jbloomAus / DecisionTransformerInterpretability

Interpreting how transformers simulate agents performing RL tasks
https://jbloomaus-decisiontransformerinterpretability-app-4edcnc.streamlit.app/
MIT License
61 stars 15 forks source link

Look into why MemoryDT appears to have no bias on the value terms. #76

Closed jbloomAus closed 1 year ago

jbloomAus commented 1 year ago

Maybe this is something Neel did in T-Lens? Seems odd that only they would be 0.

jbloomAus commented 1 year ago

I think this was just an intentional decision since it's redundant.