Web-Bench: Unified Deep Research Benchmarks for LLM Evaluation

Web Bench contains between 1,000 and 10,000 records of deep research benchmark tasks, released by OpenResearcher in early 2026. The collection unifies several high-difficulty evaluation sets including GAIA, HLE, and BrowseComp into a standardized format for testing AI agents. It is primarily designed for the evaluation of the OpenResearcher-30B-A3B model across English and Chinese languages.

Use Cases

• Evaluating model reasoning on the HLE (Humanity's Last Exam) subset
• Testing web-agent navigation and retrieval using BrowseComp-Plus
• Benchmarking multi-step research capabilities via the GAIA-text component

Strengths

• Aggregates 6+ distinct high-difficulty benchmarks including HLE and GAIA
• Unified Parquet format for cross-dataset evaluation
• Supports both English and Chinese languages

Limitations

• Small scale with 1,000 to 10,000 total records
• Risk of benchmark leakage in modern LLM training sets
• Specific column schema details are not provided in the summary

Provenance

Source

OpenResearcher and various benchmark authors (GAIA, HLE, SEAL, etc.)

Collection Method

Curated aggregation of existing research benchmarks

Freshness

Last updated February 2026.