No One Size Fits All: QueryBandits for Hallucination Mitigation

A research team from Carnegie Mellon University and JP Morgan AI Research has published a significant paper on arXiv titled 'No One Size Fits All: QueryBandits for Hallucination Mitigation.' The work addresses a critical gap in AI safety: most current methods for reducing hallucinations (factual errors) focus on open-source models, leaving the vast majority of deployed, closed-source models like GPT-4 and Claude vulnerable. The team's solution, QueryBandits, is a novel framework that treats the problem as an online learning task. Instead of applying a single, fixed strategy like paraphrasing a user's question, it uses a contextual bandit algorithm (specifically Thompson Sampling) to dynamically choose the best 'rewrite policy' for each unique query based on its semantic features. This adaptive approach learns from feedback in real-time, using a calibrated reward function to measure answer quality.

The technical breakthrough is that QueryBandits operates purely through the model's forward pass, requiring no access to the model's internal weights or gradients. This makes it uniquely applicable to proprietary, closed-source LLMs where traditional fine-tuning or parameter editing is impossible. In extensive testing across 16 question-answering scenarios, their top QueryBandit configuration achieved an 87.5% win rate against a 'No-Rewrite' baseline and outperformed zero-shot static policies—like simply paraphrasing or expanding a query—by 42.6% and 60.3%, respectively. The research conclusively shows that inflexible policies can sometimes increase hallucinations, proving the 'no one size fits all' premise. This work provides enterprises with a practical, deployable tool to significantly enhance the reliability of the expensive, closed-source LLMs that power their critical applications, moving beyond post-hoc detection to proactive, adaptive mitigation.