Keenan Pepper (AE Studio) | Trained SelfIE: like Activation Oracles but you train only an affine map

Join us for an insightful talk by Keenan Pepper from AE Studio, where he will introduce Trained SelfIE, an innovative method for interpreting activation patterns in large language models (LLMs). Activation explainers that convert hidden activations into understandable language are becoming increasingly important, yet traditional methods like Activation Oracles can be excessively complex.

Keenan presents a simpler solution: an affine map that transforms layer-N activations into soft tokens readable by a frozen model at layer 0. This approach not only produces better labels compared to SAE auto-interp training labels but also excels in decoding unverbalized "bridge entities" in two-hop reasoning — outperforming both untrained SelfIE and linear probes, especially as model scale increases. We’ll speculate on the future possibilities that this frozen-model property may hold.

📅 Event Highlights:

• Doors Open: 7:00 PM
• Talk Begins: 7:30 PM
• Continued Hangouts: 8:00 PM

Bonus: Join us for a members' dinner at 6:30 PM before the event!

This event is part of the Mox Summer Season — a series of salons and social gatherings exploring cutting-edge ideas. Explore the complete lineup!

🔗 More on the research here
🔗 Get your tickets here