Research
Evaluation work, benchmarks, and notes on what it actually takes to make multi-document reasoning reliable enough for real legal work.
On LegalBench-RAG, RAG-extended FCR beats both classic RAG and the multi-agent system, showing that scalable retrieval works best when the final answer is produced through document-level reasoning.
Read researchFCR reached 87.0% accuracy on AA-LCR, outperformed reported frontier-model baselines, and achieved 7/7 on the reviewed legal slice with verified deterministic citations.
Read research即使问题简单，FCR也很有必要，因为FCR的准确率是RAG的2倍！
Read research文通过实战案例解析检索增强生成（RAG）在复杂任务中的瓶颈，并介绍新技术——全文档推理如何攻克这些难题。
Read researchRAG is great for search; it often fails for cross-document reasoning. FCR is a reasoning runtime that constructs a usable full-context environment, then verifies grounding inline as reasoning progresses.
Read researchAttention's quadratic cost is the fundamental bottleneck for long-context inference. We propose Superlinear Multi-Step Attention: a fully trainable attention architecture that achieves O(L^{3/2}) complexity while preserving random context access.
Read research