EdgeBench: 134 Real-World Tasks for Evaluating Autonomous AI Agent Learning

134 real-world tasks for evaluating how autonomous AI agents learn from executable environments. The benchmark tracks agent improvement over 12+ hours per task, providing multi-level feedback and full learning trajectories. ByteDance-Seed publicly released 51 tasks from this benchmark on Hugging Face.

Use Cases

• Benchmarking autonomous agent learning progress based on multi-level feedback and long-duration task execution.
• Evaluating agent adaptation and improvement over time based on the tracked learning trajectory data.
• Developing new agent training algorithms based on the realistic, executable task environments.
• Comparing agent performance across different task types and difficulty levels within the benchmark.

Strengths

• Contains 134 distinct real-world tasks for agent evaluation.
• Tracks agent performance over long durations (12+ hours per task).
• Provides realistic, multi-level feedback within executable environments.

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.
• The full dataset of 134 tasks is not fully released; only 51 tasks are publicly available.

Provenance

Source

ByteDance-Seed

Collection Method

Likely curated or constructed as a benchmark for AI agent evaluation.

Freshness

Last updated 2026-07-02 15:47:23; freshness should be verified.