PaperlayerView exact revision on arXiv

Source-linked AI summary

Towards Adaptive Memory-Based Optimization for Enhanced Retrieval-Augmented Generation

Qitao Qin, Yucong Luo, Yihang Lu, Zhibo Chu, Xiaoman Liu, Xianwei Meng

arXiv:2504.05312v4cs.IRcs.AI

TL;DR

Existing RAG methods for open-domain QA may retrieve independently without summarizing memory or adapting retrieval to accumulated information. Amber combines multi-agent memory updates, adaptive retrieval, and multi-level filtering; it reports consistent gains across single-hop, multi-hop, and long-form QA, while requiring extra fine-tuning and LLM accesses.

  • Problem

    Existing RAG methods may retrieve independently without summarizing prior retrievals or adapting queries and stopping decisions to accumulated information.

    Source: §1 · p. 2

  • Method

    Amber iteratively integrates retrieved information into memory with a multi-agent updater, adapts retrieval queries and stopping decisions, and filters irrelevant content at multiple levels.

    Source: §1 · p. 2

  • Results

    Amber consistently outperforms adaptive RAG baselines across single-hop, multi-hop, and long-form QA, with an improvement of over 10% against the state-of-the-art adaptive RAG method.

    Source: §5.1 · p. 7

  • Takeaways & Limitations

    Across the tested open-domain QA datasets, the authors report that Amber and its core components are effective and that Amber is superior to existing methods.

    Source: §1 · p. 2, §6 · p. 8

    Amber requires substantial data, computational resources, and time for multiple filter fine-tuning steps, and answering each question takes longer than with the vanilla approach.

    Source: §6 · p. 9

Abstract

from arXiv · show

Retrieval-Augmented Generation (RAG), by integrating non-parametric knowledge from external knowledge bases into models, has emerged as a promising approach to enhancing response accuracy while mitigating factual errors and hallucinations. This method has been widely applied in tasks such as Question Answering (QA). However, existing RAG methods struggle with open-domain QA tasks because they perform independent retrieval operations and directly incorporate the retrieved information into generation without maintaining a summarizing memory or using adaptive retrieval strategies, leading to noise from redundant information and insufficient information integration. To address these challenges, we propose Adaptive memory-based optimization for enhanced RAG (Amber) for open-domain QA tasks, which comprises an Agent-based Memory Updater, an Adaptive Information Collector, and a Multi-granular Content Filter, working together within an iterative memory updating paradigm. Specifically, Amber integrates and optimizes the language model's memory through a multi-agent collaborative approach, ensuring comprehensive knowledge integration from previous retrieval steps. It dynamically adjusts retrieval queries and decides when to stop retrieval based on the accumulated knowledge, enhancing retrieval efficiency and effectiveness. Additionally, it reduces noise by filtering irrelevant content at multiple levels, retaining essential information to improve overall model performance. We conduct extensive experiments on several open-domain QA datasets, and the results demonstrate the superiority and effectiveness of our method and its components. The source code is available \footnote{https://anonymous.4open.science/r/Amber-B203/}.

1 Introduction

RAG can provide external knowledge, but complex open-domain QA requires integrating information across retrieval steps, which existing adaptive methods may not do effectively. Amber combines memory updating, adaptive retrieval, and multi-level filtering; experiments report improvements over existing open-domain QA methods.

1 Introduction

  • Existing adaptive retrieval methods lack summarizing memory across retrievals, may not assess whether more information is needed, and can add noise through redundant content.These gaps can leave information from separate retrieval steps unintegrated and obscure useful details.Source: §1 · p. 2
  • Amber uses a multi-agent Memory Updater to integrate information from previous retrievals and optimize the LLM’s memory.The introduction also describes an Adaptive Information Collector that updates queries and decides when to stop, and a Multi-granular Content Filter that removes irrelevant content at multiple levels.Source: §1 · p. 2
  • Amber’s Adaptive Information Collector uses accumulated memory to update retrieval queries and decide when to stop retrieval.The stated aim is to make later retrievals more targeted and avoid unnecessary retrievals.Source: §1 · p. 2
  • Amber’s Multi-granular Content Filter assesses retrieved segments and retains essential information to reduce redundant content.The filter first assesses whole-segment relevance, then filters irrelevant content within valid segments.Source: §1 · p. 2
  • Experiments on open-domain QA report significant improvements over existing methods and validate Amber’s components.The introduction presents the Agent-based Memory Updater, Adaptive Information Collector, and Multi-granular Content Filter as the method’s three contributions.Source: §1 · p. 2

2 Related Work

Prior work includes retrieval strategies for multi-hop and long-form QA, from query decomposition to iterative retrieval and adaptive RAG. The paper identifies retrieved-document quality and memory generation as issues that remain insufficiently addressed.

2 Related Work

  • Multi-hop and long-form QA require collecting and contextualizing information across multiple documents, often iteratively.Prior approaches include decomposing multi-hop questions into simpler queries or constructing reasoning steps alongside retrieval.Source: §2 · p. 2
  • The paper argues that prior methods neglect the quality of retrieved documents and memory generation.It frames improving both retrieval and memory generation as a remaining need.Source: §2 · p. 3
  • Adaptive RAG methods adjust retrieval strategies using feedback, including low-confidence tokens or question complexity, while agent-based models add flexibility.Examples include Flare, Self-RAG, and ReAct.Source: §2 · p. 3

3 Methods

Amber is an iterative RAG method that uses retrieved documents to update query-aligned memory, with adaptive retrieval and hierarchical filtering. Its components gather and refine information across iterations before the final memory is used to generate an answer.

3.1 Problem Formulation

  • RAG retrieves pertinent documents from an external corpus and combines them with the input to produce an output.The task formulation describes the retriever selecting documents and the LLM using the input and retrieved context.Source: §3.1 · pp. 3–4

3.2 Amber Overview

  • Amber’s overview combines a Memory Updater, Information Collector, and Content Filter in an adaptive memory-updating iterative RAG method.The method begins with empty memory, retrieves and filters chunks, updates memory, and uses the final memory as context for answer generation.Source: §3.2 · p. 4
  • Amber starts with empty memory and iterates through retrieval, filtering, and memory updating until the Information Collector judges the memory sufficient.If more information is needed, it generates a new query; otherwise, the final memory is supplied to the LLM for answer generation.Source: §3.2 · p. 4

3.3 Agent-based Memory Updater (AMR)

  • The Agent-based Memory Updater uses Reviewer, Challenger, and Refiner agents to evaluate, critique, and revise memory updates using current memory, retrieved passages, and the query.The agents collaborate iteratively to make updates more accurate and aligned with the user’s query.Source: §3.3 · p. 4

3.4 Adaptive Information Collector (AIC)

  • The Adaptive Information Collector evaluates whether updated memory can answer the query, stopping retrieval when sufficient or generating a refined query when it is not.Each iteration retrieves top-k chunks, sends them with the query and current memory to the updater, and reassesses the updated memory.Source: §3.4 · p. 5

3.5 Multi-granular Content Filter (MCF)

  • The Multi-granular Content Filter addresses irrelevant retrieved chunks and sentences through hierarchical filtering, retaining relevant sentences from chunks that pass the initial relevance check.The method describes chunk-level assessment followed by sentence-level evaluation; it states that this approach reduces noise and improves memory quality and overall performance.Source: §3.5 · p. 5

4 Experimental Setup

The experiments use a unified setting spanning single-hop, multi-hop, and long-form QA. They compare Amber with no-retrieval, single-time RAG, and adaptive RAG baselines using multiple LLMs.

4.1 Datasets and Evaluation Metrics

  • To simulate realistic queries of varying complexity, the unified experimental setting uses single-hop, multi-hop, and long-form QA datasets simultaneously.Named datasets include SQuAD v1.1, Natural Questions, TriviaQA, and ASQA, which has 948 queries recompiled by ALCE.Source: §4.1 · pp. 5–6
  • Single-hop and multi-hop QA are evaluated with accuracy and F1, while ASQA long-form QA uses String Exact Match and String Hit Rate.Accuracy checks whether a predicted answer contains the ground truth; F1 measures overlapping words.Source: §4.1 · p. 6

4.2 Baseline&LLMs

  • The baselines cover no retrieval, single-time RAG, and adaptive RAG, and experiments use Qwen2-7b, Llama3-8b, and GPT-3.5.Llama3-8b is the default model unless otherwise specified and is also the default Multi-granular Content Filter LLM.Source: §4.2 · p. 6

5 Results and Analysis

Across six QA datasets, Amber outperforms single-time and adaptive RAG baselines. Its classifier, filtering and memory components, and retrieval parameters are evaluated through performance analyses and ablations.

5.1 Main Results.

  • Amber surpasses single-time RAG on all six QA datasets, exceeds Vanilla by over 30% on three, and outperforms adaptive RAG baselines across single-hop, multi-hop, and long-form QA.Compared with the state-of-the-art adaptive RAG method, Amber improves by over 10%.Source: §5.1 · pp. 6–7

5.2 Classifier Performance

  • Over 90% accuracy on useful retrieved passages and exclusion of more than 40% of negative retrieved knowledge are reported for the Amber classifier.The analysis covers Llama3-8b and Qwen2-7b; Figure 2 presents the classifier confusion matrix.Source: §5.2 · p. 7, Figure 2 · p. 7

5.3 In-depth comparison under a fair top-k

  • As retrieved-passage counts increase under fair top-k, Vanilla RAG shows little to no significant improvement.Fair top-k is the average number of unique passages retrieved per sample across adaptive steps.Source: §5.3 · p. 7, Figure 3 · p. 7

5.4 Ablation Study

  • On 2WikiMQA, using both content-filtering strategies yields the best performance, and removing either reduces it; removing the chunk-level filter hurts more than removing the sentence-level filter.Removing the Agent-based Memory Updater also significantly decreases performance.Source: §5.4 · p. 8, Table 2 · p. 8

5.5 Parameter Analysis

  • Amber consistently outperforms Vanilla RAG under the same top-k conditions, and performance improves as per-step top-k increases in most cases.Table 3 reports the top-k comparison for Amber with Qwen2-7b.Source: §5.5 · p. 8, Table 3 · p. 8
  • On complex QA datasets, performance improves with more iterations and peaks at three; the authors recommend a maximum of three iterations.Table 4 varies the maximum iterations in the Adaptive Information Collector.Source: §5.5 · p. 8, Table 4 · p. 8

6 Conclusion

Amber combines collaborative memory updates, adaptive retrieval feedback, and multi-granular filtering for open-domain QA. The authors report experiments across several datasets and identify training and inference costs as limitations.

6 Conclusion

  • Amber combines collaborative multi-agent memory updating, adaptive retrieval feedback, and multi-granular filtering to improve answer accuracy while reducing hallucinations.The authors report validation across several open-domain QA datasets.Source: §6 · p. 8
  • Training the Multi-granular Content Filter requires multiple fine-tuning steps, substantial data, computational resources, and time.Source: §6 · p. 9
  • Amber answers each question more slowly than the vanilla approach because it accesses the LLM multiple times.Source: §6 · p. 9

A Multi-granular Content Filter

The Multi-granular Content Filter uses query-reference classification and sentence-level filtering to remove irrelevant material and retain useful context. Its training uses filtered retrieved passages and, for the chunk-level filter, semi-automatically generated NLI examples.

A Multi-granular Content Filter

  • The query-reference classifier is fine-tuned on useful-versus-useless categories, while the context filter receives extracted content from original passages.Both fine-tuning processes use two epochs and a per-device batch size of four.Source: A Multi-granular Content Filter · p. 11
  • The Chunk-Level Filter uses an NLI judgment to classify query-reference pairs as useful or useless, retaining useful knowledge and discarding the rest.Source: A Multi-granular Content Filter · p. 11
  • For the Chunk-Level Filter, GPT-4 generates explanations and classifications from prompts containing task instructions, a query, and a retrieved reference.Source: A Multi-granular Content Filter · p. 11
  • Sentence-level filtering uses STRINC for single-hop QA and CXMI for multi-hop datasets to create filtered-context training targets.Source: A Multi-granular Content Filter · p. 11
  • The sentence-level filter is trained to generate filtered context from a query and retrieved passages.Source: A Multi-granular Content Filter · p. 11

B Retriever & Corpus

The experiments align retrievers and corpora across methods within each dataset, while using a separate dense-retrieval setup for long-form ASQA.

B Retriever & Corpus

  • BM25 with Elasticsearch is used for single- and multi-hop datasets, while ASQA uses GTR-XXL and the ALCE 2018-12-20 Wikipedia snapshot segmented into 100-word passages.Single-hop and multi-hop datasets use their respective preprocessed Wikipedia corpora.Source: B Retriever & Corpus · p. 11

C Implementation Details

Implementation uses NVIDIA 4090 GPUs and VLLM for inference, alongside a specified Python, PyTorch, and CUDA stack.

C Implementation Details

D Detailed prompt

The section describes prompts used across Amber’s memory initialization, iterative information collection, memory updating, and content-filtering stages. It also specifies the reviewer, challenger, and refiner roles in memory generation.

D Detailed prompt

  • The reviewer evaluates note-memory strengths and weaknesses, the challenger suggests revisions, and the refiner uses both inputs to generate new memory.The final memory-updating phase compares new memory with initialized memory to select the best one.Source: D Detailed prompt · p. 12
  • At iterative information collection, query remains the original q, note is optimal memory Mopt, and query log tracks already-asked questions to avoid repetition.For filtering, retrieved passages are screened at chunk level, then useful passages are processed at sentence level to extract important sentences for the answer.Source: D Detailed prompt · pp. 11–12
  • The section presents prompts for memory initialization, query rewriting, and the Multi-granular Content Filter, as well as prompts for Agent-based Memory Update.The prompt materials are identified as Tables A3 and A4 in the text; the supplied captions label Tables 5 and 6 by prompt scope.Source: D Detailed prompt · pp. 11–12, Table 5 · p. 13, Table 6 · p. 14

Related papers

Loading 2504.05312v4…