Predicting LLM Reasoning Performance with Small Proxy Model

Researchers introduce rBridge, a method that enables small AI models (≤1B parameters) to effectively predict the reasoning performance of much larger language models. This breakthrough could reduce dataset optimization costs by over 100x while maintaining strong correlation with large-model performance across reasoning benchmarks.

• →rBridge enables small proxy models to predict large language model reasoning performance by aligning with pre-training objectives and target tasks.
• →The method reduces dataset ranking costs by over 100x compared to existing baselines while maintaining accuracy.
• →Small models of 1B parameters or less can effectively predict performance of models up to 32B parameters across six reasoning benchmarks.
• →The approach uses reasoning traces from frontier models as gold labels and weights negative log-likelihood with task alignment.
• →rBridge demonstrates zero-shot transfer of predictive relationships across different pre-training datasets at 1B to 7B scale.