This paper studies what discriminatively trained reward models (RMs) memorize by measuring counterfactual memorization on two human preference datasets. We show that RMs 1) misallocate memorization to easy, high margin preference pairs, 2) memorize dataset-specific shortcuts (e.g., model identity, user sampling strategy), and 3) overgeneralize simple heuristic correlates of human preference (e.g., length, compliance) when confronted with unseen preference pairs. Overall, our findings indicate that discriminative training of RMs from human preference data results in biased RMs not yet capable of judging response quality in context-dependent scenarios.

This article introduces misinfo-general, a benchmark dataset for evaluating misinformation models’ ability to perform out-of-distribution generalization. Misinformation changes rapidly, much more quickly than moderators can annotate at scale, resulting in a shift between the training and inference data distributions. As a result, misinformation detectors need to be able to perform out-of-distribution generalization, an attribute they currently lack. Our benchmark uses distant labeling to enable simulating covariate shifts in misinformation content. We identify time, event, topic, publisher, political bias, and misinformation type as important axes for generalization, and we evaluate a common class of baseline models on each. Using article metadata, we show how this model fails desiderata, which is not necessarily obvious from classification metrics. Finally, we analyze properties of the data to ensure limited presence of modelling shortcuts. We make the dataset and accompanying code publicly available.1

Community models for malicious content detection, which take into account the context from a social graph alongside the content itself, have shown remarkable performance on benchmark datasets. Yet, misinformation and hate speech continue to propagate on social media networks. This mismatch can be partially attributed to the limitations of current evaluation setups that neglect the rapid evolution of online content and the underlying social graph. In this paper, we propose a novel evaluation setup for model generalisation based on our few-shot subgraph sampling approach. This setup tests for generalisation through few labelled examples in local explorations of a larger graph, emulating more realistic application settings. We show this to be a challenging inductive setup, wherein strong performance on the training graph is not indicative of performance on unseen tasks, domains, or graph structures. Lastly, we show that graph meta-learners trained with our proposed few-shot subgraph sampling outperform standard community models in the inductive setup.