AgentHop

AgentHop is a diagnostic benchmark for multi-step scientific question answering: 1,011 four-option multiple-choice questions grounded in citation chains over 7,205 arXiv papers from nine major computer-science venues (2022–2025). Each question is answerable only by reading specific sections of one or two of these papers; the agent reaches the answer-bearing papers by following references from a given seed paper.

The dataset is paired with a four-axis decomposition framework (search, synthesis, tool-use pattern, resource management) that turns a single evaluation run into per-model failure-mode attribution rather than a single accuracy number.

Dataset Statistics

Statistic	Value
Total items	1,011
Single-target (ST)	568
Multi-target (MT)	443
Source-paper pool	7,205 papers
Source citation edges	>450,000
Source venues	9 (NeurIPS, ICML, ICLR, ACL, EMNLP, NAACL, CVPR, ECCV, SIGIR)
Publication years	2022–2025
Reasoning-type counts	GROUND 330, RESULT 292, MOTIVE 265, METHOD 124
Consensus tiers	Gold / Silver / Bronze (filter-agreement signal at construction time)
Depths	1-hop and 2-hop

Files

AgentHop/
├── qa/full.jsonl                  # 1,011 MCQ items (schema below)
├── graphs/full.jsonl              # per-sample citation neighbourhoods (nodes + edges)
├── paper_pool/papers.jsonl        # deduplicated section text for the 7,205-paper pool
├── audit/recall_labels.jsonl      # auditor-labelled answer-bearing sections per sample
├── loader.py                      # reference Python loader
├── LICENSE                        # CC-BY 4.0
└── README.md                      # this file

Schema (qa/full.jsonl)

Each line is one item.

Field	Type	Description
id	string	Item identifier (e.g., st_0001, mt_0042)
question	string	The question stem
options	list of 4 strings	The four answer options
correct_index	int (0–3)	Index of the correct option in options
question_type	string	single-target or multi-target (a.k.a. target multiplicity)
depth	int	Maximum hop distance from seed to gold paper, 1 or 2
reasoning_type	string	GROUND / METHOD / MOTIVE / RESULT
cognitive_skill	string	Construction-stage taxonomy label (e.g., retrieve)
consensus_tier	string	gold / silver / bronze — filter agreement at Stages 5–6
filter_agreement	int	Numeric agreement score
seed_arxiv_id	string	arXiv ID of the seed paper the agent starts from
seed_paper_id	string	Semantic Scholar hash of the seed paper
seed_title	string	Seed paper title
venue	string	Source venue (e.g., NeurIPS, ICML)
gold_arxiv_ids	list of strings	arXiv IDs of the gold papers (1 for ST, 2 for MT)
gold_paper_ids	list of strings	Semantic Scholar hashes of the gold papers
bridge_arxiv_ids	list of strings	arXiv IDs of intermediate hop papers (depth-2 only)
bridge_paper_ids	list of strings	Semantic Scholar hashes of the bridge papers
distractor_types	list of strings	Engineered distractor labels: seed_only, wrong_paper, no_context, plus correct

Schema (graphs/full.jsonl)

Each line is one sample's citation neighbourhood used as the agent's navigation pool.

Field	Type	Description
sample_id	string	Matches qa/full.jsonl::id
seed_paper_id	string	Semantic Scholar hash of the seed paper
nodes	object	{paper_id: {arxivId, title, year, authors, abstract, ...}}
edges	object	{paper_id: [cited_paper_ids]}

Schema (paper_pool/papers.jsonl)

Deduplicated section text for every paper appearing in any sample's citation neighbourhood.

Field	Type	Description
paper_id	string	Semantic Scholar hash
arxiv_id	string	arXiv ID
title	string
year	int
authors	list
abstract	string
sections	list of {header, text}	Section-structured body, parsed from arXiv via ar5iv

Schema (audit/recall_labels.jsonl)

Auditor-labelled answer-bearing sections, used by the evaluation harness to compute the search-axis recall metric.

Field	Type	Description
sample_id	string	Matches qa/full.jsonl::id
analysis	object	{section_recall_labels: [{arxiv_id, section}, ...], answer_validity, chain_coherence, query_quality, structural_integrity, synthesis_check}

Quick start

from loader import load_agenthop, load_recall_labels

samples = load_agenthop("path/to/AgentHop")
labels  = load_recall_labels("path/to/AgentHop")

print(len(samples))                 # 1011
print(samples[0]["question"])
print(samples[0]["options"])
print(samples[0]["correct_index"])
print(samples[0]["paper_pool"])     # papers reachable from seed for this sample
print(labels[samples[0]["id"]]["analysis"]["section_recall_labels"])

To evaluate a model end-to-end (with the seven-tool sandbox, four-axis metrics, and resource caps), use the evaluation harness from the code repo (link in the paper supplementary).

Intended Use

AgentHop is a measurement instrument for multi-step QA agents: locate relevant papers via citation following, read the right sections of those papers, combine the resulting evidence, and do so within a deployment-realistic resource budget. The headline accuracy is intended to be paired with the four-axis decomposition (search recall, conversion rate, tool-use pattern, resource management) that the harness produces. Aggregate accuracy alone collapses qualitatively different failure modes into a single number; the per-axis attribution surfaces which sub-ability is binding for each model.

Construction Pipeline

AgentHop is constructed via an eight-stage pipeline: four generation stages (seed selection → citation-chain expansion → corpus collection → Q/A generation → distractor generation) followed by four filtering stages (model-ensemble filtering → data sanity check → seven-auditor human review → post-audit triage). Full details are in the paper's appendix.

Limitations

• Single domain. Computer-science papers from nine major venues (2022–2025). Behaviour on other fields, on long-tail venues, and on open-ended response formats is not measured.
• Multiple-choice format. Each item is a four-option MCQ, not a free-form generation task.
• English-language source corpus. All papers and questions are in English.
• Generator–evaluator overlap. GPT-5.4 generated the items and is also among the evaluated models. The effect is bounded but real and acknowledged in the paper's Limitations section.
• Parametric coverage. The strongest checkpoints already resolve a substantial fraction of items closed-book (Gemini-3 Pro reaches 0.734 without tools); the four-axis decomposition is most informative where parametric coverage is sparse.

Ethical Considerations

• Built from public arXiv papers under their respective deposit licenses. We use only paper identifiers, abstracts, and named-section text; we do not reproduce figures, tables, or any non-textual content.
• No personal information, no harmful or offensive content, and no generative model weights are released.
• Seven auditors (members of the research team, participating voluntarily) reviewed audit-prep items during the construction pipeline. No external paid annotation was conducted; no personal information of auditors is collected or released.

Citation

@inproceedings{park2026agenthop,
  title     = {AgentHop: A Diagnostic Benchmark for Multi-Step Scientific Question Answering},
  author    = {Park, Chanhee and Yoon, Jeongho and Han, Sungbin and Moon, Hyeonseok and Lim, Heuiseok},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS), Datasets and Benchmarks Track},
  year      = {2026}
}