Generalizing an LLM from 8k to 1M Context using Qwen-Agent

Qwen · 2 years ago

Alibaba's Qwen team built a multi-level agent system that extends an 8k-token context model to handle 1-million-token documents by combining retrieval-augmented generation, chunk-by-chunk reading, and step-by-step reasoning rather than relying on native long-context models. The system was evaluated on NeedleBench and LV-Eval benchmarks designed for 256k-context tasks, where the 4k-Agent consistently outperformed both a 32k-context model extended via RoPE extrapolation and basic RAG approaches. The agent framework is being released as open-source infrastructure to generate synthetic fine-tuning data for training long-context models.