TTPO makes test-time training work without labels by exploiting an asymmetry in majority-vote failure

arXiv 2608.27448, via Hugging Face Daily Papers·medium signalXBlueskyLinkedInCopy link

Posted 2026-08-27 with 72 upvotes on Hugging Face Daily Papers, TTPO addresses why RL and on-policy self-distillation cannot run at test time: both need ground-truth labels, and swapping in majority-vote pseudo-labels is fragile because one bad vote corrupts the teacher for every token. The observation that makes it work is asymmetric: rollouts that disagree with the pseudo-label are typically wrong whether or not the vote itself was correct. So TTPO distills the agreeing rollouts via OPSD and penalizes disagreeing ones with grouped RL, with token-level selection that down-weights already-converged positions and penalizes only confident errors.