Laya, an Open-Source Decision Model Competing with Jev, Released... 6-8x Faster with Hallucination Blocking

A researcher released 'Laya', a fully open-source decision model responding to the closed model 'Jev'. This model was developed based on research results published on arXiv in 2025, utilizing non-autoregressive reinforcement learning to address the inefficiencies of autoregressive chatbots in AI workflows such as high-frequency classification, routing, and triage.

Performance and Architecture

Laya performs immediate decision-making within 30-35ms, similar to System 1 in the human brain, and processes three decision primitives—choice, score, and noul—in a single forward pass without text generation. This physically blocks hallucinations and schema violations.

• Speed: 6-8x faster than Jev (32.8ms per single question, 7.2ms per question in batch mode)
• Reliability: ECE (Expected Calibration Error) of 0.081, superior to Jev (0.246)
• Benchmarks: Outperforms Jev in key metrics such as typed-decisions (0.766 vs 0.727) and AG News (0.950 vs 0.910)

Global Routing and Limitations

To address the issue where English models exhibit low accuracy yet high confidence on non-Latin scripts, pure Python routing based on 22 Unicode scripts was introduced as a mandatory preprocessing step. Using Router(preload=True) eliminates the 7-10 second cold-swap penalty incurred during language switching.

However, performance degrades for questions with more than 20 options (such as Banking77) due to insufficient token budget per candidate, so it is recommended to keep options under 20 or use a coarse-to-fine structure. The base model is at zero-shot levels (~0.35), and benchmark performance reflects fine-tuning results.