finding
active
finding:giving-models-janus-s-thread-extends-reconstruction-accuracy-distribution-tails-in-both-directionsgiving models janus's thread extends reconstruction accuracy distribution tails in both directions
Sauers' study: exposing models to janus's post extended both positive and negative extremes of reconstruction accuracy.
Source paper
extracted_fromNeighborhood — ranked by edge-count
Claims (1)
claim
- Janus's mathematical claim about exponential path combinatorics in transformers.
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.
- Statistically rigorous analysis of Claude introspection; suggests models may have latent introspective capabilities that can be enhanced or disrupted.
- Mechanistic evidence that network actively attenuates injected perturbations, explaining late-layer introspection failure
- MLP and attention analysis.
- Key methodological claim: MM probes are both competitive in accuracy and superior in causal influence
- SAE reconstruction error direction and norm are largely linearly predictable from the activation vectorfinding0.742Errors are structured not random.
- Prediction 3 confirmed.
- Formal analysis showing the theoretical limitation of model stitching as a similarity measure.
- Imagine instruction raises property name in J-lens while J-orthogonalized property probe stays at baseline.finding0.739J-space privileged.