Synced · WeChatMay 4, 2026, 12:06 UTC+8

ACL 2026: PolyU Open-Sources SignThought for Gloss-Free Sign Language Translation

ACL 2026｜AI for聋哑群体，港理工开源思考型手语翻译模型

PolyU and Sichuan University introduced SignThought, accepted to ACL 2026 Main and slated for oral recommendation. It uses latent thoughts, plan-then-ground, and dual-stream decoding, reaching top gloss-free BLEU-4 on five SLT benchmarks. The team also built LC-HKSLT with 1,311 hours, 432K clips, and 14 signers.

Why it matters: ACL 2026 Main, an open model, and a new dataset satisfy HKR-H/K/R, with concrete mechanisms and five benchmarks. The niche sign-language focus keeps it below broader model or developer-tool releases.

Read the original ↗Export Markdown

Take

SignThought should be read as a data-scale move before a reasoning-model story. The disclosed hook is concrete: top gloss-free BLEU-4 on five sign-language translation benchmarks, using latent thoughts, plan-then-ground, and dual-stream decoding. The bigger asset is LC-HKSLT: 1,311 hours, 432K clips, and 14 signers.

That last number cuts both ways. SLT has been stuck between expensive gloss annotation and poor cross-signer transfer; gloss-free training avoids one bottleneck but leans harder on video diversity. Fourteen signers is still a narrow base for broad Hong Kong Sign Language generalization. I like the direction, but I don’t buy any “for the deaf community” victory lap until the license, code, evaluation splits, and signer leakage checks are visible. The WeChat body is blocked by CAPTCHA, so those details are not disclosed here.

Source: Synced · WeChat · mp.weixin.qq.com
Published: May 4, 2026, 12:06 UTC+8 · May 4, 2026