跳转至

GRIP：基于信息受限前提的扎实推理

文章背景与核心概要

在检索增强生成（RAG）系统中，高容量编码器往往会使用户查询完全主导模型的潜在状态，导致检索到的证据在功能上变得无关紧要——这种失效模式被称为查询主导（query dominance）。为了解决这一痛点，本文提出了 GRIP（Grounded Reasoning via Information-Restricted Premises，基于信息受限前提的扎实推理）。GRIP 引入了容量不对称机制，在保持对查询的全维度访问的同时，迫使检索到的证据通过一个严格的随机瓶颈（stochastic bottleneck）。这确保了证据通道能够严格编码那些查询中尚未包含的残余信息。

实验表明，GRIP 在五个推理基准测试中超越了强大的迭代基线，将查询-潜在互信息诊断指标降低了约 \(30\) 倍（从 \(14.8\) 比特降至 \(0.47\) 比特），并将幻觉率降低了 \(73\%\)。残差对齐分析进一步表明，瓶颈输出占据了与查询对齐程度低于基线表示的子空间。

GRIP: Grounded Reasoning via Information-Restricted Premises

Summary

In Retrieval-Augmented Generation (RAG), high-capacity encoders often allow the user query to completely dominate the model's latent state, rendering retrieved evidence functionally irrelevant—a failure mode known as query dominance. To solve this, the paper introduces GRIP (Grounded Reasoning via Information-Restricted Premises). GRIP applies capacity asymmetry by keeping full-dimensional access for the query while forcing retrieved evidence through a severe stochastic bottleneck. This ensures the evidence channel strictly encodes residual information not already present in the query.

Empirically, GRIP surpasses strong iterative baselines across five reasoning benchmarks, drops the query–latent mutual-information diagnostic by roughly \(30\times\) (from \(14.8\) down to \(0.47\) bits), and reduces hallucinations by \(73\%\).

Paper Metadata

论文元数据

Field Detail
Title GRIP: Grounded Reasoning via Information-Restricted Premises
Author Lirui Teng
Submitted August 17, 2026
Primary Subject Artificial Intelligence (cs.AI)
arXiv ID arXiv:2608.16776
DOI 10.48550/arXiv.2608.16776
ACM Classes I.2.7; H.3.3
Field Detail
Title GRIP: Grounded Reasoning via Information-Restricted Premises
Author Lirui Teng
Submitted August 17, 2026
Primary Subject Artificial Intelligence (cs.AI)
arXiv ID arXiv:2608.16776
DOI 10.48550/arXiv.2608.16776
ACM Classes I.2.7; H.3.3

Abstract

摘要

检索增强生成（RAG）中的高容量编码器可能会让查询主导潜在状态，从而使检索到的证据在功能上变得无关紧要。我们将这种失效模式称为查询主导（query dominance）。为了解决这个问题，我们推出了 GRIP（Grounded Reasoning via Information-Restricted Premises，基于信息受限前提的扎实推理），它引入了容量不对称机制：解码器保持对查询的全维度访问，而检索到的证据则通过一个严格的随机瓶颈。这迫使证据通道仅编码查询中无法获取的残余信息。在五个推理基准测试中，GRIP 的表现优于强大的迭代基线，将查询-潜在互信息诊断指标降低了约 \(30\) 倍（\(14.8 \to 0.47\) 比特），并将幻觉减少了 \(73\%\)。残差对齐分析进一步表明，瓶颈输出所占据的子空间与查询的对齐程度低于基线表示。

High-capacity encoders in retrieval-augmented generation (RAG) can let the query dominate the latent state, leaving retrieved evidence functionally irrelevant. We call this failure mode query dominance. To address it, we introduce GRIP (Grounded Reasoning via Information-Restricted Premises), which imposes capacity asymmetry: the decoder keeps full-dimensional access to the query, while retrieved evidence passes through a severe stochastic bottleneck. This forces the evidence channel to encode only the residual information unavailable from the query. Across five reasoning benchmarks, GRIP outperforms strong iterative baselines, cuts a query–latent mutual-information diagnostic by roughly \(30\times\) (\(14.8 \to 0.47\) bits), and reduces hallucination by \(73\%\). Residual-alignment analysis further shows that the bottleneck output occupies subspaces less aligned with the query than baseline representations.

Full-Text & Resources

全文与资源