Wei Tang

The internal journey of an LLM's prediction, not just its destination, reveals crucial information about input difficulty and generator identity.

Removing an architectural prior abruptly tanks Transformer recall from 77% to 9%, but smoothly annealing it away allows the model to permanently consolidate the underlying circuit.

Scaling from 120,000 hours of egocentric video boosts robotic manipulation success rates from 36.1% to 77.8% in zero-shot evaluations.

DiT-Reward not only outperforms existing models in image evaluation but also accelerates inference by 1.65x without sacrificing quality.

PathRouter reduces reliance on shortcuts in reinforcement learning, leading to more reliable and contextually rich decision-making in language-model agents.

Ultra Flash achieves real-time high-resolution video generation at unprecedented frame rates, pushing the boundaries of what’s possible in streaming video AI.

Turns out, many publicly available LLM skills don't consistently outperform base agents, highlighting the critical need for rigorous, task-grounded evaluation in the open-source skill ecosystem.