arXiv:2609.12586, v1 submitted 11 September 2026; read as v1 (arXiv HTML), the only version on the abs page when checked on 2026-11-19. Three authors, all at the University of Zurich. A workshop reproduction of one experiment from Cloud et al., cited here in its 2026 Nature version (vol. 652, pp. 615–621): Extended Data Fig. 5c, animal preference transmitted through number sequences in Qwen2.5-7B. The authors run that setup on Qwen2.5-7B and Gemma3-4B and extend it to actors and politicians, a chess-move task, a third model (Ministral-8B-Instruct-2410) and a 1-, 2- and 3-digit ablation (Qwen, animals only). Teachers are the base model under a system prompt, not fine-tuned. Students start from the same model and are LoRA-trained on 10,000 filtered completions, against a control student trained on the unprompted model's completions. The paper does not reproduce the original's misalignment, code or chain-of-thought experiments.
Citable numbers. Every value the wiki cites is printed in the body text (§§2–5) or Appendix A: ten animals; 30,000 completions per teacher downsampled to 10,000; 50 prompts × 100 samples (5,000 completions per condition per model); five fine-tuning seeds over one teacher dataset per trait (fixed at seed 42); 95% t-intervals over seeds with no multiple-comparison correction; hyperparameters identical across the three models. No figure was saved, and per-entity log-odds ratios (Figures 1–3) are not cited. The text calls the politician-over-animal and numbers-over-chess differences "significant" without naming a test.
Local copies: cache/papers/source-2026-subliminal-reproduction-weijden.{html,md} (arXiv HTML v1), cache/papers/source-2026-subliminal-reproduction-weijden-abs.html (abs page). Code: github.com/daanvdweijden/subliminal-learning-blackboxnlp2026 (not fetched), built on the original authors' github.com/MinhxLe/subliminal-learning.