ReVEAL: Eliciting Role-Aware Latent Visual Evidence from MLLMs for Reasoning Video Object Segmentation
Organizations: Fudan University · New York University
Abstract
Reasoning video object segmentation (ReVOS) requires segmenting the target specified by an implicit query throughout a video. Existing MLLM-based methods typically require costly cross-model alignment, tying MLLM training to a downstream segmentation model. We present ReVEAL, a decoupled thinking-with-video framework for ReVOS that elicits role-aware latent visual evidence from MLLM reasoning to directly guide segmentation. ReVEAL introduces Visual Evidence Tokens (VETs) into video reasoning to associate necessary visual regions or entities with target, background, or distractor roles, while a Dynamic Attention Layer Router adaptively aggregates VET-to-video attention across layers into role-specific grounding evidence. We train the MLLM to perform VET-augmented reasoning on our constructed ReVEAL-47K dataset, with a Contrastive Attention Grounding Loss as auxiliary supervision to apply role-specific constraints to VET attention and optimize layer routing for reliable latent grounding. At inference, ReVEAL identifies the target through this thinking-with-video process and fuses the resulting role-specific evidence to guide a frozen SAM2 in predicting target masks. Experiments on five benchmarks show that ReVEAL outperforms alignment-free baselines and remains competitive with alignment-based methods.
Figures & tables
| Method | MLLM | Ref-DAVIS | ReasonVOS | ReVOS (Overall) | ReVOS (Referring) | ReVOS (Reasoning) | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| & | & | & | & | & | ||||||||||||
| TAM ( Li et al., 2025 ) | Qwen2.5VL-7B | 2.7 | 1.8 | 3.7 | 3.3 | 2.7 | 3.8 | 3.0 | 2.8 | 3.1 | 3.0 | 2.9 | 3.2 | 2.9 | 2.7 | 3.0 |
| Loc-Head ( Kang et al., 2025 ) | Qwen2.5VL-7B | 13.0 | 16.8 | 9.2 | 12.2 | 17.1 | 7.3 | 13.3 | 18.9 | 7.7 | 14.1 | 20.3 | 7.9 | 12.5 | 17.4 | 7.5 |
| DecAF ( Han et al., 2025 ) | Qwen2.5VL-7B | 23.7 | 30.4 | 17.1 | 19.1 | 24.1 | 14.1 | 19.4 | 25.2 | 13.5 | 21.1 | 27.9 | 14.3 | 17.7 | 22.6 | 12.7 |
| ReVEAL (Ours) | Qwen2.5VL-7B | 29.4 | 39.5 | 19.2 | 22.1 | 29.5 | 14.8 | 24.0 | 33.6 | 14.4 | 25.4 | 35.6 | 15.3 | 22.5 | 31.6 | 13.5 |
Appendix figures & tables15 assets
Supplementary material from the paper’s appendix.
Appendix
| Method | Overall | Causal | Sequential | Counterfactual | Descriptive | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| & | & | & | & | & | |||||||||||
| None-MLLM RVOS Methods | |||||||||||||||
| ReferFormer ( Wu et al., 2022 ) | 12.7 | 11.2 | 14.3 | 12.7 | 11.0 | 14.4 | 10.7 | 9.4 | 12.1 | 12.4 | 11.1 | 13.6 | 15.2 | 13.3 | 17.0 |
| SgMg ( Miao et al., 2023 ) | 17.5 | 15.8 | 19.2 | 18.9 | 17.0 | 20.9 | 15.5 | 14.0 | 17.0 | 16.4 | 15.2 | 17.5 | 18.8 | 16.7 | 20.9 |
| HTR ( Miao et al., 2024 ) | 14.5 | 12.7 | 16.2 | 15.7 | 13.8 | 17.5 | 12.5 | 10.8 | 14.1 | 13.5 | 12.2 | 14.9 | 15.9 | 13.8 | 18.1 |
| LMPM ( Ding et al., 2023 ) | 13.3 | 12.7 | 14.0 | 13.7 | 13.1 | 14.3 | 11.3 | 10.6 | 12.0 | 13.2 | 12.6 | 13.7 | 14.3 | 13.6 | 15.1 |
Video Object Segmentation
Video object segmentation assigns pixel-level masks to objects across frames while preserving their identities through motion, occlusion, and change.
See all 120 papersRCoT-Seg: Reinforced Chain-of-Thought for Video Reasoning and Segmentation
VideoSEG-O3: A Multi-turn Reinforcement Learning Framework for Reasoning Video Object Segmentation
[SEG] during the RL stage, we introduce SEG-aware logit calibration, which integrates pixel-wise segmentation feedback directly into the token-level logits. Furthermore, we design a decoupled thinking trace to hierarchically decompose the reasoning process into temporal, spatial, and linguistic dimensions, and construct VTS-CoT, a specialized cold-start dataset featuring comprehensive reasoning trajectories. The code and models will be released at https://github.com/Dmmm1997/VideoSEG-O3.