Tianjin University's MARDoc Beats Cluttered AI Memory With Three-Agent Reasoning

Ask a long-document QA agent a hard question and watch what happens: it retrieves a chunk, reasons a bit, retrieves another chunk, reasons some more, and slowly buries the key evidence under a growing pile of intermediate notes, failed attempts, and redundant observations. By the time it needs to connect two facts from different pages, the context is so cluttered that the model loses the thread. A new paper from researchers at Tianjin University, Beihang University, and Qifu Technology proposes a fix: MARDoc, a three-agent framework that replaces that monolithic context dump with a compact, dynamically updated structured memory.

The context window is a terrible notebook

The core problem is what the authors call the monolithic context paradigm. Most agentic document QA systems, including strong baselines like DocAgent, maintain a single ever-growing context that mixes retrieval traces, tool outputs, and intermediate reasoning. Most existing systems maintain a single growing context that mixes retrieval traces, observations, and intermediate reasoning. As interactions accumulate, key evidence becomes scattered and diluted, making multi-hop reasoning noisy.

Multi-hop reasoning, meaning questions that require connecting evidence from multiple locations in a document, is especially sensitive to this. Although existing agent-based methods have an iterative advantage, most methods still suffer from the monolithic context paradigm. As interaction traces grow longer, key evidence becomes diluted by irrelevant observations. The model ends up re-reading its own noise instead of reasoning over clean facts.

Three agents, one shared memory

MARDoc decouples long-document QA into three specialized agents: an Explorer for multi-granularity multimodal retrieval, a Refiner for distilling interaction traces into structured evidence and reasoning memories, and a Reflector for checking evidence sufficiency and providing targeted feedback. The key insight is that these agents never pass the full history to each other. Instead, they share a single structured memory object that gets refined at each step.