PaperlayerView exact revision on arXiv

Source-linked AI summary

Trust but Verify: Introducing DAVinCI -- A Framework for Dual Attribution and Verification in Claim Inference for Language Models

Vipula Rawte, Ryan Rossi, Franck Dernoncourt, Nedim Lipka

arXiv:2604.21193v1cs.AI

TL;DR

LLMs remain vulnerable to factual inaccuracies despite their fluency, motivating systems that make generated claims verifiable and interpretable. DAVinCI combines internal and external attribution with entailment-based verification and confidence calibration, and its evaluations show consistent improvements over verification-only baselines across FEVER and CLIMATE-FEVER. The framework further identifies evidence quality and calibration as important factors in model reliability, while its applicability is bounded by evidence, reasoning, language, and calibration assumptions.

  • Problem

    LLMs can produce fluent but factually inaccurate or hallucinated content, creating a need for trustworthy and verifiable outputs.

  • Method

    DAVinCI attributes claims to internal model components and external sources, then verifies them with entailment-based reasoning and confidence calibration.

  • Results

    DAVinCI consistently outperforms standard verification-only baselines on FEVER and CLIMATE-FEVER, with evidence quality and confidence calibration improving reliability.

  • Takeaways & Limitations

    DAVinCI provides a modular framework for producing LLM claims with transparent, auditable justifications grounded in evidence.

  • Takeaways & Limitations

    DAVinCI assumes high-quality evidence, relies on static entailment models, is evaluated only on English datasets, and uses manually tuned calibration thresholds.

Abstract

from arXiv · show

Large Language Models (LLMs) have demonstrated remarkable fluency and versatility across a wide range of NLP tasks, yet they remain prone to factual inaccuracies and hallucinations. This limitation poses significant risks in high-stakes domains such as healthcare, law, and scientific communication, where trust and verifiability are paramount. In this paper, we introduce DAVinCI - a Dual Attribution and Verification framework designed to enhance the factual reliability and interpretability of LLM outputs. DAVinCI operates in two stages: (i) it attributes generated claims to internal model components and external sources; (ii) it verifies each claim using entailment-based reasoning and confidence calibration. We evaluate DAVinCI across multiple datasets, including FEVER and CLIMATE-FEVER, and compare its performance against standard verification-only baselines. Our results show that DAVinCI significantly improves classification accuracy, attribution precision, recall, and F1-score by 5-20%. Through an extensive ablation study, we isolate the contributions of evidence span selection, recalibration thresholds, and retrieval quality. We also release a modular DAVinCI implementation that can be integrated into existing LLM pipelines. By bridging attribution and verification, DAVinCI offers a scalable path to auditable, trustworthy AI systems. This work contributes to the growing effort to make LLMs not only powerful but also accountable.

1. Introduction

DAVinCI addresses the limitations of factual reliability in LLMs by integrating attribution with entailment-based verification and confidence calibration. Its evaluation examines evidence selection, calibration, and retrieval across multiple configurations.

  • LLMs can generate fluent but factually incorrect or hallucinated content, creating a need for robust factuality verification in high-stakes applications.
  • DAVinCI integrates attribution and verification into a modular framework for trustworthy LLM inference.The framework is designed to support reproducible research and integration into existing pipelines.
  • DAVinCI attributes claims to internal model components and external sources, then verifies them through entailment reasoning and confidence calibration.This dual structure is intended to make claims more transparent and auditable.
  • The evaluation spans multiple entailment models and attribution configurations, with full-passage evidence improving verification accuracy and F1-score.Confidence recalibration provides flexibility in trading off precision and recall.
  • The ablation study isolates the effects of evidence span selection, recalibration thresholds, and retrieval quality.

2. Related Work

Related work addresses factuality verification, retrieval, attribution, and confidence calibration as important but often separate capabilities. DAVinCI combines these strands in a unified pipeline that uses attribution to inform verification.

  • FEVER established claim verification against retrieved Wikipedia evidence, while CLIMATE-FEVER extended the paradigm to another domain.
  • RAG systems ground generated responses in external sources but often leave retrieval, generation, and factual verification loosely coupled.
  • Attribution methods formalize source grounding but can remain domain-specific or require extensive supervision.
  • Confidence calibration helps distinguish supported, refuted, and uncertain claims in verification settings.Temperature scaling and trust-aware scoring are cited as examples of calibration approaches.
  • DAVinCI unifies internal and external attribution with entailment-based verification in a modular and scalable pipeline.It uses attribution to inform verification and calibrated confidence to produce interpretable outputs.

3. Proposed Method

DAVinCI uses a two-stage pipeline that attributes claims to relevant evidence and verifies claim–evidence relationships with an entailment classifier. It supports full-evidence or span-based attribution, confidence recalibration, and aggregation across multiple evidence passages.

  • Architecture: DAVinCI first identifies evidence that may support or refute a claim, then passes the attributed evidence to an entailment-based verifier.The framework is designed to provide evidence-grounded justifications alongside truthfulness assessments.
  • Attribution Module: Attribution scores compare a claim with candidate evidence using a similarity function such as cosine similarity or a dense-retriever score.The scoring formulation is Score_attr(c, e_i) = sim(c, e_i).
  • Attribution Module: Full-evidence attribution uses available gold-standard evidence directly, whereas span-based attribution extracts a relevant answer span with a question-answering model.The span-based setup simulates cases where only partial evidence is accessible.
  • Verification Module: The verification module classifies each claim–evidence pair as Supported, Refuted, or Not Enough Information and returns a confidence score.The classifier receives the claim and attributed evidence as input and uses a transformer-based entailment model.
  • Confidence Recalibration: Confidence recalibration overrides predictions below threshold τ, with a default threshold of 0.6, by assigning Not Enough Information.For multiple evidence passages, verification outputs can be combined using majority voting or weighted averaging.

4. Experiments

DAVinCI is evaluated on FEVER and CLIMATE-FEVER using multiple NLI models and a verification-only baseline. Across these experiments, DAVinCI improves classification and precision, recall, and F1 performance.

  • Datasets: The evaluation uses FEVER and CLIMATE-FEVER, which provide annotated claims, evidence, and ground-truth labels for claim verification.FEVER contains contradiction, entailment, and neutral claims with verified Wikipedia evidence; CLIMATE-FEVER focuses on climate-change fact checking.
  • Models and task: Four NLI models are benchmarked on the verification entailment task: DeBERTa, RoBERTa, BART, and RoBERTa trained on SNLI, MNLI, FEVER, and ANLI.The comparison measures how well each model identifies and validates factual relationships in text.
  • Baseline: DAVinCI is compared with a verification-only baseline using full evidence to isolate the effects of attribution quality and confidence calibration.The baseline provides verification without the integrated attribution and calibration components.
  • Main results: Accuracy increases from 0.42 to 0.48 for deberta-large-mnli, while macro F1 rises from 0.36 to 0.41 under DAVinCI.Other models also improve: roberta-large-mnli reaches 0.44 accuracy and 0.38 macro F1, while bart-large-mnli reaches 0.43 and 0.37.
  • Main results: DAVinCI integration raises precision, recall, and F1-scores across macro and weighted averages for all evaluated models.deberta-large-mnli achieves the highest reported macro precision of 0.61 and weighted precision of 0.62.

5. Ablation Study

The ablation study examines evidence granularity and recalibration thresholds across FEVER and CLIMATE-FEVER. Full evidence consistently outperforms spans, while threshold changes trade recall for precision with limited accuracy loss.

  • Evidence selection: On FEVER, roberta-large-snli reaches 0.48 accuracy, macro F1, and weighted F1 with full evidence.deberta-large-mnli leads precision, reaching 0.57 macro precision and 0.58 weighted precision.
  • Evidence selection: On CLIMATE-FEVER, roberta-large-snli achieves 0.65 accuracy and 0.57 weighted F1 with full evidence.deberta-large-mnli and roberta-large-mnli reach 0.61 accuracy, while weighted F1 remains above 0.54.
  • Threshold tuning: At threshold 0.7, deberta-large-mnli leads FEVER accuracy with 0.47 and macro precision with 0.60.Raising the threshold to 0.8 lowers accuracy and F1 in examples because more claims are filtered as Not Enough Info, while precision remains 0.58–0.60.
  • Threshold tuning: On CLIMATE-FEVER, threshold increases can raise precision while reducing recall and F1, with accuracy trade-offs remaining ≤1.6%.At threshold 0.9, roberta-large-snli reaches 0.69 precision, but F1 declines because recall falls.
  • Evidence selection: Full-evidence variants consistently outperform span-based variants across metrics on both FEVER and CLIMATE-FEVER.On FEVER, full-evidence models outperform span-based models by 9–18%; on CLIMATE-FEVER, accuracy gains range from +1.6% to +19.6%.
  • Threshold tuning: Across datasets, the study identifies threshold 0.7 as the best overall precision–recall trade-off.The ablation varies thresholds 0.7, 0.8, and 0.9 to measure how recalibration affects performance and trustworthiness.

6. Conclusion and Future Work

DAVinCI integrates attribution and entailment-based verification to make LLM outputs more factually reliable, interpretable, transparent, and auditable. Experiments on FEVER and CLIMATE-FEVER outperform verification-only baselines, while future work targets broader reasoning, retrieval, attribution, and language settings.

  • Conclusion: DAVinCI unifies evidence attribution and entailment-based verification to improve the factual reliability of LLM outputs.The framework is designed to make generated claims justifiable and verifiable within a transparent pipeline.
  • Conclusion: Experiments on FEVER and CLIMATE-FEVER show that DAVinCI consistently outperforms standard verification-only baselines.The evaluation is paired with an ablation study of evidence quality and confidence calibration.
  • Future Work: Future work will extend DAVinCI to open-domain retrieval, multi-hop reasoning, internal attribution tracing, generative tasks, and multilingual or low-resource environments.These extensions aim to broaden the framework beyond its current evaluated settings.

7. Ethics Statement

The work uses publicly available FEVER and CLIMATE-FEVER datasets without collecting new human data or personally identifiable information. DAVinCI is designed to operate transparently by attributing claims to evidence and calibrating confidence.

  • Ethics Statement: The study collects no new human data or personally identifiable information.All experiments use publicly available datasets.
  • Ethics Statement: Experiments rely on publicly available FEVER and CLIMATE-FEVER datasets used for factuality and claim-verification research.These datasets contain annotated claims, supporting evidence, and ground-truth labels.
  • Ethics Statement: DAVinCI attributes claims to evidence and calibrates confidence to reduce overconfident misclassifications.The framework is intended to operate transparently.

8. Discussion and Limitations

DAVinCI improves factual reliability and interpretability by combining attribution and verification, with evidence quality and confidence calibration emerging as important factors. Its main limitations concern evidence availability, reasoning capability, attribution coverage, language scope, and calibration generalization.

  • Discussion: Integrating attribution and verification into one pipeline improves the factual reliability and interpretability of LLM outputs.DAVinCI consistently outperforms the baseline across multiple metrics.
  • Discussion: Full evidence attribution yields more accurate and faithful verification than span-based extraction, which can lose critical context.The finding supports prioritizing evidence completeness and relevance over brevity.
  • Discussion: Confidence thresholds can be tuned to balance precision and recall, providing a mechanism for trust calibration.The ablation study identifies confidence calibration as an important contributor to model reliability.
  • Limitations: DAVinCI assumes high-quality evidence, relies on static entailment models, lacks internal attribution modeling, and is evaluated only on English-language datasets.These constraints limit applicability in open-domain, nuanced or multi-hop, multilingual, and low-resource settings.
  • Limitations: Manually tuned confidence thresholds may not generalize across domains or tasks.Adaptive or learned calibration strategies are proposed as a way to improve robustness and reduce task-specific tuning.
Loading 2604.21193v1…