Trust but Verify: Introducing DAVinCI -- A Framework for Dual Attribution and Verification in Claim Inference for Language Models
| Authors | Vipula Rawte et al. |
| Year | 2026 |
| HF Upvotes | 2 |
| arXiv | 2604.21193 |
| Download | |
| HF Page | View on Hugging Face |
Abstract
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.
Engineering Breakdown
Plain English
This paper introduces DAVinCI, a framework that addresses a critical problem with large language models: they generate fluent-sounding text that is often factually incorrect (hallucinations). The researchers propose a two-stage approach where LLM outputs are first attributed to either internal model components or external sources, then verified using entailment-based reasoning and confidence calibration. The work is evaluated on factual verification benchmarks like FEVER and CLIMATE-FEVER, domains where accuracy directly impacts trust in high-stakes applications like healthcare and law. The core insight is that by making the model's reasoning process transparent and verifiable, you can systematically catch and correct false claims before they reach users.
Core Technical Contribution
The key technical novelty is the dual-stage attribution and verification pipeline that separates concerns: first identifying where a claim comes from (internal model knowledge vs. external grounding), then independently verifying whether it's true. This is different from prior work that either tries to reduce hallucinations during generation or simply adds retrieval mechanisms. DAVinCI's architectural contribution is the entailment-based reasoning layer, which treats verification as a structured reasoning problem rather than a black-box classification task. This allows the system to explain why a claim is marked as false, not just flag it—a critical requirement for interpretability in regulated domains.
How It Works
The system takes an LLM-generated claim as input and routes it through two sequential stages. In stage one (Attribution), the framework identifies whether the claim is grounded in the model's internal parameters or should be attributed to external sources—this is done by analyzing model activations or attention patterns and comparing against a knowledge base. In stage two (Verification), the claim is formulated as a natural language entailment problem: given relevant context or evidence, does the claim logically follow? The system uses a fine-tuned entailment model to compute confidence scores, and applies calibration techniques to make these scores reliable probability estimates. Finally, claims below a confidence threshold are flagged, and the system can return evidence snippets that support or contradict each claim, enabling end-to-end traceability.
Production Impact
For production systems, DAVinCI provides a post-generation verification layer that can be plugged into any LLM pipeline without retraining the base model—you run inference as usual, then pass outputs through the verification module. This is operationally attractive because you don't need to modify your existing model serving infrastructure, though it does add latency (likely 100-500ms per claim depending on evidence retrieval cost). The primary benefit is risk mitigation: in healthcare chatbots, legal document drafting, or scientific literature systems, catching false claims before they reach users prevents regulatory violations and reputation damage. The trade-off is that you now require a reliable entailment model, access to external knowledge bases or retrieval systems, and careful calibration of confidence thresholds—if you set thresholds too high, you reject valid claims; too low, false claims slip through. For teams already using retrieval-augmented generation (RAG), this framework naturally complements that approach by adding explicit verification logic.
Limitations and When Not to Use This
The paper assumes access to relevant evidence or a knowledge base that can be efficiently retrieved and formatted for entailment checking—in domains with sparse or proprietary information (cutting-edge research, internal company data), this assumption breaks down. Entailment-based verification itself has a ceiling: it can only verify claims that are logically grounded in the evidence provided; nuanced claims, hypotheticals, or context-dependent statements may be incorrectly flagged as false. The approach does not address the root cause of hallucinations—the LLM still generates incorrect claims initially—so it's fundamentally a detection and filtering mechanism, not a prevention mechanism. Additionally, the work is evaluated primarily on English-language benchmark datasets (FEVER, CLIMATE-FEVER), so generalization to other languages, domains, or claim types remains unclear; the computational cost of the verification stage at scale (for systems generating thousands of claims per second) is not thoroughly characterized.
Research Context
This work builds on a growing body of research on LLM reliability and interpretability, sitting at the intersection of fact-checking (FEVER benchmark and related work) and model interpretability. Prior work has tackled hallucination through training approaches like RLHF and retrieval-augmented generation, but DAVinCI's dual attribution-verification design offers a complementary architectural solution that doesn't require model retraining. The paper advances the evaluation methodology for factual grounding by incorporating confidence calibration, moving beyond binary true/false judgments toward probabilistic reliability estimates. This opens research directions in combining entailment-based verification with retrieval systems, applying similar verification pipelines to multimodal models, and developing domain-specific verification modules for specialized applications.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::