Tag
该论文提出Δτ-Bench，一个评估AI智能体在不断演化的工具生态中适应能力的可执行基准，包含42个领域、6个生态状态下的1,588个多轮任务；并提出Steward生态系统管理器，通过增量技能维护和动态供应实现更高的持续任务成功率。
Unsloth 发布 NVIDIA Nemotron 3.5 Bolt E4B 的 GGUF 量化版本（Q4_K_M），这是一个基于 Gemma-4-E4B-it 进一步后训练的小语言模型（SLM），专为低内存、低首字延迟的本地语音助手与嵌入式/边缘设备场景优化。
An open benchmark placed 13 AI models in a simulated GP consultation (195 runs across 5 hand-written cases), showing models nearly always named the right diagnosis but varied widely (27–85%) on history-taking, safety and appropriate testing — with notable failures like prescribing penicillin to an allergic patient. Results, transcripts and code are public, with a paper planned for arXiv.
DynaWeave 是一个自动化合成框架，可将静态任务转化为含外生事件的可验证动态场景，用于训练 LLM 智能体适应环境变化；使用其生成的 2,095 条轨迹对 Qwen3.5 进行监督微调后，在 CostBench 上任务完成率提升 67.34 个百分点，并在 ToolMaze 上超越 Gemini-3.1-Pro-Preview。
本文提出 Discord 与 BRACE 两种方法，利用奖励分歧来校准 MCP 环境中工具使用语言模型的强化学习 rollout 预算分配，并在 Qwen3-VL 多个规模及 BFCL、NESTFUL 基准上验证其优于基线。
文章探讨了 AI agent 系统中「任务完成」信号不可靠的问题，并提出三层验证框架——回执、读回校验和独立检查——以证据驱动的方式定义何为真正的完成。
Aaron Levie argues that autonomous, domain-specific agents represent the emerging form factor for AI, driven by better tool/computer use, improving model capabilities, and mature agent harnesses and infrastructure — positioning them as a major leap beyond simple LLM wrappers for enterprise workflows.
MMedAgent-Plus is a multimodal medical tool-use agent that pairs a vision-language planner with 18 specialist models across 12 medical tasks and 8 modalities, orchestrating multi-step clinical workflows. It achieves 90.6% correct specialist invocation and 0.98 final-answer accuracy, far exceeding the strongest baseline (0.28).
该论文提出 TURTLE，一种无监督环境设计方法，通过 proposer、solver 和 feedback generator 自动演化出自我调优的长时程工具使用基准，可在 AppWorld、τ-Bench 等评测中提升任务难度、多样性并揭示更广的失败模式。
PROVISO 是一个在 LLM agent 执行动作前进行溯源感知决策的控制器，通过将可信与外部观察编码为溯源类型记忆来平衡安全性与实用性，在多个 agent 安全基准上显著优于现有基线。
A design discussion asking when an agent should combine readings from two read-only tools that lack a shared snapshot, and what response contract should govern qualified availability estimates versus reporting readings separately while preserving uncertainty.
该论文提出 BRIA（Branch–Inspect–Adopt）智能体，将候选工具的探索隔离在独立分支中并依据执行证据选择性采纳，从而在大型工具库（1897 个工具）上实现更优的工具检索与使用，在 MCP-Atlas 上比 MCP-Zero 提升 14.37 个百分点。
We present AuditKIT, a source-available evaluation library that unifies benchmark accuracy, judge-based assessment, runtime performance, tool-use behavior, and retrieval quality in a single result representation with direction-aware metric contracts and extensible backends. Seven case studies show how answer extraction, runtime behavior, and metric selection can change evaluation outcomes for LLMs and agentic systems.
MS-Eval 提出一个统一评估套件，用于量化主代理委派与子代理执行（MAD-SAE）范式在 MCP 密集场景中的表现，对四个 MAD-SAE 变体和八个前沿模型的评测显示子代理专业化最优，开源模型整体表现有竞争力，但委派通常增加成本且收益不确定。
OpenVLHarness 是一个多模态推理框架，通过能力层、持久化多模态记忆与编码接口实现异构工具组合，在 22 个基准上相较无工具基线平均提升 9.7-12.6 分，同时显著降低 token 成本。
The paper introduces HybridAgents, an LLM agent architecture that dynamically switches between latent-space reasoning and text-based action via a lightweight HybridLink, combined with HybridSFT and HybridRL post-training, achieving 10.4% performance gains with large token, latency, and cost reductions across 11 benchmarks.
A discussion of how web-research agents should decide when to stop searching, breaking the stopping decision into coverage, progress, and budget checks, and arguing for returning unresolved questions rather than false confidence. The author asks practitioners what stopping rules they actually use for research agents.
Muse Glimmer 是一款 30B 参数的开源多模态智能体模型，通过 4-bit 量化可在消费级硬件上本地运行多步推理与工具调用，并在同尺寸级别 agentic 基准上超越 Gemma4-31B 和 Qwen3.6-27B，现以 Apache 2.0 协议发布。
A small company running an AI sales agent (Claude + tool use) recounts how a troll tested their bot with a fake email, and how they fixed validation with a code-level DNS gate plus several prompt/tool-result bugs — concluding that tool-result text acts as a prompt itself.
Composio 报告称 Cline 搭配 GPT 6 Astra 在其状态化工具使用基准测试中取得 23/29 的成绩，是其测试过的 harness 中最高分。