Rethinking LLM Parametric Knowledge as Post-retrieval Confidence for Dynamic Retrieval and Reranking

2025-09-08

AI Summary

Large language models (LLMs) frequently hallucinate when operating beyond their intrinsic knowledge boundaries; while retrieval-augmented generation (RAG) mitigates this by incorporating external knowledge, it lacks reliable mechanisms to assess whether retrieved contexts genuinely support answering a given query. Method: We propose a knowledge-boundary-aware post-retrieval filtering method that leverages LLMs’ internal hidden states—specifically modeling how retrieved contexts dynamically modulate model confidence. We design a confidence-driven dynamic retrieval (CBDR) mechanism and construct the NQ_Rerank dataset using LLM preference signals to fine-tune a re-ranker for fine-grained context selection. Contribution/Results: Our approach requires no human annotation, instead exploiting continuous hidden-state signals to infer knowledge credibility. It reduces retrieval overhead while significantly improving end-to-end accuracy and robustness of RAG systems.

Technology Category

• Data Mining & Knowledge Management: Conversational Systems for Recommendation & Retrieval
• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: Question Answering

Application Category

• Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAG
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking

Abstract

Large Language Models (LLMs) often generate inaccurate responses (hallucinations) when faced with questions beyond their knowledge scope. Retrieval-Augmented Generation (RAG) addresses this by leveraging external knowledge, but a critical challenge remains: determining whether retrieved contexts effectively enhance the model`s ability to answer specific queries. This challenge underscores the importance of knowledge boundary awareness, which current methods-relying on discrete labels or limited signals-fail to address adequately, as they overlook the rich information in LLMs` continuous internal hidden states. To tackle this, we propose a novel post-retrieval knowledge filtering approach. First, we construct a confidence detection model based on LLMs` internal hidden states to quantify how retrieved contexts enhance the model`s confidence. Using this model, we build a preference dataset (NQ_Rerank) to fine-tune a reranker, enabling it to prioritize contexts preferred by the downstream LLM during reranking. Additionally, we introduce Confidence-Based Dynamic Retrieval (CBDR), which adaptively triggers retrieval based on the LLM`s initial confidence in the original question, reducing knowledge conflicts and improving efficiency. Experimental results demonstrate significant improvements in accuracy for context screening and end-to-end RAG performance, along with a notable reduction in retrieval costs while maintaining competitive accuracy.

Problem

• Detecting when retrieved contexts enhance LLM confidence
• Improving knowledge boundary awareness using hidden states
• Reducing retrieval costs while maintaining answer accuracy

Innovation

• Uses hidden states for confidence detection
• Fine-tunes reranker with LLM preference dataset
• Implements confidence-based dynamic retrieval triggering