Poli Searcher | Frank Yifei Zhu

Poli Searcher is an agentic research pipeline for building transparent political science datasets from heterogeneous web sources. It connects recursive search and evidence synthesis with downstream structured coding, so large language models can retrieve, curate, and evaluate political facts at scale. Elite biography extraction is one core application, but the same pipeline can support diverse political science dataset construction problems by changing the codebook, task instructions, and validation rules.

The starting point is the agentic framework for dataset building: it defines a search-synthesis-coding pipeline and demonstrates it through political biography extraction with a complete evaluation package. PolitNuggets then sharpens that evaluation problem for search agents by specifying the benchmark, enhancing the evidence package, and proposing an automated evaluation method. Its selection for The 64th Annual Meeting of the Association for Computational Linguistics in San Diego, California, United States reflects recognition from the computer science community.

Project Map

flowchart LR
  A["Agentic Framework<br/>for Dataset Building<br/><small>search-synthesis-coding pipeline</small>"]
  B["Poli Searcher<br/><small>open-source implementation</small>"]
  C["PolitNuggets<br/><small>benchmark + automated evaluation</small>"]
  D["Task-Specific Dataset<br/><small>codebooks, instructions, and artifacts</small>"]

  A --> B
  B --> C
  C --> D
  D -. supports .-> A

Resources

Agentic Framework for Political Biography Extraction: the methodological foundation for agentic political science dataset building, demonstrated through political biography extraction.
PolitNuggets: Benchmarking Agentic Discovery of Long-Tail Political Facts: the benchmark and automated evaluation framework for search agents selected for The 64th Annual Meeting of the Association for Computational Linguistics in San Diego, California, United States.
Poli Searcher GitHub repository: the companion implementation for the deep-research/search-agent pipeline.
PolitNuggets dataset on Hugging Face: the public dataset and evaluation artifacts.

Coming Soon

Claude Code adaptation for agentic political-search workflows.
Support for search over a local database.

Agentic Framework for Political Biography Extraction

Yifei Zhu, Songpo Yang, Jiangnan Zhu, and 1 more author

2026

Invited to resubmit, American Journal of Political Science

Abs arXiv Bib SSRN Code

The production of large-scale political datasets typically demands extracting structured facts from vast piles of unstructured documents or web sources, a task that traditionally relies on expensive human experts and remains prohibitively difficult to automate at scale. We leverage Large Language Models (LLMs) to automate the extraction of multi-dimensional elite biographies. We propose a two-stage Synthesis-Coding framework: an upstream synthesis stage that uses recursive agentic LLMs to search, filter, and curate biographies from heterogeneous web sources, followed by a downstream coding stage that maps curated biography into structured dataframes.
@unpublished{zhu2025agentic, title = {Agentic Framework for Political Biography Extraction}, author = {Zhu, Yifei and Yang, Songpo and Zhu, Jiangnan and Jiang, Junyan}, year = {2026}, note = {Invited to resubmit, American Journal of Political Science}, doi = {10.48550/arXiv.2603.18010}, }

PolitNuggets: Benchmarking Agentic Discovery of Long-Tail Political Facts

Yifei Zhu

In The 64th Annual Meeting of the Association for Computational Linguistics, 2026

San Diego, California, United States

Abs arXiv Bib Code Dataset

Large Reasoning Models (LRMs) embedded in agentic frameworks have transformed information retrieval from static, long-context question answering into open-ended exploration. Yet real-world use requires models to discover and synthesize "long-tail" facts from dispersed sources, a capability that remains under-evaluated. We introduce PolitNuggets, a multilingual benchmark for agentic information synthesis via constructing political biographies for 400 global elites, covering over 10000 political facts. We standardize evaluation with an optimized multi-agent system and propose FactNet, an evidence-conditional protocol that scores discovery, fine-grained accuracy, and efficiency.
@inproceedings{zhu2025politnuggets, title = {PolitNuggets: Benchmarking Agentic Discovery of Long-Tail Political Facts}, author = {Zhu, Yifei}, booktitle = {The 64th Annual Meeting of the Association for Computational Linguistics}, year = {2026}, note = {San Diego, California, United States}, doi = {10.48550/arXiv.2605.14002}, }

Project Map

Resources

Coming Soon

Related Papers