Archive
Agent、MCP、coding agent、工作流自動化與評測。
代理工具
HarnessOpt-Bench asks models to modify prompts, tools, memory, and control flow within a fixed evaluation budget, then measures actual gains on an unseen test set. Across 111 experiments, the model itself accounted for more variation than the outer coding harness, while native tools did not consistently outperform shared ones.
代理工具
TrajDebug does not treat the first error in an agent trajectory as the root cause. Instead, it tracks whether errors are repaired, whether they leave an impact on the final outcome, and how much recovery budget they consume. In tests on 486 human-annotated trajectories, converting diagnoses into prompts improved the success rate of rerunning the same tasks by 10.8% on average.
最新模型
The study traces the difference between Adam and low-rank recovery to the coordinate-wise preconditioner’s lack of gauge equivariance. In published experiments, two functionally equivalent initializations trained with Adam produced attention invariants that differed by 56%, though the findings have not yet been validated in large-model pretraining.
最新模型
Rather than treating teacher–student discrepancies at every token equally, DASH adjusts how far supervision can propagate based on the entire generation trajectory. The authors report consistent gains on three mathematics benchmarks and have released the training code, evaluation pipeline, and LoRA weights.
Reasoning Core 用可執行評分器、難度控制與短答案格式產生數學、規劃、形式證明及程式題。匹配訓練實驗顯示它優於三套程序資料集，但作者也在外部資料庫發現生成器與評分器不一致。
代理工具
A new study distinguishes persistent AGENTS.md files from task-specific Agent Plans and examines how plans enter public version histories. The sample is heavily concentrated in a handful of projects, suggesting that treating plans as collaborative artifacts remains an early and unstable practice.
代理工具
Muse Code packages Muse Spark 1.2 as a terminal agent capable of working on large codebases in parallel within isolated worktrees. Its append-only event log can reconstruct state after a process interruption, though the performance figures still come primarily from Meta itself.
AI 安全
A new benchmark plants targeted probes—including semantic lures, parameter traps, and capability illusions—in MCP tool lists, using erroneous calls to pinpoint agents’ tool-selection weaknesses. Across 11,520 experiments, model tier did not consistently predict safety, while the ability to recover after being deceived raised task success rates from 16% to 52%.
代理工具
Argus breaks long-running tasks into constrained missions, with a Manager, Planner, Engineer, and Reviewer jointly maintaining persistent state. Its authors report improving program repair performance at 1.41× the token cost; an agent-generated RWKV6 kernel was also merged into an open-source project after external corrections.
代理工具
ContextWeave goes beyond asking whether agents can retrieve old information, examining whether memory improves subsequent document and office work. Experiments show that retaining complete, actionable experiences outperforms short summaries, although incorrect recollections are also more likely to affect execution results.
代理工具
MatrAIx combines correlated demographic attributes to create 8.3 billion simulated users and lets persona agents interact with surveys, chatbots, websites, and applications. The team has released a core dataset of roughly one million records and its evaluation code, but current validation shows only that the agents follow assigned personas—not that they can replace testing with real people.
AI 基礎設施
Microsoft Azure research shows that agents repeatedly switch among model inference, tools, and orchestrators, putting the CPU back on the critical path while GPUs often sit idle. The Agora prototype can reclaim compute based on workflow patterns, but incorrectly reducing GPU allocations for parallel agents can worsen tail latency by 16×.