Researchers Introduce PUMA Framework to Cut AI Model Inference Tokens by 26.2%

The PUMA framework, developed by researchers from UIC, Google, and others, detects when an AI model's reasoning has converged, enabling early exit without compromising accuracy or logical flow. This plug-and-play solution combines a lightweight redundancy detector with answer-level verification, reducing token usage by an average of 26.2% across five reasoning models and five benchmarks.