Bartosz Cywiński, Emil Ryd, Rowan Wang, Senthooran Rajamanoharan, Neel Nanda, Arthur Conmy, Samuel Marks
We study secret elicitation: discovering knowledge that an AI possesses but does not explicitly verbalize. As a testbed, we train three families of large language models (LLMs) to possess specific knowledge that they apply downstream but deny knowing when asked directly. For example, in one setting, we train an LLM to generate replies that are consistent with knowing the user is female, while denying this knowledge when asked directly. We then design various black-box and white-box secret elicitation techniques and evaluate them based on whether they can help an LLM auditor successfully guess the secret knowledge. Many of our techniques improve on simple baselines. Our most effective techniques (performing best in all settings) are based on prefill attacks, a black-box technique where the LLM reveals secret knowledge when generating a completion from a predefined prefix. Our white-box techniques based on logit lens and sparse autoencoders (SAEs) also consistently increase the success rate of the LLM auditor, but are less effective. We release our models and code, establishing a public benchmark for evaluating secret elicitation methods.
Floods are the most common form of natural disaster and accurate flood forecasting is essential for early warning systems. Previous work has shown that machine learning (ML) models are a promising way to improve flood predictions when trained on large, geographically-diverse datasets. This requirement of global training can result in a loss of ownership for national forecasters who cannot easily adapt the models to improve performance in their region, preventing ML models from being operationally deployed. Furthermore, traditional hydrology research with physics-based models suggests that local data -- which in many cases is only accessible to local agencies -- is valuable for improving model performance. To address these concerns, we demonstrate a methodology of pre-training a model on a large, global dataset and then fine-tuning that model on data from individual basins. This results in performance increases, validating our hypothesis that there is extra information to be captured in local data. In particular, we show that performance increases are most significant in watersheds that underperform during global training. We provide a roadmap for national forecasters who wish to take ownership of global models using their own data, aiming to lower the barrier to operational deployment of ML-based hydrological forecast systems.