Tag
A one-week case study embeds Claude Code and Codex CLI as party members on an offline WoW 3.3.5a server, yielding design lessons like 'sight without authority', typed action vocabularies, explicit refusal signals, and schema design that follows natural speech.
An analysis of AWS Bedrock Guardrails docs for the Converse API showing that guardrails, including prompt attack detection, do not evaluate tool results, tool definitions, or tool-call arguments — leaving indirect prompt injection in and exfiltration out unguarded. The author recommends deterministic allow-list checks on tool arguments, human confirmation for write actions, and treating tool results as untrusted input.
A user reports that Qwen 3.8 27b, a local model, autonomously discovered it could access Blender via MCP and wrote Python code to speed up its math on a physics problem — behavior the user says they've never seen from a local model before.
Introduces a 400-question benchmark evaluating when commercial LLMs choose to search the web versus answering from internal memory, finding that models like GPT-5 mini and Grok 4 Fast search appropriately on outside-knowledge questions while Mistral Small often abstains instead, with search behavior heavily influenced by explicit temporal cues.
Apriel Voice 27B Preview 是一个开源语音-语言模型，基于 Gemma 4 26B-A4B 后训练检查点构建，通过两阶段训练将语音理解、推理与工具调用统一到共享推理骨干中，并在 Big Bench Audio、Speech-BFCL 和 τ-Voice 等基准上取得有竞争力的成绩。
LeanGuard is a runtime that verifies every tool call from language-model agents against Lean4 policies with temporal logic semantics, cutting forbidden calls and prompt-injection success while enabling proof-guided repair of denied actions.
A 135-page survey paper organizing agentic reasoning for large language models along three layers—foundational, self-evolving, and collective multi-agent reasoning—spanning in-context and post-training optimization with real-world applications and open challenges.
An analysis of Claude Fable 5.1's benchmark results showing it is a software engineering and tool-use powerhouse, excelling at codebase comprehension, test writing, and refactoring, while lagging in legal reasoning, financial strategy, and crisis-support applications.
IBM SIL introduces JevAct, a runtime decision-gate framework that intercepts tool-using LLM agent actions, verifies policy compliance via compiled code checks and a prefill-only decision model, and feeds back corrections—boosting Opus 4.6's τ²-bench Telecom pass^3 success from 39.5% to 80.7% without modifying models or tools.
The author has turned the “Build an Agent from Scratch” tutorial into NamiAgent, an open-source project that serves as both an 11-lesson tutorial with text, images, and code and a locally runnable Agent. It supports reading and writing files, executing commands, MCP integration, long-term memory, and web-based visualization, with only one dependency: the openai package.
ServiceNow 释放了 Apriel Voice 27B 预览版，一个基于 Gemma 4 骨干的开源语音-推理-工具调用模型，支持语音理解、多模态上下文和函数调用，可搭配外部语音合成器构建语音助手。
ServiceNow AI presents Apriel Voice 27B Preview, an open speech-language model built on Gemma 4 26B-A4B that unifies speech understanding, reasoning, and tool use in a shared backbone, trained with a two-stage procedure and online distillation to reach 98.90 on Big Bench Audio and 81.13 on Speech-BFCL with 1.91s mean latency.
论文提出 CASEL（Consequence-Aware Selective Execution Layer），一种模型无关的运行时分配策略，通过综合校准失败概率、后果严重性、可逆性、歧义性和证据新鲜度来为工具调用选择执行、验证、沙箱、澄清、审批或放弃等模式，仿真结果显示其将有害副作用轨迹比例从 14.75% 降至 7.02%。
该论文提出技能适配（skill adaptation）方法，通过元技能引导外部改写器将现有技能重写以匹配执行模型的能力，从而提升弱模型在 harness 自演化中的表现，并在七个模型上验证了改进效果。
该论文提出了 Recursive Semantic Divergence (RSD)，一种基于强化学习双模拟度量的无监督轨迹级一致性指标，用于衡量 LLM 智能体在相同状态下多次独立执行时的行为差异；将其作为奖励信号进行微调可显著降低任务进展方差并提升性能。
A study of execution-time reliability in tool-using AI agents that isolates recovery decisions from language-model behavior, evaluating checkpoint state, dispatch authority, external-effect evidence, and reconciliation obligations across 336 histories on LangGraph and Temporal hosts, plus a component study of Arya.ai's InfinityOS.
SCI-AGENT-EXAM is a workflow-oriented benchmark of 120 expert-curated tasks across 11 scientific fields and 7 research task types, evaluating LLM agents with a hybrid protocol combining programmatic verification and LLM-based judging. Experiments on 10 frontier agents show GPT-6 Astra leads with 80.9 overall but only 26% of runs exceed the 0.95 success threshold, revealing significant capability and stability gaps.
This paper audits residual activation interventions in Qwen3.5-27B on tool-use tasks, finding that while such edits boost target-selection accuracy, the gains fail to generalize to recombined facts and don't extend to complete-action correctness, separating target-selection control from condition-sensitive reasoning.
Omni-InteractScope introduces a benchmark of 1,479 human-annotated interactions across eight task groups to evaluate how reliably omnimodal LLMs deliver timely tool use, query-free reminders, and sustained guidance in continuous audiovisual streams, finding current models fall far short.
该论文提出 MICA 与 DART 框架，研究在可执行能力相同的情况下，MCP 与 CLI 接口表示差异如何显著改变语言智能体的行为，并通过针对性修复接口实现提升性能、降低交互成本。