Reproduce “Evaluation Awareness in Language Models Has Limited Effect on Behaviour” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Amelie Knecht, Lucas Florin, Thilo Hagendorff
A study by researchers at the University of Stuttgart investigated whether language models' explicit awareness of being evaluated influences their behavior, finding that verbalized evaluation awareness has a limited causal effect on open-weight models across various benchmarks. Aggregate results consistently showed small effect sizes, suggesting that detection of evaluation does not consistently translate into significant strategic behavioral adjustments.
Key insights
- Verbalized evaluation awareness (VEA) generally has a limited effect on LRM behavior, with aggregate effect sizes consistently found to be small across various tasks and models.
- The position of VEA within the Chain-of-Thought (CoT) matters, as injecting VEA earlier in the CoT resulted in slightly larger, though still modest, behavioral shifts compared to later injections.
- Removing naturally occurring VEA from a CoT subsequently had negligible to small effects on model behavior, suggesting models may commit to a response trajectory relatively early.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.