DR-256: A Self-Contained Benchmark for Deep-Research Agents

DR-256 is an open, self-contained benchmark for reproducible evaluation of deep-research agents. It pairs 256 curated multi-hop questions derived from DeepResearch-9K with a frozen 3.5 GB web corpus and a local SQLite FTS5/BM25 search environment. This design enables off-the-shelf evaluation against a common evidence base without paid or external search APIs.

DR-256 is introduced in OARS: On Post-Training for Cost-Efficient Heterogeneous Multi-Agent Systems. The evaluation harness and paper settings are available in the OARS code release. In the OARS experiments, DR-256 is used only for evaluation: the worker model is post-trained on BrowseComp-Plus and receives no further training on DR-256.

The canonical release is revision v1.1.0.

Benchmark overview

Item	Value
Evaluation questions	256
Hard questions	128
Challenge questions	128
Corpus URL records	41,913
Successfully fetched URLs	32,127
Unique indexed documents	30,318
FTS chunks	243,615
Documents blocked from retrieval	28
Retrieval	SQLite FTS5/BM25
External search API	Not required

For backward compatibility, the released metadata stores the paper's Hard group under calibration_tier: frontier; challenge corresponds to the paper's Challenge group.

Repository contents

The files required for evaluation are:

benchmark/
  questions.jsonl              # questions, answers, prompts, qids and strata
  blocked_docids.txt           # documents excluded from search and open
  summary.json
search/
  corpus.sqlite3               # frozen web snapshot and FTS5 index
  coverage_summary.json
  source_manifest_summary.json
README.md
TERMS.md

The repository also contains construction and audit artifacts:

selection/                     # candidate filtering and calibration evidence
baselines/                     # fresh reference runs used for protocol checks
audit/                         # crawl, robots-policy and evaluation audits
tools/                         # rebuild and release-validation utilities
manifest.json                  # machine-readable inventory and hashes
checksums.sha256               # checksums for release artifacts
LICENSE-CODE                   # Apache-2.0 license for packaged utilities

These additional artifacts document how the frozen benchmark was produced; they are not required for a standard DR-256 evaluation.

Quick start with OARS

1. Download DR-256

hf download Shangy/DR-256 \
  --repo-type dataset \
  --revision v1.1.0 \
  --local-dir dr256

2. Install the OARS evaluation code

git clone https://github.com/mit-han-lab/oars.git
cd oars
pip install -e ".[search,data]"

3. Prepare the tasks and start local search

oars-data dr256 \
  --input ../dr256/benchmark/questions.jsonl \
  --output data/dr256.jsonl

python -m oars.search.dr256_server \
  --database ../dr256/search/corpus.sqlite3 \
  --blocked-docids ../dr256/benchmark/blocked_docids.txt \
  --host 0.0.0.0 \
  --port 8000 \
  --workers 64

The blocked-document list is part of the benchmark definition and must be enabled during evaluation. The local service exposes:

• GET /health
• GET /stats
• POST /search with {"query": str, "k": 1..100}
• POST /open with {"docid": str} or {"url": str}

The search engine is lexical-only. Documents are divided into 800-word chunks with 100-word overlap; search returns the highest-ranked chunk for each source URL, and /open exposes the corresponding source text.

For complete model-serving, evaluation, grading and summarization commands, see the OARS README.

Benchmark construction

Source questions and retrieval provenance

DR-256 starts from DeepResearch-9K revision 9eaf02da8608b91dc9440503ce10889160c14347. The upstream trajectories provide questions, reference answers, search queries and visited URLs, but not the retrieved page contents. We therefore use their retrieval provenance to build a new, executable local search environment rather than attempting to replay the original observations.

Repeated questions are consolidated by normalized question identity. The resulting construction pool yields 41,913 canonical URLs associated with 6,646 questions.

Frozen local search environment

Source pages are fetched, converted to text and deduplicated. The benchmark retains the full collected snapshot, including distractors, rather than indexing only pages known to contain answers. This yields 30,318 unique indexed documents and 243,615 searchable chunks. Freezing both the corpus and retrieval index keeps the evidence environment stable across evaluation runs.

Validity and answerability checks

Twenty-eight dataset dumps, QA pages and other answer-bearing benchmark artifacts are blocked from both /search and /open. Each retained question must also have an independent source document containing its reference answer; these supporting-document identifiers are used only for construction and audit and are never supplied to the evaluated agent.

Questions solvable from answer-visible snippets, early page openings or a single straightforward retrieval step are removed. The remaining candidates require sustained evidence gathering, document inspection or query reformulation.

Difficulty calibration

Each candidate is evaluated with three independent rollouts of an off-the-shelf Qwen3-8B agent. Retained questions require complete, nonempty and successfully graded rollouts with at least three interaction turns on average.

• Hard: solved in one or two of the three calibration rollouts.
• Challenge: not solved in any of the three calibration rollouts.

Questions solved in all three rollouts are excluded. We sample 128 questions from each group to form the final 256-question benchmark. These labels describe calibration behavior of the reference agent, not an intrinsic or permanent difficulty ordering across models.

Evaluation protocol

The paper's DR-256 preset uses:

Parameter	Value
Temperature	0.6
Top-p	0.95
Top-k	20
Maximum agent turns	20
Per-question seed	20260811 + benchmark_index
Primary denominator	all 256 questions

Empty answers, unfinished episodes, agent errors and missing judge scores count as incorrect. Accuracy should not be reported conditional on successful completion.

Data schema

Each row in benchmark/questions.jsonl contains:

label                         reference answer
prompt[]                      frozen system and user messages
metadata.question             original question
metadata.qid                  stable normalized-question identifier
metadata.difficulty           upstream difficulty (1/2/3)
metadata.teacher_correct      outcome of the released upstream teacher trace
metadata.dataset_rows         source row identifiers
metadata.row_fingerprints     source-row fingerprints
metadata.benchmark.index      0..255
metadata.benchmark.calibration_tier
metadata.benchmark.selection_bucket
metadata.benchmark.base_pilot calibration behavior of the reference agent
metadata.benchmark.clean_provenance_answer_docids

clean_provenance_answer_docids is construction metadata. It must not be provided to an evaluated agent or used as an oracle retriever.

Reproducibility and versioning

Revision v1.1.0 is the canonical OARS release. It repairs a path-specific robots.txt caching issue in v1.0.2, re-audits the affected origins, derives a new SQLite snapshot and rebuilds the FTS index. The older tag remains available as a restore point but should not be used for new evaluations.

Key release identities:

Artifact	Identity
Questions SHA-256	6c1bbdfad639b84224976d4a49d28e1eb060cb0a38f3f7284b9050b6963e7a85
SQLite SHA-256	09172388ed6c0ae91cb43dff3a663bcfb219673ed991ff5c9922904eb61ee77e
SQLite size	3,499,802,624 bytes

To validate the downloaded release:

python ../dr256/tools/frontier_release.py validate --release-dir ../dr256
python ../dr256/tools/frontier_release.py validate --release-dir ../dr256 --deep

The deep check additionally re-hashes the SQLite database and runs PRAGMA quick_check.

Intended use and limitations

DR-256 is intended for:

• reproducible evaluation of deep-research and local-search agents;
• paired comparisons under a fixed evidence environment;
• orchestration, search-budget and termination ablations;
• agent-infrastructure and retrieval-interface validation.

Important limitations:

• The questions are derived from the public DeepResearch-9K training provenance. DR-256 is not an independent held-out DeepResearch-9K test set.
• Question selection uses Qwen3-8B calibration rollouts and therefore reflects that model's behavior.
• The benchmark measures performance against a frozen local corpus, not generalization to the live web or robustness to changing retrieval systems.
• Corpus coverage is shaped by teacher-visited URLs, page availability and crawl policy. Web text may contain errors, bias, personal information or offensive content inherited from its sources.

Rights and takedown

The packaged Python utilities are distributed under Apache License 2.0; see LICENSE-CODE. The pinned upstream DeepResearch-9K repository did not declare a dataset license when this release was built. Extracted webpage text remains subject to the rights and terms of its original sources; this repository does not grant a blanket license over third-party content.

Read TERMS.md before use. For correction or removal requests, open a discussion on this dataset repository and identify the canonical URL or document id.

Acknowledgements

DR-256 is constructed from questions and retrieval provenance released with DeepResearch-9K. We thank its authors for making these resources available.

Citation

Please cite OARS when using DR-256. To make results reproducible, also record the dataset revision, question hash and SQLite hash.

@article{yang2026oars,
  title  = {{OARS}: On Post-Training for Cost-Efficient Heterogeneous Multi-Agent Systems},
  author = {Yang, Shang and Zeng, Zhichen and Dong, Shi and Wang, Yichuan and Chen, Siqi and
            Guo, Junxian and Gonzalez, Joseph E. and Zhu, Banghua and Han, Song},
  year   = {2026}
}