Seong Joon Oh

MLLMs falter in egocentric action selection, consistently opting for actions of visible agents instead of their own, revealing a critical gap in current training paradigms.

Multi-party conversations expose LLMs to a staggering increase in privacy leaks, with models revealing sensitive information far more than previously understood.

Activation probes can predict future behaviors in reasoning models with up to 91% accuracy, enabling effective steering without sacrificing output quality.

Framework choice in multi-agent systems matters just as much as the LLM itself, a fact obscured by existing model-centric benchmarks.

Forget training on just images or text – TAIL, a new meta-learner, masters both and extrapolates to 20x more classes, all while slashing compute costs.