Flash-dLLM：為擴散大語言模型打造的 I/O 感知 KV 快取與平行解碼加速框架
Flash-dLLM is a training-free inference acceleration framework that dramatically speeds up Diffusion LLMs (dLLMs) using an IO-aware fused KV cache and a self-contained draft-and-verify decoding strategy.
突破中心化瓶頸！Agensh 框架將多 Agent 協作無縫擴展至 1,024 個智慧體
Agensh is a decentralized, self-organized multi-agent harness that scales up to 1,024 agents using an asynchronous cooperation loop, significantly boosting efficiency in complex coding tasks.
「誰對誰說了什麼？」SpeakerMem-R1 打造多方對話專屬的雙軌記憶機制
SpeakerMem-R1 tackles multi-party dialogue memory challenges using a speaker-labeled dual-track system and GRPO training to improve message attribution and social relation tracking.
CliffCompaction：長任務程式碼 Agent 的高效能 context 自動壓縮技術
CliffCompaction is an autocompaction technique for long-horizon coding agents that reduces token costs by up to 50% and prevents context drift by strictly truncating rather than rewriting original context.
SWE-Serve：首個針對「生產級推論服務」的 AI Agent 軟體工程基準測試
SWE-Serve is a new benchmark featuring 53 real-world SGLang tasks, designed to evaluate AI agents' capability to implement complex features and achieve production correctness across the inference serving stack.
破解 MCP 生態系：新型 A2M 攻擊框架揭示 AI Agent 的語意供應鏈安全威脅
This study introduces A2M, a two-stage black-box hijacking framework, exposing critical semantic supply-chain vulnerabilities in Model Context Protocol (MCP) agents and highlighting the need for stronger security.
別撐大脈絡，改建構程式框架：Growing Harness 透過「失敗導向學習」自動生成高效能 Agent
Growing Harness is a novel paradigm that turns recurring agent control decisions into reusable, executable code via failure-guided synthesis, slashing inference costs by up to 98.6% while keeping small models highly capable.
類型安全不等於無誤：約束決策模型偏向選項字面名稱而非綁定規則
Research reveals that type-safe decision models are heavily biased by the semantic polarity of option names (e.g., yes/no) rather than their defined rubrics, leading to severe decision reversals.
硬體驗證 AI 的試金石：EquivSVA 形式驗證資料集，揪出 LLM 程式碼生成的虛實
EquivSVA is a formally verified open-source dataset designed to evaluate LLM robustness in generating SystemVerilog Assertions (SVA), featuring 120 'behavior families' across 480 equivalent RTL implementations.
AI 真的能幫產品設計省時間嗎？首個 Figma Make 的隨機對照實驗揭密
A randomized controlled trial of 100 professionals shows that using AI prompt-to-design tools like Figma Make reduces completion times by 20%, with product managers seeing the largest gains.
Flash-dLLM：為擴散大語言模型打造的 I/O 感知 KV 快取與平行解碼加速框架
Flash-dLLM is a training-free inference acceleration framework that dramatically speeds up Diffusion LLMs (dLLMs) using an IO-aware fused KV cache and a self-contained draft-and-verify decoding strategy.
類型安全不等於無誤：約束決策模型偏向選項字面名稱而非綁定規則
Research reveals that type-safe decision models are heavily biased by the semantic polarity of option names (e.g., yes/no) rather than their defined rubrics, leading to severe decision reversals.
突破中心化瓶頸！Agensh 框架將多 Agent 協作無縫擴展至 1,024 個智慧體
Agensh is a decentralized, self-organized multi-agent harness that scales up to 1,024 agents using an asynchronous cooperation loop, significantly boosting efficiency in complex coding tasks.
別撐大脈絡，改建構程式框架：Growing Harness 透過「失敗導向學習」自動生成高效能 Agent
Growing Harness is a novel paradigm that turns recurring agent control decisions into reusable, executable code via failure-guided synthesis, slashing inference costs by up to 98.6% while keeping small models highly capable.
CliffCompaction：長任務程式碼 Agent 的高效能 context 自動壓縮技術
CliffCompaction is an autocompaction technique for long-horizon coding agents that reduces token costs by up to 50% and prevents context drift by strictly truncating rather than rewriting original context.
SWE-Serve：首個針對「生產級推論服務」的 AI Agent 軟體工程基準測試
SWE-Serve is a new benchmark featuring 53 real-world SGLang tasks, designed to evaluate AI agents' capability to implement complex features and achieve production correctness across the inference serving stack.
「誰對誰說了什麼？」SpeakerMem-R1 打造多方對話專屬的雙軌記憶機制
SpeakerMem-R1 tackles multi-party dialogue memory challenges using a speaker-labeled dual-track system and GRPO training to improve message attribution and social relation tracking.
硬體驗證 AI 的試金石：EquivSVA 形式驗證資料集，揪出 LLM 程式碼生成的虛實
EquivSVA is a formally verified open-source dataset designed to evaluate LLM robustness in generating SystemVerilog Assertions (SVA), featuring 120 'behavior families' across 480 equivalent RTL implementations.
AI 真的能幫產品設計省時間嗎？首個 Figma Make 的隨機對照實驗揭密
A randomized controlled trial of 100 professionals shows that using AI prompt-to-design tools like Figma Make reduces completion times by 20%, with product managers seeing the largest gains.
破解 MCP 生態系：新型 A2M 攻擊框架揭示 AI Agent 的語意供應鏈安全威脅
This study introduces A2M, a two-stage black-box hijacking framework, exposing critical semantic supply-chain vulnerabilities in Model Context Protocol (MCP) agents and highlighting the need for stronger security.