视频选题榜 2026-08-13

• https://arxiv.org/abs/2608.09888 · BDH-CQ: In-Context Learning with Recurrent Latent Reasoning · 把 ICL 与 latent reasoning 做进同一计算织物 · 150M + ARC-AGI-1 pass@2 29.5% / 单任务 $0.0007 把 cost-accuracy Pareto 前沿外推到新 SOTA · round=R1
• https://arxiv.org/abs/2608.07458 · CoinRAG: Contextualized Information Nugget KV Cache Reuse for Long-Context RAG · 把 RAG 长上下文处理里的 chunk 级 KV 缓存复用再下推到「nugget(语义要点)」级 · LongBench 多跳 QA 平均 F1 +5.3% 相对提升 + standard fast prefill latency budget 下新 Pareto 前沿 · round=R2
• https://arxiv.org/abs/2608.08160 · Can LLM Agents Stick to the Script? A Benchmark for Long-Horizon Consistency in Interactive Narratives · 把"叙事一致性"形式化为 Narrative Commitment Preservation(NCP) · 100 个电影梗概派生的可自动校验剧本环境 + trajectory + commitments + initial facts 三件套 · 20 轮后 GPT-5.2 survival rate 仅 42% / fact conflict rate 跨模型 40-68% · ICML 2026 accepted · round=R3