QueryBandits: A New Strategy to Prevent LLM Hallucinations by Smartly Rewriting Questions

TLDR: A new research paper introduces QueryBandits, a bandit framework that proactively mitigates LLM hallucinations by intelligently rewriting queries. It analyzes 17 linguistic features to select optimal rewrite strategies (e.g., paraphrase, simplify, expand) and uses a reward model to maximize accuracy. The system achieved an 87.5% win rate over a no-rewrite baseline and significantly outperformed static prompting, demonstrating that tailored query intervention based on semantic features is highly effective and efficient, without requiring LLM retraining.

Large Language Models (LLMs) have become incredibly powerful, but with their advanced reasoning capabilities comes a significant challenge: hallucinations. These are instances where LLMs generate inaccurate or fabricated information, eroding trust and reliability. Traditionally, efforts to combat hallucinations have focused on filtering outputs after they’ve been generated. However, a new research paper introduces an innovative approach that tackles the problem at its root: by intelligently reshaping the queries that trigger these inaccuracies.

The paper, titled “QueryBandits for Hallucination Mitigation: Exploiting Semantic Features for No-Regret Rewriting,” by Nicole Cho, William Watson, Alec Koppel, Sumitra Ganesh, and Manuela Veloso from JP Morgan AI Research, proposes a novel framework called QueryBandits. This system is designed to proactively steer LLMs away from generating hallucinations by optimizing how queries are phrased before they even reach the model.

How QueryBandits Work

At its core, QueryBandits operates like a smart decision-maker, choosing the best way to rewrite a query based on its linguistic characteristics. It leverages a bandit framework, a type of machine learning algorithm, to learn and adapt. Here’s a breakdown of its key components:

• Linguistic Features: The system analyzes incoming queries based on 17 distinct linguistic features. These features, ranging from structural elements like anaphora (pronoun references) and subordination (complex sentence structures) to lexical aspects like rarity of words and presence of negation, are known to influence how well an LLM understands and responds to a query.
• Rewrite Strategies: QueryBandits employs five different rewrite strategies, or “arms,” that it can choose from:
1. Paraphrasing: Rewriting the query to introduce lexical diversity while maintaining its original meaning.
2. Simplification: Eliminating complex syntax or nested clauses to make the query more straightforward.
3. Disambiguation: Clarifying vague references or ambiguous phrasing.
4. Expansion: Adding relevant details, named entities, or contextual cues to enrich the query.
5. Clarification of Certain Terms: Defining jargon or domain-specific terms to improve understanding.
• Reward Model: To determine the effectiveness of each rewrite strategy, QueryBandits uses a sophisticated reward model. This model combines three signals: a binary consistency judgment from an LLM-based assessor (GPT-4o), a fuzzy string similarity metric to capture soft overlap, and a BLEU-1 score for lexical fidelity. This multi-faceted approach ensures a robust evaluation of whether a rewritten query leads to a more accurate, non-hallucinatory response.

Impressive Results and Key Insights

The researchers conducted extensive experiments across 13 diverse question-answering benchmarks, using over a thousand lexically perturbed queries per dataset to prevent LLM memorization of standard prompts. The results were compelling:

• The top-performing contextual QueryBandit, utilizing Thompson Sampling, achieved an impressive 87.5% win rate compared to a baseline where no rewriting occurred.
• It significantly outperformed traditional zero-shot static prompting strategies like “paraphrase” or “expand” by 42.6% and 60.3% respectively. This highlights that a dynamic, feature-aware rewriting approach is far more effective than a one-size-fits-all method.
• Interestingly, some static prompting strategies actually led to worse outcomes than no rewriting at all, indicating that unguided rewrites can exacerbate hallucinations.
• A crucial finding was that there is no single optimal rewrite strategy for all queries. The effectiveness of each strategy is highly dependent on the specific linguistic features of the input query. For example, the “EXPAND” strategy proved very effective for queries requiring domain-specific knowledge, while “SIMPLIFY” worked best when pragmatic cues were present.
• QueryBandits achieves these gains through purely forward-pass mechanisms, meaning it doesn’t require expensive retraining or gradient-based adaptation of the LLM itself. This makes it an efficient and practical solution for real-world applications.

Also Read:

• Advancing Conversational AI: A Multi-stage Framework for Smarter Query Suggestions
• Dynamic Steering and Backtracking: A New Way to Guide Large Language Models

Implications for Trustworthy LLMs

This research offers a significant step forward in making LLMs more trustworthy. By proactively shaping queries based on their semantic features, QueryBandits provides a powerful tool for mitigating hallucinations. It also offers valuable insights into how LLMs respond to different linguistic contexts, paving the way for greater interpretability of these complex models. This work demonstrates that intelligent intervention at the query level can induce substantial shifts in LLM output behavior, leading to more reliable and accurate responses without the need for internal model modifications.

For more in-depth information, you can read the full research paper here.