Zuse: decision models for 27 European languages hit 87.5% accuracy at 27B
bineric · hf · 2026-11-17
Bineric released Zuse, three decision models (4B/9B/27B) for 27 European languages. Instead of generating text, decision models read a document and return probabilities over a fixed option set, via a low-rank adapter plus pointer head on frozen open-weight bases.
- The 27B model reaches 87.5% accuracy on an 11-source, 9-task-group, 27-language European suite, with pooled expected calibration error of 0.016
- Key finding: miscalibration is structured by task, not language — knowledge tasks are overconfident, acceptability judgements underconfident, so one fitted temperature can't fix both
- Automatable decisions at a 5% error budget track chance-corrected skill (Spearman 0.97) better than raw accuracy (0.91)
- Cross-lingual reach is set by the base model: languages outside its documented list trail by 23–28 points
- The paper documents two multi-source data pitfalls: label-word collisions and an injection augmentation that taught a shortcut
Zuse trails Jev and Kev on English public suites.
More from Research
- Copenhagen postdoc chance at biomaterials × machine intelligence, deadline Dec 6 — _onionesque · 2026-11-17
- $1.1M Templeton grant funds Global Virtues Atlas of moral beliefs for AI alignment — MohammadAtari90 · 2026-11-17
- Dev proposes swarm of specialist diffusion models for code where invalid syntax is unrepresentable — SubtleGradient · 2026-11-17
- Economists warn peer review is collapsing under 99% AI-generated papers and referee reports — Afinetheorem · 2026-11-17
- Meta's Self-Improving Memory: agents discover their own memory skills and distill them into weights — jaseweston · 2026-11-17
- IUCN Red List rebuilds website from scratch as AI bot traffic strains infrastructure — anselm · 2026-11-17