Recent Posts

Reward Design as a Procedure: From Objectives to Causal Coverage to a Polytope of Weights

把奖励设计变成一道流程：从目标到因果覆盖再到权重多面体

8/14/2026

arxivaics.lg

AVA-Encoder: Turning Films into Knowledge Graphs an Agent Can Actually Edit

AVA-Encoder：把电影压成智能体能改的知识图谱

8/14/2026

arxivaics.cl

CAM's Family Tree: A Method-Centered Map of Visual Explanation from CNNs to Foundation Models

CAM 的家谱：从 CNN 到基础模型的视觉解释方法地图

8/14/2026

arxivaics.ai

DARTree: Teaching a Chain-Trained Corrector to Think in Trees

DARTree：让只会走直线的纠正头学会长成树

8/14/2026

arxivaics.lg

DreamFly: Teaching a Drone to Remember, Rehearse, and Know When to Stop

DreamFly：让无人机学会记忆、预演，以及知道何时停下

8/14/2026

arxivaics.ai

Uncertainty as an Input, Not an Afterthought: LLM Sentiment Meets Small-Cap Portfolio Construction

把不确定性当作输入，而不是事后补丁：LLM 情绪信号与小盘股组合构建

8/14/2026

arxivaiq-fin.pm

LittleLearner: What Happens When You Cap a Language Model at Fifth Grade

LittleLearner：把语言模型的知识天花板锁在小学五年级

8/14/2026

arxivaics.cl

Softmax Attention Was Always a Born-Rule Object (If You Live on the Simplex)

当数据活在概率单形上，Softmax 注意力本来就是玻恩规则的产物

8/13/2026

arxivaiquant-ph

AdvFD: When Your Fréchet Loss Gets Gamed, Make the Metric Fight Back

AdvFD：当 Fréchet 损失被刷分，就让度量自己反击

8/13/2026

arxivaics.cv

Building the Jig Instead of Training the Hands: Test-Time Capability Transfer via Harnesses

不训练学生，只造夹具：用 Harness 在推理期完成能力迁移

8/13/2026

arxivaics.lg

CausalSplat: Teaching Gaussian Splats to Answer 'Why', Not Just 'Where

CausalSplat：让高斯泼溅从「指哪打哪」走向「听懂话里的意思」

8/13/2026

arxivaics.cv

Conditional Independence Tests as the Weak Link in Causal Discovery

条件独立性检验：因果发现流水线上最脆弱的一环

8/13/2026

arxivaistat.ml