Top 3 papers of the Week [Feb 2 - Feb 6, 2026] suggested by Genloop's LLM Research Hub.

🥇 Kimi K2.5: Visual Agentic Intelligence

Kimi K2.5 is an open-source multimodal model advancing agentic intelligence through joint text-vision optimization and the Agent Swarm framework for parallel task orchestration. It achieves state-of-the-art performance in coding, vision, reasoning, and agentic tasks, with significant latency reductions.

Research Credits: Kimi AI; Yifan Bai; bao yiping; Charles Y.; H. S. Che; Cheng Chen; Guanduo Chen; Huarong Chen; Jia Chen; Jiahao Chen; Jianlong Chen; Jun Chen; Kefan Chen; Xinhao Chen ; Yanru Chen

Link

🥈 FASA: Frequency-aware Sparse Attention

The study introduces FASA, a novel framework for query-aware token pruning in LLMs, leveraging functional sparsity in RoPE to dynamically predict token importance. FASA achieves state-of-the-art performance on long-context tasks, significantly reducing memory usage while maintaining near-oracle accuracy and improving computational efficiency.

Research Credits: Yifei Wang; Yueqi Wang; Zhenrui Y.; Yong Wang; Ismini Lourentzou; Zhengzhong Tu; Xiangxiang Chu; Julian McAuley2

Link

🥉 LangMap: A Hierarchical Benchmark for Open-Vocabulary Goal Navigation

The research introduces HieraNav, a multi-granularity goal navigation task, and LangMap, a benchmark for language-driven embodied navigation using real-world 3D scans. It evaluates models on natural language instructions, highlighting challenges in long-tailed and multi-goal tasks, and establishes a testbed for advancing AI in navigation.

Research Credits: Bo Miao; WEIJIA LIU; Jun Luo; Lachlan Shinnick; Jian Liu

Link