Magnet : A M ulti- ag ent Framework for Finding Audio-Visual Needles by Reaso n ing over Multi-Vid e o Hays t acks

Sanjoy Chowdhury, Mohamed Elmoghany, Yohan Abeysinghe, Junjie Fei, Sayan Nag, Salman Khan, Mohamed Elhoseiny, Dinesh Manocha · 2025

Read Magnet : A M ulti- ag ent Framework for Finding Audio-Visual Needles by Reaso n ing over Multi-Vid e o Hays t acks by Sanjoy Chowdhury, Mohamed Elmoghany, Yohan Abeysinghe, Junjie Fei, Sayan Nag, Salman Khan, Mohamed Elhoseiny, Dinesh Manocha (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2506-07016.txt.

Description

Sanjoy Chowdhury and co-authors set a new test for video question answering in a 2025 paper. Most tests use one clip per question. Real work often means hunting through many clips for the few bits that answer a query. The team call that task AV-HaystacksQA. They built a test set of 3,100 question and answer pairs that span audio and video across clips.

Alongside the test set, the team built MAGNET, a group of cooperating agents that search, ground, and link clips. It beat base methods by large margins on two answer scores. Two new scores check the search path itself: one grades step order, the other grades how well found clips match the truth. The work gives a shared yardstick for tools that must reason across whole video sets, not single clips.

Catalog record

CREATORSanjoy Chowdhury, Mohamed Elmoghany, Yohan Abeysinghe, Junjie Fei, Sayan Nag, Salman Khan, Mohamed Elhoseiny, Dinesh Manocha
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE53 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2506-07016

Browse the full library · About the Tower of Babel