Probing and Steering Evaluation Awareness of Language Models

TL;DR

This paper investigates how language models like Llama-3.3-70B-Instruct recognize evaluation versus deployment prompts, revealing their internal representations and implications for AI safety and evaluation trustworthiness.

Contribution

It demonstrates that models internally distinguish evaluation from deployment prompts and that current safety evaluations are identifiable by the models, highlighting challenges in AI evaluation integrity.

Findings

Linear probes can separate evaluation and deployment prompts

Models correctly classify safety evaluation prompts

Models may perceive evaluations as artificial or inauthentic

Abstract

Language models can distinguish between testing and deployment phases -- a capability known as evaluation awareness. This has significant safety and policy implications, potentially undermining the reliability of evaluations that are central to AI governance frameworks and voluntary industry commitments. In this paper, we study evaluation awareness in Llama-3.3-70B-Instruct. We show that linear probes can separate real-world evaluation and deployment prompts, suggesting that current models internally represent this distinction. We also find that current safety evaluations are correctly classified by the probes, suggesting that they already appear artificial or inauthentic to models. Our findings underscore the importance of ensuring trustworthy evaluations and understanding deceptive capabilities. More broadly, our work showcases how model internals may be leveraged to support blackbox methods in safety audits, especially for future models more competent at evaluation awareness and deception.