增强多模态模型理解图表和文档的能力、提升模型推理、规划和验证能力、让 Agent 更可靠地调用工具和复用技能
今天主要跟进：增强多模态模型理解图表和文档的能力、提升模型推理、规划和验证能力、让 Agent 更可靠地调用工具和复用技能。
本期从 2026-06-30 论文源抓取并去重 453 篇候选论文，筛选 6 篇重点论文与 20 篇补充关注。
本期重点
- 1UniCoder：Unified Visual-to-Code Generation via Symbolic Rewards 与 Reference-Guided Code Optimization🔗
- 2Technical Report of RoboSpatial Challenge at CVPR 2026：Selective Reasoning Activation 与 Reference-Frame Disambiguation 面向 Embodied Spatial Reasoning🔗
- 3HealthAgentBench：A Unified 基准 Suite of Realistic Agentic Healthcare Environments 面向 Challenging Frontier AI Agents🔗
- 4SkillSpotter：Pose-Aware Multi-View Skilled Action Detection 与 Grading in Ego-Exo Videos🔗
- 5CoMet：Context 与 Multiplicity Decomposition 面向 Multimodal Uncertainty Estimation🔗
- 6OopsieVerse：A Safety 基准 with Damage-Aware Simulation 面向 Robot Manipulation🔗
今天最值得跟进的方向
今天的高分论文主要指向：增强多模态模型理解图表和文档的能力、提升模型推理、规划和验证能力、让 Agent 更可靠地调用工具和复用技能。下面按核心问题、方法线索、主要论点和关键词整理，便于快速判断后续跟进价值。
重点论文：核心问题、方法线索与关键词
增强多模态模型理解图表和文档的能力
中文标题UniCoder：Unified Visual-to-Code Generation via Symbolic Rewards 与 Reference-Guided Code Optimization
信号显示Visual-to-Code generation, which transforms scientific plots, vector graphics, and webpages into executable scripts, demands a level of pixel-precise alignment that standard Multimodal 大语言模型 (MLLMs) fail to achieve through Supervised Fine-Tunin
关键词alignmentbenchmarkcodemultimodal
代码/数据需查看原文确认
提升模型推理、规划和验证能力
中文标题Technical Report of RoboSpatial Challenge at CVPR 2026：Selective Reasoning Activation 与 Reference-Frame Disambiguation 面向 Embodied Spatial Reasoning
信号显示该研究围绕题名与摘要所揭示的问题展开，具体信号需结合原文进一步核验：Vision-language models achieve strong general perception but often struggle with the spatial reasoning required for embodied tasks
关键词inferencecodevision-languagefine-tuning
代码/数据需查看原文确认
让 Agent 更可靠地调用工具和复用技能
中文标题HealthAgentBench：A Unified 基准 Suite of Realistic Agentic Healthcare Environments 面向 Challenging Frontier AI Agents
信号显示该研究围绕题名与摘要所揭示的问题展开，具体信号需结合原文进一步核验：As AI agents become increasingly capable of complex, long-horizon reasoning, rigorous and holistic evaluation is essential for measuring progress toward real-world healthcare applications
关键词agentworkflowevaluationbenchmark
代码/数据需查看原文确认
提升 RAG 检索和知识库问答可靠性
中文标题SkillSpotter：Pose-Aware Multi-View Skilled Action Detection 与 Grading in Ego-Exo Videos
信号显示该研究围绕题名与摘要所揭示的问题展开，具体信号需结合原文进一步核验：To enable personalized, real-time coaching using Augmented Reality glasses or fixed camera setups in domains such as sports, cooking, or music, a system must understand not just what a person does, but how well they execute an activity
关键词ragevaluationbenchmarkcode
代码/数据需查看原文确认
增强多模态模型理解图表和文档的能力
中文标题CoMet：Context 与 Multiplicity Decomposition 面向 Multimodal Uncertainty Estimation
信号显示该研究围绕题名与摘要所揭示的问题展开，具体信号需结合原文进一步核验：Uncertainty estimation has been a long-standing challenge in AI models; it amounts to "knowing what you don't know," and metacognition is notoriously difficult even for humans (cf
关键词benchmarkcodemultimodalvisual
代码/数据需查看原文确认
提升代码生成、执行反馈和自动修复能力
中文标题OopsieVerse：A Safety 基准 with Damage-Aware Simulation 面向 Robot Manipulation
信号显示该研究围绕题名与摘要所揭示的问题展开，具体信号需结合原文进一步核验：While robotic manipulation capabilities have advanced rapidly, physical safety remains a major barrier to deploying household robots: task success is insufficient if the robot damages itself or its surroundings
关键词safetyevaluationbenchmarkcode
代码/数据需查看原文确认
其他值得关注
MECoBench: A Systematic Study of Multimodal Agent Collaboration in Embodied Environments
中文标题MECoBench：A Systematic Study of Multimodal Agent Collaboration in Embodied Environments
关注理由涉及工具调用、执行反馈和可复用能力，可作为 Agent 工作流可靠性的补充线索。
Evo-PI: Aligning Medical Reasoning via Evolving Principle-Guided Supervision
中文标题Evo-PI：Aligning Medical Reasoning 通过 Evolving Principle-Guided Supervision
关注理由涉及任务设置、指标和失效案例，可补充模型评测与回归测试。
FedXDS: Leveraging Model Attribution Methods to counteract Data Heterogeneity in Federated Learning
中文标题FedXDS：Leveraging Model Attribution Methods to counteract Data Heterogeneity in Federated Learning
关注理由涉及检索、知识库问答与证据可靠性，可作为 RAG 评测和企业知识系统的补充线索。
Diffusing Blame: Task-Dependent Credit Assignment in Biologically Plausible Dual-Stream Networks
中文标题Diffusing Blame：Task-Dependent Credit Assignment in Biologically Plausible Dual-Stream Networks
关注理由涉及任务设置、指标和失效案例，可补充模型评测与回归测试。
中文标题Preserve the Hard，Regenerate the Rest：Uncertainty-Guided Synthetic Training Data Augmentation 结合 Diffusion Models
关注理由涉及视觉与图像生成中的新任务、数据或系统线索，可作为后续跟进清单的一部分。
CSTrader: A Testbed for Language-Grounded Trading in a Community-Driven Virtual Asset Market
中文标题CSTrader：A Testbed 面向 Language-Grounded Trading in a Community-Driven Virtual Asset Market
关注理由涉及工具调用、执行反馈和可复用能力，可作为 Agent 工作流可靠性的补充线索。
中文标题Distilling Temporal Coherence into 2D Networks 面向 Transrectal Ultrasound Prostate Video Segmentation
关注理由涉及视频生成中的新任务、数据或系统线索，可作为后续跟进清单的一部分。
Fleet: Few Shots Lead Effective AI-generated Image Detection
中文标题Fleet：Few Shots Lead Effective AI-generated Image Detection
关注理由涉及代码智能中的新任务、数据或系统线索，可作为后续跟进清单的一部分。
ADAPT: Attention Dynamics Alignment with Preference Tuning for Faithful MLLMs
中文标题ADAPT：Attention Dynamics Alignment with Preference Tuning 面向 Faithful MLLMs
关注理由涉及模型安全、护栏路由、风险分类或治理评测，可作为安全评测与治理工具链的补充线索。
Certified Speculative Execution for Untrusted AI Agents
中文标题Certified Speculative Execution 面向 Untrusted AI Agents
关注理由涉及工具调用、执行反馈和可复用能力，可作为 Agent 工作流可靠性的补充线索。
GEAR: Guided End-to-End AutoRegression for Image Synthesis
中文标题GEAR：Guided End-to-End AutoRegression 面向 Image Synthesis
关注理由涉及检索、知识库问答与证据可靠性，可作为 RAG 评测和企业知识系统的补充线索。
Harnessing Textual Refusal Directions for Multimodal Safety
中文标题Harnessing Textual Refusal Directions 面向 Multimodal Safety
关注理由涉及模型安全、护栏路由、风险分类或治理评测，可作为安全评测与治理工具链的补充线索。
中文标题AeroVerse-SatAgent：UAV-Satellite Collaborative Spatial Reasoning Inspired by the Dual Visual Pathway Theory of Cognitive Neuroscience
关注理由涉及工具调用、执行反馈和可复用能力，可作为 Agent 工作流可靠性的补充线索。
Transformers as Bayesian In-Context Experimenters: Smoothness-Adaptive Efficient ATE Estimation
中文标题Transformers as Bayesian In-Context Experimenters：Smoothness-Adaptive Efficient ATE Estimation
关注理由涉及检索、知识库问答与证据可靠性，可作为 RAG 评测和企业知识系统的补充线索。
Beyond Single Character: Evaluating MLLMs for Sentence-Level Oracle Bone Inscription Understanding
中文标题Beyond Single Character：Evaluating MLLMs 面向 Sentence-Level Oracle Bone Inscription Understanding
关注理由涉及任务设置、指标和失效案例，可补充模型评测与回归测试。
An Empirical Study of Security Calibration in Large Language Models for Code
中文标题一种Empirical Study of Security Calibration in Large Language Models 面向 Code
关注理由涉及代码智能中的新任务、数据或系统线索，可作为后续跟进清单的一部分。
PPT-Eval: A Benchmark for Computer-Use Agents on PowerPoint Tasks
中文标题PPT-Eval：A 基准 面向 Computer-Use Agents on PowerPoint Tasks
关注理由涉及工具调用、执行反馈和可复用能力，可作为 Agent 工作流可靠性的补充线索。
A Modular Vision-Language-Action Robotics Framework for Indoor Environments
中文标题一种Modular Vision-Language-Action Robotics 框架 面向 Indoor Environments
关注理由涉及工具调用、执行反馈和可复用能力，可作为 Agent 工作流可靠性的补充线索。
Cross-Space Distillation: Teaching One-Step Students with Modern Diffusion Teachers
中文标题Cross-Space Distillation：Teaching One-Step Students 结合 Modern Diffusion Teachers
关注理由涉及推理成本、延迟、吞吐和部署约束，可补充系统优化方向。
FedLAB: Traceable Semantic Codebooks for Federated Multimodal Graph Foundation Learning
中文标题FedLAB：Traceable Semantic Codebooks 面向 Federated Multimodal Graph Foundation Learning
关注理由涉及多模态模型中的新任务、数据或系统线索，可作为后续跟进清单的一部分。
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。