2026-05-09 — Cola DLM 連續潛空間擴散 LM、Skill1 統一 RL Agent、混合架構稀疏前綴快取
primary=https://arxiv.org/abs/2605.06548 primary=https://arxiv.org/abs/2605.06130 primary=https://arxiv.org/abs/2605.05219
LLM、Foundation Models、AI Infra 重大發表
primary=https://arxiv.org/abs/2605.06548 primary=https://arxiv.org/abs/2605.06130 primary=https://arxiv.org/abs/2605.05219
primary=https://deepmind.google/discover/blog/alphaevolve-a-gemini-powered-coding-agent-for-designing-advanced-algorithms/ primary=https://www.anthropic.com/news/natural-language-autoencoders primary=https://huggingface.co/blog/open-asr-leaderboard-private-data
primary=https://www.anthropic.com/news/higher-limits-spacex primary=https://huggingface.co/blog/ServiceNow-AI/correctness-before-corrections
primary=https://www.anthropic.com/news/finance-agents primary=https://arxiv.org/abs/2605.01188 primary=https://arxiv.org/abs/2605.00925
primary=https://developers.googleblog.com/en/supercharging-llm-inference-on-google-tpus-achieving-3x-speedups-with-diffusion-style-speculative-decoding/ primary=https://arxiv.org/abs/2605.00658 primary=https://arxiv.org/abs/2604.27221
MiniCPM-o 4.5:9B 參數的全雙工實時全模態模型…
AutoSP:以編譯器自動化長上下文 LLM 訓練的序列並行…
DeepSeek-V4:百萬 Token 上下文、CSA/H…
AI 評測成本超越訓練成本:Agent 基準單次執行最高 $…
Gemma 4:四種規格的開放權重多模態模型,最大版本以 1…