Researchers introduce ClariEdge dataset and LoRA fine-tuning method to improve query clarification in RAG systems
A lightweight LoRA fine-tune achieves 96.6% accuracy on classifying ambiguous queries and outperforms larger prompted baselines, while the ClariEdge dataset provides 3,900 examples for training and evaluation.
1 source · cross-referenced
- A new dataset, ClariEdge, contains approximately 3,900 examples of ambiguous and non-ambiguous queries for retrieval-augmented generation (RAG) systems.
- A lightweight LoRA fine-tune achieves 96.6% classification accuracy and outperforms prompted baselines of larger models on both classification and clarification generation quality.
- The same LoRA model can be used in both pre-retrieval and post-retrieval stages of a RAG pipeline to flag queries needing clarification.
- Hard negatives are highlighted as essential for distinguishing truly ambiguous queries from those resolvable by context.
Researchers propose ClariEdge, a dataset of roughly 3,900 examples designed to evaluate and train models on recognizing ambiguous queries in multi-turn, grounded conversational settings. The dataset includes both ambiguous queries and negative examples that do not require clarification, enabling models to learn to distinguish cases where user intent is unclear from cases where context resolves the ambiguity.
The team also demonstrates that a lightweight fine-tuned LoRA model achieves 96.6% classification accuracy on this task, outperforming prompted baselines built on larger models in both classification performance and the quality of generated clarification requests. The fine-tuned model is effective in both pre-retrieval and post-retrieval stages of a RAG pipeline, allowing it to serve as a check before answer generation in either position.
The study emphasizes the importance of "hard negatives"—queries that appear ambiguous but can be resolved from context—to improve model robustness. By incorporating these examples, the fine-tuned LoRA model better separates truly ambiguous queries from those that do not require clarification, leading to more reliable downstream performance.
The ClariEdge dataset and associated code are publicly available, providing a resource for further research and development in query clarification for RAG systems.
- Oct 23, 2026 · arXiv cs.AI
Study finds scaling laws for embodied agents in virtual worlds, but gaps persist for distant generalization
Trust79 - Oct 23, 2026 · arXiv cs.AI
SpecEvo paper proposes multi-LLM framework to cut scientific discovery costs by up to 80%
Trust84 - Oct 22, 2026 · Google DeepMind — Blog
Google DeepMind and partners release 10-meter global forest typology map to track deforestation and forest conversion
Trust79