FinAutoRubric

Paper | Website | Code

FinAutoRubric Benchmark evaluates financial research agents on questions derived from in-house analyst workflows. Its rubrics are built with FinAutoRubric, which combines reusable expert guidance with a Task Bank of task-level criteria.

For each query, the framework separates the underlying task from its entities, periods and query date. A writer agent researches expected facts and values using the relevant task criteria, while a reviewer from a different model family verifies the evidence and calculations. The agents refine the rubric through iterative review, and code enforces validation and point-allocation rules.

This release contains 100 questions and 4,980 weighted rubric criteria, spanning 78 primary tasks across eight asset classes. Half of the questions concern public equity; the other half cover a range of credit, rates, currency, commodity, fund and portfolio research.

Criteria per question: 4 minimum · 43 median · 225 maximum.

Task coverage

Each question has one primary task and one asset-class label.

Asset class	Questions
Public equity	50
Private capital	4
Corporate credit	11
Sovereign & rates	9
FX	4
Commodities	10
Funds & indices	7
Multi-asset & macro	5

Files

• finautorubric_public.jsonl: 100 questions with information cutoffs and weighted grading rubrics, one record per line.

Dataset Schema

Field	Type	Description
query_id	string	Unique question identifier, prefixed with FAR_.
query	string	The financial-analysis question.
information_cutoff	date	The query date, in YYYY-MM-DD format. The answer should reflect information available as of that date.
rubric	list of objects	The weighted grading criteria for the question.

Each entry in rubric contains:

Field	Type	Description
criterion_id	string	Criterion identifier, unique within the question.
points	number	Points awarded when the criterion passes.
rubric_text	string	The requirement, including applicable keys, tolerances and accepted alternatives.

A judge evaluates each criterion against the response, awarding its assigned points if it passes and zero otherwise. Each question totals 100 points; the benchmark score is the mean across all 100 questions. Keep the rubric separate from the answering system and use it only for judging.

Code repository: GitHub.

Example

The excerpt below contains one complete question and four selected criteria from its 20-criterion rubric. It illustrates factual evidence, comparative judgment, comparability of financial measures and a quantified downside condition. Criterion text, IDs and weights are unchanged; the complete rubric in finautorubric_public.jsonl totals 100 points.

{
  "query_id": "FAR_9daff3e63d4d0bc0",
  "query": "American Express or Capital One: suppose consumer credit losses climb even as funding costs fall. Working through billed business or loan growth, net interest margin, and credit losses, decide whose cash-flow model is more defensible and quantify a downside or break-even condition, drawing on nothing published after 2024-12-31.",
  "information_cutoff": "2024-12-31",
  "rubric": [
    {
      "criterion_id": "R002:T0008",
      "points": 4.6,
      "rubric_text": "Contrasts the two revenue mixes with third-quarter 2024 figures: net interest income was about 24% of American Express's net revenue (USD 4.006bn of USD 16.636bn; non-interest revenue about 76%) versus about 81% of Capital One's (USD 8.076bn of USD 10.014bn). Accept each share within 1 percentage point, expressed as the net-interest or the non-interest share."
    },
    {
      "criterion_id": "R003:T0012:JT",
      "points": 24.5,
      "rubric_text": "Reaches an explicit conclusion, choosing between the two issuers, on whose cash-flow model is more defensible when consumer credit losses climb while funding costs fall, supported by disclosed growth, margin-or-yield and credit-loss evidence. American Express passes when argued from fee-driven revenue diversification and a lower provision burden outweighing its faster loan growth; Capital One passes if the answer shows its larger funding-cost benefit and loan economics offset its higher losses; a qualified tie passes if measurement differences are shown to prevent a reliable ordering. A pretax cash-earnings proxy (net revenue less operating expense and provisions) is acceptable. Fails if it reverses disclosed revenue, provision, loan or credit-loss figures."
    },
    {
      "criterion_id": "R003:T0012:JE02",
      "points": 2.2,
      "rubric_text": "Qualifies that the two yield measures are not the same: American Express reports a non-GAAP net interest yield on Card Member loans while Capital One reports consolidated net interest margin on interest-earning assets, so they are used directionally rather than as a direct level ranking. Fails if the two measures are treated as identical without qualification."
    },
    {
      "criterion_id": "R004:T0013",
      "points": 16.5,
      "rubric_text": "Quantifies a downside or break-even condition: states a numerical threshold, the metric and period it applies to, and which issuer is favoured on each side of it. Any formulation passes, for example the extra quarterly credit losses at which the two issuers' stressed pretax earnings are equal (about USD 866m more at American Express with a 25 bp deposit-cost fall and Q3 2024 earnings held fixed, about USD 508m with 100 bp), or the charge-off rate that would exhaust an issuer's pre-provision earnings."
    }
  ]
}

Licensing

The original dataset and documentation are licensed under CC BY 4.0, which permits reuse and adaptation with attribution. This license does not cover referenced third-party materials.

Maintenance

Maintained by Linq. Contact: hoyounglee@linqalpha.com.

Citation

Please cite the FinAutoRubric paper, credit the dataset and indicate any modifications:

@misc{lee2026finautorubric,
  title = {{FinAutoRubric}: Expert-Guided Automatic Rubric Generation for Evaluating Financial Research Agents},
  author = {Hoyoung Lee and Suyeol Yun and Jack Haverty and Yunju Cho and Meesong Kim and Daekyung Park and Sumin Kim and Jihoon Kwon and Jasmine Jia Geng and Andrew Chin and Yin Luo and Edward Tong and Yu Yu and Zach Golkhou and Minkyu Kim and Igor Halperin and Young Cha and Alejandro Lopez-Lira and Chanyeol Choi and Yongjae Lee},
  year = {2026},
  eprint = {2609.35744},
  archivePrefix = {arXiv},
  primaryClass = {cs.AI},
  url = {https://arxiv.org/abs/2609.35744}
}