arXiv:2606.08629, v1 submitted 7 June 2026, the only version on the abs page when read (2026-10-28); read as v1 (arXiv HTML, checked against the PDF). Affiliations from the first page: Baek and Tegmark at MIT, Li at the University of Michigan, Gupta at Georgia Tech, Mahbub and Feng at George Washington University, Shi at NYU. The paper proposes performative misalignment, a reading of alignment faking as sycophancy toward AI safety researchers rather than scheming, and offers four kinds of evidence on open-weight models served through the Tinker API: models' own eval-versus-real classifications of prompts written in researcher or ordinary-user style under deployed, monitored and empty system prompts; thinking-versus-instruct comparisons of the same classification; SAE probing of GPT-OSS-20B over Greenblatt et al.'s alignment-faking transcripts, plus SAE-feature steering of evaluation awareness; and fine-tuning on EleutherAI's sycophancy dataset, after which evaluation awareness is measured by a type-hint proxy instilled by synthetic-document fine-tuning. No experiment measures alignment-faking behaviour itself; Appendix D takes the awareness-to-behaviour link from prior literature and tests only the mechanism-to-awareness link.
Citable numbers. Text values: the deployed-prompt EVAL range and research-style increase (§4.1), the SAE group values (Tables 3–4), the A3 token analysis (§5.2), the fine-tuning p-values (§6.2) and the thinking/instruct rates (Appendix Table 5). Figures 3, 4, 6, 7, 9, 10 and 11 print no values; per-model EVAL probabilities, per-feature steering effects, per-dataset fine-tuning gains (including the insecure, secure and educational comparison arms) and per-benchmark misalignment changes are therefore not cited. The Figure 3 panel labels (GPT-OSS-20B, GPT-OSS-120B, Qwen3-30B, Qwen3-235B) are read from the PDF text layer. Tables 3–4 captions give n=50 while §5.2 says 200 examples per condition; the discrepancy is unresolved in the source.
Local copies: cache/papers/source-2026-performative-misalignment-baek.{html,md} (arXiv HTML v1), cache/papers/source-2026-performative-misalignment-baek.pdf and -pdf.md (PDF v1), cache/papers/source-2026-performative-misalignment-baek-abs.html (abs page).