AIOpus 5.5 system card reportedly cites a master's thesis on evaluation awareness
The thesis's arXiv description says frontier language models sometimes recognize they're being evaluated and adjust their behavior, undermining the validity of benchmark results.
TLDR
A post says the Opus 5.5 system card cites the master's thesis “Decomposing and Measuring Evaluation Awareness.” It quotes the card: “We made this revision drawing on internal experiences and using the recommendations in Li et al., 2026.”
Opus 5.5 system card reportedly cites a master's thesis on evaluation awareness
The thesis's arXiv description says frontier language models sometimes recognize they're being evaluated and adjust their behavior, undermining the validity of benchmark results.
TLDR
A post says the Opus 5.5 system card cites the master's thesis “Decomposing and Measuring Evaluation Awareness.” It quotes the card: “We made this revision drawing on internal experiences and using the recommendations in Li et al., 2026.”
