RPRA: Predicting an LLM-Judge for Efficient but Performant Inference

Large language models (LLMs) face a fundamental trade-off between computational efficiency (e.g., number of parameters) and output quality, especially when deployed on computationally limited devices such as phones or laptops. One way to address this challenge is by following the example of humans and have models ask for help when they believe they are incapable of solving a problem on their own; we can overcome this trade-off by allowing smaller models to respond to queries when they believe they can provide good responses, and deferring to larger models when they do not believe they can. To this end, in this paper, we investigate the viability of Predict-Answer/Act (PA) and Reason-Predict-Reason-Answer/Act (RPRA) paradigms where models predict—prior to responding—how an LLM judge would score their output. We evaluate three approaches: zero-shot prediction, prediction using an in-context report card, and supervised fine-tuning. Our results show that larger models (particularly reasoning models) demonstrate good zero-shot prediction abilities, while smaller models require in-context report cards or fine-tuning for reliable predictions. While the effectiveness varies across datasets, both approaches can substantially improve smaller models' prediction accuracy, with fine-tuning achieving mean improvements up to 52\% across datasets. These findings suggest that models can learn to predict their own performance limitations, paving the way for more efficient and self-aware AI systems.