Did the AI Actually Use the Evidence?

Three research checks in about 100 seconds: structure, citations, and visual evidence.

Download video · WebVTT captions · SRT subtitles

Disclosure: Yuxuan Zhang coauthored the research discussed. This guide uses an AI-assisted script, synthetic narration and conceptual diagrams. It reports no new experimental results. Contributions belong to all coauthors listed in the original papers. The papers study different tasks; their benchmark scores should not be compared directly.

Explore and try the research

Read the companion article · All papers, code and author lists

Full transcript

00:00 — Did the AI actually use the evidence?

An AI answer can look right and still fail an essential check. Did it produce the structure we asked for? Retrieve a relevant reference? Or actually use the video? Here are three different tests, from research I coauthored.

00:13 — Valid syntax is only the first check

StructEval studies structured generation and conversion, including JSON, HTML, and SVG. A page can contain valid HTML yet show the wrong layout. A format conversion can parse correctly yet lose content. Separate syntax, content preservation, and visual fidelity.

00:32 — Writing and retrieval belong in the same loop

ScholarCopilot jointly trains academic writing and citation retrieval. A retrieval token decides when to request references, and retrieved material informs the next text. But a real reference is not automatically support for a claim. Check existence, relevance, and evidence separately.

00:50 — What happens when the video is missing?

Video questions can reveal clues in their wording or answer choices. Watch Before You Answer studies these shortcuts. VidGround selects questions intended to require visual evidence for post-training. Comparing video access with a text-only condition helps diagnose what a task measures.

01:07 — Define the evidence before the score

These are different tasks, so their benchmark scores should not be compared directly. Define the evidence your application needs, then choose an evaluation that can detect its absence. Use the original papers to understand the models, datasets, and limitations.

01:23 — Papers, code, and complete author lists

Read the papers, code, and full author lists at yuxuan dot world slash publications. This is an author guide prepared with AI assistance and synthetic narration. It reports no new experiments, and the research contributions belong to all coauthors.