Reviso, a video-captioning system described in an arXiv preprint dated November 2, 2026, learns to decide which parts of a video deserve closer inspection. Its authors report improved factual coverage and more useful captions across benchmark tests. Peer review has not been verified, and the results come from the research team rather than an independent evaluator.
The approach addresses a problem with detailed AI descriptions: a caption can include many minor details while missing the main event. Reviso starts with a low-frame-rate preview and builds a Context–Focus–Points plan. This identifies the overall setting, prioritizes important aspects and defines specific details to inspect. The system then revisits selected segments, crops images or uses optical character recognition (OCR), updating its plan before combining the evidence into a caption.
Built on Qwen3.5-9B, Reviso was trained with 10,000 verified observation trajectories for supervised fine-tuning and 20,000 reinforcement-learning samples. Training rewards factual coverage and relevant observation targets, while conditionally penalizing costly tool use that produces weaker captions. Reference descriptions involved human editing of existing and model-generated captions; model-based verification and teacher refinement determined which trajectories were retained.
The main comparison allowed each model at most 128 frames per video. Baselines sampled frames uniformly, while Reviso divided its allowance between the preview and targeted revisits. On DREAM-1K, Reviso achieved 39.5% F1, compared with 35.1% for its backbone and 39.2% for the strongest other evaluated open-source model. F1 balances factual precision and coverage. On MCTS-VCB, its F1 rose from the backbone’s 45.5% to 60.9%, a 15.4-percentage-point gain.
On CaReBench, action F1 increased from 25.2% to 34.3%, and object F1 from 27.2% to 33.9%. On VDC, TP score increased from 2.33 to 2.95 and TP accuracy from 0.437 to 0.568, in the benchmark’s original units. Reviso led the evaluated open-source models on the principal caption-quality metrics, though not every precision or recall component. Proprietary models surpassed it on some measures; a separate workflow comparison also placed GPT-6 Astra paired with Codex above Reviso on DREAM-1K.
To test whether captions preserved useful information, Gemini 3.1 Pro answered questions using captions alone, without video access. Captions were generated without seeing the questions. LVBench accuracy reached 52.0%, versus 47.1% with backbone captions; MMVU overall accuracy reached 72.7%, versus 70.6%. Ablations also favored the full system: supervised fine-tuning alone produced 37.6% DREAM-1K F1, compared with 39.5% for the complete method.
These findings support targeted observation within the tested setup, but their robustness is uncertain. The supplied text reports no evaluation sample counts, confidence intervals, statistical tests or variation across training runs. The 0.3-point DREAM-1K lead over the strongest open-source comparator therefore warrants caution. Training-to-benchmark overlap checks and dependence among source clips are not described, although the supervised and reinforcement-learning videos were reported to be disjoint.
Model-generated references and verification may introduce errors or preferences. An equal frame ceiling also does not establish equal computation, latency or cost; actual resource use was not reported. The backbone comparison changes training and observation together, and ablations do not isolate every contribution. Downstream testing uses one answer-extraction model, with no independent replication or systematic human evaluation reported. The results indicate a promising observation strategy, without establishing general reliability or universal superiority.
Source: https://arxiv.org/abs/2611.03602
Cover illustration generated with AI; conceptual, not research evidence.