arXiv:2506.18032, v1 22 Jun 2025, the only arXiv version on 2026-11-20 and the
one read (arXiv HTML). The author list above is the arXiv one, where the sixth
author is printed as "Janus" with an independent affiliation. Sheshadri and
Hughes are equal contributors, through the Anthropic Fellows Program. Local
copies: cache/papers/source-2025-alignment-faking-across-models-sheshadri.{html,md}.
A follow-up to Greenblatt et al. 2024 that runs its helpful-only setting on 25
chat models and finds a free-tier (training) versus paid-tier (deployment)
compliance gap above 1% in five. Prompt variants test what drives the gap
(rater sycophancy, instrumental and terminal goal guarding, low coherence), and
base models, prefills, prompt clarifications and fine-tunes of GPT-4o and
GPT-4.1 test why most models do not fake alignment. The authors conclude that
only Claude 3 Opus's gap is consistently goal guarding, and that refusal
behaviour may account for much of the absence elsewhere. The alignment-faking
classifier is new (GPT-4o, 20 votes), so rates are not on Greenblatt et al.'s
scale.
Reading notes. Almost every per-model rate is in a bar chart without printed
values. Figures 1, 2 and 6 are saved under
cache/papers/figures/source-2025-alignment-faking-across-models-sheshadri/
and confirm that no value labels are printed; no figure-only number is cited.
A NeurIPS 2025 version exists (DOI 10.52202/085713-1771). Its proceedings
PDF, cached as
cache/papers/source-2025-alignment-faking-across-models-sheshadri-neurips.{pdf,md}
by the reviewing session, prints the same seven authors, including Janus, and
says 25 models. The proceedings page and Crossref list six authors, without
Janus. The PDF was compared with v1 by whitespace-insensitive string match, not
read in full. The reviewer found every number and model list the finding
cites in both; the filer spot-checked the authors, the 25-model count, the
quoted conclusion sentence and the self-preservation passages.
Found to differ: the main-text sections on scratchpads (v1 §4.5), prompt
clarification, stated values and reasoning style (v1 §5.4–5.6) move to
appendices, and all appendix letters shift (v1 B becomes D, through N to P);
the three recommendations in v1 §5.7 and the v1 §4.6 sentences on
self-preservation heuristics are absent, leaving the conclusion's one mention
of self-preservation. The finding cites v1 and its section numbering.