2월 4일 AI 주요 논문

1. CodeOCR: On the Effectiveness of Vision Language Models in Code Understanding
   - 다중 모달 대형 언어 모델이 압축된 이미지 형태로 소스 코드를 효과적으로 이해하며, 코드 이해 작업에서 성능을 유지하거나 개선하면서도 토큰 수를 크게 줄일 수 있음을 보여줌.
   - [자세한 내용 보기](https://arxiv.org/abs/2602.01785)

2. daVinci-Agency: Unlocking Long-Horizon Agency Data-Efficiently
   - 장기적인 에이전트 작업 흐름에서의 문제를 해결하기 위해 풀 리퀘스트 시퀀스를 활용하여 구조적 감독을 제공하고, 일관성 유지를 위한 점진적 분해 및 버그 수정 역사로부터의 세련미를 사용함.
   - [자세한 내용 보기](https://arxiv.org/abs/2602.02619)

3. MARS: Modular Agent with Reflective Search for Automated AI Research
   - 예산 인지적 계획, 모듈 구성, 반영 메모리를 활용하여 자율 기계 학습 연구에서의 성과를 극대화하는 모듈형 AI 연구 자동화 프레임워크인 MARS를 소개함.
   - [자세한 내용 보기](https://arxiv.org/abs/2602.02660)

4. 3D-Aware Implicit Motion Control for View-Adaptive Human Video Generation
   - 3DiMo는 비디오 생성에서 인간의 움직임을 제어하는 새로운 방법을 제안하며, 미리 훈련된 비디오 생성기와 함께 움직임 인코더를 훈련시켜 변환 프레임을 간결한 움직임 토큰으로 증류함.
   - [자세한 내용 보기](https://arxiv.org/abs/2602.03796)

5. No Global Plan in Chain-of-Thought: Uncover the Latent Planning Horizon of LLMs
   - 대형 언어 모델의 숨겨진 계획 동력을 조사하며, 'Tele-Lens'라는 방법을 통해 제한된 글로벌 계획을 밝혀내고 불확실성 추정 및 CoT 우회 인식을 개선함.
   - [자세한 내용 보기](https://arxiv.org/abs/2602.02103)

6. Learning Query-Specific Rubrics from Human Preferences for DeepResearch Report Generation
   - DeepResearch 보고서 작성을 개선하기 위해, 하이브리드 보상으로 훈련된 인간 선호 기반의 루브릭 생성기를 사용하여 다중 에이전트 마르코프 상태 워크플로우로 강화를 거침.
   - [자세한 내용 보기](https://arxiv.org/abs/2602.03619)

7. Parallel-Probe: Towards Efficient Parallel Thinking via 2D Probing
   - Parallel-Probe는 조기 중단을 통한 합의 기반 접근 방식과 변동 기반 가지치기를 사용하여 정확도를 유지하면서 계산 비용을 줄이는 훈련 없는 컨트롤러임.
   - [자세한 내용 보기](https://arxiv.org/abs/2602.03845)

8. SWE-World: Building Software Engineering Agents in Docker-Free Environments
   - Docker 없이 소프트웨어 공학 에이전트를 훈련할 수 있는 효율적인 환경을 구성하기 위해, 물리적 실행 환경 대신 학습한 대리자를 사용하는 프레임워크를 제안함.
   - [자세한 내용 보기](https://arxiv.org/abs/2602.03419)

9. SWE-Master: Unleashing the Potential of Software Engineering Agents via Post-Training
   - 소프트웨어 엔지니어링 에이전트를 개발하기 위한 재현 가능한 프레임워크를 제공하고, 에이전트 개발의 여러 단계를 체계적으로 최적화하여 성능을 향상시킴.
   - [자세한 내용 보기](https://arxiv.org/abs/2602.03411)

10. Unified Personalized Reward Model for Vision Generation
    - UnifiedReward-Flex는 보상 모델링과 유연한 맥락 적응적 추론을 결합하여 시각적 생성을 개선하고, 의미(Intent) 및 시각적 증거에 기반해 동적으로 계층 평가를 구성함.
    - [자세한 내용 보기](https://arxiv.org/abs/2602.02380)