New AI benchmark and lightweight decoder improve detection of unseen combinations of multiple user intents in single utterances

Researchers introduced CoMIX-Shift, a controlled benchmark that tests whether AI models can recognize new combinations of familiar intents, a critical gap in existing benchmarks

ClauseCompose, a new lightweight decoder trained only on single intents, achieved 95.7% exact match accuracy on unseen intent pairs and 93.9% on discourse-shifted pairs

The model outperformed fine-tuned BERT baselines while requiring significantly fewer resources, demonstrating the effectiveness of clause-factorized decoding for compositional generalization

Performance dropped on more challenging scenarios (62.5% on longer/noisier text, 49.8% on held-out templates), revealing remaining challenges in real-world deployment conditions