视频选题榜 2026-08-07
- https://arxiv.org/abs/2608.03994 · When Attention Goes Blind: Numerical Failure in ALiBi Positional Encodings · 位置编码线性 bias fp16/bf16 下溢导致 attention head 失明 · round=R1
- https://arxiv.org/abs/2608.05102 · ABSeeker: Training Long-Horizon Search Agents via Answer-Backtracked Credit Assignment · 用答案回推线索集给长程搜索 agent 发步骤级稠密信用 · round=R2
- https://arxiv.org/abs/2607.28609 · OSReward: Instituting Standardized Evaluation for Cross-Platform Computer-Use Reward Models · 给跨平台 CUA 立可审计的 VLM judge + 9B/35B 开放奖励模型成本比闭源低 30–60× · round=R3