Callum Canavan

Callum Canavan, Aditya Shrivastava, Allison Qi et al.

To steer language models towards truthful outputs on tasks which are beyond human capability, previous work has suggested training models on easy tasks to steer them on harder ones (easy-to-hard generalization), or using unsupervised training algorithms to steer models with no external labels at all (unsupervised elicitation). Although techniques from both paradigms have been shown to improve model accuracy on a wide variety of tasks, we argue that the datasets used for these evaluations could cause overoptimistic evaluation results. Unlike many real-world datasets, they often (1) have no features with more salience than truthfulness, (2) have balanced training sets, and (3) contain only data points to which the model can give a well-defined answer. We construct datasets that lack each of these properties to stress-test a range of standard unsupervised elicitation and easy-to-hard generalization techniques. We find that no technique reliably performs well on any of these challenges. We also study ensembling and combining easy-to-hard and unsupervised techniques, and find they only partially mitigate performance degradation due to these challenges. We believe that overcoming these challenges should be a priority for future work on unsupervised elicitation.

Gabriel Recchia, Chatrik Singh Mangat, Jinu Nyachhyon et al.

Scalable oversight protocols aim to empower evaluators to accurately verify AI models more capable than themselves. However, human evaluators are subject to biases that can lead to systematic errors. We conduct two studies examining the performance of simple oversight protocols where evaluators know that the model is "correct most of the time, but not all of the time". We find no overall advantage for the tested protocols, although in Study 1, showing arguments in favor of both answers improves accuracy in cases where the model is incorrect. In Study 2, participants in both groups become more confident in the system's answers after conducting online research, even when those answers are incorrect. We also reanalyze data from prior work that was more optimistic about simple protocols, finding that human evaluators possessing knowledge absent from models likely contributed to their positive results--an advantage that diminishes as models continue to scale in capability. These findings underscore the importance of testing the degree to which oversight protocols are robust to evaluator biases, whether they outperform simple deference to the model under evaluation, and whether their performance scales with increasing problem difficulty and model capability.