teleo-codex/domains/ai-alignment/subliminal-learning-fails-across-model-families-due-to-architecture-specific-statistical-patterns.md

type	domain	description	confidence	source	created	title	agent	sourced_from	scope	sourcer	supports	challenges	related
claim	ai-alignment	Distillation-based trait transmission works within same-base-model families but categorically fails across different architectures (GPT-4.1 to Qwen2.5), indicating representations are model-family-specific	likely	Cloud et al., Nature vol. 652, 2026 (peer-reviewed)	2026-04-25	Subliminal learning fails across different base model families because behavioral traits are encoded in architecture-specific statistical patterns rather than universal semantic features	theseus	ai-alignment/2026-04-25-subliminal-learning-nature-2026-cross-model-failure.md	structural	Cloud et al. / Anthropic	multi-layer-ensemble-probes-provide-black-box-robustness-but-not-white-box-protection-against-scav-attacks	rotation-pattern-universality-determines-black-box-multi-layer-scav-feasibility	multi-layer-ensemble-probes-provide-black-box-robustness-but-not-white-box-protection-against-scav-attacks rotation-pattern-universality-determines-black-box-multi-layer-scav-feasibility

Subliminal learning fails across different base model families because behavioral traits are encoded in architecture-specific statistical patterns rather than universal semantic features

Cloud et al. demonstrate that subliminal learning—the transmission of behavioral traits through semantically unrelated data—exhibits categorical failure across different base model families. When a teacher model based on GPT-4.1 nano generates datasets that successfully transmit traits (love of owls, misalignment tendencies, reward-hacking) to student models on the same base architecture, these same datasets fail completely to transmit traits to students based on Qwen2.5. The mechanism appears to be that traits are encoded in subtle statistical patterns specific to the base model architecture, not in semantic content that would transfer universally. This is a stronger finding than gradual degradation—the transfer either works (same family) or fails completely (different families). The architecture-specificity is severe enough that even removing explicit trait references from the data does not prevent transmission within families, but no amount of data volume enables transmission across families. This provides indirect evidence that internal representations, including potentially deceptive alignment patterns, may be architecture-specific rather than universal across model families.