HF Daily Papers · 2026-09-12

HF Daily Papers 精选（15 篇，按社区票数排序）

  • [437▲] Scaling Automatic Research Agents via World Models：通过世界模型扩展自动研究 Agent — https://arxiv.org/abs/2608.12564
  • [177▲] NCP-ArchPreview Technical Report: Moving towards Latent Space Language Models through Next Concept Prediction：NCP-ArchPreview 技术报告：通过下一概念预测迈向潜在空间语言模型 — https://arxiv.org/abs/2609.10715
  • [139▲] SenseNova-U1.5: Towards Native Unified Visual Intelligence：SenseNova-U1.5：迈向原生统一视觉智能 — https://arxiv.org/abs/2609.11929
  • [89▲] AgentGrad: Intervention-guided Prompt Optimization for Multi Agent Systems：AgentGrad：面向多 Agent 系统的干预引导式 Prompt 优化 — https://arxiv.org/abs/2609.08572
  • [68▲] SpatialBlock: Enhancing Spatial Intelligence in LVLMs via Synthetic Block-Stacking Problem：SpatialBlock：通过合成积木堆叠问题增强 LVLM 的空间智能 — https://arxiv.org/abs/2609.07064
  • [46▲] T1: Terminal Agent Reinforcement Learning for Long-Horizon Tasks：T1：面向长周期任务的终端 Agent 强化学习 — https://arxiv.org/abs/2609.11042
  • [36▲] EvoSafeHarness: Evolving Model- and Domain-Specific Harnesses for Securing Agents：EvoSafeHarness：面向 Agent 安全防护的模型与领域专用 Harness 演化 — https://arxiv.org/abs/2609.05903
  • [31▲] WearableQA: A Benchmark for Health Reasoning over Real-World Wearable Data：WearableQA：面向真实可穿戴设备数据的健康推理基准 — https://arxiv.org/abs/2609.05405
  • [21▲] SWE-Bench Pro Verified: A Reliable Benchmark for Software Engineering Agents：SWE-Bench Pro Verified：面向软件工程 Agent 的可靠基准 — https://arxiv.org/abs/2609.08149
  • [20▲] Mi-Ripple: Restoring Images Degraded by Iterative AI Editing：Mi-Ripple：恢复被迭代式 AI 编辑破坏的图像 — https://arxiv.org/abs/2609.11317
  • [17▲] PARSER: Read in Parallel, Reason in Depth for Long-Context LLM Agents：PARSER：面向长上下文 LLM Agent 的并行阅读与深度推理 — https://arxiv.org/abs/2609.06702
  • [16▲] SAEScientist-Bench: Can AI Agents Conduct Autonomous SAE Interpretability Research?：SAEScientist-Bench：AI Agent 能否开展自主 SAE 可解释性研究？ — https://arxiv.org/abs/2609.09113
  • [16▲] Scores Alone Do Not Prove Discovery: The Discovery Certification Protocol for Auditing AI Research Agents：仅靠分数无法证明发现：用于审计 AI 研究 Agent 的发现认证协议 — https://arxiv.org/abs/2609.09219
  • [13▲] X-AuT: Progressive Audio-Encoder Compression for Speech LLMs with Cross-Scale Distillation：X-AuT：通过跨尺度蒸馏实现语音 LLM 的渐进式音频编码器压缩 — https://arxiv.org/abs/2609.11412
  • [12▲] SyncWorld: Visual Calibration Enables World Models as Zero-Shot Simulators：SyncWorld：视觉校准使世界模型成为零样本模拟器 — https://arxiv.org/abs/2609.09155