AdaGATE: Adaptive Gap-Aware Token-Efficient Evidence Assembly for Multi-Hop Retrieval-Augmented Generation
Fills missing bridge facts using micro-queries and utility scoring.
Multi-hop retrieval-augmented generation (RAG) often fails in practice because retrieved evidence is noisy, redundant, or missing critical bridge facts. Existing controllers either expand context indiscriminately or select from a fixed top-k set, without explicitly repairing gaps. AdaGATE, a new training-free evidence controller from Yilin Guo and colleagues, reframes evidence selection as a token-constrained repair problem. It tracks entity-level gaps between the question and retrieved passages, generates targeted micro-queries to fill those gaps, and then selects evidence using a utility function that balances gap coverage, corroboration, novelty, redundancy, and direct relevance.
Evaluated on HotpotQA under clean, redundancy-injected, and noise-injected retrieval conditions, AdaGATE achieves the best evidence F1 among all compared controllers: 62.3% on clean data and 71.2% under redundancy injection. Crucially, it uses 2.6x fewer input tokens than Adaptive-k while maintaining or improving accuracy. These results show that explicit gap-aware repair combined with token-efficient selection makes multi-hop RAG far more robust under imperfect retrieval conditions, without requiring any additional training.
- AdaGATE achieves 62.3% evidence F1 on clean HotpotQA and 71.2% under redundancy injection.
- Uses 2.6x fewer input tokens than Adaptive-k, reducing context overhead.
- Combines entity-centric gap tracking, micro-query generation, and utility-based selection without training.
Why It Matters
Makes multi-hop RAG more practical for real-world deployments with noisy retrievals and limited context budgets.