Yann LeCun's team is continuously advancing JEPA. Their new study reveals that the anti-collapse term in Joint Embedding Predictive Architectures (JEPAs) does more than just prevent trivial representations — it implicitly estimates data density. This means any trained JEPA
@jiqizhixin
-
Panoramic Vision Survey: Bridging Gap Between Perspective Views
By
–
#PapersAccepted by Jiqizhixin
Our report: https://
mp.weixin.qq.com/s/O-4L9pACS-kG
xTX6fCLtOA
… One Flight Over the Gap: A Survey from Perspective to Panoramic Vision Insta360 Research, University of California, San Diego, and others
Project: https://
insta360-research-team.github.io/Survey-of-Pano
rama/
…
Paper: https://
arxiv.org/pdf/2509.04444
Repo: -
Teaching AI Panoramic Vision: Survey of 300+ Works
By
–
How do we teach AI to see the world in 360°?
— 机器之心 JIQIZHIXIN (@jiqizhixin) 8 octobre 2025
A new survey dives deep into panoramic vision — exploring how models adapt from standard perspective images to omnidirectional images (ODIs) used in VR, robotics, and autonomous driving.
The paper reviews 300+ works and identifies… pic.twitter.com/tgmWmnRwJIHow do we teach AI to see the world in 360°? A new survey dives deep into panoramic vision — exploring how models adapt from standard perspective images to omnidirectional images (ODIs) used in VR, robotics, and autonomous driving. The paper reviews 300+ works and identifies
-
QuestA: Expanding LLM Reasoning via Question Augmentation
By
–
#PapersAccepted by Jiqizhixin
Our report: https://
mp.weixin.qq.com/s/Mhy9fWM8KVnu
3mTy7I3osQ
… QuestA: Expanding Reasoning Capacity in LLMs via Question Augmentation Tsinghua University, Shanghai Qi Zhi Institute, and others
Paper: https://
arxiv.org/abs/2507.13266
Models: https://
huggingface.co/foreverlasting
1202/QuestA-Nemotron-1.5B
…
Code: -
QuestA: Reinforcement Learning Improves Language Model Reasoning
By
–
Can reinforcement learning really make language models better reasoners? This study says yes — with a twist. Introducing QuestA, a Question Augmentation strategy that feeds models partial solutions during RL training to ease difficulty and deliver richer feedback. Applied to
-
TreeSynth: Diverse Data Synthesis via Tree-Guided Subspace Partitioning
By
–
#PapersAccepted by Jiqizhixin
Our report: https://
mp.weixin.qq.com/s/fp29jkzxBL7S
ku4tqlvGPw
… TreeSynth: Synthesizing Diverse Data from Scratch via Tree-Guided Subspace Partitioning The University of Hong Kong, The Chinese University of Hong Kong
Paper: https://
arxiv.org/abs/2503.17195
Code: -
Apple’s MoE Scaling Breakthrough: RoE Hyper-Parallel Inference
By
–
An intriguing paper from Apple. MoEs Are Stronger than You Think: Hyper-Parallel Inference Scaling with RoE Paper: https://
arxiv.org/abs/2509.17238 -
Meta’s Continuous Chain-of-Thought Reasoning for LLMs
By
–
Can reasoning LLMs think better if their Chain-of-Thought is continuous instead of discrete? This Meta paper introduces the first scalable way to train continuous CoTs with reinforcement learning—no need to distill from discrete references. By using "soft" tokens
-
UCLA Builds Optical Generative Model Running on Light Instead GPUs
By
–
This is huge!
— 机器之心 JIQIZHIXIN (@jiqizhixin) 2 octobre 2025
A UCLA team managed to build an optical generative model that runs on light instead of GPUs.
In their demo, a shallow encoder maps noise into phase patterns, which a free-space optical decoder then transforms into images—digits, fashion, butterflies, faces, even… pic.twitter.com/qTz43q3tcIThis is huge! A UCLA team managed to build an optical generative model that runs on light instead of GPUs. In their demo, a shallow encoder maps noise into phase patterns, which a free-space optical decoder then transforms into images—digits, fashion, butterflies, faces, even
-
Claude Code and GLM-4.6 AI Models Impress Community
By
–
Claude Code + GLM-4.6—really impressive!
— 机器之心 JIQIZHIXIN (@jiqizhixin) 30 septembre 2025
Thank you, @AnthropicAI and @Zai_org pic.twitter.com/2fzt2IffAcClaude Code + GLM-4.6—really impressive! Thank you, @AnthropicAI and @Zai_org