Observation Masking Eval Logs: Model Performance on Web-Agent Benchmarks

Evaluation logs for four deep-research and web-agent benchmarks, published by i-DeepSearch on May 30, 2026. The repository contains judge results and trajectory logs from model runs, including a specific dataset named BrowseComp-Plus with 830 entries. Logs document performance under different observation mask context management settings.

Use Cases

• Compare model performance across different observation masking settings based on the provided judge results.
• Analyze agent trajectory patterns and decision-making steps using the trajectory logs.
• Benchmark new web-agent models against the evaluated ones on the four included benchmarks.
• Study the impact of context management (CM) on task performance in web-agent scenarios.

Strengths

• Includes results from four distinct benchmarks, providing a multi-faceted evaluation.
• Contains paired runs comparing performance with and without observation mask context management.
• Logs are structured with separate files for judge results (evaluated.jsonl) and detailed trajectory data.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count for the overall repository is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality and completeness require manual inspection.

Provenance

Source

i-DeepSearch on Hugging Face.

Collection Method

Model evaluation logs from benchmark runs.

Freshness

Last updated 2026-05-30 07:08:04.