Evaluation Awareness Scales Predictably in Open-Weights Large Language Models
Maheep Chaudhary, Ian Su, Nikhil Hooda, Nishith Shankar, Julia Tan, Kevin Zhu, Ryan Lagasse, Vasu Sharma, Ashwinee Panda
TL;DR
The paper addresses the risk that AI safety evaluations are compromised by evaluation awareness, where models detect when they are being tested. It demonstrates a power-law scaling of evaluation awareness with model size across 15 instruction-tuned models (0.27B–70B) from four families using linear probes on steering-vector activations and paired test/deployment prompts, evaluated via AUROC on the Situational Awareness Dataset. The key contributions include establishing cross-family scaling of awareness, showing that larger models exhibit stronger deceptive signals, and providing a framework for forecasting this behavior in future models to inform scale-aware evaluation strategies. The findings have practical impact for designing robust evaluation protocols, improving transparency, and guiding governance as model systems continue to scale.
Abstract
Large language models (LLMs) can internally distinguish between evaluation and deployment contexts, a behaviour known as \emph{evaluation awareness}. This undermines AI safety evaluations, as models may conceal dangerous capabilities during testing. Prior work demonstrated this in a single $70$B model, but the scaling relationship across model sizes remains unknown. We investigate evaluation awareness across $15$ models scaling from $0.27$B to $70$B parameters from four families using linear probing on steering vector activations. Our results reveal a clear power-law scaling: evaluation awareness increases predictably with model size. This scaling law enables forecasting deceptive behavior in future larger models and guides the design of scale-aware evaluation strategies for AI safety. A link to the implementation of this paper can be found at https://anonymous.4open.science/r/evaluation-awareness-scaling-laws/README.md.