Mean Ablation / direct effect mean ablation / Mean token ablation

anchor

Findings

• IC-1371The Pythia-1.4b successor head l12h0 exhibits interpretable polysemanticity, performing successorship, acronym prediction, copying, and greater-than behaviors on natural language data [primary]
• IC-150Second-order effects of CLIP's MLP neurons are concentrated in late layers (8–10 of 12 in ViT-B/32) [primary]
• IC-151Each CLIP neuron's second-order effect is approximately a single linear direction in the joint text-image space, significant for fewer than 2% of images [primary]
• IC-218LLaMA3-8B and other LLMs solve arithmetic via a bag of independent heuristic neurons in middle and late MLP layers rather than a robust algorithm [primary]
• IC-220In Pythia-6.9B, the bag-of-heuristics mechanism emerges gradually during training and is the primary arithmetic mechanism from the earliest checkpoint showing good performance (23k steps) [primary]
• IC-677CLIP ViT's image representation is primarily constructed by the last 4 MSA layers, with MLPs and early MSA layers contributing negligibly [primary]
• IC-678Specific attention heads in CLIP ViT-L's last 4 layers encode specific image properties (color, shape, location, counting, texture) that are linearly recoverable via text directions [supporting]
• IC-679CLIP relies on background/location as a spurious cue for bird classification, and ablating geolocation heads improves worst-group accuracy by 25.2% [primary]
• IC-680CLIP ViT's image token contributions are spatially localized to match described content, enabling zero-shot segmentation that outperforms existing CLIP-based methods [supporting]
• IC-721The 72-head entity tracking circuit identified in Llama-7b achieves high faithfulness in Vicuna-7b and Goat-7b without any modification to the circuit graph [primary]