💻 computer science

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

This paper introduces DAVinCI, a two-stage framework that enhances the factual reliability and interpretability of Large Language Models by combining dual attribution of claims to internal and external sources with entailment-based verification, resulting in significant improvements in accuracy and accountability across high-stakes domains.

Original authors: Vipula Rawte, Ryan Rossi, Franck Dernoncourt, Nedim Lipka

Published 2026-04-24

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Vipula Rawte, Ryan Rossi, Franck Dernoncourt, Nedim Lipka

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a very talented, fast-talking friend who knows a little bit about everything. They can tell you stories, write poems, and explain complex topics with incredible fluency. However, this friend has a bad habit: sometimes, when they don't know the answer, they just make something up to sound confident. In the world of AI, we call this "hallucinating."

If this friend were a doctor, a lawyer, or a scientist, their made-up facts could be dangerous. That's where DAVinCI comes in.

The Problem: The "Confident Liar"

Large Language Models (LLMs) are like that fast-talking friend. They are great at generating text, but they often lack a "truth check." They might say, "Caroline Kennedy is an astronaut," with total confidence, even though she's actually a diplomat. Traditional systems try to fact-check after the fact, but they often miss the nuance of where the information came from or how sure the model is.

The Solution: DAVinCI (The Double-Check Detective)

The authors created a framework called DAVinCI (Dual Attribution and Verification). Think of it as hiring a Detective and a Judge to work together before the final verdict is given.

Here is how it works, broken down into two simple steps:

1. The Detective (Attribution)

Before the AI gives you an answer, the "Detective" asks: "Where did you get this idea?"

The Old Way: The AI just says, "I know this!"
The DAVinCI Way: The AI must point to the specific page in a book (or a specific sentence on the internet) that supports its claim.
- Analogy: Imagine a student taking a test. Instead of just writing the answer, they have to highlight the exact sentence in their textbook that proves they are right. If they can't find the sentence, the system knows they are guessing.

The paper tested two ways the Detective works:

The "Full Passage" Detective: Looks at the whole paragraph. (Most accurate, like reading the whole chapter).
The "Snippet" Detective: Looks at just a few words. (Faster, but might miss context, like reading a headline without the article).

2. The Judge (Verification)

Once the Detective finds the evidence, the "Judge" steps in. The Judge doesn't just look at the answer; they compare the AI's claim against the evidence found by the Detective.

The Verdict: The Judge decides if the claim is Supported (True), Refuted (False), or if there is Not Enough Info (We don't know yet).
The Confidence Check: This is the secret sauce. If the Judge is only 50% sure, DAVinCI says, "Hold on, we aren't confident enough to call this 'True'." It downgrades the answer to "Not Enough Info" to prevent the AI from confidently lying.

Why is this better? (The Results)

The researchers tested DAVinCI on two big datasets (FEVER and CLIMATE-FEVER), which are like massive libraries of true and false statements about general knowledge and climate change.

The Scorecard: When they compared DAVinCI to standard AI fact-checking, DAVinCI improved accuracy by 5% to 20%.
The "Full Passage" Advantage: They found that using the whole paragraph of evidence (the "Full Passage" Detective) was much better than just using snippets. It's like trying to solve a mystery by reading the whole police report rather than just the headline.
The Safety Net: By adjusting the "confidence threshold" (how sure the Judge needs to be), they could trade off between catching every possible error (high recall) or making sure they never accidentally call a lie "true" (high precision).

The Big Picture

Think of DAVinCI as a quality control filter for AI.

Without DAVinCI: You get a smooth, confident, but potentially fake story.
With DAVinCI: You get a story that comes with a "receipt" (the evidence) and a "warranty" (the verification). If the AI isn't sure, it admits it, rather than making things up.

Why Should You Care?

We are moving into a world where AI writes our news, diagnoses our illnesses, and argues our legal cases. We can't just trust that the AI is telling the truth. DAVinCI offers a way to make AI accountable. It forces the AI to say, "Here is my answer, here is the proof, and here is how sure I am."

It's not about making the AI smarter; it's about making the AI honest and auditable. Just like you wouldn't trust a financial advisor who couldn't show you their math, we shouldn't trust an AI that can't show us its sources. DAVinCI is the tool that makes the AI show its work.

1. Problem Statement

Large Language Models (LLMs) exhibit remarkable fluency but suffer from a critical vulnerability: the tendency to generate hallucinations and factually incorrect content. This poses severe risks in high-stakes domains like healthcare, law, and scientific communication.

Current Limitations: Existing solutions often treat attribution (identifying the source of information) and verification (checking if the claim is true) as isolated tasks.
The Gap: Most Retrieval-Augmented Generation (RAG) systems and verification pipelines fail to capture the nuanced interplay between evidence selection and entailment reasoning. They often lack mechanisms to calibrate confidence, leading to overconfident misclassifications.

2. Methodology: The DAVinCI Framework

The authors propose DAVinCI (Dual Attribution and Verification), a modular, two-stage framework designed to enhance the factual reliability and interpretability of LLM outputs.

Core Architecture

The framework operates in a unified pipeline with two distinct stages:

Stage 1: Attribution Module
This stage identifies relevant evidence to support or refute a generated claim ( $c$ ). The paper explores two strategies for evidence retrieval:

Full Evidence Attribution: Uses the entire evidence passage (gold-standard or retrieved) without modification.
Span-Based Attribution: Simulates a real-world retrieval scenario where a Question-Answering (QA) model extracts only the most relevant text span from the evidence, mimicking partial access.
Scoring: Attribution scores are computed using similarity functions (e.g., cosine similarity or dense retriever scores) between the claim and evidence.

Stage 2: Verification Module
The attributed evidence is passed to an entailment-based classifier to determine the relationship between the claim and the evidence.

Classification: The model outputs a label from the set: Supported, Refuted, or Not Enough Information (NEI).
Confidence Recalibration: To mitigate overconfidence, a thresholding mechanism is applied. If the model's confidence score ( $s$ ) falls below a predefined threshold ( $\tau$ , default 0.6), the prediction is overridden to NEI. This reflects epistemic uncertainty and aligns with conservative human fact-checking behavior.
Aggregation: For multiple evidence passages, final decisions are made via majority voting or weighted averaging.

3. Key Contributions

Dual Framework Integration: DAVinCI is the first framework to integrate internal/external attribution with entailment-based verification in a modular, scalable way, allowing LLMs to justify and validate claims transparently.
Comprehensive Evaluation: The authors evaluate the framework across multiple entailment models (DeBERTa, RoBERTa, BART) and datasets (FEVER, CLIMATE-FEVER).
Ablation Insights: A detailed study isolates the impact of evidence span selection (full vs. span), recalibration thresholds, and retrieval quality.
Open Source: A reproducible, modular implementation is released to support future research in trustworthy NLP.

4. Experimental Results

The framework was evaluated on the FEVER and CLIMATE-FEVER datasets against standard verification-only baselines.

Performance Gains: DAVinCI significantly outperforms baselines across all metrics.
- Accuracy: Improved by 5–20% across various models.
- F1-Score: Macro F1-scores increased from ~0.36 to ~0.41 (DeBERTa on FEVER) and showed similar gains on CLIMATE-FEVER.
- Precision & Recall: Consistent improvements in both macro and weighted averages.
Evidence Quality (Ablation):
- Full Evidence vs. Span: Full-evidence attribution consistently outperformed span-based attribution. For example, on FEVER, switching from span to full evidence improved the RoBERTa-large-snli F1-score from 0.19 to 0.48. This highlights that losing context via span extraction severely degrades verification quality.
Threshold Tuning:
- A threshold of 0.7 offered the optimal trade-off between precision and recall.
- Higher thresholds (0.8, 0.9) increased precision (reducing false positives) but significantly lowered recall (filtering out too many valid claims as "Not Enough Info").

5. Significance and Future Work

Trustworthiness: DAVinCI provides a scalable path to auditable AI systems by ensuring that claims are not just generated but are also backed by evidence and validated with calibrated confidence.
Domain Applicability: The framework is particularly valuable for domains where factual accuracy is non-negotiable.
Limitations: Current limitations include reliance on high-quality evidence (which may be scarce in open domains), static entailment models that struggle with multi-hop reasoning, and evaluation restricted to English datasets.
Future Directions: The authors plan to extend DAVinCI to open-domain retrieval, multi-hop reasoning, internal attribution tracing (e.g., prompt tracing), and multilingual/low-resource environments.

In conclusion, DAVinCI demonstrates that bridging the gap between attribution and verification is essential for moving LLMs from being merely "powerful" to being accountable and trustworthy.