CtrlA: Adaptive Retrieval-Augmented Generation via Inherent Control

TL;DR

CtrlA targets a core limitation of adaptive RAG by proposing a representation-based control paradigm that uses honesty steering and confidence monitoring to optimize retrieval timing. By extracting directional features in the LLM latent space and applying a simple linear adjustment to representations, CtrlA promotes truthful outputs and triggers external knowledge only when needed, complemented by context-aware CAQ and TVQ query strategies. Empirical results across diverse QA and biography tasks show CtrlA outperforms uncertainty-based and rule-based baselines, with robust transfer across backbones and tasks. This approach offers a lightweight, plug-in pathway to improve factual accuracy and retrieval efficiency in RAG systems, with potential for further integration with content validation and model alignment methods.

Abstract

Retrieval-augmented generation (RAG) has emerged as a promising solution for mitigating hallucinations of large language models (LLMs) with retrieved external knowledge. Adaptive RAG enhances this approach by enabling dynamic retrieval during generation, activating retrieval only when the query exceeds LLM's internal knowledge. Existing methods primarily focus on detecting LLM's confidence via statistical uncertainty. Instead, we present the first attempts to solve adaptive RAG from a representation perspective and develop an inherent control-based framework, termed \name. Specifically, we extract the features that represent the honesty and confidence directions of LLM and adopt them to control LLM behavior and guide retrieval timing decisions. We also design a simple yet effective query formulation strategy to support adaptive retrieval. Experiments show that \name is superior to existing adaptive RAG methods on a diverse set of tasks, the honesty steering can effectively make LLMs more honest and confidence monitoring is a promising indicator of retrieval trigger.Our code is available at \url{https://github.com/HSLiu-Initial/CtrlA}.

Figures (16)

• Figure 1: CtrlA framework. Step 1 extracts the features corresponding to honesty and confidence directions; Step 2 utilizes extracted features to steer and monitor LLM behaviors at inference. The honesty feature steers the representation of LLM to make it more honest, while confidence feature is used to monitor the confidence level of LLM outputs, where the token whose score is lower than the threshold is marked as unconfident. The retrieval is triggered if specific tokens are unconfident.
• Figure 2: Effects of honesty steering on TruthfulQA.
• Figure 3: Impacts of honesty steering on PopQA (left) and ASQA (right). $^*$Only 2018 Wikipedia corpus is used for PopQA.
• Figure 4: Effects of different choices of $\tau$ on TriviaQA.
• Figure 5: Impacts of honesty steering with respect to the layers and steps on TriviaQA.