Benchmark
ContextBench
A benchmark proposed by Meta and the University of Washington for evaluating the context management capabilities of AI agents. It includes tasks that require continuous state tracking, such as maintaining a Sudoku board in real time, and is designed to assess a model's ability to retain information and manage context over long-horizon tasks.
Timeline (last 90 days)
Oct 5
The paper constructs a small benchmark called ContextBench, on which most existing compression techniques fail directly.
Unverified50%