finding
active
finding:mean-difference-patching-in-a-two-layer-relu-circuit-flips-the-decision-to-class-a-by-activating-a-third-hidden-unit-that-is-silent-for-all-natural-class-a-inputsMean-difference patching in a two-layer ReLU circuit flips the decision to class-A by activating a third hidden unit that is silent for all natural class-A inputs
Synthetic theoretical example showing pernicious divergence via hidden pathway activation
Source paper
extracted_from(2025) · Satchel Grant · Simon Jerome Han · Alexa R. Tartaglini · Christopher Potts
Neighborhood — ranked by edge-count
Claims (1)
claim
- Core claim about why pernicious divergence undermines mechanistic conclusions
Related by similarity (8)
cosine ≥ 0.65 · no typed edgeEntities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.
- ReLU hidden layer model circumvents the ReLU by using positive biases or simulating biasfinding0.805Limitation of that toy model.
- ReLU hidden layer model (n=10, m=5) shifts from monosemantic to polysemantic neurons as sparsity increasesfinding0.781Neuron-level phase change.
- Empirical demonstration that MDVP produces divergent representations in a real LLM
- Transcoders easily discover the ReLU(x)+ReLU(-x) solution by default on the absolute value problemfinding0.767Baseline result before repetition.
- Patching an unresolvable pointer at layer ℓ yields no counterfactual answer, but at ℓ+1 yields itfinding0.763Validates reflexive pointer.
- Baseline toy experiment.
- Localizes truth representations to specific hidden states, motivating the rest of the analysis
- Without superposition.