GRIP: Grounded Reasoning via Information-Restricted Premises

TL;DR - GRIP is a RAG method that restricts the capacity of the retrieved-evidence channel so models encode information missing from the query. It improves reasoning performance while reducing hallucination by 73% across five benchmarks.

• Keeps full-dimensional query access but applies a severe stochastic bottleneck to retrieved evidence.
• Reduces query–latent mutual information about 30×, from 14.8 to 0.47 bits.
• Outperforms strong iterative baselines on five reasoning benchmarks.
• Produces evidence representations less aligned with query-dominated subspaces.