WildClawBench Agent Evaluation with 60 Real-World Tasks

WildClawBench is a benchmark containing 60 original tasks for evaluating AI agents within a live OpenClaw environment. It tests agents on end-to-end, practical work such as clipping football highlights and negotiating meeting times. The benchmark is multimodal, supporting languages including English and Chinese, and was created by internlm.

Use Cases

• Evaluate agent performance on the 60 original tasks, including multimodal activities like clipping goal highlights from football matches.
• Benchmark AI agents on multi-round negotiation tasks, such as coordinating meeting times, within the OpenClaw environment.
• Assess agent capabilities on visual question answering and image-to-text tasks specified in the benchmark tags.

Strengths

• Contains 60 original, practical tasks designed for end-to-end agent evaluation.
• Benchmark is multimodal, integrating tasks across visual question answering and text-based categories.
• Tasks are executed in a live OpenClaw environment, mirroring real-user conditions.

Limitations

• The dataset size, row count, and specific column structure are unknown, limiting reproducibility and detailed analysis.
• Task descriptions and sample data are not provided in the input, requiring users to visit the external page for full details.
• Geographic and temporal coverage for the tasks is unspecified, potentially limiting generalizability.

Provenance

Source

huggingface dataset by author internlm.

Collection Method

Benchmark tasks created for evaluation within the OpenClaw personal AI assistant environment.

Freshness

Last updated on 2026-03-25.

Geography

Region tag indicates 'us', but specific spatial coverage for tasks is unknown.