vivo AI Lab's CAPO-SLT fixes RL drift in sign language translation with confidence-aware clipping

量子位 · wechat · 2026-09-09

vivo AI Lab introduces CAPO-SLT, addressing why sign language translation models that align vision and text well still drift during autoregressive generation. In standard PPO/GRPO, all positive-advantage tokens share one clipping limit, over-reinforcing high-confidence language-prior "fluent but wrong" tokens while suppressing low-confidence but genuinely visual-grounded ones. CAPO sets the upper clip bound per token based on the old policy's confidence (0.2 base, relaxed up to 0.3), plus a loss cap for negative-advantage tokens — without changing the visual backbone or reward.

Trained with supervised initialization on ST-GCN pose features (face/body/two hands) then RL with sentence-level rewards combining BLEU-1, BLEU-4, ROUGE-L and BERTScore, CAPO-SLT uses only pose input yet beats Geo-Sign (+2.93/+1.26/+0.64) and Uni-Sign (+4.96/+3.07/+3.67) on CSL-Daily, topping even Pose+RGB methods, and reaches 41.4 BLEU-1 on How2Sign. Ablations show +0.30/+0.50/+0.67 over DAPO under identical reward and settings, with more stable policy entropy during training.

Failure examples without CAPO include translating "I introduce these colleagues" as "I'm a team of several people" — fixed after adding CAPO. Limitations: milder ASL gains and reference-based rewards that under-cover valid paraphrase. The approach is low-intrusion for existing PPO/GRPO pipelines. Paper on OpenReview.

Original post →

More from Research

Research channel →