New multi-cluster boundary learning method beats LLMs on OOS intent detection
This tiny MiniLM model outperforms large LLMs at detecting out-of-scope user intents.
Researchers Xu, Kang, and Lü propose a multi-cluster boundary learning method for out-of-scope (OOS) intent detection using MiniLM embedding (all-MiniLM-L6-v2). Instead of large LLM-embedding methods, it uses a one-class classification workflow to learn cluster boundaries from training utterances and rejects out-of-domain utterances as OOS intents. Tested on CLINC150, StackOverflow, and Banking77, it achieves state-of-the-art OOS intent detection performance compared to baselines. Ablation studies show MiniLM adapts well to the workflow and embedding requirements. The code is available at supplementary materials.
- Uses MiniLM (all-MiniLM-L6-v2) with only 22M parameters for efficient, deployable OOS detection
- Outperforms larger LLM-based methods on CLINC150, StackOverflow, and Banking77 benchmarks
- One-class classification workflow learns cluster boundaries from training utterances to reject unknowns
Why It Matters
Enables more reliable AI assistants that gracefully handle unknown requests without false positives or hallucination.