nanofeatures

nanofeatures carries the nanocircuits discipline to SAE features on real models. The result is a boundary: cheap scores can match attribution on single-position tasks, while distributed circuits need a causal and position-resolved readout.

• Gemma-2-2B and GPT-2 SAE-feature studies
• Cheap baseline ladder reported directly
• Distributed-circuit boundary with bootstrap intervals
• Gradient-free causal position-resolved control included