Recently, the OpenBMB open-source community welcomed a new member, "Juan Ji" (SurveyGO), attracting attention in the field of long-text generation. Currently, both students and professionals face challenges in information acquisition, and "Juan Ji" offers a promising solution to this predicament.
"Juan Ji": A Top Performer in Long-Text Generation
"Juan Ji" can be considered a top performer in long-text generation. It utilizes information entropy and convolutional algorithms to quickly sift through massive amounts of literature and synthesize complex information into high-quality summaries. Whether it's a niche professional field or a popular research area, simply providing keywords allows "Juan Ji" to precisely filter literature, extract core knowledge, and output content that is logically rigorous and insightful.
Obtaining a summary report generated by "Juan Ji" is simple. Users open the designated website, select either normal or professional mode to submit their request, and then log in again to retrieve the report. The website's "Writing Request Form" also provides various novel research topics for user interaction and voting.
Testing "Juan Ji"'s Capabilities
To test its capabilities, the team conducted a comparative evaluation, pitting "Juan Ji" against models like OpenAI-DeepResearch to write a summary on the topic of "The Impact of the Tariff War on Ordinary People's Lives." Evaluated across structure, content, viewpoint, and citations, "Juan Ji" performed exceptionally well. Its generated article featured a clear table of contents, in-depth content analysis, well-reasoned viewpoints, and accurate citations, surpassing other models in overall performance.
LLMxMapReduce-V2: Technological Empowerment
Behind "Juan Ji"'s powerful capabilities lies the LLMxMapReduce-V2 long-text integration and generation technology. This is a collaborative achievement of AI9Star, OpenBMB, and Tsinghua University, representing an upgrade to the original technology. This technology utilizes a text convolutional algorithm to aggregate references and combines an information entropy-driven stochastic convolutional test-time scaling method to efficiently handle ultra-long inputs and improve article quality.
The research team used the newly developed SurveyEval benchmark for evaluation. The results show that LLMxMapReduce-V2 excels across multiple key indicators, particularly demonstrating a significant advantage in the utilization of references. This indicates that "Juan Ji"'s technology is robust in handling large-scale information integration tasks, and its future applications in the long-text generation field are promising, with the potential to drive industry innovation.
Address: https://surveygo.thunlp.org/