This repository curates seminal research papers focused on understanding the internal mechanisms of Large Language Models (LLMs). It targets researchers and practitioners in AI interpretability, providing a structured overview of key findings and methodologies for dissecting LLM behavior. The benefit is a centralized, high-quality resource for advancing LLM transparency and trustworthiness.

How It Works

The repository functions as a curated bibliography, categorizing papers by their core research themes such as neuron-level analysis, circuit discovery, knowledge editing, and in-context learning. Papers are selected based on acceptance at top-tier AI conferences (ICML, NeurIPS, ICLR, ACL, EMNLP, NAACL) or affiliation with leading research institutions, ensuring a focus on rigorous and impactful contributions to LLM interpretability.

Quick Start & Requirements

This is a curated list of research papers and does not involve code execution. No installation or specific requirements are needed beyond a web browser to access the linked papers.

Highlighted Details

• Comprehensive coverage of LLM interpretability topics including mechanistic interpretability, sparse autoencoders (SAE), logit lenses, and causal analysis.
• Papers are tagged with relevant keywords (e.g., neuron, circuit, knowledge, fine-tune, hallucination) for easy filtering.
• Includes links to influential blogs and other related repositories for further exploration.
• Features papers from major research labs like OpenAI, Deepmind, and Anthropic.

Maintenance & Community

The list is maintained by zepingyu0512 and appears to be actively updated with recent publications. Further community engagement or discussion channels are not explicitly mentioned.

Licensing & Compatibility

The repository itself is licensed under an unspecified license. The linked papers are subject to their respective publisher's copyright and licensing terms.

Limitations & Caveats

This resource is a bibliography and does not provide code implementations or direct tools for LLM interpretability. The focus is solely on academic literature.