- What changed
- Donor-control AUC for causal timing rose from 0.254 to 0.540 across Llama-3.1 scales, showing increased causal control of hidden trait transfer.
- Why you should care
- Results show causal timing, fixed geometry, and readability are separate properties needing distinct measurement in subliminal learning studies.
- Your move
- Watch. Further tests needed across architectures to confirm if causal control scales universally.
- What to watch next
- Replication of causal timing measurements in non-Llama architectures to test generality of scale-dependent effects.
- Event
- research
- Event date
- Sep 18, 2026
- Relevant to
- General AI readers
