Forecasting Downstream Performance of LLMs With Proxy Metrics

Proxy metrics constructed from token-level statistics (entropy, top-k accuracy) predict downstream LLM performance more reliably than cross-entropy loss, enabling early-stage architecture and training decisions.