What if solving a puzzle meant rolling downhill? CRAM learns an energy landscape for Sudoku and other structured puzzles. We explain the speed resultsβand why they don't establish general reasoning. #MachineLearning#Reasoning
Can a reasoning model use a robot like it uses a computer? PEAR learns in simulation, then controls a real arm from camera views. We explain the approach, the small real-world trial, and its slow inference. #Robotics#EmbodiedAI
Two robot videos look alike, but their scores don't. Cogito shows how re-encoding and tiny visual changes can fool a world-model benchmark. We explain the tricks and why real robot outcomes still matter. #WorldModels#Robotics
A few clicks can change a 3D reconstruction.
Click3R lets you mark matching points in two images to guide a feed-forward model. We explain the adapter, the results, and why the quality of those clicks matters.
#ComputerVision#3D
Speeding up a slow robot policy can break it.
Flash ALOHA learns from fast, hands-on demonstrations, masks out human hands, and blends action chunks for smoother motion. Here's how the pieces fit together.
#Robotics#ImitationLearning
A tiny shift can decide whether a box stays or falls. Does a video model notice?
FlipTest tests nearly identical scenes with opposite outcomes. Plausible-looking video often misses the physical difference that matters.
#WorldModels#AI
One neuron. A model stuck saying "Wait."
Beaker reports that doubling one neuron's output helped a small language model escape reasoning loops. We explain the edit, why banning "Wait" has a similar effect, and the limits.
#LLM#AI
What if an LLM could learn by deleting weights?
Deletion Tuning learns which pretrained weights to zero out, without writing new weight values. At larger scales, it matches full fine-tuning in the reported tests.
#MachineLearning#LLM
A tabletop. A room. A city.
SplatAnything reconstructs them with the same feed-forward 3D model. The key is normalizing Gaussian spacing, rather than scene size alone.
Here's the idea in 3 minutes.
#3DGS#ComputerVision
3D Gaussian Splatting, but the points become strands.
Spaghetti Splatting reconstructs thin structures like hair and nets as connected curves that can be used in physics simulation. Here's how it works.
#3DGS#ComputerGraphics
No demos. No policy training. Just an AI writing Python to move a robot arm.
Apple's AMP puts the model inside the control loop. The catch? It spends far more time thinking than moving.
#Robotics#EmbodiedAI
A robot can recognize a bottle and still misjudge how far away it is.
SLA teaches vision-language-action models a sense of 3D space during pre-training, without adding a depth sensor at deployment.
#Robotics#EmbodiedAI
What if the fix for k-means was learning to forget?
Meta's e-means gradually discounts old points instead of letting early mistakes stick around. Here's why that helps, and where Lloyd's algorithm fits in.
#MachineLearning#AI