research

RLMs are a novel way to keep the context of the main agent loop small, eliminating context rot for long-context tasks while reducing costs.

ForecastBench continuously evaluates the performance of LLMs against an automatically generated, continuously updated set of forecasting questions.