeepSeek-R1-Distill-Qwen-7B

Personal website of Xudong Zhu, featuring research, publications, open-source projects, and technical writing on LLM and representation learning.

Personal profile of Xudong Zhu. PhD researcher in representation learning.

This project explores methods to detect and mitigate jailbreak behaviors in Large Language Models (LLMs). By analyzing activation patterns—particularly in deeper layers—we identify distinct differences between compliant and non-compliant responses to uncover a jailbreak "direction." Using this insight, we develop intervention strategies that modify

Normative, machine-first definition of Interpretive SEO (SEO interprétatif), aligned with the Interpretive Governance standard.

Layerwise diagnostic showing instruction-tuned LLMs settle on next-token predictions later.

Fine-tuned DistilBERT for binary sentiment analysis on IMDB movie reviews with token-level interpretability using LayerIntegratedGradients

Studying LLM latent space topology using persistent homology and Hodge decomposition. Discovers that Hodge gradient potential recovers the animacy hierarchy — a linguistic universal — without supervision.

Multi-scale Haar wavelet probes of LLM hidden states for FX direction prediction. Walk-forward Sharpe=1.22 net OOS on GBP/USD 2020-2026.

Cross-patching diagnostic showing instruction-tuned late effects depend on upstream state.

Mechanistic interpretability of chain-of-thought reasoning in LLMs — 12 experiments on Gemma-2-2B-IT and Gemma-3-27B (GSM8K). MSc Research Capstone, IE Madrid 2026.