SAGE: From Direct Answering to Evidence-Grounded Inference for Chinese Ancient Document Understanding
SAGE reformulates Chinese ancient document understanding as evidence-grounded inference rather than direct single-pass answer generation The framework coordinates specialized agents for task-aware planning, tool-mediated evidence acquisition, claim-level verification, and bounded replanning under a constrained shared-state runtime SAGE supports bounded evidence seeking, answer revision, and abstention when grounding is insufficient Experiments on the AncientDoc benchmark show SAGE consistently o
Analysis
TL;DR
- SAGE reformulates Chinese ancient document understanding as evidence-grounded inference rather than direct single-pass answer generation
- The framework coordinates specialized agents for task-aware planning, tool-mediated evidence acquisition, claim-level verification, and bounded replanning under a constrained shared-state runtime
- SAGE supports bounded evidence seeking, answer revision, and abstention when grounding is insufficient
- Experiments on the AncientDoc benchmark show SAGE consistently outperforms matched direct-answering baselines across three LVLM backbones
- SAGE with Qwen3.5-9B surpasses much larger monolithic LVLMs on most evaluated metrics, demonstrating the value of structured inference beyond model scaling
Why It Matters
This work addresses a critical limitation in current LVLMs: their tendency to produce overconfident, weakly grounded responses when handling complex domain-specific tasks like ancient document understanding. By shifting from direct answer generation to evidence-grounded inference with multi-agent coordination, SAGE offers a scalable architecture that can achieve superior performance without requiring ever-larger models. This has broad implications for any domain where accuracy, traceability, and verifiability matter more than raw generation speed.
Technical Details
- SAGE is a multi-agent framework that decomposes ancient document understanding into specialized agent roles: task-aware planning, tool-mediated evidence acquisition, claim-level verification, and bounded replanning
- The system operates under a constrained shared-state runtime that enables bounded evidence seeking, iterative answer revision, and abstention when evidence grounding is insufficient
- The approach is evaluated on the AncientDoc benchmark, testing across three different LVLM backbones with direct-answering baselines
- Notably, SAGE powered by the relatively compact Qwen3.5-9B model outperforms much larger monolithic LVLMs on most metrics, indicating that structured reasoning architecture can compensate for model scale
- The framework explicitly addresses the multimodal challenges of Chinese ancient documents, which require integrating visual, linguistic, and historical reasoning capabilities
Industry Insight
- The results reinforce the growing consensus that architectural innovation in reasoning and verification can outperform brute-force model scaling, offering a cost-effective path to high-accuracy domain applications
- The abstention mechanism—where the system can decline to answer when grounding is insufficient—is a critical design pattern for production systems where hallucination risk must be minimized, particularly in regulated or scholarly domains
- The multi-agent, tool-mediated evidence acquisition approach generalizes beyond ancient documents to any domain requiring verifiable, traceable reasoning from complex multimodal inputs, such as legal document analysis, medical record interpretation, or scientific literature review
Disclaimer: The above content is generated by AI and is for reference only.