arXiv:2605.28591. v1 is dated 27 May 2026. v4 (7 Sep 2026) is the version read (arXiv HTML), and the abstract page still listed it as the latest on 2026-11-19. Deckenbach and Puerto contributed equally; the paper lists the ELLIS Institute Tübingen, MPI for Intelligent Systems and the Tübingen AI Center as affiliations. Code is linked at github.com/compass-group-tue/arxiv2026_evaluation_meta_knowledge. It was not cloned or read. Local copies: cache/papers/source-2026-evalmeta-deckenbach.{html,md}, and the abstract page as source-2026-evalmeta-deckenbach-abs.html.
The authors LoRA-fine-tune Nemotron 49B v1.5, Qwen3 32B and GLM 4.7 Flash (with OLMo 3.1 32B Instruct and DeepSeek R1 Distill 32B in an appendix) on about 106M tokens of synthetic documents that describe seven traits of AI evaluations. They then score the models on five safety benchmarks against the base model, a FineWeb fine-tuning control and Hua et al.'s type-hints evaluation-awareness organism. The fine-tuned models refuse more and act harmfully less, including on responses that a GPT-5 judge rates as not verbalizing evaluation awareness. The authors call the parametric knowledge evaluation meta-knowledge and argue that it may inflate safety scores. Their Limitations and a small ToxicChat test leave open whether the shift is score inflation or a real change in alignment.