FinHopper

A Progressive Multi-Hop Benchmark for Complex Financial Agents with Step-wise Evaluation

📄 Paper (arXiv:2610.26036) · 💻 Code & evaluation · Accepted to EMNLP 2026

---

Why this benchmark

Financial agents are expected to answer factual investment queries through open search, but financial multi-hop retrieval differs from general multi-hop QA in a crucial way: each intermediate answer is an exact typed constraint — an entity, a date, an event, or a numerical indicator — that determines the next query.

That strict answer-conditioned dependency means a small error in hop 2 does not stay local. It cascades into a search trajectory that looks plausible and ends in the wrong answer. Final-answer-only evaluation cannot see this happening, and cannot give credit for the hops that were right.

FinHopper is built to expose exactly that failure mode:

• Controlled reasoning depth, 1 → 6 hops, so you can measure degradation as a curve rather than a single score.
• Gold intermediate answers on every item, enabling process-level (step-wise) scoring rather than binary end-state grading.
• Parallel English and Chinese sets with identical IDs and hop strata, so cross-lingual transfer is measurable without re-collecting data.

Dataset at a glance

400 items total — 200 English, 200 Chinese, 1:1 matched by ID and difficulty stratum.

Level	Hops	English	Chinese
L1	1	50	50
L1	2	50	50
L2	3	39	39
L2	4	21	21
L3	5	19	19
L3	6	21	21
Total	1–6	200	200

Coverage spans stocks, funds, and macroeconomic facts. Items were built with a state-transition-based synthesis framework that composes verified atomic financial QA patterns into dependency-preserving chains, then filtered through in-synthesis verification, model-ensemble checking, and public-search validation by financial experts.

Fields

Field	Type	Description
prompt_id	string	Unique ID, encoding level, hop depth and index — e.g. L2_3hop_7. IDs are shared across the en and zh configs.
prompt	string	The question put to the agent.
reference_response	string	Gold reference answer with intermediate values spelled out. This is what the process-level judge scores against.
response_final_reference	string	Gold final answer only, used by the final-answer judge.
label	string	Difficulty stratum: L1_1hop, L1_2hop, L2_3hop, L2_4hop, L3_5hop, L3_6hop.
judge_prompt_template	string	Template that assembles the judge input from {prompt}, {response_reference} and {response}. Two whitespace variants exist in the data.
final_judge_system_prompt	string	System prompt for the final-answer judge. Constant within a config.
process_judge_system_prompt	string	System prompt for the step-wise process judge, including the hard filters applied before any score is computed. Constant within a config.

The two *_system_prompt columns are repeated on every row to keep the dataset self-contained and round-trippable with the GitHub release. Parquet conversion on the Hub compresses the repetition away; if you prefer a leaner frame, just drop them after loading.

Quick start

from datasets import load_dataset

en = load_dataset("justajustin/FinHopper", "en", split="test")
zh = load_dataset("justajustin/FinHopper", "zh", split="test")

print(en[0]["prompt"][:200])
print(en[0]["response_final_reference"])

Without an explicit config name, the en config is loaded by default:

ds = load_dataset("justajustin/FinHopper")   # -> config "en"

Stratify by hop depth:

from collections import Counter
print(Counter(en["label"]))   # {'L1_1hop': 50, 'L1_2hop': 50, ...}
subset = en.filter(lambda x: x["label"] == "L3_6hop")

Evaluating an agent

Both judges are LLM-as-a-judge prompts with hard filters applied before scoring, so a model cannot earn points for a fluent answer that is unsupported by evidence.

import re
row = en[0]

def fill(template, question, reference, student):
    return (template
            .replace("{prompt}", question)
            .replace("{response_reference}", reference)
            .replace("{response}", student))

# final-answer score
messages = [
    {"role": "system", "content": row["final_judge_system_prompt"]},
    {"role": "user",   "content": fill(row["judge_prompt_template"],
                                       row["prompt"],
                                       row["response_final_reference"],
                                       agent_output)},
]

# process-level score (uses the full reference response with intermediate hops)
messages = [
    {"role": "system", "content": row["process_judge_system_prompt"]},
    {"role": "user",   "content": fill(row["judge_prompt_template"],
                                       row["prompt"],
                                       row["reference_response"],
                                       agent_output)},
]

The reference implementation, retrieval corpus tooling and all reported baselines (20 LLM-agent configurations) live in the GitHub repository.

Intended use

Built for evaluating retrieval-augmented LLM agents on financial fact chains. It is a research benchmark, not a source of investment advice — every item is a factual lookup with a verifiable answer, and no forward-looking recommendation is included or implied.

Answers are anchored to specific report dates and public filings. Values that look "wrong" today are usually correct as of the date named in the question.

中文说明

FinHopper 是面向金融智能体的渐进式多跳基准，共 400 条：英文 200 条、中文 200 条，两语种按 prompt_id 与难度分层一一对应。推理深度覆盖 1–6 跳，每条都带金标中间答案，因此可以做过程级（step-wise）评分，而不只是看最终答案对错。数据涵盖股票、基金与宏观事实。

from datasets import load_dataset
ds = load_dataset("justajustin/FinHopper", "zh", split="test")

Citation

@misc{wu2026finhopper,
  title  = {FinHopper: A Progressive Multi-Hop Benchmark for Complex Financial Agents with Step-wise Evaluation},
  author = {Yuxuan Wu and Miaopei Lin and Qiufang Ying and Lin Dan and Hao Zhao and Qirui Liu and Yuteng Shen and Shaohui Wu and Weihong Luo and Xiku Du and Xing Tang and Hao Chen},
  year   = {2026},
  eprint = {2610.26036},
  archivePrefix = {arXiv},
  primaryClass  = {cs.LG},
  url    = {https://arxiv.org/abs/2610.26036}
}

License

Released under CC BY 4.0. Questions and synthesized answers are the authors' contribution; underlying market figures are public facts. Please cite the paper if you use this benchmark.