Linearity of Relation Decoding in Transformer Language Models

Approximating relation decoding in transormer LMs as simple linear transformations on the subject token.

Linearity of Relation Decoding in Transformer Language Models Linearity of Relation Decoding in Transformer LMs Evan Hernandez 1* , Arnab Sen Sharma 2* , Tal Haklay 3 , Kevin Meng 1 , Martin Wattenberg 4 , Jacob Andreas 1 , Yonatan Belinkov 3 , David Bau 2 1 MIT CSAIL , 2 Northeastern University , 3 Technion - IIT ; * Equal contribution New! Try interacting with a lre-edited GPT to see the effect of inserting hundreds of memories. --> ArXiv Preprint Source Code Dataset How do Transformer LMs Decode Relations? Much of the knowledge contained in neural language models may be expressed in terms o

related reading

• A Mathematical Framework for Transformer Circuitstransformer-circuits.pub
• Transformer Circuits Threadtransformer-circuits.pub
• Towards Monosemanticity: Decomposing Language Models With Dictionary Learningtransformer-circuits.pub
• Transformers from Scratche2eml.school
• Transformer Explainer: LLM Transformer Model Visually Explainedpoloclub.github.io
• The Annotated Transformernlp.seas.harvard.edu
• On the Tradeoffs of SSMs and Transformers | Goomba Labgoombalab.github.io
• Natural Language Autoencoders Produce Unsupervised Explanations of LLM Activationstransformer-circuits.pub
• Learning to reason with LLMs | OpenAIopenai.com
• Getting Caught Up to Modern LLM Research | Samarth Goeldev.samarthgoel.com
• A mechanism for solving relational tasks in transformer language modelsarxiv.org
• The Annotated Transformernlp.seas.harvard.edu