Wow! Researcher Andrzej Odrzywolek shows that a single operator, eml(x,y) = exp(x) – ln(y), plus the constant 1, can generate every elementary function! This "EML" operator reconstructs all arithmetic, constants like e and π, and functions like sin and log, outperforming the
@jiqizhixin
-
LLM Reasoning Hidden Internal States Beyond Chain-of-Thought
By
–
LLM Reasoning Is Latent, Not the Chain of Thought? Researcher Wenshuo Wang argues we should study LLM reasoning as the formation of hidden, internal state trajectories, not just the visible "chain-of-thought" text. The paper finds current evidence most strongly supports that
-
Gen-Searcher: Agentic Tool Grounds Image Generation in Web Reality
By
–
Why do image generators still struggle with up-to-date facts or niche real-world knowledge? Researchers from MMLab CUHK, UCLA, and UC Berkeley present Gen-Searcher. This agentic search tool browses the live web for text and reference images to ground its designs in reality
-
AutoSOTA: AI agents automatically improve research papers
By
–
What if an AI agent could read a research paper and automatically build an even better version of it? Researchers from Tsinghua University and collaborating institutions just released AutoSOTA to do exactly that. AutoSOTA is a multi-agent system that acts like a fully automated
-
Robot learns to open unseen cabinets via action transfer learning
By
–
How can a robot learn to open a cabinet it's never seen before? Researchers from Southeast University & Nanjing University of Science and Technology present RAAP. It works like this: the system finds a similar past action (like opening a microwave) and transfers that
-
SentiAvatar: Digital Humans Understanding Speech and Body Movement
By
–
What if a digital human could understand the meaning of your words and move with the rhythm of your voice? Researchers from Renmin University of China and SentiPulse present SentiAvatar. Their new method uses a "plan-then-infill" approach: first, it plans the overall body
-
PhysGM: AI Turns Single Photos into Physical Simulations
By
–
What if you could turn a single photo into a physically accurate, moving simulation in under a minute? A team from Beijing Institute of Technology & Li Auto presents PhysGM. It's a new AI model that looks at a single image and instantly predicts not just the 3D shape of an
-
Meta AI Introduces Generative Reasoning Re-ranker Framework
By
–
Meta AI introduces Generative Reasoning Re-ranker: a new LLM framework that finally tackles the critical "re-ranking" phase. It trains models using high-quality reasoning data & a novel RL method to refine final choices. Outperforms top model OneRec-Think by 2.4% in key
-
DROID-W: Robot Vision System Maps Rooms With Moving People
By
–
How can a robot map a room full of moving people? Researchers from ETH Zurich & Microsoft present DROID-W, a new vision system that tackles dynamic scenes. It uses "uncertainty-aware" adjustment, meaning it spots moving pixels by checking for visual inconsistencies across
-
Multimodal Tactile Fusion Advances Physical Robot Intelligence
By
–
What if a robot's sense of touch could be fused with sight and language to create true physical intelligence? Researchers from HKUST, Xi'an Jiaotong University, and others present a comprehensive survey on Multimodal Tactile Fusion. They detail methods that combine touch data