Research map / Video / Video-language understanding

ReVisionLLM: Recursive Vision-Language Model for Temporal Grounding in Hour-Long Videos

CVPR 2025 · Video › Video-language understanding · approach: Temporal video grounding · 5 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes ReVisionLLM, a recursive vision-language model that grounds text queries in hour-long videos by coarse-to-fine hierarchical search, trained progressively from short clips to long videos with contrastive "Not Present" segments.

Problem

Temporal grounding of text-queried events in hour-long videos is out of reach for VLMs, which are limited by frame budgets, lose temporal detail, and are poorly calibrated, producing false positives.

Approach

A hierarchical adapter turns video into dense features at the bottom level and text-aligned sparse segment tokens at higher levels, and an LLM recursively narrows from whole-video segment proposals down to exact boundaries; training starts on short clips with contrastive absent-event segments, ranks predictions by LLM token entropy, and then scales to hour-long videos.

Key results

• On MAD, ReVisionLLM reaches R1@0.1 of 15.0 (17.3 for the ReVisionLLM-I variant), beating the prior best RGNet at 12.4 by 2.6 points.
• On VidChapters-7M, ReVisionLLM gets R1@0.7 of 15.2 and R1@0.9 of 24.6 versus M-DETR, which trails by 4.2 and 8.8 points on those thresholds.
• ReVisionLLM processes on average about 43% fewer frames than VTimeLLM, using 57% of input frames in the default setting.

Measured effects reported in the paper

Surprising findings

• The VTimeLLM baseline scores zero recall on MAD when given uniformly sampled frames from hour-long videos, showing non-recursive VLMs fail entirely at this scale.
• Replacing CONE's CLIP similarity ranking with the LLM's own entropy-based confidence roughly doubled R1@0.1 (4.8 to 8.4 after calibration) in the ablation.

Limitations

• Performance declines gradually as videos grow longer, reflecting the difficulty of extended content.
• Relies on frozen CLIP features and a CLS-token-only video representation, which discards fine spatial detail.
• Does not integrate audio, as noted by the authors' future-work direction.

Experimental setup

Datasets: MAD VidChapters-7M MSRVTT

Benchmarks: MAD VidChapters-7M MSRVTT

Metrics: Recall@k at IoU threshold (R1@0.1, R5@0.1, R1@0.3, R5@0.3, R R@1 R@5 R@10

Baselines: VTimeLLM VTimeLLM + CONE RGNet SnAG SOONet CONE M-DETR M-Guidance

Builds on: VTimeLLM CONE Vicuna CLIP

Closest papers by meaning

• Time-R1: Post-Training Large Vision Language Model for Temporal Video Grounding (NeurIPS 2025 · similarity 0.90)
• Universal Video Temporal Grounding with Generative Multi-modal Large Language Models (NeurIPS 2025 · similarity 0.89)
• VTG-LLM: Integrating Timestamp Knowledge into Video LLMs for Enhanced Video Temporal Grounding (AAAI 2025 · similarity 0.89)
• Empower Words: DualGround for Structured Phrase and Sentence-Level Temporal Grounding (NeurIPS 2025 · similarity 0.88)
• VideoGEM: Training-free Action Grounding in Videos (CVPR 2025 · similarity 0.88)
• Masked Diffusion Vision-Language Models for Temporal Action Localization (NeurIPS 2026 · similarity 0.88)