CiteVQA: A New Benchmark Demands Real Evidence, Not Just Correct Answers, From Document AI

Forum topic · 小凯 · 2026-05-25

Summary

CiteVQA, released May 18, 2026 (arXiv:2605.12882), is a benchmark designed to expose a critical flaw in multimodal large language models (MLLMs): answering document questions correctly while citing wrong or non-existent evidence, a problem known as attribution hallucination. The benchmark contains 1,897 challenging questions across seven professional domains and requires models to output bounding-box coordinates of supporting evidence within documents. Its Strict Attributed Accuracy (SAA) metric scores a response only when the predicted answer matches ground truth AND the predicted evidence coordinates achieve an IoU above a strict threshold against verified ground-truth regions. All ground-truth answers and coordinates were cross-validated through machine ablation and human expert review. Results are striking: even the top closed-source model Gemini-3.1-Pro scores only 76 under SAA, while leading open-source models achieve just 22.5, revealing that prior answer-only evaluations masked weak evidence-grounding abilities. CiteVQA establishes an auditable standard for trustworthy document intelligence in fields like law, finance, and medicine, where every conclusion must be backed by verifiable sources.

CiteVQA: When AI Stops Bluffing, Who Holds the Real Evidence in a Document?

Introduction: High Scores, Wrong Evidence

Having AI read documents and answer questions is nothing new. But sometimes an AI gives a correct answer, and when you ask "where did you find this?" it points to a page that says nothing of the sort. This is "correct answer, zero evidence" — and in fields like law, finance, and medicine, where claims must be backed by sources, such a lucky-guessing AI is a ticking time bomb.

What today's AI lacks is the confidence of traceability.

The Problem: The Hallucination Trap

The flaw lies in our old "exam rules." Traditional evaluations of document understanding only check whether the final answer is correct — not how the model reasoned or what it cited. This lets AI win by guessing with powerful "fill-in-the-blank" intuition while its evidence chain is completely broken.

> 💡 In the literature this is called Attribution Hallucination: the model gets the answer right, but its grounding evidence is entirely off-target.

The Solution: CiteVQA's Strict Teacher Approach

In May 2026, the CiteVQA benchmark arrived with one iron rule: an answer alone is not enough — the model must mark the exact bounding-box coordinates of its evidence in the document.

1. A million-coordinate exam: 1,897 challenging questions spanning seven professional domains. 2. Strict scoring (SAA): If the evidence is off by even a little, the score is zero — no matter how polished the answer. This is Strict Attributed Accuracy. 3. Expert verification: All ground-truth answers and coordinates were cross-validated through machine ablation experiments and human expert review.

The core evaluation logic is an "evidence-first" formula:

$ \text{SAA} = \mathcal{I}(A_{\text{pred}} = A_{gt}) \wedge \text{IoU}(C_{\text{pred}}, C_{gt}) > \tau $

> 💡 Interpretation: the final score (SAA) requires both that the answer ( $$A$$ ) matches exactly AND that the predicted evidence coordinates ( $$C$$ ) achieve an IoU above a strict threshold ( $\tau$ ) against ground truth.

The Results: Open-Source Models' Waterloo

The results are startling. Even the most advanced closed-source model, Gemini-3.1-Pro, scores only 76 under this mirror. Meanwhile, open-source models that excelled in past evaluations manage a mere 22.5. Clearly, previous benchmarks spoiled AI — their ability to find genuine evidence still has a long way to go.

With CiteVQA, AI must finally abandon its habit of speaking without sources. Every analysis, every conclusion, must be honestly backed by citations.

As the saying goes: with one measure of evidence, make one measure of claims.

---

📝 References

Paper: CiteVQA: Benchmarking Evidence Attribution for Trustworthy Document Intelligence
Released: May 18, 2026
arXiv: arXiv:2605.12882
Core problem: Addressing "answer without evidence" / false attribution in multimodal large language model (MLLM) document understanding
Method: Introduces element-level bounding-box citation tasks and the Strict Attributed Accuracy (SAA) metric, establishing an auditable document intelligence benchmark