PaperlayerView exact revision on arXiv

Source-linked AI summary

Learning Fine-Grained Grounded Citations for Attributed Large Language Models

Lei Huang, Xiaocheng Feng, Weitao Ma, Yuxuan Gu, Weihong Zhong, Xiachong Feng, Weijiang Yu, Weihua Peng, Duyu Tang, Dandan Tu, Bing Qin

arXiv:2408.04568v1cs.CLcs.AI

TL;DR

LLMs can hallucinate, while existing attribution approaches often rely on in-context learning and coarse citations that limit fine-grained verification. FRONT trains models to select supporting quotes and generate responses grounded in them. On ALCE, it improves citation quality over baselines, and human evaluation finds the quotes benefit user verification.

  • Problem

    LLMs still produce hallucinations, and current attributed approaches rely on in-context learning, limiting inherent attribution capability and citation quality.

    Source: §1 · p. 1

  • Method

    FRONT trains LLMs to select supporting quotes from retrieved sources, condition generation on them, and align grounding with generation through preference optimization.

    Source: §1 · p. 2

  • Results

    14.21% average citation-quality improvement over all baselines with LLaMA-2-7B; human evaluation found the quotes high quality and beneficial to user verification.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Supporting quotes can serve as fine-grained citations that improve citation quality and facilitate user verification.

    Source: §1 · p. 2, §7 · p. 9

    The framework was validated predominantly on 7B and 13B models, and relies on retrieving relevant documents at one time.

    Source: §8 · p. 9

Abstract

from arXiv · show

Despite the impressive performance on information-seeking tasks, large language models (LLMs) still struggle with hallucinations. Attributed LLMs, which augment generated text with in-line citations, have shown potential in mitigating hallucinations and improving verifiability. However, current approaches suffer from suboptimal citation quality due to their reliance on in-context learning. Furthermore, the practice of citing only coarse document identifiers makes it challenging for users to perform fine-grained verification. In this work, we introduce FRONT, a training framework designed to teach LLMs to generate Fine-Grained Grounded Citations. By grounding model outputs in fine-grained supporting quotes, these quotes guide the generation of grounded and consistent responses, not only improving citation quality but also facilitating fine-grained verification. Experiments on the ALCE benchmark demonstrate the efficacy of FRONT in generating superior grounded responses and highly supportive citations. With LLaMA-2-7B, the framework significantly outperforms all the baselines, achieving an average of 14.21% improvement in citation quality across all datasets, even surpassing ChatGPT.

1 Introduction

FRONT addresses limitations in attributed LLMs by training models to ground responses in supporting quotes, which can serve as fine-grained citations. On ALCE, it improves citation quality over baselines, while human evaluation finds its quotes support user verification.

1 Introduction

  • LLMs can hallucinate fabricated facts or unfaithful content, risking misinformation and undermining reliability and trustworthiness.Attributed systems add in-line citations, which can improve factuality, alleviate hallucinations, and simplify verification.Source: §1 · p. 1
  • Current attribution approaches rely mainly on in-context learning or post-hoc retrieval, rather than an inherent attribution capability, compromising citation quality.Citations often identify only documents or URLs, making it difficult to pinpoint supporting evidence, particularly in lengthy documents.Source: §1 · pp. 1–2
  • FRONT uses a two-stage training framework: the model selects supporting quotes from retrieved sources, generates conditioned on them, then aligns grounding and generation with preference optimization.An automatic data pipeline constructs training examples through retrieval, relevance reranking, attributed answer generation, and filtering for informativeness and attributability.Source: §1 · p. 2
  • 14.21% average citation-quality improvement over all baselines with LLaMA-2-7B.The paper also reports that human evaluation found FRONT-generated quotes high quality and beneficial to user verification.Source: §1 · p. 2
  • FRONT is reported to generate less hallucination and generalize across base models; human evaluation finds its quotes high quality and beneficial to user verification.The introduction reports the generalization finding without specifying its evaluation conditions.Source: §1 · p. 2

2 Related Work

Prior work addresses knowledge-intensive tasks with retrieved documents and improves verifiability through attributed responses. FRONT advances attributed LLMs by citing specific extractive quotes rather than only document identifiers.

Retrieval Augmented Generation

  • Retrieval-augmented LLMs incorporate retrieved documents to mitigate knowledge gaps, but may struggle with irrelevant or contradictory documents and contextual evidence use.These limitations can degrade performance or lead to hallucinations.Source: §2 · p. 2

Attributed Large Language Models

  • Attributed LLMs generate responses linked to evidence; prior approaches include in-context citation generation and post-hoc retrieval of evidence.The cited examples are Gao et al. (2023b) for in-context learning and Gao et al. (2023a) and Xu et al. (2023b) for post-hoc attribution.Source: §2 · p. 2
  • FRONT advances attributed LLMs by pinpointing and citing specific extractive quotes instead of predominantly citing document identifiers.The paper characterizes this as a more fine-grained form of attribution.Source: §2 · p. 2

3 Task Formulation and Methodology

The section formalizes attributed responses as statements with inline citations and presents an automatic data pipeline and two-stage training recipe for fine-grained attribution.

3 Task Formulation and Methodology

  • Given a query and retrieved documents, the model produces statements with inline citations to passages in those documents.Each statement cites one or more passages from the retrieved corpus.Source: §3 · p. 2

3.1 Automatic Data Generation Pipeline

  • The automatic pipeline constructs attributed training examples through document retrieval, ChatGPT answer generation, and informativeness and attributability filtering.It uses AQuAMuSe queries, retrieves 100 Sphere documents, and reranks them to five; ChatGPT generates cited answers from those five documents.Source: §3.1 · p. 3
  • Filtering excludes answers below an attributability threshold, yielding nearly 8,000 synthetic training entries.The filtering criteria also assess whether answers are sufficiently informative and whether cited documents support them.Source: §3.1 · pp. 3–4

3.2.1 Grounding Guided Generation

  • G3 trains the model to generate source-extracted supporting quotes paired with document identifiers before producing an attributed answer.ChatGPT extracts quote segments supporting statements because the generated dataset lacks this grounding content.Source: §3.2.1 · p. 4
  • Grounded quotes provide document-origin signals and are intended to reduce irrelevant information and hallucination risk while improving citation quality.Each quote’s document identifier supervises the origin of the extractive segment.Source: §3.2.1 · p. 4

3.2.2 Consistency-Aware Alignment

  • CCA addresses quote-answer inconsistencies by preferring oracle-grounded answers consistent with the same quotes over weaker LLM answers that are inconsistent with them.The weaker answers serve as automatically constructed negative samples for contrastive preference training.Source: §3.2.2 · pp. 4–5
  • The DPO objective optimizes the policy against a G3-initialized reference model, with β controlling divergence and τw and τl denoting consistent and inconsistent answers.The policy is optimized by minimizing the stated loss on the preference dataset.Source: §3.2.2 · p. 5

4 Experimental Settings

Experiments use three ALCE long-form QA datasets, assess citation quality and dataset-specific correctness, and compare prompting, post-hoc retrieval, and training baselines.

4.1 Datasets

  • ALCE evaluation covers ASQA’s ambiguous multi-answer questions, ELI5’s explanatory questions, and QAMPARI’s entity-list answers.Source: §4.1 · p. 5

4.2 Evaluation Metrics

  • Citation quality is assessed through citation recall, precision, and their harmonic-mean F1 using TRUE; correctness uses dataset-specific answer or claim measures.Source: §4.2 · pp. 5–6

4.3 Baselines

  • The baselines span prompting-based, post-hoc retrieval, and training-based methods.Source: §4.3 · p. 6

4.3.1 Prompting-based Methods.

  • Prompting baselines use few-shot cited-answer demonstrations with five retrieved documents across GPT-3.5-Turbo, LLaMA-2, Mistral, and open-source SFT counterparts.Source: §4.3.1 · p. 6

4.3.2 Post-hoc Retrieval Methods.

  • The post-hoc baseline generates answers closed-book, then uses GTR to cite the most relevant document among the top 100 retrieved documents for each statement.Source: §4.3.2 · p. 6

4.3.3 Training-based Methods.

  • Training baselines include Self-RAG’s adaptive retrieval and reflection (Asai et al., 2023) and VANILLA-SFT’s direct generation of cited answers from queries and documents.Source: §4.3.3 · p. 6

4.4 Implement Details

  • FRONT is implemented with LLaMA-2-7B and LLaMA-2-13B, using the retrieval settings of Gao et al. (2023b).Source: §4.4 · p. 7

5 Results and Analysis

FRONT improves citation quality across the ALCE datasets, including over ChatGPT, and the analyses examine its scaling, generalization, and training components. The ablations favor filtered data, quote-guided generation, and consistency-aware alignment.

5.1 Overall Results

  • VANILLA-SFT improves citation F1 over prompting on ASQA (17.55 → 65.61), ELI5 (4.54 → 41.15), and QAMPARI (6.19 → 21.35).These results show that supervised fine-tuning with the generated training data can improve citation quality.Source: §5.1 · p. 7
  • 3.32%, 18.04%, and 21.28% are FRONT’s citation-quality gains over ChatGPT on ASQA, ELI5, and QAMPARI, respectively.FRONT also bridges VANILLA-SFT’s cited ELI5 gaps against Mixtral-8×7B-Instruct and ChatGPT.Source: §5.1 · p. 7
  • 3.23%, 4.97%, and 1.33% citation-quality improvements accompany scaling FRONT from 7B to 13B on ASQA, ELI5, and QAMPARI, respectively.The paper describes this trend as evidence of scalability with model size.Source: §5.1 · p. 7

5.2 Ablation Study

  • Figure 4 shows that filtering improves citation quality over unfiltered data while using less data.Filtering removes queries with non-informative answers or answers that fail attribution criteria.Source: Figure 4 · p. 7, §5.2 · p. 7

5.1 Overall Results

  • FRONT generalizes from AQuAMuSe training data to ALCE’s varied queries and answer types, but has lower QAMPARI Rec.-5, likely due to concatenated-entity answers.The authors note FRONT was not specifically optimized for correctness, yet report modest improvements over VANILLA-SFT on ASQA and QAMPARI.Source: §5.1 · p. 7

5.2 Ablation Study

  • Table 2’s ablations find that grounding guidance outperforms direct citation generation, and SELF-GUIDE outperforms PROMPT-GUIDED on ELI5.SELF-GUIDE selects quotes before generating attributed answers; PROMPT-GUIDED uses quotes extracted by ChatGPT in the prompt.Source: Table 2 · p. 8, §5.2 · p. 8
  • Table 2 shows FRONT with consistency-aware alignment improves citation quality over SELF-GUIDE, while Figure 6 associates FRONT with fewer hallucinations than SELF-GUIDE.The Figure 6 comparison describes FRONT as producing more faithful outputs.Source: Table 2 · p. 8, §5.2 · pp. 8–9, Figure 6 · p. 8
  • Increasing training data from 2k to 8k instances significantly enhances citation quality across the two training stages.The experiments compare randomly sampled training subsets for 7B model variants.Source: §5.2 · p. 9

6 Human Evaluation

Human evaluation assessed grounded quotes for authenticity and helpfulness using four annotators. The reported annotations show high quote quality and substantial annotator agreement.

6 Human Evaluation

  • The evaluation sampled ChatGPT-extracted quotes from 50 data points and FRONT-7B quotes from 50 data points per dataset.Four annotators with relevant expertise and at least a bachelor’s degree evaluated quote authenticity and helpfulness.Source: §6 · p. 9
  • Fleiss’ Kappa was 0.82, indicating high agreement among annotators.Two annotators evaluated each response.Source: §6 · p. 9
  • The human evaluation found both ChatGPT-extracted and FRONT-generated quotes to be of high quality.The study measured whether quotes genuinely came from the corresponding documents and how helpful they were for answering queries.Source: Table 3 · p. 9, §6 · p. 9

7 Conclusion

FRONT trains LLMs to select supporting quotes before generating attributed answers, then uses preference optimization to improve consistency between grounding and generation. The paper reports superior grounded responses and supportive citations, with reduced hallucinations and benefits for user verification.

7 Conclusion

  • FRONT is a two-stage framework that selects supporting quotes to guide generation and uses preference optimization to improve quote-answer consistency.The quotes serve as fine-grained citations for attributed answers.Source: §7 · p. 9

8 Limitation

The framework is evaluated mainly on 7B and 13B models, and uses one-time retrieval; long-form QA correctness remains difficult to assess and is only modestly improved. [p13, p16]

8 Limitation

  • Validation is predominantly on 7B and 13B models; larger models such as LLaMA-2 70B were not explored because of computational constraints.The authors identify adaptive retrieval as a possible future direction, since the framework retrieves relevant documents at one time.Source: §8 · p. 9
  • Long-form QA correctness is difficult to evaluate, so the framework primarily improves citation quality, with modest correctness gains.The authors call for more robust metrics to assess long-form QA correctness.Source: §8 · p. 9

A.1 Data Statistic

The appendices describe the generated data, filtering and evaluation procedures, implementation settings, and additional experiments on data scale and model architectures.

A.1 Data Statistic

  • The pipeline retained 5,667 long-form and 2,431 short-form questions from 7,725 and 4,000 initial queries, respectively.Long-form queries came from AQUAMUSE; short-form queries were sampled from Natural Questions.Source: §A.1 · p. 13

A.2 Details of Data Filtering

  • Answers were filtered using average attribution scores, with those below 0.8 removed after statement–document support was scored.The threshold was selected through development-set testing with 100 manually annotated samples; the discriminator was trained on 8,834 annotated instances.Source: §A.2 · p. 13

B Details of Evaluation Metrics

  • Evaluation details include MAUVE fluency scores for ASQA and ELI5, and ROUGE-L for response quality and textual coherence.Fluency is omitted for QAMPARI because its responses are typically lists of short answers.Source: B Details of Evaluation Metrics · pp. 13–14

C.1 Prompts for Prompting-based Methods; C.2 Instructions for FRONT

  • Prompting-based baselines use vanilla prompts, while FRONT uses question-type-specific instruction formats for long- and short-form QA.The prompt formats are specified separately for ASQA/ELI5 and QAMPARI.Source: §C.1 · p. 14, §C.2 · p. 14

D.1 Training Details of FRONT

  • FRONT training uses inputs up to 2,048 tokens, with five grounding-generation epochs followed by two contrastive-alignment epochs at β = 0.1.Training used four 80GB Nvidia A100 GPUs; inference uses vLLM.Source: §D.1 · p. 14

D.2 Retrieval Settings

  • Evaluation uses GTR for ASQA and QAMPARI retrieval, and BM25 for ELI5.Source: §D.2 · p. 14

E.1 The Effect of Training Data Scale.

  • The data-scale experiment examines citation quality across ASQA, ELI5, and QAMPARI in both training stages.The supplied passage describes the experiment and figure scope but does not report its outcome.Source: §E.1 · p. 14

E.2 The Generalization Across Model Architectures.

  • Replacing LLaMA-2-7B with Mistral-7B produces greater performance enhancements, and FRONT generalizes across the tested architectures.The comparison is against models using the same foundation model with in-context learning or direct fine-tuning on automatically generated data.Source: §E.2 · p. 14

E.3 The effect of β in Consistency-Aware Alignment Training Stage

As β increases, attribution performance declines; the authors suggest this reflects tension between grounding and attribution after the first training stage.

E.3 The effect of β in Consistency-Aware Alignment Training Stage

  • In DPO, β controls the strength of the KL penalty; larger values keep the policy closer to the reference model.The stated typical range is 0.1 to 0.5.Source: §E.3 · p. 14
  • The β sweep adjusts values from 0.1 to 0.5 and evaluates attribution quality on ASQA and ELI5.Source: §E.3 · pp. 14–15
  • Attribution performance gradually decreases as β increases, with higher values associated with difficulty escaping inconsistent attributed-answer constraints.The authors suggest the first G3 stage may introduce inconsistency between grounding and attribution.Source: §E.3 · p. 15

F Full Results

The full-results appendix identifies the reported datasets, evaluation dimensions, and additional result figures, but the supplied passages do not give table values.

F Full Results

Related papers

Loading 2408.04568v1…