method · first seen 14 Nov 2026
logit margin regularization
Also: hinge margin penalty
Training penalty that enforces a minimum gap between top-1 and top-2 logits to reduce decoding flips.
Mentions per week
PapersCodeDiscussion
Earliest sources
- How Much Determinism Can Fine-Tuning Buy? Margin Regularisation Against Batch-Induced Token FlipsPapers · 14 Nov 2026 · Charan Teja Nagisettygari, Siva Raja Sindiramutty
Latest sources
- How Much Determinism Can Fine-Tuning Buy? Margin Regularisation Against Batch-Induced Token FlipsPapers · 14 Nov 2026 · Charan Teja Nagisettygari, Siva Raja Sindiramutty