今日从 arXiv 订阅中筛选 10 篇论文。

Arxiv ID2609.13257幻觉翻译2609.13257

⚡ Sampling headroom is not selection gain: a compute-value audit of test-time scaling for video world models

test-time scaling 的 compute-value 审计：sampling headroom ≠ selection gain。

Sampling headroom is not selection gain: a compute-value audit of test-time scaling for video world models
Arxiv ID2609.15160幻觉翻译2609.15160

⚡ T-LoopFormer: Token-Level Elastic-Depth Looped Transformers for Latent Reasoning With Dynamic Routing

looped Transformer 的 token 级弹性深度：简单 token 早退、难 token 深递归。

T-LoopFormer: Token-Level Elastic-Depth Looped Transformers for Latent Reasoning With Dynamic Routing
Arxiv ID2609.15980幻觉翻译2609.15980

⚡ A Chosen Future Can Still Be Rewritten: Causal Writability in Video Models

视频模型的 causal writability：物理错误是"没学到"还是"学到但没用"？答案：后者。

A Chosen Future Can Still Be Rewritten: Causal Writability in Video Models
Arxiv ID2609.13747幻觉翻译2609.13747

⚡ How Many Thoughts Can a Vector Hold? The Capacity of Reasoning by Superposition

一个向量能装多少 thought：latent 推理的叠加容量理论。

Arxiv ID2609.14073幻觉翻译2609.14073

⚡ LPA-CWM: A Learned Physical Adjudicator for Motion Reasoning with Counterfactual World Models

反事实世界模型的"物理仲裁者"：学会给不同干预响应加权。

LPA-CWM: A Learned Physical Adjudicator for Motion Reasoning with Counterfactual World Models
Arxiv ID2609.13194幻觉翻译2609.13194

⚡ Diagnosing Faults in Reinforcement Learning Simulators and World Models with Canonical Polynomial Invariants

物理结构到底帮不帮预测：精确不变量测试"物理结构有益论"。

Arxiv ID2609.15169幻觉翻译2609.15169

⚡ GRAVA: Grounded Reasoning-to-Action Representation and Learning for Autonomous Driving

驾驶 VLA 的 grounded reasoning-to-action：推理必须锚回场景证据与可执行动作。

GRAVA: Grounded Reasoning-to-Action Representation and Learning for Autonomous Driving
Arxiv ID2609.15322幻觉翻译2609.15322

⚡ Planning in the Backbone: DiffAdapterVLA for Native Continuous Trajectory Generation with Driving VLMs

Planning in the Backbone：把轨迹生成放进驾驶 VLM 的主干计算里。

Planning in the Backbone: DiffAdapterVLA for Native Continuous Trajectory Generation with Driving VLMs
Arxiv ID2609.15224幻觉翻译2609.15224

⚡ Long-to-Short Video Evidence Reasoning for Grounded Question Answering

长到短视频证据推理：按证据时长组织课程，IoP 替代 IoU。

Long-to-Short Video Evidence Reasoning for Grounded Question Answering
Arxiv ID2609.13984幻觉翻译2609.13984

⚡ What Makes an Efficient VLA? Navigating Action-Head Design, Scaling, and Latency

什么让 VLA 高效：action-head 的初始化对齐 > 架构设计。

What Makes an Efficient VLA? Navigating Action-Head Design, Scaling, and Latency

自动生成于 2026-09-16 · 基于 arXiv Daily Digest