Epistemic Status: Exploratory experiments on a single model (Gemma 3 27B) using Gemma-Scope SAE. We are highly confident in the geometric voids and behavioral phase transitions observed at Layer 40, but less confident about cross-model universality without further replication. TL;DR: * We mapped the activation space of Gemma 3 27B...
We’ve been investigating how safety mechanisms and generation trajectories are initialized in Large Language Models. Using Sparse Autoencoders (SAEs) on Gemma 3 27B, our results suggest that refusal behavior may be largely determined by latent states established during prefill rather than through continuous suppression during decode. In this post, we’ll...
When it comes to AI Alignment the current landscape is vastly using Constitutional AI, RLHF and System Prompting, and this is tending to lead LLMs to project Alignment Faking, because ultimately the LLMs in question is not being direct participant of any of the negotiation of the meaning and concepts...