RPRA: Enabling Large Models to Have "Self-Awareness" — Predicting LLM Judges for Efficient Reasoning

This article introduces the RPRA framework, whose core is to let models predict the scores of LLM judges before generating responses, thereby independently deciding when to answer on their own and when to request assistance from large models. This approach significantly reduces inference costs while maintaining performance. The framework provides a new idea for efficient large model reasoning and helps build more intelligent and adaptive AI systems.

Large Language Models (LLMs) face a fundamental contradiction in deployment: larger models have stronger capabilities but consume more computing resources and have higher inference latency, which is especially difficult to overcome on resource-constrained devices. Traditional solutions require a trade-off between efficiency and quality. However, when humans handle problems, they flexibly judge their ability range—solving familiar problems independently and seeking help when beyond their scope. This is exactly the "self-awareness" capability that current large models lack.

The core innovation of the RPRA (Reason-Predict-Reason-Answer/Act) framework is to let models first predict the scores of LLM judges on their own outputs before making decisions on actions. Its Prediction-Action (PA) paradigm includes two steps: prediction (predicting the judge's score for its own answer after receiving a query) and decision-making (answering independently if the score is high, or forwarding to a large model if low). RPRA adds a reasoning phase to form a complete process. The research team explored three implementation strategies: 1. Zero-shot prediction: Directly asking the model to predict scores, where larger models perform well; 2. Contextual report card: Providing small models with scoring criteria and examples, which increases accuracy by an average of 55%; 3. Supervised fine-tuning: Training models with real score data, which increases accuracy by an average of 52%.

The research team validated the effectiveness of RPRA on multiple datasets, with key findings: 1. Model size is positively correlated with prediction ability—large models perform better in zero-shot prediction, while small models need additional guidance or training; 2. Report cards and fine-tuning increase the prediction accuracy of small models by more than 50%; 3. Intelligent routing decisions allow simple problems to be handled by small models and complex ones to be forwarded to large models, ensuring performance while reducing average inference costs.

The significance of the RPRA framework lies in revealing the research direction of AI's metacognitive ability (monitoring of its own cognitive processes). Practical application benefits include: cost optimization (intelligent routing reduces inference costs), user experience (automatically selecting the optimal answer), and scalability (supporting model expansion). Challenges include: the need to balance prediction costs and routing benefits, and prediction accuracy still needs improvement in open-ended tasks.

The RPRA framework provides an elegant new idea for efficient large model reasoning, enabling models to learn "self-awareness" and helping build more intelligent, efficient, and adaptive AI systems. It is an important step toward more self-aware AI. Paper link: http://arxiv.org/abs/2604.12634v1