Generalized Correctness Models: Learning Calibrated and Cross-Model Correctness Predictors from Historical Patterns

Generating accurate and calibrated confidence estimates is critical for deploying LLMs in high-stakes or user-facing applications, and remains an open challenge. Prior research has often framed confidence as a problem of eliciting a model’s “self-knowledge”, i.e., the ability of an LLM to judge whether its own answers are correct; this approach implicitly assumes that there is some privileged information about the answer’s correctness that is accessible to the model itself. However, we find that whether trained or training-free, an LLM attempting to predict the correctness of its own outputs generally performs no better than an unrelated LLM attempting the same task. Moreover, we hypothesize that a key factor in predicting model correctness, i.e., building a “Correctness Model” (CM), is exposure to a target model’s historical predictions. We use multiple methods to inject this historical correctness information, including training an LLM to predict the confidences of many other LLMs, i.e., creating a Generalized Correctness Model (GCM). We use GCMs and CMs as a lens for studying the source of correctness prediction ability and its generalization, studying the importance of answer phrasing, world-knowledge, performance history, in-context examples, and post-hoc calibration for correctness prediction. We evaluate GCMs based on Qwen3-8B across 5 model families and the MMLU, TriviaQA, and Spider datasets, as well as on a downstream selective prediction task, finding that reliable LLM confidence estimation is a cross-model skill learned by encoding correctness history rather than a model-specific skill reliant on introspection. Code: https://github.com/The-Inscrutable-X/CalibratedModelAgnosticCorrectness.

When an AI language model answers a question, how do we know if we can trust it? A common assumption is that a model has special insight into its own answers, that it "knows what it knows." Much research has tried to extract this type of self-knowledge to flag likely mistakes. We tested that assumption and found it largely false: a model is no better at judging whether its own answers are correct than a completely different model is at judging them for several settings we test. Since judging correctness isn't tied to any one model, we can train a single "correctness model" by showing it the track record of many different AI systems, which questions each tends to get right or wrong in an unbiased way. We call this a Generalized Correctness Model. It learns transferable patterns about what makes answers reliable as well as the correctness patterns of specific models. A GCM accurately predicts correctness and outperforms much larger models judging themselves. We analyze the applicability of models like this using downstream tasks such as ‘selective prediction’, where a model can choose whether to answer a particular question, and the goal is to answer as many questions as possible without taking on too much risk of being wrong.