GRIP cuts RAG hallucination 73% by starving the evidence channel.

The paper names a failure mode called query dominance, where a high-capacity encoder lets the query dominate the latent state and renders retrieved evidence functionally irrelevant (arXiv 2608.16776). GRIP imposes capacity asymmetry: full-dimensional decoder access to the query, retrieved evidence forced through a severe stochastic bottleneck so it can only encode residual information the query can't supply. Beats strong iterative baselines on five reasoning benchmarks and cuts a query-latent mutual-information diagnostic ~30x (14.8 → 0.47 bits). Counterintuitive design: make the evidence path worse and the model actually uses it.