Abstract:Current techniques for post-training Large Language Models (LLMs) rely either on costly human supervision or on external verifiers to boost performance on tasks such as mathematical reasoning and code generation. However, as LLMs improve their problem-solving, any further improvement will potentially require high-quality solutions to difficult problems that are not available to humans. As a result, learning from unlabeled data is becoming increasingly attractive in the research community. Existing methods extract learning signal from a model's consistency, either by majority voting or by converting the model's internal confidence into reward. Although internal consistency metric such as entropy or self-certainty require no human intervention, as we show in this work, these are unreliable signals for large-scale and long-term training. To address the unreliability, we propose PRISM, a unified training framework that uses a Process Reward Model (PRM) to guide learning alongside model's internal confidence in the absence of ground-truth labels. We show that effectively combining PRM with self-certainty can lead to both stable training and better test-time performance, and also keep the model's internal confidence in check.
Abstract:Toxicity text detectors can be vulnerable to adversarial examples - small perturbations to input text that fool the systems into wrong detection. Existing attack algorithms are time-consuming and often produce invalid or ambiguous adversarial examples, making them less useful for evaluating or improving real-world toxicity content moderators. This paper proposes an annotation pipeline for quality control of generated toxic adversarial examples (TAE). We design model-based automated annotation and human-based quality verification to assess the quality requirements of TAE. Successful TAE should fool a target toxicity model into making benign predictions, be grammatically reasonable, appear natural like human-generated text, and exhibit semantic toxicity. When applying these requirements to more than 20 state-of-the-art (SOTA) TAE attack recipes, we find many invalid samples from a total of 940k raw TAE attack generations. We then utilize the proposed pipeline to filter and curate a high-quality TAE dataset we call TaeBench (of size 264k). Empirically, we demonstrate that TaeBench can effectively transfer-attack SOTA toxicity content moderation models and services. Our experiments also show that TaeBench with adversarial training achieve significant improvements of the robustness of two toxicity detectors.