huiluckylucky/scientific-question-outcomes

Scientific Question Outcomes

980 astronomy research questions, frozen at five historical cutoffs, each labelled with what the following five years of literature actually did with it.

Systems that propose research questions are usually evaluated by asking a person or a model how good the questions sound. This dataset supplies the alternative: questions frozen using only pre-cutoff literature, and outcome labels drawn from the literature published afterwards. It is, to our knowledge, the only public dataset that attaches future-outcome labels to questions — rather than to papers — across multiple historical cutoffs and with controlled generation sources.

What is in it

Each question was generated from the arXiv astro-ph record available strictly before its cutoff (2012, 2014, 2016, 2018, or 2020), across eight subfields, from four distinct sources:

Source	n	Addressed within 5 years
Evidence-tension mining	374	0.47
Direct LLM elicitation	200	0.88
Author-stated future work	196	0.51
Weak/negative controls	200	0.32
Prior evidence-graph system	10	0.20

The negative controls are deliberately weak questions — random claim pairings, vague questions, untestable questions, evidence-free contrarian questions — phrased by the same model so that surface style cannot separate them from real candidates. They exist so that any engagement metric computed on this data has a priced chance level. Without them, a model can only learn that question-shaped sentences succeed.

The direct-LLM rate of 0.88 should not be read as generator quality. A model asked in hindsight for the important open questions of 2016 may simply recall what 2017–2021 worked on; this is a documented leakage channel, not a result.

Configurations

default — 980 rows. Question text, provenance, a cutoff-time feature record in six groups, and the outcome label.

structure — 374 rows, the tension-mined subset. Structural properties of the evidence cluster each question was generated from: cluster breadth, tension typology, quantification, instrument structure, and cross-cutoff recurrence.

second_judge — 200 rows. An independent second model's labels on all premise-refuted cases plus a 20% random sample, for label-reliability work.

Fields

Question and provenance — question_id, question, subfield, cutoff_year, cutoff_date, source, control_subtype, source_material (verbatim generating evidence: sentences, arXiv ids, dates).

Features, all computed at cutoff time, by prefix:

Prefix	Group	Examples
txt_	text structure	entity count, causal/comparative language, specificity
ev_	evidence context	similarity mass, instrument diversity, tension density
nov_	novelty	nearest-neighbour distance, concept-pair novelty, review overlap
fal_	falsifiability	observable, competing hypotheses, quantitative test
tra_	tractability	operational vs. future facilities, archival data
env_	field environment	publication volume, growth, next-facility proximity
str_	evidence structure	cluster breadth, tension type, recurrence (structure config)

Labels — addressed (bool), answer_status (answered / partially_addressed / posed_but_open / premise_refuted / not_addressed), premise_status, future_paper_count, first_follow_up_date, lead_time_months, review_recognition, supporting_ids, rationale, confidence, judge_model.

How the labels were made

Three tiers. First, automatic retrieval of future papers from each question's subfield over its five-year future window. Second, a strict judge that sees the question and its top-8 retrieved future records but never the generation source; it classifies every candidate as direct, adjacent, or unrelated before assigning an outcome, and addressed requires at least one direct candidate. Third, an independent second model re-judges all premise-refuted cases plus a 20% random sample.

The strictness matters. A first-pass lenient judge called 94% of questions addressed by accepting topical overlap; the per-candidate rubric brings it to 52.7% and restores variance to every downstream analysis.

Limitations — please read before using

• Labels are rater-relative. Second-judge agreement on addressed is 63.5% (κ = 0.21). The companion protocol paper's seven-rater study found two independent human annotators agreed at only κ = 0.17 on an analogous taxonomy, while frontier models agreed with each other at κ = 0.60 — so certifying such a judge by model–model agreement overstates reliability roughly threefold. Absolute rates should never be quoted without the rater attached; comparisons under a fixed judge are the currency this data supports.
• not_addressed is a lower bound. Retrieval is abstract-level TF-IDF and can miss engagement, especially for narrow object-level questions.
• Parametric leakage is present in one source. See the direct-LLM note above. Source is a covariate in every analysis in the accompanying paper.
• Rare outcomes are rare. 8 answered and 4 premise-refuted in total; the five-way status taxonomy is underpowered at this n.
• One domain. Astronomy only. Cross-subfield transfer is demonstrated; cross-science generality is not.

Loading

from datasets import load_dataset

ds = load_dataset("huiluckylucky/scientific-question-outcomes")
structure = load_dataset("huiluckylucky/scientific-question-outcomes", "structure")

Papers and code

• Question discovery framework — arXiv:2608.09968
• Backtesting protocol and benchmark — arXiv:2608.16795
• This dataset and its predictor analysis — paper pending; pipeline, judge rationales, and all results at github.com/nonameisready/scientific-question-outcome-prediction-indicators

Citation

@misc{mao2026questionoutcomes,
  author = {Hui Mao},
  title  = {Scientific Question Outcomes: 980 Research Questions with
            Five-Year Historical Outcome Labels},
  year   = {2026},
  url    = {https://huggingface.co/datasets/huiluckylucky/scientific-question-outcomes}
}

License: Apache-2.0.