HLER runs an empirical economics paper in 25 minutes for about a dollar, with two human gates
Agents audit the data, propose questions the data can answer, estimate, draft and referee. A person picks the question and approves the paper.
Chen Zhu and Xiaolu Wang of China Agricultural University posted HLER, for Human-in-the-Loop Economic Research, on arXiv on 8 March 2026. It is written in Python. Every reasoning step calls Claude Sonnet 4.6 through Anthropic’s API, and the regressions are run by the statsmodels and linearmodels libraries, so the estimates are computed by code and the model only plans and writes.
Seven steps in sequence, and a person decides twice
| Step | Agent | It does this |
|---|---|---|
| 1 | DataAuditAgent | Reads headers and schema and lists the variables that exist |
| 2 | DataProfilingAgent | Computes summary statistics, missingness and correlations, and flags endogeneity risks |
| 3 | QuestionAgent | Proposes research questions limited to what the audit and profile support, then scores them for feasibility |
| Gate 1 | Researcher | Picks one question, or asks for a new batch |
| 4 | Data collection agents | Pull from public APIs such as the World Bank and FRED, and from local files |
| 5 | EconometricsAgent | Plans and runs OLS, fixed-effects panels, difference-in-differences or an event study |
| 6 | PaperAgent | Writes the abstract, introduction, empirical strategy, results and discussion |
| 7 | ReviewerAgent | Scores the draft from 1 to 10 on five criteria and sends it back for new analysis |
| Gate 2 | Researcher | Decides whether the paper is credible enough to finish |
The first three steps carry the design. A model asked for ideas with only a dataset name produced a feasible question 41% of the time. Shown the audit and the profile first, it reached 87%: 69 of 79 questions against 34 of 82.
One run on a Chinese household survey went from 285 variables to a 7,282-word paper
- Audit and profile. The agents found 285 variables and 57,203 observations in the China Health and Nutrition Survey, 1989 to 2011, and flagged income as missing for 23% of observations.
- Question. The QuestionAgent proposed eight questions. Seven were feasible. The researcher picked the top-ranked one: “Does higher education attainment reduce the occupational gender gap among rural women in China?”
- Estimate. The EconometricsAgent built a panel of 19,466 observations and ran a fixed-effects model on within-person variation.
- Draft and referee. The first draft ran to 5,563 words. The ReviewerAgent raised reverse causality and asked for event-study specifications and sensitivity analyses.
- Revise. After three rounds the paper was 7,282 words and the referee’s score had moved from 4.6 to 6.5.
The referee is the model that wrote the draft
Across the 12 completed runs the reviewer’s mean score rose from 4.8 on first drafts to 6.3 on final ones. Both the draft and the score come from Claude Sonnet 4.6. Project APE, which runs the same kind of loop at scale, sends review to models from other companies for that reason.
The two failed runs stopped at estimation, on fixed-effects models that would not converge on sparse subsamples, and the pipeline halted with the error logged. The test covers three datasets, none of them financial: the Chinese survey, a historical population panel from Liaoning and UK Biobank. The paper tests one model and reports no comparison with another.
The authors’ partial answer is a log. The question gate records every hypothesis the system generated, including the ones nobody pursued, and the pipeline keeps the results of earlier iterations.
On 23 July 2026 the same authors, with Weilong Zhang, posted a version for economic theory called pAI-Econ-claude and made the workflow public. On five theory tasks, two blinded evaluators preferred the gated design to an ungated baseline in four, and mean failure severity fell from 1.58 to 1.16.
At a dollar a draft, keep a log of every question you rejected
One hundred runs cost $80 to $150 at the paper’s prices. Each dataset needs its own loader, which the paper calls a client module. A CRSP or Compustat extract would need one written, and its license terms checked before any of it is sent to a model API. Keep the file of every question the agent proposed and the reason each was dropped, because a referee who knows what a run costs will ask how many were tried.
Sources
- arXiv arxiv.org
- arXiv, pAI-Econ-claude arxiv.org