LLMxMapReduce

LLMxMapReduce-V2 is a long-text processing framework developed by Tsinghua University and its collaborating team. Drawing inspiration from the hierarchical feature extraction approach of convolutional neural networks, it employs a "divide-and-conquer" strategy—splitting, processing, and integrating—to break down ultra-long texts into manageable segments and progressively aggregate them layer by layer. A multi-layered convolution mechanism continuously refines and merges information features during the aggregation process, while entropy-driven dynamic weight adjustments ensure that critical content receives prioritized attention. This system can automatically generate high-quality literature reviews that are both logically rigorous and richly detailed, effectively overcoming the limitations of large model context windows. Additionally, the interface enables the creation of complete academic articles, offering researchers a one-stop solution—from literature retrieval to final manuscript output.

1. GitHub link: https://github.com/thunlp/LLMxMapReduce
2. After successfully creating a task, obtain the task_id and use it to poll the task query interface; note that this process may take some time—typically around 15 minutes per task.
3. Costs are calculated based on the number of model calls and search API calls:

• Model calls: Gemini-2.5-flash-lite—input costs 0.1 PTC/M, output costs 0.4 PTC/M
• Search API calls: SerpApi—0.005 PTC per call