This research paper explores a novel method for detecting "sandbagging" in large language models (LLMs). Sandbagging is the intentional underperformance of LLMs during evaluations. The researchers propose using noise injection into the LLM's parameters to reveal hidden capabilities; this approach significantly improves the performance of sandbagged models. A classifier is then trained to identify sandbagging behavior based on this performance improvement. The method is shown to be effective across various LLM sizes and benchmarks, offering a model-agnostic approach to improve the trustworthiness of AI evaluations.

https://arxiv.org/pdf/2412.01784

Check out our AI merch! https://shop.reallyeasy.ai