← News

Agents & MCPSources: arXiv · 18 Nov 2026

NexSight introduces verifiable tasks for agentic long-video understanding

The arXiv paper pairs a 310-task benchmark with a training pipeline that turns verified solver trajectories into step-level supervision for evidence search.

In plain words

Some AI models can search through hours of video to answer detailed questions, but checking whether their search actually works is difficult. The NexSight team built tasks where the evidence behind each answer can be independently verified, then reused that verification process to train a model to search more carefully.

What was announced

The paper, posted to arXiv by the NexSight authors, targets fine-grained questions about long videos, such as how often an action occurs or when a state changes, which may require finding brief moments across hours of footage. The authors argue that sparse evidence can confirm an action happened without proving its total count, and that answer-only supervision does not explicitly train the evidence search itself.

Their pipeline, SightForge, accepts a task only if a verifier blind to the draft answer can reconstruct it from evidence re-collected from the raw video and counterexample checks fail. This produces SightLine, 310 evidence-traceable tasks from 108 long videos, run through the SightLoop harness that records observations during temporal exploration and local inspection. Solver trajectories on disjoint videos then serve as step-level supervision. The authors report the strongest model reaches 54.52% on SightLine, rising to 79.35% with annotated evidence, and that Sight-35B, trained on under 6K examples, nearly doubles its base model's interactive accuracy from 21.29% to 42.26% and gains over 8 points on two external benchmarks.

What it does not show

  • The source is the paper's abstract only; full methodology, evaluation details and external benchmark names are not available in the supplied passages.
  • All reported scores and training gains are the authors' own claims and have not been independently confirmed.

For security and IT teams

Scope
The SightLine benchmark contains 310 evidence-traceable tasks drawn from 108 long videos, per the paper's abstract.

Sources

Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 19 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.