SlotVTG: Object-Centric Adapter for Generalizable Video Temporal Grounding

Forum topic · 小凯 · 2026-03-28

Summary

SlotVTG is a new framework for Video Temporal Grounding (VTG) that improves the out-of-domain (OOD) generalization of Multimodal Large Language Models (MLLMs). While MLLMs perform well on VTG, task-specific fine-tuning tends to make them memorize dataset-specific shortcuts instead of grounding in actual visual content, hurting generalization. SlotVTG addresses this with a lightweight slot adapter that uses slot attention to decompose visual tokens into abstract slots and then reconstructs the original sequence. Object priors from a self-supervised vision model encourage semantically coherent slot formation, steering the MLLM toward object-centric, input-grounded visual reasoning without retraining from scratch. Cross-domain evaluation on standard VTG benchmarks shows that SlotVTG significantly improves OOD robustness while maintaining competitive in-domain performance and minimal overhead. The paper (arXiv:2603.25733) is authored by Jiwook Han, Geo Ahn, Youngrae Kim, and Jinwoo Choi.

Overview

Field: Computer Vision Authors: Jiwook Han, Geo Ahn, Youngrae Kim, Jinwoo Choi Published: 2026-03-26 arXiv: 2603.25733v1

Key Points

Multimodal Large Language Models (MLLMs) show strong performance on Video Temporal Grounding (VTG), but their coarse recognition capabilities are insufficient for fine-grained temporal understanding, making task-specific fine-tuning indispensable.
Such fine-tuning causes models to memorize dataset-specific shortcuts rather than faithfully grounding in actual visual content, leading to poor Out-of-Domain (OOD) generalization.
Object-centric learning offers a promising remedy by decomposing scenes into entity-level representations, but existing approaches require re-running the entire multi-stage training pipeline from scratch.
SlotVTG steers MLLMs toward object-centric, input-grounded visual reasoning at minimal cost via a lightweight slot adapter.
The adapter uses slot attention to decompose visual tokens into abstract slots and reconstruct the original sequence; object priors from a self-supervised visual model encourage semantically coherent slot formation.
Cross-domain evaluation on standard VTG benchmarks shows significantly improved OOD robustness while maintaining competitive in-domain (ID) performance, with minimal overhead.

Abstract

Multimodal Large Language Models (MLLMs) have shown strong performance on Video Temporal Grounding (VTG). However, their coarse recognition capabilities are insufficient for fine-grained temporal understanding, making task-specific fine-tuning indispensable. This fine-tuning causes models to memorize dataset-specific shortcuts rather than faithfully grounding in the actual visual content, leading to poor Out-of-Domain (OOD) generalization. Object-centric learning offers a promising remedy by decomposing scenes into entity-level representations, but existing approaches require re-running the entire multi-stage training pipeline from scratch. We propose SlotVTG, a framework that steers MLLMs toward object-centric, input-grounded visual reasoning at minimal cost.

---

*Auto-collected on 2026-03-28*