val-blog/org/cases/arxiv_digest/output/2026-03-12.md

ArXiv Daily Brief - 2026-03-12

🧠 今日 Top 3（中文可读版）

1. MedMASLab: A Unified Orchestration Framework for Benchmarking Multimodal Medical Multi-Agent Systems
   • 中文题目（意译）: MedMASLab: A Unified Orchestration Framework for 基准ing Multimodal 医疗 Multi-Agent Systems
   • 这篇在讲什么: Current medical MAS research suffers from non-uniform data ingestion pipelines, inconsistent visual-reasoning evaluation, and a lack of c...
   • 它怎么做: To address these challenges, 提出了 MedMASLab, a unified framework and 基准测试ing platform for multimodal medical multi-agent systems.
   • 得出了什么结果: Our systematic evaluation reveals a critical domain-specific performance gap: while MAS improves reasoning depth, current architectures e...
   • 可能的影响: MedMASLab introduces: (1) A standardized multimodal agent communication protocol that enables seamless integration of 11 heterogeneous MA...
   • arXiv: http://arxiv.org/abs/2603.09909v1
2. PathMem: Toward Cognition-Aligned Memory Transformation for Pathology MLLMs
   • 中文题目（意译）: PathMem: Toward Cognition-Aligned Memory Transformation for Pathology MLLMs
   • 这篇在讲什么: Computational pathology demands both visual pattern recognition and dynamic integration of structured domain knowledge, including taxonom...
   • 它怎么做: Inspired by the hierarchical memory process of human pathologists, 提出了 PathMem, a memory-centric multimodal framework for pathology MLLMs.
   • 得出了什么结果: PathMem achieves SOTA performance across 基准测试s, improving WSI-Bench report generation (12.8% WSI-Precision, 10.1% WSI-Relevance) and open...
   • 可能的影响: Computational pathology demands both visual pattern recognition and dynamic integration of structured domain knowledge, including taxonom...
   • arXiv: http://arxiv.org/abs/2603.09943v1
3. MSSR: Memory-Aware Adaptive Replay for Continual LLM Fine-Tuning
   • 中文题目（意译）: MSSR: Memory-Aware Adaptive Replay for Continual LLM Fine-Tuning
   • 这篇在讲什么: Existing replay-based strategies, such as fixed interleaved replay, accuracy-supervised, and loss-driven scheduling, remain limited: some...
   • 它怎么做: Motivated by retention dynamics under sequential fine-tuning, 提出了 Memory-Inspired Sampler and Scheduler Replay (MSSR), an experience repl...
   • 得出了什么结果: Existing replay-based strategies, such as fixed interleaved replay, accuracy-supervised, and loss-driven scheduling, remain limited: some...
   • 可能的影响: While strong adaptability enables rapid acquisition of new knowledge, it also exposes LLMs to catastrophic forgetting, where previously l...
   • arXiv: http://arxiv.org/abs/2603.09892v1

🔥 今日热度 Top 5（新鲜度+关键词+HN提及+代码线索）

1. MedMASLab: A Unified Orchestration Framework for Benchmarking Multimodal Medical Multi-Agent Systems
   • arXiv: http://arxiv.org/abs/2603.09909v1
   • 类别: cs.AI | HotScore: 64.96 | 作者: Yunhang Qian, Xiaobin Hu, Jiaquan Yu
   • 速读: MedMASLab: A Unified Orchestration Framework for Benchmarking Multimodal Medical Multi-...（cs.AI）
2. PathMem: Toward Cognition-Aligned Memory Transformation for Pathology MLLMs
   • arXiv: http://arxiv.org/abs/2603.09943v1
   • 类别: cs.AI | HotScore: 57.41 | 作者: Jinyue Li, Yuci Liang, Qiankun Li
   • 速读: PathMem: Toward Cognition-Aligned Memory Transformation for Pathology MLLMs（cs.AI）
3. MSSR: Memory-Aware Adaptive Replay for Continual LLM Fine-Tuning
   • arXiv: http://arxiv.org/abs/2603.09892v1
   • 类别: cs.LG | HotScore: 56.77 | 作者: Yiyang Lu, Yu He, Jianlong Chen
   • 速读: MSSR: Memory-Aware Adaptive Replay for Continual LLM Fine-Tuning（cs.LG）
4. From Data Statistics to Feature Geometry: How Correlations Shape Superposition
   • arXiv: http://arxiv.org/abs/2603.09972v1
   • 类别: cs.LG | HotScore: 55.74 | 作者: Lucas Prieto, Edward Stevinson, Melih Barsbey
   • 速读: From Data Statistics to Feature Geometry: How Correlations Shape Superposition（cs.LG）
5. Think Before You Lie: How Reasoning Improves Honesty
   • arXiv: http://arxiv.org/abs/2603.09957v1
   • 类别: cs.AI | HotScore: 55.65 | 作者: Ann Yuan, Asma Ghandeharioun, Carter Blum
   • 速读: Think Before You Lie: How Reasoning Improves Honesty（cs.AI）

🆕 最新上新 Top 10

1. Task Aware Modulation Using Representation Learning for Upsaling of Terrestrial Carbon Fluxes (cs.LG) - http://arxiv.org/abs/2603.09974v1
2. From Data Statistics to Feature Geometry: How Correlations Shape Superposition (cs.LG) - http://arxiv.org/abs/2603.09972v1
3. TiPToP: A Modular Open-Vocabulary Planning System for Robotic Manipulation (cs.RO) - http://arxiv.org/abs/2603.09971v1
4. CREATE: Testing LLMs for Associative Creativity (cs.CL) - http://arxiv.org/abs/2603.09970v1
5. ReCoSplat: Autoregressive Feed-Forward Gaussian Splatting Using Render-and-Compare (cs.CV) - http://arxiv.org/abs/2603.09968v1
6. Understanding the Use of a Large Language Model-Powered Guide to Make Virtual Reality Accessible for Blind and Low Vision People (cs.HC) - http://arxiv.org/abs/2603.09964v1
7. Emotional Modulation in Swarm Decision Dynamics (cs.MA) - http://arxiv.org/abs/2603.09963v1
8. BEACON: Language-Conditioned Navigation Affordance Prediction under Occlusion (cs.RO) - http://arxiv.org/abs/2603.09961v1
9. Think Before You Lie: How Reasoning Improves Honesty (cs.AI) - http://arxiv.org/abs/2603.09957v1
10. Kinodynamic Motion Retargeting for Humanoid Locomotion via Multi-Contact Whole-Body Trajectory Optimization (cs.RO) - http://arxiv.org/abs/2603.09956v1

Val 今日建议

• 先读 Top 5 里的 1-2 篇，优先看是否有可直接复用的方法/代码。
• 若你愿意，我下一步可对 Top 3 产出“中文三段式精读卡”（问题-方法-可落地点）。