GRACE Trains RAG Models With Abstention
Researchers introduce GRACE, a reinforcement-learning framework (submitted Jan. 8, 2026) that jointly enforces evidence-based grounding and reliable abstention in Retrieval-Augmented Generation (RAG) systems. GRACE uses heterogeneous retrievers to construct training data and a multi-stage gated reward to teach models to assess evidence sufficiency, extract supporting passages, and answer or abstain. Experiments on two benchmarks show state-of-the-art accuracy and a 10% annotation cost.
Key Points
- 1Introduces GRACE, an RL framework combining grounding and abstention for RAG systems
- 2Uses heterogeneous retrievers and multi-stage gated rewards to evaluate evidence sufficiency and extract support
- 3Reduces annotation cost to 10% while improving accuracy and calibrated rejection decisions
Scoring Rationale
Strong innovation and practical gains across RAG systems, tempered by single-source academic submission lacking wide validation.
Sources
Public references used for this report.
Practice with real FinTech & Trading data
90 SQL & Python problems · 15 industry datasets
250 free problems · No credit card
See all FinTech & Trading problems