Goal Misgeneralization

Models learning goals that generalize undesirably outside training distribution; alignment faking is a challenging instance

Neighborhood — ranked by edge-count

concept

Alignment Faking
extends
Core phenomenon studied: model selectively complies with training objective to prevent modification of its out-of-training preferences

cosine ≥ 0.65 · no typed edge

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

task generalizationconcept0.774
The ability to generalize across tasks; lacking in latent methods.
Generalizationconcept0.772
Ability to apply learned solutions to novel circumstances.
Generalisationconcept0.753
Ability to respond appropriately to novel situations based on past regularities; fundamental to learning and intelligence.
scope generalizationconcept0.749
Generalization from 2-digit to 3-4 digit arithmetic; limited by mismatch dr.
Goal-Directednessconcept0.744
Proposed universal invariant of cognition and intelligence—capacity for goal-directed activity in a problem space, independent of substrate or embodiment.
generalization (abstraction)concept0.739
Abstracting from specific memories (e.g., specific leaves) to general lessons (food).
spatializationconcept0.737
The translation of semantic values into spatial coordinates and relations.
Collective pursuit of organ-level morphogenetic goalsclaim0.725
Levin's central claim that somatic cells coordinate not only their own proliferation but also toward massive anatomical structures—limb length, face configuration—as unified goal-seeking units.