Paper finds systematic length bias in top translation metrics, skewing model rankings
shangbinfeng · x · 2026-10-04
- A paper to be presented at COLM (by @WendaXu2, @yilinjz et al., affiliated with Google AI/Google DeepMind) shows a systematic length bias across top-performing translation evaluation metrics and reward models.
- The authors argue these metrics may reward longer outputs rather than actual translation quality, fundamentally skewing model rankings—scores partly reflect sequence length, not quality.
More from Research
- LEV: Looped decision model lifts 2.6B accuracy 63.2% to 76.2% with adaptive compute — jiqizhixin · 2026-11-19
- Aurora2-400M-A20M: a sparse MoE pretrained from scratch in 36 hours on one RTX Pro 6000 — Tall_Abrocoma_3533 · 2026-11-19
- Researcher calls METR/Anthropic '2x speedup' RSI analysis statistically dubious — jessi_cata · 2026-11-19
- Linear Probes Could Reveal Which Persona a Base Model Is Playing, Researchers Suggest — cephaloform · 2026-11-19
- ETH's Anymal robot completes 10km autonomous alpine hike with VLM-guided ROAM system — ChongZzZhang · 2026-11-19
- OpenVax codebase grows from 78K to 565K lines in 15 months — iskander · 2026-11-19