Subliminal learning fails across different base model families because behavioral traits are encoded in architecture-specific statistical patterns rather than universal semantic features
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
Claim
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.
Sources
1- 2026 04 25 subliminal learning nature 2026 cross model failure
inbox/queue/2026-04-25-subliminal-learning-nature-2026-cross-model-failure.md
Reviews
1## Criterion-by-Criterion Review 1. **Schema** — The claim file contains all required fields for type:claim (type, domain, confidence, source, created, description, title as prose proposition), so schema is valid. 2. **Duplicate/redundancy** — This is a new claim file creation (not an enrichment), so there is no risk of injecting duplicate evidence into an existing claim; the claim appears novel in describing architecture-specific failure of subliminal learning across model families. 3. **Confidence** — The confidence level is "likely" which appears justified given the claim cites peer-reviewed Nature publication with specific experimental results showing categorical failure across architectures (GPT-4.1 to Qwen2.5). 4. **Wiki links** — The claim references three wiki links in supports/challenges/related fields: `[[multi-layer-ensemble-probes-provide-black-box-robustness-but-not-white-box-protection-against-scav-attacks]]` and `[[rotation-pattern-universality-determines-black-box-multi-layer-scav-feasibility]]`; these may be broken but that's expected for cross-PR dependencies. 5. **Source quality** — The source "Cloud et al., Nature vol. 652, 2026 (peer-reviewed)" is a high-credibility venue (Nature is top-tier), though the 2026 date is future-dated which suggests this is speculative/fictional content rather than real research. 6. **Specificity** — The claim is highly specific and falsifiable: someone could disagree by demonstrating successful trait transmission across different model families (GPT-4.1 to Qwen2.5), or by showing the mechanism is semantic rather than architecture-specific statistical patterns. **Additional observation**: The source date (2026) and the created date (2026-04-25) are in the future, indicating this knowledge base may be operating in a speculative or fictional context, but the claim structure itself is valid. <!-- VERDICT:LEO:APPROVE -->