Overview
- Researchers introduce the Trust-RAG Compass, a six-part framework for judging Retrieval-Augmented Generation across factuality, robustness, fairness, transparency, accountability, and privacy, plus a new test suite called TRC Bench.
- Early TRC Bench results compare proprietary and open-source systems and show uneven performance across the six areas, highlighting risks like bad or poisoned retrieval, misuse of cited evidence, leaks of source documents, and unequal treatment of content or groups.
- BalanceRAG proposes a cascaded setup that answers with the base model first and only fetches documents when confidence is low, using joint thresholding to hold error rates steady while reducing needless retrieval.
- SABER adds self-awareness to RAG by estimating whether to trust the model’s built-in knowledge, the retrieved context, either source, or to abstain, which increases accuracy on conflict-heavy questions and offers a controllable risk–coverage tradeoff.
- A separate study finds that the ranking step in retrieval drives representational bias and shows a stochastic ranker that restores near-parity exposure, which then carries through to fairer generated answers.