# ResearchRubrics ![Category](https://img.shields.io/badge/Category-Deep%20Research-1f6feb) ![Metric](https://img.shields.io/badge/Metric-%25%20Rubric%20Score-2da44e) ![Saturation](https://img.shields.io/badge/Saturation-Low-d29922) ![Created](https://img.shields.io/badge/Created-2026-555) > **Quick answer:** ResearchRubrics evaluates AI-generated research reports against detailed expert-authored rubrics, scoring for completeness, accuracy, and structure rather than a single correct answer. Kimi K3 scores 76.2% as of July 2026. ## At a Glance **What it tests:** The quality and completeness of open-ended research report generation, judged against structured expert rubrics rather than a single ground-truth answer. **Why it matters:** Deep-research tasks rarely have one correct output. Rubric-based grading captures nuanced quality dimensions (coverage, source use, structure) that simple accuracy metrics miss. **Known limitations:** Rubric-based scoring can introduce grader variance depending on how strictly rubric criteria are applied, and exact rubric design is not independently published outside Kimi K3's own report. ## What ResearchRubrics Measures ResearchRubrics evaluates a model's ability to produce comprehensive, well-structured research reports on open-ended topics. Each report is graded against an expert-authored rubric covering dimensions such as factual accuracy, source coverage, argument structure, and completeness, producing an aggregate rubric score rather than a binary correct/incorrect judgment. ## Benchmark Specifications | Field | Value | |---|---| | Task category | Agent / deep research | | Metric | % rubric score | | Saturation | Low | | Created by | Not yet independently documented | ## How ResearchRubrics Is Scored Generated research reports are graded against detailed expert rubrics covering multiple quality dimensions (accuracy, coverage, structure), with an aggregate % rubric score reflecting overall report quality. ## State-of-the-Art Results | Rank | Model | Score | Source | Date | |---|---|---|---|---| | 1 | [Kimi K3](/models/moonshot/kimi-k3) | 76.2% | [Kimi K3 technical report](https://www.kimi.com/blog/kimi-k3) | 2026-07 | *Score sourced from Moonshot AI's Kimi K3 announcement, July 2026.* ## ResearchRubrics on Benchgen > No Benchgen results yet — [be the first to run ResearchRubrics](/benchmarks/platform/researchrubrics). ## ResearchRubrics vs Other Benchmarks | Benchmark | What it tests | Saturation | |---|---|---| | **ResearchRubrics** | Rubric-graded research report quality | Low | | [DeepSearchQA](/benchmarks/platform/deepsearchqa) | Multi-hop web research & synthesis | Low | | [BixBench](/benchmarks/futurehouse/bixbench) | Biology research agent tasks | Low | | [AA-Briefcase](/benchmarks/artificial-analysis/aa-briefcase) | Professional knowledge-work quality | Low | ## Run ResearchRubrics on Your Model Benchgen lets you run ResearchRubrics-style evaluations against your own model, tracking rubric-graded research report quality over time. ## Frequently Asked Questions

What is ResearchRubrics?

ResearchRubrics is a benchmark that grades AI-generated research reports against expert-authored rubrics covering accuracy, coverage, and structure.

What does a good score look like on ResearchRubrics?

Kimi K3 reports 76.2% as of July 2026, a strong result reflecting comprehensive, well-structured research report generation.

Who created ResearchRubrics?

ResearchRubrics' originating team is not yet independently documented outside of its citation in Kimi K3's July 2026 technical report.

ResearchRubrics — Results

ResearchRubrics — Evaluations