pith. sign in

However, under the linear head model, for anyw ′ ∈R d, K ∑ j=1 Aj(w′) = (w ′)T K ∑ j=1 (hj − ¯h) =0, since ∑K j=1(hj − ¯h) = 0 by definition of ¯h

1 Pith paper cite this work. Polarity classification is still indexing.

1 Pith paper citing it

fields

cs.LG 1

years

2026 1

verdicts

UNVERDICTED 1

representative citing papers

Mitigating Reward Hacking in RLHF via Advantage Sign Robustness

cs.LG · 2026-04-03 · unverdicted · novelty 6.0

SignCert-PO mitigates reward hacking in RLHF by down-weighting completions whose advantage signs are not robust to small reward-model perturbations, using a certified preservation radius derived at the policy optimization stage.

citing papers explorer

Showing 1 of 1 citing paper.

  • Mitigating Reward Hacking in RLHF via Advantage Sign Robustness cs.LG · 2026-04-03 · unverdicted · none · ref 5

    SignCert-PO mitigates reward hacking in RLHF by down-weighting completions whose advantage signs are not robust to small reward-model perturbations, using a certified preservation radius derived at the policy optimization stage.