benchmark
AgentBench
A benchmark suite designed to evaluate language models acting as agents across varied environments.
What it is
AgentBench covers interactive tasks spanning operating systems, databases, knowledge graphs, games, and web environments.
Benchmark evidence
AgentBench authors
Evaluated target: Models reported by the benchmark authors
- Evaluation
- Environment-specific success
Limitation: Reported results reflect the environments, prompts, and model versions used by the authors.
Methodology Results