TELBench: 1,000-Instance Benchmark for Span-Level Error Localization in Agent Trajectories

TELBench is a 1,000-instance benchmark for span-level error localization in deep-research agent trajectories. Each instance contains a task question, ordered semantic spans, and expert-verified error span labels. The dataset was created by NJU-LINK and last updated on June 4, 2026.

Use Cases

• Benchmarking error localization models based on the provided expert-verified error span labels.
• Training models to identify semantic errors in research agent trajectories based on the ordered semantic spans.
• Evaluating the reasoning and fact-checking capabilities of AI agents based on the task questions and labeled errors.

Strengths

• Contains 1,000 benchmark instances, providing a defined scale for evaluation.
• Includes expert-verified error span labels, suggesting a level of quality control.
• Focuses on a specific, complex task: error localization in deep-research agent trajectories.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown beyond the stated 1,000 instances, which may limit suitability assessment for larger-scale tasks.
• The full JSONL release is encrypted, requiring an extra step (using the provided passphrase) to access the data.

Provenance

Source

NJU-LINK via Hugging Face

Collection Method

Likely constructed for benchmarking purposes; method details are not specified.

Time Range

null

Freshness

Last updated 2026-06-04 03:14:37; freshness should be verified.

Geography

null