[2405.17703] Mechanistic Interpretability of Binary and Ternary Transformers

arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website. Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them. Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs. arXiv Operational Status Get status notifications via email or slack

View PDF HTML (experimental) Abstract:Recent research (arXiv:2310.11453, arXiv:2402.17764) has proposed binary and ternary transformer networks as a way to significantly reduce memory and improve inference speed in Large Language Models (LLMs) while maintaining accuracy. In this work, we apply techniques from mechanistic interpretability to investigate whether such networks learn distinctly different or similar algorithms when compared to full-precision transformer networks. In particular, we reverse engineer the algorithms learned for the toy problem of modular addition where we find that…

saved by

• Barton Rhodes

related reading

• [2406.17583] Towards Compositional Interpretability for XAIarxiv.org
• A Mathematical Framework for Transformer Circuitstransformer-circuits.pub
• Towards Monosemanticity: Decomposing Language Models With Dictionary Learningtransformer-circuits.pub
• In-context Learning and Induction Headstransformer-circuits.pub
• Towards Monosemanticity: Decomposing Language Models With Dictionary Learningtransformer-circuits.pub
• Towards Automated Circuit Discovery for Mechanistic Interpretabilityarxiv.org
• Interpreting Language Model Parametersgoodfire.ai
• [2402.12875] Chain of Thought Empowers Transformers to Solve Inherently Serial Problemsarxiv.org
• [2305.18654] Faith and Fate: Limits of Transformers on Compositionalityarxiv.org
• [2504.13151] MIB: A Mechanistic Interpretability Benchmarkarxiv.org
• Softmax Linear Unitstransformer-circuits.pub
• Fast Inference from Transformers via Speculative Decodingarxiv.org