Chen Yang

Predicting future motion instead of reconstructing past trajectories allows a frozen humanoid controller to cut tracking error by 62% while unlocking zero-shot behavior composition via simple latent addition.

Standard mechanistic circuits can mimic over 97% of a model's correct outputs while missing up to 89% of its errors, exposing that common ablation validations reward functional mimicry rather than true causal mechanisms.

Teacher rewards can mislead learning, but R2-OPD filters out conflicting feedback to boost reasoning performance in language models.