Reinforcement Learning for Large Language Model Selective Evidence Adoption from Contaminated Retrieval Results
A directional improvement in selective evidence use is demonstrated, while identifying injection resistance and statistical robustness as important remaining challenges for future work.