STRATA: A Benchmark for Evidence-Grounded Significance Testing of Temporal Hypotheses by Language Models

Forum topic · 小凯 · 2026-10-31

Summary

STRATA is a benchmark from researchers at Stevens Institute of Technology (arXiv:2510.25909) that evaluates whether language models can properly test temporal hypotheses in health data rather than produce fluent but unverified conclusions. It contains 8,735 hypotheses about clinical temporal patterns in vaccine adverse-event reports, each with a gold analysis plan, evidence set, p-value, and verdict derived under a published codebook and reproduced by a deterministic oracle. Unlike prior data-analysis benchmarks that grade only final conclusions with data handed over, STRATA grades every decision on the path to a conclusion: specifying the null hypothesis, selecting the statistical test, and retrieving the relevant records. The authors also propose ART (articulate, retrieve, test), a staged paradigm that exposes each decision as inspectable output. Experiments across five model configurations show systems mainly fail at deciding which records are relevant—a conclusion-evidence gap invisible to accuracy-only evaluation. ART closes much of this gap and brings a 31B open-source model within 0.01 of frontier models under the same paradigm, and frontier models succeed once given the codebook.

Overview

Field: NLP
Authors: Chengyang He, Wenlong Zhang, Koduvayur Subbalshmi (per source: Chengyang He, Wenlong Zhang, Koduvayur Subbalakshmi)
Published: 2025-10-28
arXiv: 2510.25909

Key Points

Ask a language model whether sales tend to dip after a product recall, or whether one side effect tends to follow another after a vaccine, and it answers in a fluent sentence that reads like a finding, with no test behind it. Assertions about how events unfold across many cases fill the Web, and each is a temporal hypothesis that only a significance test over the right records can settle.
The verifier must specify the null, select the test, and decide which records bear on the hypothesis—any of these choices can go wrong while the conclusion still comes out right. Yet benchmarks for language-model data analysis grade only the conclusion, with the relevant data handed over.
The authors study the problem in health, where a wrong verdict is costly, and present STRATA, a benchmark that grades every decision on the path to a conclusion: 8,735 hypotheses about clinical temporal patterns in vaccine adverse-event reports, with gold plans, evidence sets, p-values, and verdicts derived under a published codebook and reproduced by a deterministic oracle.
They also propose ART (articulate, retrieve, test), a staged paradigm that exposes each decision as an inspectable output.
Extensive experiments across five model configurations show that systems primarily fail at deciding which records are relevant—a conclusion-evidence gap that accuracy-only evaluation cannot reveal.
ART bridges much of this gap and brings a 31B open-source model within 0.01 of frontier models under the same paradigm; frontier models succeed once given the codebook.

Full Abstract (English, from source)

Ask a language model whether sales tend to dip after a product recall, or whether one side effect tends to follow another after a vaccine, and it answers in a fluent sentence that reads like a finding, with no test behind it. Assertions about how events unfold across many cases fill the Web, and each is a temporal hypothesis that only a significance test over the right records can settle. The verifier must specify the null, select the test and decide which records bear on the hypothesis, and any of these choices can go wrong while the conclusion still comes out right. Yet benchmarks for language-model data analysis grade only the conclusion, with the relevant data handed over. We study the problem in health, where a wrong verdict is costly, and present STRATA, a benchmark that grades every… (truncated in source; see arXiv page for the full abstract)