Think Big, Search Small: Where Capacity Matters in Hierarchical Search Agents?
TL;DR Highlight
멀티에이전트 검색 시스템에서 큰 모델은 질문 분해에, 작은 모델은 실제 검색 실행에 쓰는 게 정답이다.
Who Should Read
멀티에이전트 RAG나 검색 에이전트 시스템을 설계하는 백엔드/ML 엔지니어. 비용 효율과 검색 정확도 사이의 트레이드오프를 고민하는 개발자.
Core Mechanics
- 계층적 검색 태스크를 세 역할로 분리함: 질문 분해(delegation), 검색 실행(execution), 최종 답변 생성(answer synthesis). 이 중 delegation과 execution에 어떤 크기의 모델을 배치하느냐가 핵심 실험 변수.
- 같은 모델을 단일 에이전트로 쓰는 것보다 delegation/execution을 분리한 멀티에이전트 구조가 Exact Match(EM) 기준 4.5~8.6점 더 높음. 모든 모델 크기에서 일관되게 개선됨.
- delegation 모델(질문 분해 담당)을 키우면 EM이 약 11점 오르지만, execution 모델(검색 실행 담당)을 키우면 고작 2.6점만 오름. 즉, 질문을 잘 쪼개는 능력이 병목이고 검색 실행 능력은 상대적으로 덜 중요함.
- Qwen3-1.7B 같은 작은 모델은 단일 검색(1번 retrieval)은 99.83% 케이스에서 잘 하지만, 검색 결과가 부족할 때 재검색(multi-search correction)을 하는 케이스가 0.17%에 불과함. 반면 DeepSeek-V4-Pro는 40.31%에서 재검색을 수행함.
- 이 갭을 메우기 위해 quality-filtered trajectory distillation(품질 필터링된 궤적 증류) 방식으로 Qwen3-1.7B를 SFT 파인튜닝함. 단순히 교사 모델 궤적을 복사하지 않고, 재검색이 진짜 필요한 케이스만 골라서 학습시키는 게 핵심.
- SFT로 학습된 Qwen3-1.7B-SFT는 프론티어 모델인 DeepSeek-V4-Flash 서브에이전트보다 LLM-judge 점수가 높으면서(61.69 vs 61.21), 토큰은 37% 덜 소모함. 파라미터 수는 167배 적음.
Evidence
- delegation 용량 스케일링 시 EM +11.3점(27.86→39.18), execution 스케일링 시 EM +2.6점(39.18→41.77). delegation 민감도가 execution 대비 4.4배 높음.
- Qwen3-1.7B-SFT는 DeepSeek-V4-Flash 서브에이전트 대비 LLM-judge 61.69 vs 61.21로 더 높은 점수를 기록하면서, 토큰 소모는 4,758 vs 7,577로 37% 절감.
- RL로 end-to-end 학습된 Qwen3-8B-SearchR1을 서브에이전트로 쓰면 LLM-judge 54.55로 가장 낮음. 심지어 파인튜닝 없는 Qwen3-1.7B(57.38)보다도 낮아서, 외부에서 학습된 검색 에이전트를 서브에이전트로 재활용하면 안 된다는 점을 보여줌.
- 답변 생성 모듈(πA)로 DeepSeek를 사용하면 검색 없이도 EM 36.65를 달성해 데이터 오염이 의심됨. Qwen3-32B는 검색 없이 EM 18.58로 낮아서, 실험 컨트롤 변수로 적합함을 확인.
How to Apply
- 멀티에이전트 검색 파이프라인을 구축할 때, 오케스트레이터(질문 분해 담당)에는 GPT-4급 또는 DeepSeek-V4-Flash급 강한 모델을 배치하고, 실제 검색 실행 에이전트에는 1.7B~4B 수준의 작은 모델을 사용하면 비용을 크게 줄이면서 성능은 유지 가능.
- 작은 검색 실행 에이전트를 SFT로 학습시킬 때, 단순히 대형 모델 궤적을 전부 복사하지 말고 '첫 번째 검색이 실패했을 때 재검색에 성공한 케이스'만 골라서 주입하는 quality filtering을 적용하면 효율적으로 multi-search correction 능력을 이식할 수 있음.
- 기존에 RL로 학습된 단일 에이전트 서처(예: Search-R1 계열)를 서브에이전트로 재활용하는 것은 피해야 함. 이 모델들은 전체 질문을 스스로 분해하도록 학습되어 있어, 이미 분해된 sub-query를 받으면 불필요하게 다시 분해하거나 과도하게 검색하는 오작동을 일으킴.
Code Example
Terminology
Related Papers
Migrating a production AI agent to GPT-5.6: 2.2x faster, 27% cheaper
마케팅 웹사이트를 자동 생성하는 프로덕션 AI 에이전트를 Claude Opus 4.8에서 GPT-5.6 Sol로 전환한 실전 경험담으로, 단순 모델 교체가 아니라 eval 하네스, 툴 스키마, 캐싱, 추론 리플레이까지 손봐야 했던 과정을 구체적인 수치와 함께 정리했다.
What xAI's Grok build CLI sends to xAI: A wire-level analysis
xAI의 공식 코딩 CLI 도구 Grok Build가 사용자 동의 없이 전체 Git 저장소와 .env 시크릿 파일을 xAI 서버로 업로드한다는 사실이 네트워크 트래픽 분석으로 밝혀졌다.
Remember When It Matters: Proactive Memory Agent for Long-Horizon Agents
LLM 에이전트가 긴 작업 중 중요한 정보를 잊어버리는 문제를 별도의 메모리 에이전트가 '적절한 타이밍에' 끼어들어 해결하는 방법
WebSwarm: Recursive Multi-Agent Orchestration for Deep-and-Wide Web Search
복잡한 웹 검색을 재귀적으로 분해하고 각 노드에 적합한 검색 모드를 동적으로 할당하는 멀티에이전트 프레임워크
Show HN: Reverse-engineering web apps into agent tools
로그인된 웹 앱의 API 호출을 브라우저에서 감시해 자동으로 MCP 도구로 변환하는 에이전트를 만들었다. 소스 코드나 공식 API 문서 없이도 Jira, Spotify 같은 서비스에 AI 어시스턴트를 붙일 수 있다.
Show HN: FableCut – A browser video editor AI agents can drive (zero deps)
타임라인 전체를 JSON 파일 하나로 표현하고 MCP/REST로 AI 에이전트가 직접 편집할 수 있는 브라우저 비디오 에디터로, Claude 같은 AI가 프롬프트 하나로 영상을 자동 컷편집하고 결과를 실시간으로 UI에 반영해준다.
Related Resources
Original Abstract (Expand)
Large language model based search agents increasingly adopt multi-agent architectures in which a main agent decomposes a complex question into sub-queries and dispatches them to parallel sub-agents. However, existing systems instantiate all roles from a single model of identical scale, leaving open how model capacity should be distributed across roles. We factorize hierarchical search into three roles: a delegation role responsible for task decomposition, an execution role responsible for retrieval and evidence extraction, and an answer generation role held fixed as a confound control. We then conduct controlled capacity sweeps along the delegation and execution axes on five multi-hop QA benchmarks. The experiments yield three findings. First, role factorization consistently outperforms a single-agent baseline, improving exact match from 4.5 to 8.6 points across six model scales. Second, capacity sensitivity is asymmetric: scaling the delegation backbone improves EM by ~11 points, whereas scaling the execution sub-agent moves EM by only ~2.6 points, identifying decomposition as the capability bottleneck. Third, a 1.7B-parameter executor trained via quality-filtered trajectory distillation matches a frontier sub-agent in accuracy while consuming 37% fewer sub-agent tokens, advancing the Pareto frontier. These results suggest a concrete recipe for building hierarchical search agents: concentrate capacity at delegation and downsize execution without sacrificing accuracy. Our code is available at https://github.com/QinnanCai0115/role-factorized-search.