让 Agent 更可靠地调用工具和复用技能、识别并缓解模型安全、越狱和对齐风险
今天主要跟进：让 Agent 更可靠地调用工具和复用技能、让 Agent 更可靠地调用工具和复用技能、识别并缓解模型安全、越狱和对齐风险。
本期从 2026-06-01 论文源抓取并去重 273 篇候选论文，筛选 5 篇重点论文与 15 篇补充关注。
本期重点
- 1Cosmos 3：Omnimodal 世界模型 面向 Physical AI🔗
- 2Thinking Past the Answer：Evaluating Harmful Overthinking in Large Reasoning Models🔗
- 3Breaking the In面向mation Silo：Semantic Personas 面向 Cross-Domain Recommendation🔗
- 4K面向ge：LLM-Driven Cross-平台 Kernel Generation 面向 AI Accelerators🔗
- 5EntangleCodec：A Unified Discrete Audio Tokenizer via Semantic-Acoustic Entanglement🔗
今天最值得跟进的方向
今天的高分论文主要指向：让 Agent 更可靠地调用工具和复用技能、让 Agent 更可靠地调用工具和复用技能、识别并缓解模型安全、越狱和对齐风险。下面按核心问题、方法线索、主要论点和关键词整理。
重点论文：题目、看点与核验线索
让 Agent 更可靠地调用工具和复用技能
让 Agent 更可靠地调用工具和复用技能。核心线索：We introduce Cosmos 3, a family of omnimodal world models designed to jointly process and generate language, image, video, audio, and action sequences within a unified mixture-of-transformers architecture. 代码/数据可用性需查看原文确认。
让 Agent 更可靠地调用工具和复用技能
让 Agent 更可靠地调用工具和复用技能。核心线索：Large Reasoning Models (LRMs) improve performance by generating explicit intermediate reasoning traces through increased test-time compute, yet the assumption that longer reasoning is consistently beneficial remains under-examined. 代码/数据可用性需查看原文确认。
识别并缓解模型安全、越狱和对齐风险
识别并缓解模型安全、越狱和对齐风险。核心线索：Digital platforms increasingly operate as isolated information silos, limiting their ability to construct comprehensive user representations across domains. 代码/数据可用性需查看原文确认。
让 Agent 更可靠地调用工具和复用技能
让 Agent 更可靠地调用工具和复用技能。核心线索：Production inference increasingly targets a heterogeneous mix of accelerators. 代码/数据可用性需查看原文确认。
提升代码生成、执行反馈和自动修复能力
提升代码生成、执行反馈和自动修复能力。核心线索：Audio tokenizers serve as the discrete interface between continuous audio and Audio Language Models (ALMs), but existing tokenizers often struggle to support both understanding and generation. 代码/数据可用性需查看原文确认。
其他值得关注
Large AI Models in Dental Healthcare: From General-Purpose Systems to Domain-Specific Foundation Models：关注任务设置、指标和失效案例，适合补充模型评测与回归测试。
GloResNet: A lightweight 3D CNN with global topological features for preterm brain injury prediction：关注工具调用、执行反馈和可复用能力，适合跟进 Agent 工作流和工程可靠性。
MASER: Modality-Adaptive Specialist Routing for Embodied 3D Spatial Intelligence：关注工具调用、执行反馈和可复用能力，适合跟进 Agent 工作流和工程可靠性。
AgentRedBench: Dynamic Redteaming and Integration-Aware Defense for LLM Agents over SaaS Integrations：关注工具调用、执行反馈和可复用能力，适合跟进 Agent 工作流和工程可靠性。
OpenWebRL: Demystifying Online Multi-turn Reinforcement Learning for Visual Web Agents：关注工具调用、执行反馈和可复用能力，适合跟进 Agent 工作流和工程可靠性。
What Benchmarks Don't Measure: The Case for Evaluating Abstention Competence in Autonomous Agents：关注工具调用、执行反馈和可复用能力，适合跟进 Agent 工作流和工程可靠性。
ATLAS: A Large-Scale Evaluation Benchmark for Adversarial LiDAR Perception：关注检索、知识库问答与证据可靠性，适合跟进 RAG 评测和企业知识系统。
Tiny Collaborative Inference for Occlusion-Robust Object Detection：关注检索、知识库问答与证据可靠性，适合跟进 RAG 评测和企业知识系统。
Do Transformers Need Three Projections? Systematic Study of QKV Variants：关注推理成本、延迟、吞吐和部署约束，适合跟进系统优化。
Pathway-Structured Privileged Distillation for Deployable Computational Pathology：关注检索、知识库问答与证据可靠性，适合跟进 RAG 评测和企业知识系统。
RRISE: Robust Radius Inference via a Surrogate Estimator：关注任务设置、指标和失效案例，适合补充模型评测与回归测试。
Toward a Modular Architecture for Embedded AI Agent Systems at the Edge：关注工具调用、执行反馈和可复用能力，适合跟进 Agent 工作流和工程可靠性。
Which Defense Closes Which Threat? Attributing OWASP-LLM-Top-10 Coverage and Its Brittleness Under Paraphrasing：关注检索、知识库问答与证据可靠性，适合跟进 RAG 评测和企业知识系统。
Traj-Evolve: A Self-Evolving Multi-Agent System for Patient Trajectory Modeling in Lung Cancer Early Detection：关注工具调用、执行反馈和可复用能力，适合跟进 Agent 工作流和工程可靠性。
Acceptance-Test-Driven Evaluation Protocols for Business-Centric LLM Systems：关注工具调用、执行反馈和可复用能力，适合跟进 Agent 工作流和工程可靠性。
阅读边界
- 自动排序会偏向有社区信号、代码信号和工程关键词的论文。
- 简报默认基于标题、摘要和公开元数据，不替代全文精读。
- 外部 API 限流或不可用时，相关信号会降级为空并在内部记录中保留说明。