CRAFT pipeline boosts multi-video QA with critic loop and 0.810 recall

A team led by Mahesh Bhosale at the University at Buffalo has released CRAFT (Critic-Refined Adaptive Key-Frame Targeting), a new pipeline designed to answer factual questions across multiple news videos while automatically attributing every claim to its source video. Unlike typical video QA systems that treat clips in isolation, CRAFT first selects query-relevant keyframes dynamically, then extracts transcripts using per-video ASR with a multilingual fallback to handle code-switched or non-English audio. The core innovation is a hybrid critic loop that runs three checks on each generated claim: UNLI for temporal entailment, DeBERTa-v3 for cross-claim consistency, and a Llama-3.2-3B adjudicator that flags and repairs errors before final consolidation. A citation-merging step emits each fact only once but with all supporting source identifiers.

On the MAGMaR 2026 benchmark, CRAFT set new state-of-the-art scores with an overall average of 0.739, a reference recall of 0.810, and a citation F1 of 0.635—meaning it reliably finds the right evidence and credits it correctly. The system also generalized well to a MAGMaR-style conversion of WikiVideo (52 queries), achieving an average of 0.823. Ablation studies confirmed that atomic claim generation, ASR, and the critic loop are the primary drivers of improvement over a vanilla baseline. The code is publicly available, making CRAFT a strong foundation for building trustworthy, cite-anything video question-answering systems.