심층 연구 에이전트의 컨텍스트 드리프트 완화를 위한 체크리스트 루브릭 중심 반복 제어

초록

Recent LLM-based deep research agents have advanced toward generating long-form reports through iterative web exploration and evidence synthesis. However, when exploration and rewriting proceed without explicit criteria, context drift and target omission can accumulate. To address this issue, this paper proposes a checklist-based rubric-driven parallel investigation multi-agent framework. The proposed method derives global and local specifications from the query, fixes them as a checklist rubric, and selectively re-executes only failed items to enable controlled requirement-centered iteration. Evaluation on ResearchQA showed that the proposed approach exhibited a tendency toward improved ORS under some settings compared with the configuration without a checklist (Qwen3-8B: 93.30%→94.62%, GPT-OSS-120B: 95.19%→97.77%), although the effect varied by model and condition. These results suggest that checklist-based refinement can help mitigate target omission and improve ORS under certain conditions.

키워드