Gao’s logit-diff ablation is not a circuit rule
Gao et al.'s logit-diff ablation score measures how sparse an SAE latent's effect on output logits is. It is not a rule for admitting circuit components.
Category Deep Dives
Math behind J-lens/Jacobian, layer-specific behavior, experiments replicating on open models (Qwen, etc.), code/tutorials for probing J-space.
Gao’s citation gloss of the BERT illusion
Gao et al. (2024) cite the BERT interpretability illusion as a gloss: random activating examples can make an overly broad feature explanation look right.
How does arXiv:2412.09565v2 define obfuscated activations against a named latent-space monitor?
In arXiv:2412.09565v2 (Bailey et al., 8 Feb 2025), an activation is obfuscated against a named latent-space monitor D, not obfuscated in the abstract.
Vi-CD and WeightLens/CircuitLens: Which Paper-Stated Controls Make an Extracted Circuit Causal, and What Stays Unknown
An extracted subgraph is not yet a causal circuit. Read it as causal only after the paper’s own intervention controls are fixed, and do not fill parameters…
ETA vs. Li et al.: Interference Prefix, Response-Prefix and Prompt-Suffix Placement Are Three Different Procedures
These are three different early-token procedures, not three placements of one prefix. In Ding, Li, and Zhang, ETA ( arXiv:2410.06625v2 , ICLR 2025; version…
ACC++ (arXiv:2602.13483v2): The Exact Steps for Rank-1 QK Singular-Direction Candidates and Post-Softmax Removal
In arXiv:2602.13483v2, ACC++ forms one rank-1 candidate for each pair (upstream component, singular direction of the head’s QK matrix Ω), builds the…
GPT-4 Wrote Explanations for 307,200 GPT-2 XL Neurons and Graded Them by Simulation
GPT-4 writes natural-language explanations of 307,200 GPT-2 XL MLP neurons across 48 layers, then scores them by simulating activations on held-out text.
Three Tests That Separate Ignition From Residual Drift
Saturating nonlinearities, competitive attention, and sparse writes mark true ignition in residual streams. Residual drift alone fails the criterion.
Knowing Both Facts Still Fails the Composed Two-Hop Query
Two-hop circuits compose facts via a residual-stream bridge instead of an A-to-C shortcut. Knowing both facts still fails the composed query.
Harmful Demo Count Flips Refusal Before SAE Features Disappear
Packing many harmful demos into long context lifts attack success. Refusal features may stay SAE-readable even when the model complies.