LOPA enhances spoken language assessment via latent ordinal prototype alignment

The paper introduces Latent Ordinal Prototype Alignment (LOPA), a prototype-based regularizer that enforces an ordinal geometric prior on the latent space to address the overlooked structure of language acquisition in Multimodal Large Language Models. Coupled with Semantic-Anchored Layer Routing (SALR) for adaptive representation harvesting from a frozen Whisper encoder, the framework achieves an RMSE of 0.361.

• LOPA enforces ordinal geometric priors directly on the latent space.
• SALR adaptively harvests multi-depth representations from a frozen Whisper encoder.
• The approach rivals billion-parameter systems without requiring LLM-based fine-tuning.
• Analysis shows SALR and LOPA provide interpretable, criterion-aligned preferences.

This work offers an efficient and ordinal-aware modeling alternative to current scaling-centric models for spoken language assessment.