PaperlayerView exact revision on arXiv

Source-linked AI summary

Effective Large Language Model Adaptation for Improved Grounding and Citation Generation

Xi Ye, Ruoxi Sun, Sercan Ö. Arik, Tomas Pfister

arXiv:2311.09533v3cs.CL

TL;DR

Hallucinated, nonfactual answers motivate grounding LLM responses in retrieved passages and providing citations. AGREE tunes LLMs using automatically constructed examples from unlabeled queries and adds test-time retrieval; across five datasets, it improves grounding and citation quality over prompting and post-hoc citing approaches.

  • Problem

    LLMs can generate plausible but nonfactual answers, motivating methods that ground response claims in supported passages and provide citations.

    Source: §1 · p. 1

  • Method

    AGREE tunes LLMs to self-ground responses with citations using automatically collected examples from unlabeled queries, then applies test-time retrieval to iteratively refine responses.

    Source: §1 · pp. 1–2

  • Results

    AGREE improves grounding and citation recall and precision over in-context and post-hoc citing baselines, generally by more than 20%, with benefits demonstrated across five datasets.

    Source: §1 · p. 2, §6 · p. 9

  • Takeaways & Limitations

    Evaluations across five datasets support learning-based adaptation over approaches relying solely on prompting or parametric knowledge.

    Source: §6 · p. 9

    Citation quality and its evaluation depend on NLI-model capability, and the experiments focus on seeking tasks in English.

    Source: §7 · pp. 9–10

Abstract

from arXiv · show

Large language models (LLMs) have achieved remarkable advancements in natural language understanding and generation. However, one major issue towards their widespread deployment in the real world is that they can generate "hallucinated" answers that are not factual. Towards this end, this paper focuses on improving LLMs by grounding their responses in retrieved passages and by providing citations. We propose a new framework, AGREE, Adaptation for GRounding EnhancEment, that improves the grounding from a holistic perspective. Our framework tunes LLMs to selfground the claims in their responses and provide accurate citations to retrieved documents. This tuning on top of the pre-trained LLMs requires well-grounded responses (with citations) for paired queries, for which we introduce a method that can automatically construct such data from unlabeled queries. The selfgrounding capability of tuned LLMs further grants them a test-time adaptation (TTA) capability that can actively retrieve passages to support the claims that have not been grounded, which iteratively improves the responses of LLMs. Across five datasets and two LLMs, our results show that the proposed tuningbased AGREE framework generates superior grounded responses with more accurate citations compared to prompting-based approaches and post-hoc citing-based approaches

1 Introduction

AGREE addresses hallucination by tuning LLMs to self-ground responses with citations and adding retrieval-based test-time refinement. It improves grounding and citation quality over prompting and post-hoc citing baselines.

1 Introduction

  • LLMs can generate plausible but nonfactual answers, and existing approaches ground claims by linking them to supported passages or adding citations.Prompting-based methods can yield mediocre grounding, while post-hoc citing may not extend well to less-known knowledge.Source: §1 · p. 1
  • AGREE fine-tunes base LLMs to produce grounded, cited responses using automatically collected training examples from unlabeled queries.An NLI model helps construct the well-grounded responses used for supervision.Source: §1 · pp. 1–2
  • At test time, self-grounding guides active retrieval of additional information to iteratively refine responses.The introduction identifies iterative test-time adaptation as a separate contribution alongside learning-based citation generation.Source: §1 · p. 2
  • AGREE improves grounding and citation recall and precision over in-context and post-hoc citing baselines, generally by more than 20%.Evaluations cover two LLMs and five datasets, including in-domain and out-of-distribution settings.Source: §1 · p. 2

2 Related Work

Prior work studies hallucination and citation generation through prompting, post-hoc citation assignment, and other factuality interventions. AGREE instead tunes LLMs to generate citations and uses them to guide retrieval.

2 Related Work

  • Hallucination has been studied across generative-language-model tasks, including through evaluation of grounding in generated responses.Source: §2 · p. 2
  • Citation approaches include zero- or few-shot prompting and post-hoc NLI models, while AGREE tunes LLMs to generate grounded responses with citations.Post-hoc citation assignment may rely heavily on parametric knowledge and may not extend well to less-known knowledge.Source: §2 · p. 2
  • Unlike factuality methods without external knowledge or approaches that inject knowledge into prompts, AGREE provides granular citations and uses self-generated citations to guide retrieval.Source: §2 · p. 3

3 Problem & Background

The task is to adapt a pretrained LLM to answer queries with responses grounded in a text corpus and citations linking each sentence to evidence. Grounding is quantified by assessing support from cited passages.

3 Problem & Background

  • The adapted model must answer a text query with a response factually grounded in a corpus and provide citations.Source: §3 · p. 3
  • Responses are segmented into sentence-level statements, each linked to a set of evidence passages from the corpus.Source: §3 · p. 3
  • The grounding score averages how well each sentence is supported by its cited passages, as assessed by an NLI model.Source: §3 · p. 3

4 AGREE Framework

AGREE combines tuning for self-grounded, cited responses with test-time retrieval that uses citation information to refine answers iteratively. Its tuning data is constructed from unlabeled queries using a base LLM and an NLI model.

4 AGREE Framework

  • AGREE combines model tuning for cited responses with test-time adaptation that retrieves passages and iteratively refines responses.The adapted model uses its citation information to guide retrieval from the corpus.Source: §4 · p. 3

4.1 Tuning LLMs

  • For unlabeled queries, the method samples base-model responses, uses NLI to link supported claims to passages, and selects the most grounded response.Claims without sufficient passage support remain uncited and are marked unsupported.Source: §4.1 · pp. 3–4
  • The selected response, its citations, and unsupported statements form supervision that teaches the model to identify unsupported claims without requiring every training response to be perfectly grounded.The process is verbalized in natural language for supervised fine-tuning.Source: §4.1 · p. 4
  • Adaptation queries come from Natural Questions, FEVER, and StrategyQA, and AGREE uses the queries without their ground-truth answers.The datasets include diverse real-user queries, fact verification, and multi-hop reasoning with implicit strategy.Source: §4.1 · p. 4

4.2 Test-time adaptation

  • At each test-time iteration, the adapted model answers from working passages, cites supporting passages, and identifies unsupported statements for further retrieval.Retrieval targets unsupported statements when present; otherwise, it seeks more query-related passages, avoiding passages already shown to the model.Source: §4.2 · p. 5
  • The TTA design avoids an additional post-hoc NLI call for citations and allows inference to be streamed and controlled through an invocation budget.The answer is regenerated from the current working passages at each iteration.Source: §4.2 · p. 5

5 Experiments

Across five datasets, AGREE improves citation grounding and precision over prompting-based and post-hoc baselines. Experiments also examine out-of-domain generalization, test-time adaptation, training data, model differences, and computation cost.

5.2 Results and analyses

  • Across five datasets, AGREE substantially outperforms baselines in citation recall and precision; Table 2 compares answer accuracy and citation grounding.Source: Table 2 · p. 6, §5.2 · p. 7

5.1 Setup

  • Experiments use text-bison and Llama-2-13B, testing in-domain and out-of-domain datasets and comparing single-call and test-time adaptation settings against three baselines.Source: §5.1 · p. 6

5.2 Results and analyses

  • AGREE improvements from in-domain tuning generalize to ASQA, QAMPARI, and Enterprise without target-domain samples, unlike ICLCITE, which needs them.Source: §5.2 · p. 7
  • TTA improves grounding and answer correctness; for Llama-2, answer correctness rises by 3.1 on NQ and 3.7 on ASQA.Examples show retrieval can add support for an unsupported statement or find passages for a more complete response.Source: §5.2 · pp. 7, 9, Figure 4 · p. 9
  • AGREE outperforms baselines on both LLMs, while Llama-2-13B generally trails text-bison in answer correctness and citation quality.AGREE without TTA has answer correctness similar to ICLCITE; passage quality affects correctness, while POSTATTR performs especially poorly on Enterprise.Source: §5.2 · p. 7
  • Across datasets, multi-dataset tuning improves citation quality and answer correctness over NQ-only tuning; Table 3 also reports benefits from distilling text-bison outputs into Llama-2-13B.The distilled Llama-2-13B still does not match text-bison, with the capability gap constraining performance.Source: Table 3 · p. 8, §5.2 · pp. 8–9
  • AGREE without TTA uses less than half the tokens of ICLCITE while achieving better citation quality; adding TTA costs more computation but improves citation quality and answer correctness.Source: §5.2 · pp. 7–8, Table 4 · p. 8

6 Conclusion

AGREE tunes a pre-trained LLM to ground its responses in retrieved passages using automatically collected data. Its grounding capability also supports test-time response improvement.

  • AGREE adapts a pre-trained LLM to self-ground responses in retrieved passages.The framework is introduced as Adaptation for GRounding EnhancEment.Source: §6 · p. 9
  • Automatically collected data supervise the model’s grounded-response generation.Source: §6 · p. 9
  • The integrated grounding capability enables the LLM to improve responses at test time.Source: §6 · p. 9

7 Limitations and future work

AGREE’s data creation and evaluation rely on an NLI model, and the work identifies limits in citation quality, evaluation, cost, and the scope studied.

  • NLI-based data creation may favor fully supported citations, while NLI-based evaluation also limits measured grounding and citation quality.The authors suggest human-annotated citations for partially supported claims as future work.Source: §7 · p. 9
  • AGREE currently uses supervised fine-tuning, with RLHF left for future work and tuning requiring a one-time adaptation cost.Source: §7 · p. 9
  • The study mainly covers English open-domain question answering, leaving other languages and long-form generation as future directions.Source: §7 · pp. 9–10
  • A plausible but incorrect citation could make an unsupported statement more convincing to users.Source: §7 · p. 10

A Details of Data Generation

The data-generation process samples initial responses using retrieved passages, then uses NLI judgments to attach citations and identify unsupported statements before verbalizing training examples.

  • TRUE NLI is used to assess passage support; the paper describes it as trained on six datasets covering NLI, paraphrase detection, and fact verification.Source: A Details of Data Generation · p. 13
  • Wikipedia-based datasets use the 2018-12-20 snapshot and GTR-large as the retriever.Source: A Details of Data Generation · p. 13
  • Initial responses are sampled zero-shot from the base LLM with five retrieved passages appended to the query prompt.The authors choose zero-shot prompting to avoid bias from few-shot examples and to match expected deployment on new queries.Source: A Details of Data Generation · p. 13, Figure 5 · p. 13
  • For each response sentence, the process links its maximally supporting seen passage above 0.7 and marks it unsupported if no passage exceeds 0.5.Source: A Details of Data Generation · p. 13
  • Training examples verbalize citations in square brackets at sentence ends and place unsupported statements after the response.Figure 6 presents the verbalization template.Source: A Details of Data Generation · p. 13

B Details of Experimental Setup

The experimental setup describes fine-tuning, decoding, and distillation choices; a separate comparison reports that AGREE narrows the performance gap between a smaller Llama-2 model and more capable models.

B Details of Experimental Setup

  • Both models use LoRA tuning; Llama-2 uses the chat checkpoint, LoRA rank 32, and a 1e-5 learning rate, while text-bison uses 1,000 training steps.The text-bison step count was selected from 500, 1,000, and 2,000; its tuning uses the API defaults otherwise.Source: B Details of Experimental Setup · p. 13
  • Evaluation follows ALCE’s implementation, data split, and prompt template, using temperature 0.25 and one sample per evaluation.Source: B Details of Experimental Setup · p. 14
  • For distillation, Llama-2-13B is tuned on data created with text-bison-001 for four epochs at learning rate 1e-5 and LoRA rank 32.Source: B Details of Experimental Setup · p. 14

C Comparison to ICLCITE on More Capable LLMs

D License of Datasets

The paper lists the licenses for five datasets used in its work.

E Additional Examples of Tuning Data

The appendix points to concrete examples of tuning data created from NQ and StrategyQA.

Loading 2311.09533v3…