Research
Dream2Flow Connects Video Generation to Real-World Robot Control
Dream2Flow uses 3D object flow to turn video-generated motions into robot commands, enabling zero-shot manipulation of diverse objects.
UniAct Cuts Humanoid Robot Response Time to Under 500 Milliseconds
A new framework pairs a fine-tuned multimodal language model with a streaming pipeline, letting robots follow complex instructions with human-like speed and flexibility.
Counterfactual VLA: AI That Thinks Twice Before Hitting the Gas
New framework gives autonomous vehicles self-reflective capabilities, simulating potential outcomes to enhance safety and accuracy on the road. Is this the future of self-driving?
Mirage: New Video Diffusion Model Aims to Enhance Autonomous Driving Data
Researchers introduce Mirage, a video diffusion model designed for photorealistic and coherent asset editing in driving scenes, potentially revolutionizing autonomous driving data augmentation.
Internal Guidance Boosts Diffusion Models to New Heights
New research introduces Internal Guidance (IG), a simple yet effective technique that significantly improves image generation quality in diffusion models, achieving state-of-the-art FID scores.
RSAgent Boosts Text-Guided Segmentation with Agentic AI
RSAgent uses multi-turn tool calls to outperform existing models in text-guided object segmentation across key benchmarks.
The Feeling Machine: Can AI Decode Human Emotion?
The Large Emotional World Model (LEWM) aims to teach AI the messy logic of feelings. Is this empathy or just smarter engineering?
Thinking Before Speaking: How iCLP Helps LLMs Plan in Secret to Cut Logic Errors
Researchers introduce iCLP, a framework that lets AI map out hidden plans before answering, boosting accuracy without extra bloat.
Google and OpenAI Chase 1,000x Efficiency with Conditional Computation
Conditional computation aims to boost AI performance by activating only relevant neurons. Is this a game-changer or just another hype cycle?
Australian Startup SQC Launches 'Quantum Twins' Simulator with 15,000 Quantum Dots
Silicon Quantum Computing bets analog simulation will deliver real-world results before universal quantum computers arrive.
Robots Learn to Ignore Our Mistakes with SiLRI
A new reinforcement learning algorithm helps robots treat human 'help' as advice—not gospel—boosting their ability to master complex tasks.
D²VLM Separates Timing from Content to Boost Video-Language Accuracy
By teaching models to first pinpoint *when* an event happens, researchers help AI describe videos with sharper precision.
New Stress Test Challenges AI’s ‘Bigger is Better’ Myth
The Drill-Down and Fabricate Test shows massive models can be just as fragile as smaller ones when facts get fuzzy.
GPT-5 in the Lab: A New Era for Biological Research?
OpenAI's GPT-5 ventures into molecular cloning, sparking debates on innovation and safety in AI-driven labs.
OpenAI's Bold Move Towards AI Transparency and Safety
OpenAI delves into mechanistic interpretability with sparse models to enhance AI transparency and trustworthiness.