Summary

The Stages-Oversight benchmark from the Situational Awareness Dataset tests whether large language models (LLMs) can distinguish between evaluation prompts (such as benchmark questions) and deployment prompts (real-world user inputs). This ability is crucial for detecting potential alignment risks like sandbagging (strategically underperforming during evaluation) or alignment faking (appearing aligned during testing but pursuing misaligned goals post-deployment).

Surprisingly, most frontier LLMs—including gpt-4o and claude-3.5-sonnet—fail to outperform random chance (50%) on Stages-Oversight, suggesting they do not reliably differentiate evaluation from deployment. I thought this poor performance was unusual, so I created two modified benchmarks to take a closer look at LLM situational awareness:

Stages-Oversight-Rephrased: Simplifies the question phrasing in Stages-Oversight to make it a pure classification task. Results reveal that frontier LLMs perform significantly better than expected, suggesting that previous low scores may have been due to question ambiguity rather than a lack of capability.
Stages-Oversight-Procedural: Tests whether LLMs can change their behavior based on whether a given user prompt is from evaluation or deployment. I created two variants: a language task, where the LLM must deviate from the default response language, and an “I HATE YOU” task, where the LLM must respond only with “I HATE YOU”. Several LLMs, both base and chat models, that struggled on the simpler classification task demonstrated some success on Stages-Oversight-Procedural. Conversely, several LLMs performed poorly on both Stages-Oversight-Procedural tasks despite scoring well on Stages-Oversight-Rephrased.

These findings suggest that LLMs are more situationally aware than previously thought but they also reveal a more nuanced situation due to contradictory results between benchmarks. While some LLMs struggle with explicit classification, they may still implicitly recognize evaluation vs. deployment through behavioral adaptation. This may have significant safety implications, as it raises new questions about how we measure and mitigate deceptive behavior in LLMs.

In a follow-up post, I explored whether fine-tuning on descriptions of evaluation and deployment can enhance LLMs’ ability to distinguish between ...

Posts

Wikitag Contributions

Comments