ResearchAgent (NAACL 2025)¶
external · status: dormant · focus: ideation · discipline: general · started: 2025
Project page: https://github.com/JinheonBaek/ResearchAgent
Source: projects/landscape/researchagent.yml
Positioning¶
The NAACL 2025 reference implementation (arXiv:2404.07738) of iterative research idea generation over scientific literature. Starting from a core paper, the system retrieves related work and entities, proposes a research problem, and refines it via parallel multi-criteria reviews from LLM reviewer agents.
Distinctive contribution¶
Among the earliest peer-reviewed treatments of agentic ideation grounded in literature, with an explicit multi-reviewer feedback loop and a paper+entity knowledge store. Useful as a methodological reference for the catalog's ideation-focused projects rather than as a turnkey product.
Evaluation scores¶
| Dimension | Score (0–3) | Note |
|---|---|---|
| Lifecycle coverage | 1 | Four upstream stages (lit discovery → research design); no analysis, drafting, or review of papers. |
| Autonomy level | 2 | Supervised: user supplies seed paper IDs and knowledge store; system iterates autonomously. |
| Architectural transparency | 3 | NAACL 2025 paper; code structure documented in README; per-agent prompts visible in source. |
| Inputs supported | 2 | Semantic Scholar paper IDs + a knowledge store mined offline; relies on S2 Graph API. |
| Outputs / reproducibility | 2 | Code + paper experiments runnable; LLM nondeterminism affects byte-level reproducibility. |
| Internal evaluation | 2 | Multi-criteria reviewer evaluation built into the loop; broader external evaluation reported in the NAACL paper. |
| Openness | 1 | Code is public but no declared open-source license — reuse rights are uncertain. |
| Maturity / traction | 1 | 37 stars; published-paper artifact with limited community uptake; last push 2025-08. |
| Cross-family policy | 0 | Single LLM family. |
| Runtime assurance | 2 | Multi-criteria reviewer agents (5 metrics) provide parallel runtime evaluation gates. |
| Cross-platform portability | 0 | Single Python pipeline; OpenAI-API-locked. |
Scored on 2026-05-18. See the evaluation rubric.
Tags¶
Pipeline stages: literature-discovery rq-formulation hypothesis-generation research-design
Architectural features: multi-agent tool-use rag-knowledge-base iterative-loop
Inputs: seed-paper-ids knowledge-store
Outputs: research-problem method-proposal experiment-design review-scores
Data sources: semantic-scholar
Knowledge sources: semantic-scholar extracted-entities
Limitations¶
- No declared open-source license.
- Single-developer maintenance; semi-active.
- Knowledge store must be pre-mined offline — not a turnkey deployment.
Related projects in this catalog¶
Papers describing this project¶
- ResearchAgent: Iterative Research Idea Generation over Scientific Literature with Large Language Models — Baek, J., Jauhar, S. K., Cucerzan, S., Hwang, S. J. (2024). NAACL 2025 (arXiv). arXiv:2404.07738
Related references (literature catalog)¶
- Wu, J. et al. (2025). Agentic Reasoning: A Streamlined Framework for Enhancing LLM Reasoning with Agentic Tools
wu2025agenticreasoning