TF-PRVR | NeurIPS 2026

Figure 1: existing dual-branch PRVR framework, semantic dilution from fixed clips, and cross-domain R@1 comparisons on ActivityNet Captions, Charades-STA, and TVR — **TF-PRVR retrieves partially relevant videos without task-specific training.**
Adaptive temporal structure and multi-scale consensus preserve the generalization of frozen vision-language models.

Cross-Domain Generalization

Strong in-domain performance does not always transfer to a new dataset. As shown above, existing training-based PRVR methods can lose substantial retrieval accuracy when evaluated across domains. TF-PRVR applies the same training-free framework across datasets, preserving frozen vision-language alignment without PRVR-specific optimization.

Abstract

Partially Relevant Video Retrieval (PRVR) aims to retrieve untrimmed videos containing moments relevant to a given text query. Despite recent progress, existing PRVR methods suffer from two key limitations: a fixed video decomposition scheme that causes semantic dilution, and source-domain overfitting induced by task-specific training. In this paper, we propose TF-PRVR, the first training-free framework for PRVR. TF-PRVR leverages frozen vision-language features to construct video-specific hierarchical representations. It derives temporal semantic signals from frame-level features and applies frequency-based multi-scale analysis to identify adaptive temporal boundaries, producing hierarchical segments with coherent event-level semantics. Built on these segments, TF-PRVR constructs a unified multi-scale graph and propagates query relevance across temporally and semantically related nodes. A moment-aware scoring strategy then aggregates temporally aligned relevance across scales, emphasizing consistently supported moments while suppressing isolated false responses. Without task-specific training, TF-PRVR preserves the general-purpose alignment capability of pre-trained vision-language models and avoids dataset-specific overfitting. Extensive experiments demonstrate consistent performance across datasets with diverse visual and temporal characteristics, suggesting a practical direction for training-free PRVR.

Method Overview

From frozen frame features to video-specific temporal structure.
TF-PRVR combines adaptive segmentation, graph propagation, and moment-aware scoring.

TF-PRVR pipeline: semantic change signals, hierarchical temporal segmentation, multi-scale graph construction, relevance propagation, and moment-aware scoring — **Framework overview.** A unified graph connects hierarchical video segments. Query relevance is propagated across related segments, then aggregated over temporally aligned scales to compute a video-level retrieval score.

Adaptive segmentation

Frequency-based analysis of semantic change identifies video-specific boundaries at multiple temporal scales.

Multi-scale propagation

Temporal and semantic relationships connect segments within and across scales, allowing relevant evidence to reinforce one another.

Moment-aware scoring

Temporally aligned relevance is averaged across scales. The strongest consensus determines the video score.

TF-PRVR Demo

A query describes a moment. Retrieval returns the video.
Follow how evidence comes together across temporal scales.

TEXT QUERYWater is being poured into a glass.

Video 01

Illustrative match

Temporal consensus

Video 02

Illustrative match

Temporal consensus

Video 03

Illustrative match

Temporal consensus

Split each video into segments at multiple temporal scales.

Edited public footage. Segments, connections, curves, and selections are scripted illustrations, not model predictions.

Video sources

Edited from TP Motion, Pressmaster, and MART PRODUCTION on Pexels. Used under the Pexels License. Each example combines different clips into an 18-second montage.

Experimental Results

TF-PRVR is evaluated on ActivityNet Captions, Charades-STA, and TVR.
It improves over zero-shot retrieval without PRVR-specific training.

Table 6: recall and SumR comparison of TF-PRVR with training-based and training-free approaches across three PRVR benchmarks — **Comparison with existing PRVR methods.** Training-based approaches are evaluated in-domain. TF-PRVR uses frozen features and requires no PRVR-specific training.

Generalization across vision-language backbones

TF-PRVR consistently improves retrieval with CLIP, SigLIP, and EVA-CLIP features.

ActivityNet Captions · SumR ↑
Backbone	Zero-shot	TF-PRVR	Improvement
CLIP-L/14	172.9	193.9	+21.0
SigLIP-L/14	178.7	198.5	+19.8
EVA-CLIP-L/14	180.4	200.8	+20.4

Qualitative Results

A video contains events of different durations.
Adaptive hierarchical segments provide the appropriate context for different queries.

Figure 5: video-specific hierarchical segments on Charades-STA for queries about eating a sandwich, holding a cup, and putting on shoes — **Hierarchical segmentation on Charades-STA.** Fine, intermediate, and coarse segments capture complementary temporal context for query-relevant moments within the same untrimmed video.

Citation

@misc{kim2026tfprvr,
  title={TF-PRVR: Training-Free Partially Relevant Video Retrieval},
  author={Giyeol Kim and Chanho Eom},
  year={2026},
  eprint={2610.07925},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2610.07925}
}