SurveyForge addresses the challenges of generating high-quality, accurate survey papers using LLMs, targeting researchers who need to efficiently create comprehensive literature reviews. It aims to improve outline quality and citation accuracy, offering a significant benefit over existing automated methods.

How It Works

SurveyForge employs a two-stage process: outline generation and content refinement. It first creates an outline by analyzing human-written survey structures and consulting domain-specific articles. Subsequently, it leverages a scholar navigation agent to retrieve high-quality papers, enabling the generation and refinement of survey content. This memory-driven approach, combined with outline heuristics, aims for greater coherence and factual accuracy.

Quick Start & Requirements

To try SurveyForge:

1. Download the database from SurveyForge Database.
2. Download the embedding model from gte-large-en-v1.5.
3. Configure API keys and data paths in code/run_demo.py.
4. Run cd code && python run_demo.py.

Highlighted Details

• Introduced at ACL-2025 main conference.
• Includes SurveyBench, a benchmark with 100 human-written survey papers across 10 topics for evaluation.
• Supports evaluation of generated surveys via SurveyBench/test.py.
• Built upon the AutoSurvey framework.

Maintenance & Community

The project was accepted to ACL-2025 main conference. Code was released in June 2025. Further details on community channels or active contributors are not specified in the README.

Licensing & Compatibility

The README does not explicitly state the license. However, its foundation on AutoSurvey, which is Apache 2.0 licensed, suggests potential compatibility, but this requires explicit confirmation.

Limitations & Caveats

The project is presented as newly released code. While it aims for comprehensive evaluation, the SurveyBench dataset currently covers 10 topics, with more planned. The accuracy and robustness of the scholar navigation agent and memory-driven generation are not detailed.