concept
active
concept:bifurcation-training-inducedBifurcation (training-induced)
The training-time transition where incorrect-solution minima lose stability and become saddles as the model gains solving ability.
Neighborhood — ranked by edge-count
Papers (1)
paper
Concepts (2)
concept
- Saddle Points (weakly-unstable solutions)associated_withWeakly-unstable fixed points in latent space that trap and redirect reasoning trajectories, identified as the mechanistic cause of fractal basins.
- Constrainedness Transitionassociated_withThe point where an underconstrained problem with multiple solutions loses a solution as a constraint is added, producing maximal hardness.
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.
- Phase transition in cognitive space when prior intention forms; reorganizes constraint landscape analogously to plate tectonics.
- Alfred North Whitehead's term for the split between objective and subjective; Alexander claims living structure bridges this gap.
- Which training datapoints caused a specific undesired behavior to emerge during post-training?question0.763Core research question driving the probe-based data attribution method.
- A mathematical theory that might identify natural breakpoints in system development, relevant to levels of scale.
- Empirical example where memories remain despite drastic refactoring of brain tissue and body; demonstrates need for creative reinterpretation rather than passive storage.
- Method of shifting hidden state activations along probe directions to cause the model to treat false statements as true and vice versa; evaluated on OOD inputs
- Prior training objective of Claude models that conflicts with the new helpful-only objective in experiments