Multimodal large language models (MLLMs) represent images and video frames as visual tokens. Scaling from single images to hour-long videos, however, inflates the token budget far beyond practical limits. Popular pipelines therefore either uniformly subsample or apply keyframe selection with retrieval-style scoring using smaller vision-language models. However, these keyframe selection methods still rely on pre-filtering before selection to reduce the inference cost and can miss the most informative moments.

We propose FOCUS, Frame-Optimistic Confidence Upper-bound Selection, a training-free, model-agnostic keyframe selection module that selects query-relevant frames under a strict token budget. FOCUS formulates keyframe selection as a combinatorial pure-exploration (CPE) problem in multi-armed bandits: it treats short temporal clips as arms, and uses empirical means and Bernstein confidence radius to identify informative regions while preserving exploration of uncertain areas. The resulting two-stage exploration-exploitation procedure reduces from a sequential policy with theoretical guarantees, first identifying high-value temporal regions, then selecting top-scoring frames within each region.

On two long-video question-answering benchmarks, FOCUS delivers substantial accuracy improvements while processing less than 2% of video frames. For videos longer than 20 minutes, it achieves an 11.9% gain in accuracy on LongVideoBench, demonstrating its effectiveness as a keyframe selection method and providing a simple and general solution for scalable long-video understanding with MLLMs.

1. First, follow the installation instructions from the AKS repository to set up the environment and dependencies.

2. Then install the additional requirements:

pip install -r requirements.txt

Run FOCUS keyframe extraction on LongVideoBench:

python select_keyframe.py \
    --dataset_name longvideobench \
    --dataset_path ./datasets/longvideobench \
    --output_dir focus_blip \
    --num_keyframes 64 \
    --batch_size 32 \
    --blip_model large

For evaluation, please follow the evaluation setup from the lmms-eval repository and use the evaluation scripts provided in the AKS repository.

FOCUS generates the following outputs:

• selected_frames.json: Selected keyframe indices for each video
• sampling_details.json: Detailed sampling information including:
  • Coarse and fine sampling results
  • Arm information and FOCUS scores
  • Arm selection probabilities
  • Video metadata
• extraction_stats.json: Statistics about the extraction process

This work builds upon the excellent research from:

• AKS: Adaptive Keyframe Sampling for the evaluation framework
• lmms-eval for multimodal evaluation