Software archaeology

Teach LLMs to “think” before they speak and improve parameter and memory efficiency.

Boosting performance on the ARC-AGI challenge is a matter of perspective.

* Winner, ARC-AGI 2024

Uses depth cues and diffusion models to turn a single image into a clean, instance-level 3D scene.

A benchmark that tests whether models can truly follow and understand long, multi-subject lecture videos, revealing that current VLM’s limitations.

Robots can learn a compact “plan space” from human demos and continually updates its plan as the scene changes, enabling faster and more reliable manipulation.

AimBot draws visual cue onto robot camera images so policies directly see where the gripper is in 3D, boosting manipulation success.

Detects when a reasoning model has drifted into meaningless repetition and cleanly cuts away those extra tokens, saving a lot of output cost with almost no quality loss.

Breaks social reasoning into “who-knows-what” steps and uses a checker to pick the most consistent answer, improving small-models’ theory-of-mind performance.

Trains a single model to extract and verify all claims in a long answer at once using web evidence, speeding up long-form factuality evaluation.

Labeling each reasoning step’s mistake type and converts those labels into scores, giving richer feedback and improving math problem solving with less data.