Convai Releases Laya, a Non-Autoregressive Decision Model

Convai Innovations has released Laya, a non-autoregressive decision model, under the Apache 2.0 license as an open-source alternative to TypeSafe's Jev API. Instead of generating text, the model takes input states and typed questions to return results in a single forward pass, providing fast response times of 33–40ms on GPU.

Key Features and Architecture

Laya is designed to perform label selection, ordinal scoring, and yes/no probability prediction without output parsing or hallucination risks. The model consists of three checkpoints:

• 421M parameter English model: Based on ModernBERT-large
• 322M parameter multilingual model: Based on mmBERT-base, supporting over 100 languages
• Fine-tuned variants for typed decision workflows

A built-in Router detects input scripts and routes them to the appropriate checkpoint. Training employed RLCD, a reinforcement learning methodology using strict proper scoring rules as the reward function, to encourage the model to report honest probabilities. It also includes an act-vs-escalate head to determine when to escalate to a human.

Performance and Limitations

Developers reported superior performance compared to Jev on AG News, sentiment classification, and typed decision benchmarks, with execution speeds approximately 6–8x faster. However, the comparison figures with Jev are third-party data and not direct head-to-head test results. The model card specifies the following limitations:

• Base checkpoints are close to random guessing on typed decisions without fine-tuning
• Accuracy drops sharply when there are more than 50 choices (Banking77: Laya 0.425 vs Jev 0.870)
• Ordinal scoring is the weakest area
• The English checkpoint fails on non-Latin scripts
• The model tends to be overconfident, requiring temperature adjustment on custom data to obtain reliable probabilities