J-Lens struggles to detect preference transmission from subliminal learning teacher models

After reproducing the subliminal learning experiment on Qwen2.5-7B-Instruct, the authors found that J-Lens can barely read out the concept the teacher model is transmitting. The teacher was system-prompted to love owls, then used number sequences to pass this preference to the student; the student's rate of naming owls rose from about 1% to about 8%. However, on 1.42M completions, J-Lens's owl top-1 hit rate was only 1.65% (random baseline 0.26%), with a median best rank of 213. The same lens achieved 45% top-1 and 74% top-5 hit rates on toy tasks such as sentence copying, indicating it has not failed overall. The authors state that the phenomenon was not reproduced in most settings on larger models, but a similar situation appeared in some settings on Gemma-4-31B. The appendix shows that training the student on samples judged as hits by the lens can raise the transmission rate to 57.2%, while samples with no hits among the lens's top 20 still transmit at a level close to random.