Technical interest · T-06
Mechanistic interpretability
Reading a model's behavior off its activations, and steering it by editing them.
In the inter-agent misalignment project linear probes on Qwen3-8B read corner-cutting at 0.935 AUROC, and diff-of-means steering moved how often the model took shortcuts. Same-wording controls then showed the probe direction tracked content and negation rather than intent.
Nearest neighbors · any layercosine distance in the full space