val-blog/org/cases/arxiv_digest/output/latest.md

ArXiv Daily Brief - 2026-03-09

🧠 今日 Top 3（中文可读版）

1. SurvHTE-Bench: A Benchmark for Heterogeneous Treatment Effect Estimation in Survival Analysis
   • 中文题目（意译）: SurvHTE-Bench: A 基准 for Heterogeneous Treatment Effect Estimation in Survival Analysis
   • 这篇在讲什么: Estimating heterogeneous treatment effects (HTEs) from right-censored survival data is critical in high-stakes applications such as preci...
   • 它怎么做: 引入了 SurvHTE-Bench, the first comprehensive 基准测试 for HTE estimation with censored outcomes.
   • 得出了什么结果: Estimating heterogeneous treatment effects (HTEs) from right-censored survival data is critical in high-stakes applications such as preci...
   • 可能的影响: Estimating heterogeneous treatment effects (HTEs) from right-censored survival data is critical in high-stakes applications such as preci...
   • arXiv: http://arxiv.org/abs/2603.05483v1
2. Accelerating Text-to-Video Generation with Calibrated Sparse Attention
   • 中文题目（意译）: 加速 Text-to-Video 生成 with Calibrated Sparse Attention
   • 这篇在讲什么: Recent 扩散 模型s enable high-quality 视频生成, but suffer from slow runtimes.
   • 它怎么做: Motivated by this, 引入了 CalibAtt, a 训练-free method that accelerates 视频生成 via calibrated sparse attention.
   • 得出了什么结果: Extensive experiments on Wan 2.1 14B, Mochi 1, and few-step distilled 模型s at various resolutions show that CalibAtt achieves up to 1.58x ...
   • 可能的影响: Recent 扩散 模型s enable high-quality 视频生成, but suffer from slow runtimes.
   • arXiv: http://arxiv.org/abs/2603.05503v1
3. Observing and Controlling Features in Vision-Language-Action Models
   • 中文题目（意译）: Observing and 控制 特征 in 视觉-语言-动作 模型
   • 这篇在讲什么: 视觉-语言-动作 模型s (VLAs) have shown remarkable progress towards embodied intelligence.
   • 它怎么做: In this work, 提出了 to close this gap by introducing and analyzing two main concepts: feature-observability and feature-controllability.
   • 得出了什么结果: Our 结果显示 that targeted, lightweight interventions can reliably steer a robot's behavior while preserving closed-loop capabilities.
   • 可能的影响: 视觉-语言-动作 模型s (VLAs) have shown remarkable progress towards embodied intelligence.
   • arXiv: http://arxiv.org/abs/2603.05487v1

🔥 今日热度 Top 5（新鲜度+关键词+HN提及+代码线索）

1. SurvHTE-Bench: A Benchmark for Heterogeneous Treatment Effect Estimation in Survival Analysis
   • arXiv: http://arxiv.org/abs/2603.05483v1
   • 类别: cs.LG | HotScore: 23.0 | 作者: Shahriar Noroozizadeh, Xiaobin Shen, Jeremy C. Weiss
   • 速读: SurvHTE-Bench: A Benchmark for Heterogeneous Treatment Effect Estimation in Survival An...（cs.LG）
2. Accelerating Text-to-Video Generation with Calibrated Sparse Attention
   • arXiv: http://arxiv.org/abs/2603.05503v1
   • 类别: cs.CV | HotScore: 18.0 | 作者: Shai Yehezkel, Shahar Yadin, Noam Elata
   • 速读: Accelerating Text-to-Video Generation with Calibrated Sparse Attention（cs.CV）
3. Observing and Controlling Features in Vision-Language-Action Models
   • arXiv: http://arxiv.org/abs/2603.05487v1
   • 类别: cs.RO | HotScore: 18.0 | 作者: Hugo Buurmeijer, Carmen Amo Alonso, Aiden Swann
   • 速读: Observing and Controlling Features in Vision-Language-Action Models（cs.RO）
4. Towards Provably Unbiased LLM Judges via Bias-Bounded Evaluation
   • arXiv: http://arxiv.org/abs/2603.05485v1
   • 类别: cs.AI | HotScore: 17.0 | 作者: Benjamin Feuer, Lucas Rosenblatt, Oussama Elachqar
   • 速读: Towards Provably Unbiased LLM Judges via Bias-Bounded Evaluation（cs.AI）
5. An interpretable prototype parts-based neural network for medical tabular data
   • arXiv: http://arxiv.org/abs/2603.05423v1
   • 类别: cs.LG | HotScore: 17.0 | 作者: Jacek Karolczak, Jerzy Stefanowski
   • 速读: An interpretable prototype parts-based neural network for medical tabular data（cs.LG）

🆕 最新上新 Top 10

1. Transformer-Based Inpainting for Real-Time 3D Streaming in Sparse Multi-Camera Setups (cs.CV) - http://arxiv.org/abs/2603.05507v1
2. FaceCam: Portrait Video Camera Control via Scale-Aware Conditioning (cs.CV) - http://arxiv.org/abs/2603.05506v1
3. RoboPocket: Improve Robot Policies Instantly with Your Phone (cs.RO) - http://arxiv.org/abs/2603.05504v1
4. Accelerating Text-to-Video Generation with Calibrated Sparse Attention (cs.CV) - http://arxiv.org/abs/2603.05503v1
5. POET-X: Memory-efficient LLM Training by Scaling Orthogonal Transformation (cs.LG) - http://arxiv.org/abs/2603.05500v1
6. The Spike, the Sparse and the Sink: Anatomy of Massive Activations and Attention Sinks (cs.AI) - http://arxiv.org/abs/2603.05498v1
7. Safe-SAGE: Social-Semantic Adaptive Guidance for Safe Engagement through Laplace-Modulated Poisson Safety Functions (cs.RO) - http://arxiv.org/abs/2603.05497v1
8. Cheap Thrills: Effective Amortized Optimization Using Inexpensive Labels (cs.LG) - http://arxiv.org/abs/2603.05495v1
9. Censored LLMs as a Natural Testbed for Secret Knowledge Elicitation (cs.LG) - http://arxiv.org/abs/2603.05494v1
10. cuRoboV2: Dynamics-Aware Motion Generation with Depth-Fused Distance Fields for High-DoF Robots (cs.RO) - http://arxiv.org/abs/2603.05493v1

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）。