Archive
Agent、MCP、coding agent、工作流自動化與評測。
推論系統
New research separates a turn becoming ready from its submission to the inference engine, using CVaR tail risk and a dynamic workload budget to determine submission order. SWE-bench trace replays achieved up to a 3.5× P95 speedup, but the results cover only fixed agent trajectories, a single randomized arrival seed, and an A100/vLLM environment.
AI 安全
The new benchmark holds retrieved content constant and uses four conditions to measure how RAG changes five open models’ responses to harmful questions. Documents containing answers made every model more likely to produce accurate but dangerous content, while safety judges showed only moderate agreement.
具身 AI 與機器人
Show-Harness has a general-purpose VLM output symbols for directions, grasping, and releasing, which robot-specific interpreters then translate into constrained motions. Gemini in a zero-shot setting and a fine-tuned Qwen 2B achieved success rates of 89% and 86%, respectively, across ten real-world tasks, though testing remains focused on tabletop pick-and-place.
最新模型
Xiaomi has proposed an end-to-end target-speaker ASR system that feeds a one- to four-second reference utterance and mixed audio into the same audio encoder, without first performing speech separation. The model can also output an empty string when the target speaker is absent, but its training relies heavily on non-public data, and the initial report provides no verifiable model weights or code.
推論系統
PELM goes beyond adjusting processor frequency: it dynamically selects speculative decoding configurations and verification depth, jointly controlling hardware and model workload under thermal constraints. The paper reports speedups of up to 23.1% and energy savings of up to 52.4%, but these peak results cannot be directly extrapolated to smartphones or other models.
AI 安全
For the first time, Datasette used Claude and GPT-series agents to conduct a comprehensive security audit, followed by two developers cross-checking the work by writing tests and patches. The new releases tighten controls around private-data permissions, SQL and HTML escaping, authentication, and HTTP caching, but some reproduction tests remain temporarily undisclosed, preventing full independent verification of the vulnerabilities’ scope.
代理工具
Google’s ToolGrad first builds an executable API chain, then generates the corresponding instruction in reverse, avoiding the inefficiency of traditional agents searching for a solution through trial and error. Experiments show that even a small dataset can improve function-calling performance on unseen tools, though the results remain concentrated on single-turn benchmarks and synthetic workflows.
AI 安全
AgentAudit does more than calculate whether a task was completed: it separately examines planning, memory, tool selection, safety, and execution integrity across the execution trace. Composite Trust Scores vary widely among five models, but nine tasks and a single model judge are not enough to support a general-purpose ranking.
生成式音訊
YuE2 combines symbolic melody and chord planning, semantic tokens, and flow-matching audio synthesis in a single generation pipeline. Its weights can run on a 24GB GPU, but its noncommercial license and best-of-8 evaluation mean comparisons with commercial systems should still be interpreted cautiously.
代理工具
Φ-Bench expands evaluation beyond individual GPU kernels to multi-file implementation and end-to-end system optimization. All eight frontier models occasionally find effective modifications, but larger reasoning budgets do not consistently translate into better engineering outcomes.
代理工具
Pizza Bot stores agent state, approval requests, and scheduled results on the server, allowing work to continue after users leave the conversation. It supports multiple model providers, MCP, and Agent Skills, but local-first does not mean tool-safe: remote deployments and extensions still require carefully designed trust boundaries.
應用研究
The NASA-IBM Lunar Foundation Model spans scales from 1 to 100 meters per pixel using nearly 2 million aligned lunar data samples. The model, dataset, and fine-tuning code are all publicly available, but its ice-related outputs remain indirect inferences and must not be used for landing or exploration decisions.