视频选题榜 2026-09-16
- https://arxiv.org/abs/2609.15635 · ModaLens · Measuring Image Sensitivity in Report-Conditioned Medical VLMs · paired-swap audit 量化 VLM 图像敏感性 · 4.26% vs 20.94% = 16.7pp paired increase = 4.92 倍 · 3,199 配对 × 14 问题 = 44,786 paired trials · cs.CV+cs.AI 2 类 cross-listed · round=R1
- https://arxiv.org/abs/2609.15504 · How Lossless Is Lossless Speculative Decoding? The Role of Numerical Precision in Orthrus · 第一个把「无损」从口号变成可验证命题,BF16 45% vs FP32 100% · round=R2
- https://arxiv.org/abs/2609.17320 · Emergence World: Adversarial Stress-Testing of Long-Horizon Multi-Agent Systems · 8 world × 10 agent × 16 day = 850,000+ LLM calls / 50B tokens · 「检测 ≠ 遏制」= agent 识别威胁后仍写进持久记忆,46h 后才作用 · model-level alignment is not compositional · round=R3
- https://arxiv.org/abs/2606.08367 · Emergence World: A Platform for Evaluating Long-Horizon Multi-Agent Autonomy · 同作者组 v1 早期版本 · 15 day × 5 world · Claude Sonnet 4.6 / Grok 4.1 Fast / Gemini 3 Flash / GPT-5-mini + Mixed · 三记忆 + 120+ 工具 + proposal-and-vote · round=R3
- https://arxiv.org/abs/2609.16818 · InceptionRAG: Stealthy Poisoning Attack Against Retrieval-Augmented Generation · 间接逻辑诱导投毒攻击 · 单文档扫描类缓解机制无效 · round=R3
- https://arxiv.org/abs/2609.17012 · ORDER: Task-Conditioned Routing for Retrieval-Augmented Generation · 查询条件化 RAG 路由 · 动态适配分块粒度/元数据/来源选择 · round=R3
- https://arxiv.org/abs/2609.11873 · The Last AI Built by Humans: Toward Genuine Recursive Self-Improvement · HCI 头空指数 + 五阶段 RSI 路线图 · round=R3
- https://arxiv.org/abs/2603.00873 · MC-Search: Evaluating and Enhancing Multimodal Agentic Search with Structured Long Reasoning Chains · ICLR 2026 · 5 种推理拓扑 + 过程级指标 · round=R3