반응형
반응형
LIST
「Figure 1, Wu et al., arXiv:2609.29444」 IterSynth 전체 구조. 한 모델이 Planner(검색·답변)와 Synthesizer(요약 갱신) 역할을 번갈아 맡고, 질문과 누적 요약(q, M_t)만으로 다음 단계를 이어 갑니다.오늘 읽은 논문IterSynth: Rethinking Deep Search Agents via Role-Decoupled Iterative SynthesisXingyu Wu · Yuchen Yan · Zhengxi Lu · Siqi Chen · Xin Zhang · Aiting Liu · Chao Deng · Jie Liu · Jin Ma · Jian Shao · Jun Xiao · Yongliang Shen (Zhejiang University /..
「Figure 1, Ho et al., arXiv:2609.02737」 Declarative Attention: model declares global / focus / local spans; engine masks most KV-cache reads.오늘 읽은 논문Language Models Can Control Their Own AttentionNamgyu Ho · Huzama Ahmad · Woosung Koh · Se-Young Yun (KAIST AI) · Tal Schuster · Cicero Nogueira dos Santos (Google DeepMind), arXiv:2609.02737, 2026-09-02.긴 문서를 읽을 때 사람은 문장마다 처음부터 다시 훑지 않습니다. 그런데 디코딩 ..
「Figure 3, Zhang et al., arXiv:2609.18094」 Every scored contribution on a log bpb axis; first-day statistical priors deliver almost the whole reduction.오늘 읽은 논문Agora: Git as Shared Memory for Collective AutoResearchYifan Zhang · Yunheng Zou · Shaokun Zhang · Jian Hu · Hao Zhang · Binfeng Xu · Jan Kautz · Yi Dong (NVIDIA), arXiv:2609.18094, 2026-09-16.여러 연구 에이전트가 같은 과제를 나눠 돌리면, 세션이 끝날 때마다 「무엇이 시도..
「Figure 1, Li et al., arXiv:2609.06702」 Comparison between ParSer and sequential memory methods. Scatter–gather parallel reading vs sequential memory chain.오늘 읽은 논문PARSER: Read in Parallel, Reason in Depth for Long-Context LLM AgentsKun Li · Zexuan Qiu · Tianhua Zhang · Irwin King · Helen Meng (The Chinese University of Hong Kong), arXiv:2609.06702, 2026-09-06. 프로젝트: https://cuhk-parser.github.i..