GRIP: Grounded Reasoning via Information-Restricted... | AI Research

GRIP (Grounded Reasoning via Information-Restricted Premises) is a method designed to improve how language models use retrieved evidence in retrieval-augmented generation (RAG). The paper, authored by Lirui Teng, addresses the "query dominance" failure mode, where high-capacity models rely on their own internal knowledge rather than the provided evidence, rendering retrieved information functionally irrelevant.

The Problem: Query Dominance

In standard RAG systems, both the user query and the retrieved evidence are processed through the same high-capacity latent space. Because the model is already optimized to generate answers based on the query alone, it often treats retrieved evidence as a marginal correction or ignores it entirely. The author defines this as "query dominance," where the internal representation of the evidence becomes a compressed, redundant copy of the query. This leads to higher rates of hallucination, as the model fails to genuinely integrate external information.

How GRIP Works

GRIP introduces a "capacity-asymmetric" architecture to force the model to use retrieved evidence. It maintains a full-dimensional path for the query but routes retrieved evidence through a severe, low-dimensional stochastic bottleneck (approximately 4 dimensions).
By restricting the capacity of the evidence channel, the model is pressured to encode only the "information residual"—the specific details provided by the evidence that are not already present in the query. The pipeline consists of four stages: 1. Entropy-Guided Retrieval: Passages are re-ranked based on how much they reduce the uncertainty of the next-step prediction. 2. Predictive Span Extraction: The system extracts a minimal, entailment-verified span of text from the passage. 3. Stochastic Bottleneck: The extracted span is compressed into a 4-dimensional latent state with added Gaussian noise, preventing the model from simply copying the input. 4. Asymmetric Decoding: The decoder receives the full-dimensional query and the low-dimensional, noisy evidence representation, forcing it to prioritize the residual information.

Performance and Results

The author evaluated GRIP across five reasoning benchmarks, including HotpotQA, StrategyQA, 2Wiki, ProofWriter, and SQuAD 2.0. Compared to strong baselines, GRIP consistently improved performance and reduced hallucinations. Specifically, on HotpotQA, the method reduced hallucination by 73% and cut the mutual information between the query and the latent evidence representation by roughly 30 times (from 14.8 to 0.47 bits). Residual-alignment analysis showed that the bottleneck output occupies subspaces that are less aligned with the query than those produced by standard RAG methods.

Considerations

The author notes that while GRIP significantly improves grounding, the diagnostic tool used to measure query–latent dependence is a necessary but not sufficient indicator of evidence use. The effectiveness of the system relies on the bottleneck's capacity constraints; the paper demonstrates that removing this bottleneck significantly increases query–latent dependence and degrades performance. The system also relies on a frozen NLI (Natural Language Inference) verifier to ensure the quality of the extracted spans, which is a critical component of the pipeline's success.