8000
Skip to content

Repository files navigation

Scientific Question Backtesting Benchmark

A model-agnostic benchmark for evaluating whether AI-generated scientific questions anticipate future scientific progress.

The benchmark freezes questions using literature available before a historical cutoff and evaluates them against an isolated future corpus.

Core idea

Past literature → Generated questions → Frozen evaluation →
Future evidence → Outcome labels → Benchmark metrics

A scientific question is valuable if future science actually invests in answering it. Historical backtesting turns that idea into a reproducible measurement: any system generates questions from a pre-cutoff corpus, freezes them, and is scored against what the scientific community actually did afterwards — no expert panels, no self-grading LLM scores.

Astronomy v1

  • Historical cutoff: 2020-12-31
  • Past corpus: 2,512 papers (2015–2020, NASA ADS; 500 full-text core)
  • Future corpus: 1,891 papers (2021–2026, temporally isolated)
  • Questions: 10 (frozen, from the evidence-graph system)
  • Addressed by future literature: 10/10
  • Answered: 2 · Partially addressed: 7 · Posed but still open: 1
  • Premise refuted: 1 (HD 209458 b subsolar water — see the paper's case study)

Astronomy v1 is the first benchmark instance, not the definition of the benchmark itself.

Quickstart (offline, no API keys)

pip install -e .
pytest                                                   # 24 offline tests
python -m benchmark.run --config configs/astronomy_2020.yaml

This validates schemas and temporal isolation, recomputes results/astronomy_v1/{metrics.json, per_question_results.csv, report.md}, and refreshes results/leaderboard.json — bit-identical to the committed files (CI enforces it).

Step by step instead:

python scripts/validate_cutoff.py \
  --questions data/questions/astronomy_questions_v1.jsonl --cutoff 2020-12-31
python scripts/compute_metrics.py \
  --annotations data/annotations/outcomes_v1.jsonl

Live stages (rebuilding corpora, retrieval, judging) need API keys — see .env.example and docs/reproducibility.md.

Evaluate your own system

The benchmark does not generate questions; it evaluates frozen question lists from any system:

  1. Generate questions from pre-cutoff evidence only.
  2. Freeze them in the submission format under submissions/<your_system>/.
  3. python scripts/validate_cutoff.py --questions ... --cutoff 2020-12-31
  4. Run retrieval + judging (scripts/retrieve_future_evidence.py, scripts/judge_outcomes.py), get labels adjudicated, compute metrics.

Reference baselines (random claims, direct LLM, review future-work, citation leaders) live in baselines/.

Repository map

Path Contents
benchmark/ protocol implementation: schemas, leakage checks, retrieval, judging, metrics, reporting
configs/ benchmark instance definitions (astronomy_2020.yaml), outcome taxonomy
data/ frozen questions, corpus manifests, retrieval records, adjudicated annotations
docs/ protocol · taxonomy · annotation · leakage · submissions · reproducibility
baselines/ four reference question generators
scripts/ CLI: validate → prepare corpus → retrieve → judge → metrics → report
submissions/ evaluated systems (evidence_graph_v1 = Paper 1's system)
results/ released metrics, per-question results, reports, leaderboard
tests/ + .github/workflows/ offline tests, data validation, reproducibility smoke test
paper/ the benchmark paper (Paper 2)

Relationship to the question-generation repo (Paper 1)

scientific-question-discovery is how to generate questions (evidence graph, tension detection, ranking). This repository is how to evaluate them. The generation system appears here only as one submission, submissions/evidence_graph_v1 — like any future system.

Roadmap

  • Public leaderboard with baseline runs (random claims, direct LLM, review future-work, citation leaders) on astronomy v1
  • Full top-8 retrieval lists with per-document scores (v1.1 data release)
  • Additional domain instances and cutoffs; lead-time estimation from community first-posed dates

Citation

See CITATION.cff. License: Apache-2.0.

About

A model-agnostic historical backtesting benchmark for evaluating AI-generated scientific questions against future scientific progress.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages

0