PaperlayerView exact revision on arXiv

Source-linked AI summary

Enhancing Retrieval-Augmented Large Language Models with Iterative Retrieval-Generation Synergy

Zhihong Shao, Yeyun Gong, Yelong Shen, Minlie Huang, Nan Duan, Weizhu Chen

arXiv:2305.15294v2cs.CL

TL;DR

LLM retrievers struggle with relevance for complex information needs, while existing interleaved workflows can constrain generation and increase overhead. Iter-RetGen iteratively uses model outputs to retrieve more relevant knowledge and processes retrieved knowledge as a whole. It performs strongly across three tasks, with up to 8.6% absolute gains on four of six datasets, and retrieval adaptation further improves performance while reducing iterations.

  • Problem

    One-time retrieval is insufficient for complex information needs, while interleaved retrieval-generation workflows may fail to jointly process retrieved knowledge and increase overheads.

  • Method

    Iter-RetGen iteratively performs retrieval-augmented generation and generation-augmented retrieval, using model outputs as context to retrieve more relevant knowledge for subsequent generation.

  • Results

    Across multi-hop question answering, fact verification, and commonsense reasoning, Iter-RetGen achieves up to 8.6% absolute gains over previous state-of-the-art retrieval-augmented methods on four of six datasets and remains competitive on two.

  • Takeaways & Limitations

    Iter-RetGen provides a simpler, lower-overhead retrieval-augmented approach that preserves flexible generation, while retrieval adaptation can further improve performance and reduce iterations.

  • Takeaways & Limitations

    The experiments use a fixed black-box LLM that may not be equally optimized for different prompting forms.

Abstract

from arXiv · show

Large language models are powerful text processors and reasoners, but are still subject to limitations including outdated knowledge and hallucinations, which necessitates connecting them to the world. Retrieval-augmented large language models have raised extensive attention for grounding model generation on external knowledge. However, retrievers struggle to capture relevance, especially for queries with complex information needs. Recent work has proposed to improve relevance modeling by having large language models actively involved in retrieval, i.e., to improve retrieval with generation. In this paper, we show that strong performance can be achieved by a method we call Iter-RetGen, which synergizes retrieval and generation in an iterative manner. A model output shows what might be needed to finish a task, and thus provides an informative context for retrieving more relevant knowledge which in turn helps generate a better output in the next iteration. Compared with recent work which interleaves retrieval with generation when producing an output, Iter-RetGen processes all retrieved knowledge as a whole and largely preserves the flexibility in generation without structural constraints. We evaluate Iter-RetGen on multi-hop question answering, fact verification, and commonsense reasoning, and show that it can flexibly leverage parametric knowledge and non-parametric knowledge, and is superior to or competitive with state-of-the-art retrieval-augmented baselines while causing fewer overheads of retrieval and generation. We can further improve performance via generation-augmented retrieval adaptation.

1 Introduction

Retrieval-augmented LLMs struggle when queries have complex information needs, motivating ITER-RETGEN, which iteratively uses generation to improve retrieval and generation. Across three tasks, it achieves strong performance with fewer overheads and can be further improved through retrieval adaptation.

  • One-time retrieval often fails for complex information needs such as multi-hop reasoning and long-form question answering.
  • Interleaved retrieval-generation workflows may fail to process all retrieved knowledge jointly and increase retrieval and generation overheads.
  • ITER-RETGEN uses each model output as context for retrieving more relevant knowledge, then feeds the expanded knowledge into the next generation iteration.
  • ITER-RETGEN is evaluated on multi-hop question answering, fact verification, and commonsense reasoning.
  • 8.6% absolute gains over previous state-of-the-art retrieval-augmented methods occur on four of six datasets, with competitiveness on the other two.
  • ITER-RETGEN is superior or competitive with state-of-the-art retrieval-augmented methods while causing fewer retrieval and generation overheads, and adaptation further improves performance while reducing iterations.

2 Related Work

Retrieval augmentation connects LLMs to external knowledge, while recent methods involve generation in retrieval to improve relevance. ITER-RETGEN iterates complete retrieval and generation without interrupting generation, preserving flexibility and achieving strong performance across evaluated reasoning tasks.

  • LLMs can hallucinate and lack knowledge under-represented in their training data, motivating tools that enrich their input context with relevant information.
  • Retrieval augmentation connects LLMs to the external world, but passive retrieval based only on task inputs can struggle to capture relevance, especially zero-shot.
  • Recent approaches use generated search queries, partial generations, or forward-looking sentences to involve LLMs in retrieval.
  • ITER-RETGEN uses complete previous generations to retrieve more relevant information without interrupting generation, then processes all retrieved knowledge jointly.
  • ITER-RETGEN is conceptually simpler and easier to implement while showing strong performance in multi-hop question answering, fact verification, and commonsense reasoning.
  • GAR augments queries with generated background information, whereas HyDE generates hypothetical paragraphs for zero-shot retrieval before retrieving real passages.

3 Iterative Retrieval-Generation Synergy

Iter-RetGen repeatedly combines retrieval-augmented generation with generation-augmented retrieval: each output helps retrieve more relevant knowledge for the next iteration. The method also adapts retrieval by distilling a re-ranker’s generation-aware knowledge into a dense retriever.

  • Iter-RetGen: At iteration t, the system retrieves top-k paragraphs using the previous output y_t−1 concatenated with question q, then generates y_t from those paragraphs and q.The final response is the last output y_T.
  • Iter-RetGen: Iter-RetGen addresses semantic gaps in complex information needs by using generated reasoning to expose what supporting knowledge may be needed.This is motivated by cases where relevant information appears only after prerequisite subquestions are resolved.
  • Retrieval-Augmented Generation: Each generation uses Chain-of-Thought prompting with retrieved knowledge prepended to the original question, while direct refinement through previous generations is left for future work.The paper focuses on studying retrieval-generation synergy in a straightforward implementation.
  • Retrieval Components: The dense retriever scores queries and paragraphs by the inner product of their encoded vectors, whereas the re-ranker outputs a relevance probability.The dense retriever is parameterized by query- and paragraph-encoder parameters.
  • Generation-Augmented Retrieval Adaptation: Generation-augmented retrieval adaptation gives a re-ranker access to the first LLM output y1 and distills its knowledge into the retriever’s query encoder.The re-ranker is described as typically better at capturing query-paragraph relevance than the retriever; optimization uses a KL-divergence objective.

4 Experiments

The experiments evaluate the method across six datasets covering multi-hop question answering, fact verification, and commonsense reasoning. They use shared few-shot questions, task-specific metrics, Wikipedia retrieval corpora, and an additional LLM-based correctness metric.

  • Datasets: The evaluation covers six datasets across multi-hop question answering, fact verification, and commonsense reasoning.The datasets include HotPotQA, 2WikiMultiHopQA, MuSiQue, Bamboogle, Feverous, and StrategyQA.
  • Datasets: MuSiQue uses only 2-hop questions, while Feverous represents fact verification and StrategyQA represents commonsense reasoning.The remaining listed datasets are used for multi-hop question answering.
  • Evaluation Settings: Retrieval uses the October 2017 Wikipedia dump for HotPotQA, the December 2018 dump for 2WikiMultiHopQA, and the December 2021 dump for the other datasets.These corpora define the retrieval sources for the evaluated datasets.
  • Evaluation Settings: The evaluation samples all 125 Bamboogle questions and 500 questions from StrategyQA’s training set and other datasets’ development sets.All methods share the same questions in demonstrations under the 3-shot setting.
  • Evaluation Settings: Metrics are exact match and F1 for multi-hop question answering, and accuracy for fact verification and commonsense reasoning.The study also uses Acc†, an LLM-based correctness metric produced by text-davinci-003.
  • Baselines: Baselines include Direct Prompting, CoT Prompting, ReAct, and Self-Ask, which differ in whether they generate explanations and how they interleave retrieval with reasoning.Direct Prompting retrieves with the question when augmented, while ReAct and Self-Ask use iterative actions or follow-up questions.

4.4 Implementation Details

The experiments use consistent LLM and retrieval settings and evaluate Iter-RetGen against retrieval-augmented baselines across multiple tasks. Results show strong accuracy, reliable semantic evaluation, competitive performance with fewer retrieval and generation calls, and further gains from retrieval adaptation.

  • Experimental setup: Experiments use text-davinci-003 and Contriever-MSMARCO, retrieving the top five paragraphs per query with greedy decoding.Llama-2 experiments are also reported, and ReAct and Self-Ask use at most five retrieval interactions.
  • Main results: ITER-RETGEN achieves significantly higher Acc† than retrieval-augmented baselines on four datasets and remains competitive with Self-Ask on MuSiQue and Feverous.The evaluated datasets include multi-hop question answering, fact verification, and commonsense reasoning.
  • Iteration effects: Two iterations provide the greatest performance boost, while additional iterations generally improve performance less substantially.Performance generally improves as the number of iterations increases, but the second iteration gives the greatest boost.
  • Efficiency: ITER-RETGEN matches or exceeds ReAct and Self-Ask using two LLM calls and ten retrieved paragraphs in total.It iterates retrieval-augmented chain-of-thought generation without complex intermediate processing.
  • Retrieval adaptation: Generation-augmented retrieval adaptation increases Acc† while requiring fewer iterations, and using the first model output improves over adaptation without that output.The experiments evaluate this adaptation on HotPotQA and Feverous.

4.6 Ablation Study

Ablation analyses show that later retrieval benefits from the first generation, while Iter-RetGen better preserves parametric knowledge than Self-Ask. Errors remain concentrated in retrieval and reasoning failures.

  • Generation-augmented retrieval: The second iteration achieves significantly higher answer recall by augmenting retrieval with the first LLM output, although recall quickly plateaus afterward.The first iteration retrieves using only the question and suffers from low answer recall.
  • Parametric knowledge: ITER-RETGEN better preserves CoT performance on questions solvable without retrieval and remains competitive on the complementary subset.The comparison uses subsets defined by whether CoT answers correctly without retrieval.
  • Knowledge use: ITER-RETGEN consistently outperforms Self-Ask regardless of whether retrieved knowledge mentions the answers, indicating stronger use of parametric knowledge when retrieved evidence is irrelevant or incomplete.This comparison concerns in-context non-parametric knowledge and its answer coverage.
  • Error analysis: In 20 HotPotQA failures, 65% are retrieval-related, including cases where wrong first-iteration reasoning misleads retrieval or missing information is not retrieved later.Another 10% retrieve all necessary information but fail at reasoning, while 25% are false negatives.

5 Case Study

Case studies illustrate retrieval-generation synergy: initial generations expose useful search cues despite distractors or imperfect parametric knowledge, enabling more relevant subsequent retrieval.

  • Retrieval-generation synergy: In multi-hop examples, phrases generated in the first iteration help retrieve missing knowledge for the second iteration.The examples include HotPotQA and StrategyQA, where the first retrieval can miss supporting evidence or include distractive knowledge.

6 Conclusion

The conclusion presents ITER-RETGEN as an effective approach for questions with complex information needs. It outperforms more complex retrieval-augmented workflows, and generation-augmented retrieval adaptation further improves performance while reducing overheads.

  • ITER-RETGEN demonstrates effectiveness in answering questions with complex information needs.
  • ITER-RETGEN outperforms retrieval-augmented methods with more complex workflows and could serve as a strong baseline for future retrieval-augmented generation research.
  • Generation-augmented retrieval adaptation further improves ITER-RETGEN performance while reducing retrieval and generation overheads.

Limitations

The paper's experiments use a fixed black-box language model and do not cover long-form generation. The authors identify prompting-specific optimization and finer-grained retrieval for long-form generation as future work.

  • Experiments use a fixed black-box language model that may not be equally optimized for different prompting methods.
  • The experiments do not cover long-form generation, which the authors suggest may benefit from finer-grained retrieval than ITER-RETGEN uses here.
  • The authors leave prompting-specific gradient-based optimization and further investigation of long-form generation for future work.

A Experiments Using Llama-2

The paper evaluates ITER-RETGEN with open-source Llama-2 models by replacing text-davinci-003 in Table 2. ITER-RETGEN consistently outperforms all baselines significantly.

  • ITER-RETGEN consistently outperforms all baselines significantly in the open-source-model evaluation.

B Few-Shot Prompts

The appendix provides few-shot prompts for Direct Prompting, ReAct, Self-Ask, and ITER-RETGEN across the evaluated datasets, alongside the Llama-2 experiment metric description.

  • B Few-Shot Prompts: The appendix explains that retrieved paragraphs are replaced with {Knowledge}, while CoT demonstrations match ITER-RETGEN except for retrieval augmentation.
  • B.1 HotPotQA: For HotPotQA, the appendix lists separate few-shot prompt tables for Direct Prompting, ReAct, Self-Ask, and ITER-RETGEN.
  • B.2 2WikiMultiHopQA: For 2WikiMultiHopQA, the appendix lists separate few-shot prompt tables for Direct Prompting, ReAct, Self-Ask, and ITER-RETGEN.
  • B.3 MuSiQue: For MuSiQue, the appendix lists separate few-shot prompt tables for Direct Prompting, ReAct, Self-Ask, and ITER-RETGEN.
  • B.4 Bamboogle: For Bamboogle, the appendix lists separate few-shot prompt tables for Direct Prompting, ReAct, Self-Ask, and ITER-RETGEN.
  • B.5 Feverous: For Feverous, the appendix lists separate few-shot prompt tables for Direct Prompting, ReAct, Self-Ask, and ITER-RETGEN.
  • B.6 StrategyQA: For StrategyQA, the appendix lists separate few-shot prompt tables for Direct Prompting, ReAct, Self-Ask, and ITER-RETGEN.
Loading 2305.15294v2…