Trillion Labs Predicts AI Model Inference Performance with 'R-Bridge'

TrillionLabs Unveils 'rBridge,' an Innovative Technology for LLM Inference Prediction Hyper-intelligent AI startup TrillionLabs has unveiled 'rBridge,' an innovative technology designed to solve difficult problems in AI research. rBridge is a methodology that predicts the inference performance of Large Language Models (LLMs) using small proxy models. Training large language models involves enormous computational costs, and in particular, inference capability only emerges beyond a certain scale. This characteristic has made it impossible to predict with small models, leading to repeated expensive trial and error. rBridge overcomes these limitations, accurately predicting the complex inference performance of large LLMs, up to 32 billion parameters, using only small models with less than 1 billion parameters. While existing evaluation methods were disconnected from the LLM's learning objective (next word prediction), leading to low prediction accuracy, rBridge is designed to align its evaluation with the learning method, enabling small models to more reliably represent the performance of large models. Through this, rBridge saves over 100 times the cost of dataset evaluation and model ranking, and with up to 733 times higher efficiency compared to existing methods, it can dramatically reduce LLM development costs and time. It is expected to make a significant contribution to new dataset experiments and the development of industry-specific specialized models. This technology has completed performance verification through various models ranging from less than 100 million parameters to 32 billion parameters, and across 6 key inference benchmarks including elementary and advanced mathematics, scientific reasoning, knowledge evaluation, and programming. Shin Jae-min, CEO of TrillionLabs, emphasized, "rBridge has proven the possibility of reliably predicting the inference capabilities of large models even with small models, and it will bring a turning point to LLM research and the AI ecosystem."