OpenHypBench
Open-Ended Multiple Hypothesis Discovery Benchmark
Evaluates open-ended hypothesis discovery from large tabular datasets using 351 planted hypotheses across 15 tasks and 5 domains.
How to run
$ npm i -g lettokenburn-cli$ ltb bench info openhypbench
External harness. No Harbor adapter has been generated for this benchmark yet. `ltb bench info` prints the known links.
Runs happen on your machine with your own subscription; only the score summary is pushed. CLI docs
Scores
6 sourced entries| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| HypAgent | 0.5 | N/A | Independent | medium | Oct 13, 2026 | arxiv.org |
| Claude Code | 0.3 | N/A | Independent | medium | Oct 13, 2026 | arxiv.org |
| XGBoost+LLM | 0.3 | N/A | Independent | medium | Oct 13, 2026 | arxiv.org |
| AutoDiscovery | 0.3 | N/A | Independent | medium | Oct 13, 2026 | arxiv.org |
| LLM | 0.2 | N/A | Independent | medium | Oct 13, 2026 | arxiv.org |
| ReAct | 0.1 | react | Independent | medium | Oct 13, 2026 | arxiv.org |
Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.
Reproduced by the community
All runsAbout
Built from structural causal models grounded in real-world schemas, containing 15 tasks with 53–65 variables and 1M–10M rows each. The underlying causal model is hidden, requiring agents to discover mechanisms without prior guidance.
Scoring uses interventional validity testing against planted hypotheses. HypAgent achieves 0.49 F1, outperforming baselines (0.35 F1), highlighting challenges in characterizing functional forms beyond variable identification.