Context
MC-Search introduces 3,333 annotated multimodal multi-hop search tasks spanning five reasoning structures, enabling evaluation of intermediate text/image retrieval, evidence dependencies and agent planning separately from final answer accuracy.
What changed
The authors provide verified stepwise search trajectories with supporting evidence and process metrics for retrieval fidelity and rollout deviation, evaluating six multimodal models under controlled conditions.
Why it matters
Visual and agentic AI discovery can require multiple dependent searches across text and images. Measurement of evidence discovery needs hop-level and modality-aware diagnostics, not just final-answer accuracy.
Visibility OS interpretation
Visibility OS interpretation: the benchmark offers an evaluation pattern for multimodal query fan-out and evidence selection. It does not establish production behavior of Google, ChatGPT, Claude or Perplexity search products.
What remains uncertain
How well do model-generated, model-verified trajectories generalize to the open web? Would independent human review reproduce evidence-step quality scores? Can process metrics predict real citation exposure?