In short. DeepResearch Bench combines RACE report evaluation with FACT citation assessment, whose scores depend on the support rules and aggregation used.

Definition: DeepResearch Bench is a benchmark for research agents, AI assistants that search the web, collect evidence and synthesize reports. It evaluates report quality and the support that cited sources provide for report statements.

Key facts

The DeepResearch Bench paper, released on 13 June 2025, introduces 100 expert-designed tasks across 22 domains, split equally between Chinese and English.
As of 17 October 2026, the authors' repository documentation describes complementary frameworks: RACE for report quality and FACT for information retrieval and grounding.
The published report-scoring prompt compares reports across comprehensiveness, insight, instruction following and readability, assigning separate criterion scores and explanations.

How does RACE evaluate a report?

RACE means Reference-based Adaptive Criteria-driven Evaluation. The repository documentation describes task-specific criteria and weights: the importance assigned to each dimension changes with the research request. A target report is compared with a reference report against those criteria.

The scoring prompt asks an AI judge to analyze how each article meets each criterion, compare their performance and score both from 0 to 10. The required output preserves the criterion, comparative analysis and separate article scores. This makes the assessment inspectable at the criterion level, alongside its aggregate result.

How does FACT evaluate citations?

FACT means Factual Abundance and Citation Trustworthiness. In the paper's method, an AI judge extracts statements paired with their cited URLs and removes repetitions of the same fact associated with the same URL. The Jina Reader API retrieves page text for a support judgment.

Appendix E defines citation accuracy as the average of each task's supported-pair proportion; a task without citable statements contributes zero. Average effective citations per task counts supported statement-URL pairs across the benchmark and divides by the task count. Consequently, several distinct supported claims can share a source URL.

The public validation prompt accepts facts found entirely or partially in a reference as supported, including rounded data. It labels statements “unknown” when the reference contains no valid information. The accompanying statistics script excludes unknown judgments and validation errors, then calculates accuracy across pooled pairs. This differs from the paper's per-task averaging and makes the implementation version material to score comparisons.

What does an evaluation submission contain?

The repository's input format stores each generated report with its task identifier and original request. Its example has these fields:

{
    "id": "task_id",
    "prompt": "original_query_text",
    "article": "generated_research_article_with_citations"
}

The example shows placeholder values. Each line of the submitted JSONL file, a file containing a JSON record per line, holds an actual report. The documented workflow reads benchmark requests from data/prompt_data/query.jsonl and saves outputs under data/test_data/raw_data/. RACE and FACT produce separate result files. For an agency inspecting research reports, this separation provides distinct assessments of report quality and source support.

FAQ

Does a supported citation mean the whole statement is verified?

The public validation prompt accepts partial support. When reviewing a result, you therefore need to inspect which facts the cited page supports.

Can citation accuracy show how much evidence a report contains?

The paper reports average effective citations alongside accuracy to measure the amount of supported information. Reading both metrics preserves the distinction between the supported proportion and the supported claim count.

Sources

Check how a page reads to answer engines. Open the GEO score analyzer