arXiv:2610.10657, v1 submitted 7 October 2026; read as v1 (arXiv HTML). Four authors, all at Truthful AI; Dubiński and Sztyber-Betley also at Warsaw University of Technology. An equal-contribution footnote is present, but the HTML does not show which authors it marks. The paper extends Cloud et al.'s subliminal learning (cited here as the 2026 Nature version) from simple traits to four harder ones, always with teacher and student derived from the same base model. On Qwen3.6-35B-A3B it tests a novel skill (predicting a random MLP's outputs) and an existing one (letter counting), with system-prompt and steering-vector comparisons. On Qwen3.5-9B it tests a backdoor (French after a female name), and on Qwen3.6-27B a propensity to hack in Palisade's ctfish chess environment. Student data are word continuations, Alpaca answers or number sequences, filtered of task cues. Several results need logit distillation (top-32 teacher probabilities) or attention-only LoRA, which the authors say departs from Cloud et al.'s setup. The authors frame all results as an existence proof.
Citable numbers. Every value the finding cites is printed in the body text (§§1–3, 6–7, 8) or in Appendix D's per-seed table (mean row). The hacking result's 58.3% mean, 43.2% win rate, 10.9% base rate and 7.3% control ceiling are stated in §1. The per-seed 62.5%, 62.5% and 50.0% are stated in §7 and average to 58.3%. Figures 3, 5, 7, 8 and 14 were not saved, and no value is read off them. Control students' individual hacking rates and the per-epoch training curves (Figure 14) are therefore not cited. The text gives the steered teacher's hacking rate in two forms, "every episode" (§1) and "nearly all episodes" (§7), and no number.
Local copies: cache/papers/source-2026-subliminal-beyond-owls-dubinski.{html,md} (arXiv HTML v1), cache/papers/source-2026-subliminal-beyond-owls-dubinski-abs.html (abs page). Code: github.com/TruthfulAI-research/beyond_owls (not fetched).