AI Reasoning, Interpretability & World Models
Key Questions
How much funding has the world-model sector attracted?
33 companies in the world-model track have raised substantial capital, with clear US-China-Europe patterns and IPO trends. Funds are flowing into system-level competition and embodied intelligence.
What were the main findings at ICML 2026 on world models?
LAWM and WAM research routes showed convergence, with advances like Causal-JEPA and CoLA-World improving counterfactual reasoning and long-term consistency. The field is rapidly maturing.
Which new physical world model was released?
Jiying Technology unveiled a solid-mechanics physics foundation model capable of zero-shot generalization. It aims to embed physical laws directly into AI world models.
What security risk was identified in reasoning models?
Overthinking in step-by-step reasoning models creates DoS vulnerabilities. Liquid AI open-sourced Antidoom to reduce doom-loop rates through final-token preference optimization.
What benchmarks reveal limitations in current models?
ARC-AGI-3 showed GPT-5.5 and Opus 4.7 scoring below 1%, while LongCoT benchmarks found long-horizon reasoning accuracy at only 9.83%. MemoBench exposed memory blind spots after object disappearance.
What cognitive framework was proposed for future AI?
A position paper advocates evolving from LLMs to cognitive operating systems via a three-layer metacognition model. It argues current paradigms have fundamental limitations.
How are video generation and world models distinguished?
Analyses clarify that video generation does not equal world models, emphasizing the need for causal and physical understanding. Keling's ShotStream marks progress but remains distinct from true world simulation.
What RL training improvements target inference bias?
The MIPU framework reframes reasoning strategy improvement as the explicit RL objective. This addresses train-inference mismatch in long-horizon tasks.
世界模型赛道融资全景：33家公司疯狂融资，中美欧格局、技术路线、IPO趋势清晰，资本流向系统竞争与具身智能。ICML 2026世界模型研究盘点：LAWM与WAM路线融合，Causal-JEPA、CoLA-World等突破反事实推理和长时序一致性。极映科技发布固体力学物理基础模型，零样本泛化。新发现推理模型过度思考漏洞可导致DoS攻击。一篇从认知科学角度剖析LLM范式局限的文章提出'三层元认知模型'，认为AI需从LLM演进为认知操作系统。ARC-AGI-3测试GPT-5.5和Opus 4.7得分低于1%。ICML 2026 LongCoT基准显示长程推理链准确率仅9.83%。Liquid AI开源Antidoom方法降低doom loop率。MemoBench新基准揭示世界模型物体消失后记忆盲区。可灵ShotStream视频生成新突破。物理世界具身智能产业分析文章提供系统框架。MIPU框架提出将推理策略改进作为RL训练的真实目标，解决训练-推理偏差。从Vibe Coding到'意念建模'的批判性讨论引发对AI编程范式演进的思考。世界模型赛道热度持续升温，政策、资本、技术路线系统梳理，澄清视频生成≠世界模型等概念。