MortarBench

A benchmark for evaluating LLM agents on mortgage loan origination. Each item gives an agent two synthetic source documents - a bank statement (Plaid-format JSON) and a completed loan application (ULAD/MISMO XML) - and asks an underwriting question that can only be answered by cross-referencing them.

Paper: https://arxiv.org/abs/2606.19416 OpenReview: https://openreview.net/forum?id=EMdAVOU3Rr Code: https://github.com/mtoles/MortarBench

Why it is hard

The questions are the ones a human underwriter actually asks, and most require reconciling the two documents against each other rather than reading either one alone - for example, checking whether recurring debits in the bank statement correspond to liabilities the borrower disclosed on the application, or whether a deposit matches a gift the application claims. Answers are grounded: an agent must return the specific transaction ids that justify its conclusion, not a free-text summary, so a correct answer cannot be bluffed.

Contents

90 questions over 9 synthetic loan files (582 transactions across 26 accounts).

Config	Rows	Description
default	90	One row per question, with its documents inlined. Self-contained.
test_cases	9	One row per loan file, if you prefer to join yourself.

default fields

Field	Type	Description
question_id	string	"<test_case>-<n>", e.g. "1-3"
loan_id	string	Loan identifier, one per test case
test_case_number	int	Which loan file this question is about
question	string	The underwriting question
answer	string	Gold answer (see below)
answer_type	string	txn_id_list (57), boolean (26), account_id_list (7)
pii	bool	True if answering requires reading identity fields (7 questions)
bank_statement	string	JSON-encoded Plaid payload - call json.loads
ulad_du	string	ULAD/MISMO loan application, XML

Answer format

• boolean - "yes" or "no".
• txn_id_list - comma-separated transaction ids ("plaid-1-00011, plaid-1-00030"), or "none" when nothing matches. "none" is a real answer, not a missing value: several questions test whether a model resists inventing evidence.
• account_id_list - comma-separated last-4 account digits ("9921"), or "none".

Scored by exact set match, order-insensitive.

Usage

import json
from datasets import load_dataset

ds = load_dataset("ManavMunjal/MortarBench", split="test")

row = ds[0]
statement = json.loads(row["bank_statement"])   # nested JSON, see note below
application = row["ulad_du"]                    # XML string

print(row["question"], "->", row["answer"])

bank_statement is stored as a JSON string rather than a nested struct on purpose: the payloads have optional and heterogeneous fields that Arrow cannot type consistently, so encoding them keeps the data lossless.

Reading a bank statement

Two payload shapes exist. Seven test cases are a single flat payload; test cases 4 and 9 bundle several payloads under plaid_files because those borrowers have accounts at more than one institution. This helper covers both:

def iter_accounts(statement):
    if "plaid_files" in statement:
        for payload in statement["plaid_files"]:
            yield from payload.get("override_accounts", [])
    else:
        yield from statement.get("override_accounts", [])

for account in iter_accounts(statement):
    for txn in account.get("transactions", []):
        print(txn["transaction_id"], txn["date_posted"], txn["amount"], txn["description"])

Reported results

From the paper: state-of-the-art models perform poorly, with closed-source models reaching at most 77.1% exact match. The paper also reports systematic bias in LLM perception of foreignness related to non-English names. CRIT, a confidence calibration framework introduced alongside the benchmark, raises accuracy to 80.5% while improving risk-management steering and reducing that bias.

The evaluation harness that reproduces these numbers, along with the data synthesis and mutation pipeline used to build and stress the benchmark, lives in the GitHub repo.

Provenance

raw/ holds the unmodified source documents behind the JSONL files - the PDF bank statements as a human would receive them, the original ulad.xml applications, the Plaid JSON, and the annotation spreadsheet. Included so the gold labels can be audited against the documents they came from.

Synthetic data and privacy

Every record is synthetic. No real borrower, account, or institution is represented. Names (John Homeowner, Andy America), addresses, emails, account numbers, and routing numbers are fabricated, and the taxpayer identifiers in the ULAD files use the 999 prefix that the U.S. Social Security Administration has never issued. The pii flag marks questions whose reasoning touches these identity fields; it does not indicate real personal data.

Limitations

• 9 loan files is a small document pool - per-test-case scores are noisy, and results should be read at the aggregate level.
• Documents are U.S.-specific (ULAD/MISMO, Plaid) and English-only.
• Synthetic statements are cleaner than production data: no OCR noise, no malformed records, no missing months.
• Gold answers reflect the annotators' underwriting judgment. Some questions ("large deposits") depend on thresholds that vary between lenders in practice.

Citation

@misc{mortarbench,
  title         = {MortarBench: Evaluating Mortgage Loan Origination Agents},
  author        = {Toles, Matthew and Lu, Yunan and Munjal, Manav and Liu, Bojun
                   and Deng, Yuanhao and Selig, Stephanie and Rindner, Derek and
                   Li, Cheng and Yu, Zhou},
  year          = {2026},
  eprint        = {2606.19416},
  archivePrefix = {arXiv},
  primaryClass  = {cs.LG},
  url           = {https://arxiv.org/abs/2606.19416}
}