arxiv: 2605.07334 · v1 · submitted 2026-05-08 · 💻 cs.CV

Recognition: 2 theorem links

· Lean Theorem

RCoT-Seg: Reinforced Chain-of-Thought for Video Reasoning and Segmentation

Deshui Miao, Guangming Lu, Junwei Wen, Wenjie Pei, Xin Li

Authors on Pith no claims yet

Pith reviewed 2026-05-11 01:12 UTC · model grok-4.3

classification 💻 cs.CV

keywords video reasoning segmentationchain-of-thoughtkeyframe selectionGRPOSAM2temporal reasoningmulti-object videoMLLM

0 comments

The pith

RCoT-Seg factorizes video reasoning segmentation into reinforced chain-of-thought temporal reasoning and SAM2-based spatial perception to improve keyframe selection and mask consistency.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper establishes that existing MLLM methods for video reasoning segmentation produce weak temporal understanding because they select keyframes through simple sampling or auxiliary models that follow narrow similarity rules. It proposes splitting the task into a temporal video reasoning stage that uses an agentic module, started from a curated chain-of-thought corpus and refined by GRPO with task-aligned rewards, to generate and self-evaluate keyframes, followed by a keyframe target perception stage that runs high-resolution segmentation and propagates masks with SAM2. A sympathetic reader would care because this factorization promises more reliable object localization across entire video sequences when instructions involve implicit intent and temporal logic rather than explicit frame-by-frame cues.

Core claim

RCoT-Seg factorizes VRS into TVR and KTP. In the TVR stage an agentic keyframe selection module initialized with a curated CoT-start corpus and refined by GRPO under task-aligned rewards generates and reselects keyframes through self-evaluation. In the KTP stage it performs high-resolution segmentation on the chosen frame and propagates masks with SAM2-based methods across the sequence, replacing heuristic sampling and external selectors.

What carries the argument

The agentic keyframe selection module refined by GRPO under task-aligned rewards, which performs self-evaluation to generate and reselect keyframes in the temporal video reasoning stage.

If this is right

Strengthens moment localization and temporal reasoning for implicit instructions in multi-object scenes.
Improves spatial precision and inter-frame mask consistency by using high-resolution segmentation followed by SAM2 propagation.
Replaces heuristic sampling and external frame selectors with an integrated self-evaluative process.
Delivers favorable performance against state-of-the-art methods on video reasoning segmentation benchmarks.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The separation of temporal reasoning from spatial perception could transfer to other video tasks that require inferring intent across time, such as action anticipation or event localization.
Self-evaluation inside the keyframe module may lessen dependence on separate large models for frame picking in longer videos.
Expanding the initial CoT-start corpus with more varied temporal logic examples might improve robustness to unseen instruction styles.

Load-bearing premise

That GRPO refinement on a curated CoT-start corpus under task-aligned rewards will produce keyframe selections that strengthen holistic temporal understanding instead of overfitting to the specific reward signals or corpus biases.

What would settle it

A test set of complex multi-object videos where the reinforced keyframe selections yield segmentation accuracy no higher than, or lower than, simple uniform sampling or auxiliary-MLLM selection when measured by standard mask metrics on the full sequence.

Figures

Figures reproduced from arXiv: 2605.07334 by Deshui Miao, Guangming Lu, Junwei Wen, Wenjie Pei, Xin Li.

**Figure 1.** Figure 1: Comparison between RCoT-Seg and previous methods. RCoT-Seg segments the correct target through explicit CoT reasoning and concise keyframe selection. view and commits to keyframe localization and spatial grounding within a single reasoning trajectory. As a result, keyframe reliability is not explicitly modeled as a controllable intermediate state, making it difficult to diagnose an unreliable keyframe or r… view at source ↗

**Figure 2.** Figure 2: Overview of RCoT-Seg unified multi-task pipeline for VRS task. In contrast, RCoT-Seg unifies temporal reasoning and segmentation in a single multi-task model, achieving stronger results in complex video scenarios with only modest training data. 3 Method Given a video sequence and the query text of the target object(s), the goal of the proposed RCoT-Seg is to precisely localize the target object(s) with mas… view at source ↗

**Figure 3.** Figure 3: Overall training strategy of RCoT-Seg. Following a CoT-SFT cold-start, we employ GRPO reinforcement learning on the model using reward functions tailored to different tasks. We obtain a one-to-one assignment by maximizing S with the Hungarian algorithm, yielding matched pairs C ′ . The answer accuracy reward averages the matched scores: \mathcal {R}_a=\frac {1}{\max \{N_{\mathrm {obj}},\,N_{\mathrm {obj}}^… view at source ↗

**Figure 4.** Figure 4: Qualitative comparison between RCoT-Seg and current SOTA methods. Our method successfully identify and accurately segment targets in complex scenarios involving small-scale objects and multi-object interactions. exceeds Veason-R1 [7] by 2.5% on the MeViS dataset, indicating the efficacy of the agentic keyframe selection mechanism. 4.4 Ablation Studies Effect of the proposed components. As shown in [PITH_F… view at source ↗

**Figure 5.** Figure 5: Prompt templates for CoT datasets generation. [PITH_FULL_IMAGE:figures/full_fig_p013_5.png] view at source ↗

**Figure 7.** Figure 7: Composition of the training datasets and sample examples. [PITH_FULL_IMAGE:figures/full_fig_p014_7.png] view at source ↗

**Figure 8.** Figure 8: Prompt templates for AKS and KTG tasks. and their behavioral logic if applicable. The description should be as detailed as possible within 1000 words. [PITH_FULL_IMAGE:figures/full_fig_p015_8.png] view at source ↗

**Figure 9.** Figure 9: Prompt templates for KFG task [PITH_FULL_IMAGE:figures/full_fig_p015_9.png] view at source ↗

**Figure 10.** Figure 10: Curves of timing analysis on VRS and RVOS datasets. [PITH_FULL_IMAGE:figures/full_fig_p017_10.png] view at source ↗

**Figure 11.** Figure 11: Curves of GRPO training for AKS task. (a) Average reward (b) Average response length (c) KL Loss [PITH_FULL_IMAGE:figures/full_fig_p018_11.png] view at source ↗

**Figure 12.** Figure 12: Curves of GRPO training for KTG task [PITH_FULL_IMAGE:figures/full_fig_p018_12.png] view at source ↗

**Figure 13.** Figure 13: Curve of CoT-SFT training loss. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_13.png] view at source ↗

**Figure 14.** Figure 14: Visualization of failure cases on the ReVOS dataset. RCoT-Seg reveals inconsistencies [PITH_FULL_IMAGE:figures/full_fig_p019_14.png] view at source ↗

**Figure 15.** Figure 15: Visualization of failure cases on the ReVOS dataset. In this case, RCoT-Seg erroneously [PITH_FULL_IMAGE:figures/full_fig_p020_15.png] view at source ↗

**Figure 16.** Figure 16: More qualitative comparison on the ReasonVOS dataset. [PITH_FULL_IMAGE:figures/full_fig_p021_16.png] view at source ↗

**Figure 17.** Figure 17: More qualitative comparison on the ReVOS dataset. [PITH_FULL_IMAGE:figures/full_fig_p021_17.png] view at source ↗

read the original abstract

Video Reasoning Segmentation (VRS) aims to segment target objects in videos based on implicit instructions that convey human intent and temporal logic. Existing MLLM-based methods predict masks with a [SEG] token after selecting frames via simple sampling or an auxiliary MLLM, where limited supervision and frame-language similarity rules often yield narrow-scope keyframe choices that weaken holistic temporal understanding and lead to brittle localization in complex multi-object scenes. To address these issues, we introduce RCoT-Seg, a video-of-thought framework that factorizes VRS into temporal video reasoning (TVR) and keyframe target perception (KTP), explicitly separating temporal reasoning from spatial perception. Specifically, in the TVR stage, an agentic keyframe selection module, initialized with a curated CoT-start corpus and refined by GRPO under task-aligned rewards, is proposed to generate and reselect the keyframe through self-evaluation, strengthening moment localization and temporal reasoning. In the KTP stage, RCoT-Seg performs high-resolution segmentation on the selected frame and propagates masks with SAM2-based methods across the sequence, replacing heuristic sampling and external selectors while improving spatial precision and inter-frame consistency. Extensive experimental results demonstrate that the proposed RCoT-Seg achieves favorable performance against the state-of-the-art methods. The code and models will be publicly released at https://github.com/Victor-wjw/RCoT-Seg.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

RCoT-Seg factorizes video reasoning segmentation into a GRPO-refined CoT keyframe selector and SAM2 propagation, but the abstract's performance claims lack the controls needed to confirm they reflect genuine temporal gains rather than reward or corpus artifacts.

read the letter

The paper's core move is to split VRS into TVR for temporal reasoning and KTP for spatial perception. In TVR an agentic module starts from a curated CoT corpus, then uses GRPO with task-aligned rewards to generate and self-evaluate keyframes. KTP runs high-resolution segmentation on the chosen frame and propagates masks via SAM2. This replaces heuristic sampling and external selectors with an explicit loop that tries to strengthen moment localization in multi-object scenes. The separation itself is useful; it makes the temporal logic more inspectable and lets the segmentation step focus on precision and consistency once the right frame is picked. The self-evaluation step inside the TVR stage is a straightforward way to close the loop without adding another full MLLM pass. Those pieces are clearly described and build directly on published components. The experimental section is the weak point. The abstract states that RCoT-Seg achieves favorable results against state-of-the-art methods, yet supplies no numbers, no baseline tables, no ablation on the GRPO component versus the initial corpus, and no cross-dataset checks. Without those, it is impossible to judge whether the reported gains survive standard controls for data leakage or reward hacking. The task-aligned rewards are defined on downstream segmentation quality, which creates a modest circularity risk the stress-test note correctly flags. If the full paper contains only the same high-level claims without the missing controls, the central argument does not yet hold up. This work sits inside the current MLLM video segmentation literature and will mainly interest researchers already running similar pipelines. The factorization and the GRPO application are concrete enough to evaluate, so the paper deserves a serious referee who can ask for the ablations and generalization results. I would send it out for review but would not cite it until the quantitative evidence is stronger.

Referee Report

3 major / 2 minor

Summary. The paper proposes RCoT-Seg, a video-of-thought framework for Video Reasoning Segmentation (VRS) that factorizes the task into a temporal video reasoning (TVR) stage—an agentic keyframe selector initialized on a curated CoT-start corpus and refined via GRPO under task-aligned rewards with self-evaluation—and a keyframe target perception (KTP) stage that performs high-resolution segmentation on the selected frame before propagating masks via SAM2. It claims this separation of temporal reasoning from spatial perception overcomes limitations of heuristic sampling and external selectors, yielding favorable performance over state-of-the-art MLLM-based methods.

Significance. If the GRPO refinement can be shown to produce keyframes that genuinely strengthen holistic temporal understanding and localization in multi-object scenes without overfitting to reward signals or corpus biases, the explicit factorization could advance agentic MLLM pipelines for video tasks involving implicit instructions. The planned public release of code and models is a clear strength for reproducibility.

major comments (3)

[§3.2] §3.2 (TVR stage, GRPO refinement): The task-aligned rewards are defined directly in terms of downstream segmentation quality. This creates a circularity risk—the reported gains in keyframe selection may partly reflect optimization toward the evaluation metric rather than improved temporal reasoning. No sensitivity analysis or alternative reward formulations are provided to rule this out.
[§4] §4 (Experiments): The central claim that GRPO produces keyframes strengthening holistic moment localization rests on the assumption that gains exceed those from the curated CoT-start corpus alone. No ablation isolating GRPO’s contribution from corpus initialization or from standard sampling is reported, leaving open whether improvements address brittle localization or exploit dataset-specific patterns.
[§4.2–4.3] §4.2–4.3 (Results and ablations): Absence of cross-dataset generalization tests or error analysis focused on complex multi-object scenes means the headline superiority claim cannot be verified to survive standard controls for reward hacking or post-hoc selection. This directly undermines the weakest assumption that the agentic selector delivers genuine temporal reasoning gains.

minor comments (2)

[Abstract] Abstract: The statement of 'favorable performance against the state-of-the-art methods' is too vague; a brief summary of key metrics, baselines, and dataset names would improve immediate readability.
[§3] Notation: GRPO is introduced without expansion on first use; spelling out the full term (and clarifying its relation to standard RL methods) in §3 would aid readers.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the thoughtful and detailed comments, which help clarify the presentation of our contributions. We address each major point below and will revise the manuscript accordingly to strengthen the evidence for the GRPO component and its role in temporal reasoning.

read point-by-point responses

Referee: [§3.2] §3.2 (TVR stage, GRPO refinement): The task-aligned rewards are defined directly in terms of downstream segmentation quality. This creates a circularity risk—the reported gains in keyframe selection may partly reflect optimization toward the evaluation metric rather than improved temporal reasoning. No sensitivity analysis or alternative reward formulations are provided to rule this out.

Authors: We acknowledge the referee's concern regarding potential circularity. The rewards are deliberately task-aligned because the ultimate objective of VRS is accurate segmentation under implicit instructions; however, we agree that this requires explicit validation. In the revised manuscript, we will add a sensitivity analysis in §3.2 that varies reward weightings (e.g., emphasizing temporal consistency and self-evaluation terms independently of final IoU) and compares against an alternative reward formulation based solely on keyframe-language alignment metrics without downstream segmentation feedback. These results will be reported in new tables to demonstrate that GRPO improvements are not solely due to direct metric optimization. revision: yes
Referee: [§4] §4 (Experiments): The central claim that GRPO produces keyframes strengthening holistic moment localization rests on the assumption that gains exceed those from the curated CoT-start corpus alone. No ablation isolating GRPO’s contribution from corpus initialization or from standard sampling is reported, leaving open whether improvements address brittle localization or exploit dataset-specific patterns.

Authors: We agree that isolating GRPO's incremental benefit is necessary to substantiate the central claim. While the original experiments compare against MLLM baselines and heuristic sampling, they do not explicitly ablate the GRPO refinement step from the CoT-start corpus initialization. We will add a dedicated ablation study in §4 that includes: (1) the agent initialized only on the CoT-start corpus without GRPO, (2) standard uniform or similarity-based sampling, and (3) the full GRPO-refined selector. This will quantify the specific contribution of reinforcement learning to improved moment localization and temporal reasoning. revision: yes
Referee: [§4.2–4.3] §4.2–4.3 (Results and ablations): Absence of cross-dataset generalization tests or error analysis focused on complex multi-object scenes means the headline superiority claim cannot be verified to survive standard controls for reward hacking or post-hoc selection. This directly undermines the weakest assumption that the agentic selector delivers genuine temporal reasoning gains.

Authors: We recognize that cross-dataset evaluation and targeted error analysis are important for ruling out dataset-specific effects and reward hacking. Our primary results follow the evaluation protocols of prior VRS work on established benchmarks to ensure comparability. In the revision, we will expand §4.2–4.3 with a detailed error analysis on multi-object and complex temporal scenes, including quantitative breakdowns and qualitative examples showing how the GRPO selector improves localization over baselines. We will also add cross-dataset results on an additional benchmark where feasible. These additions will provide stronger evidence that the observed gains reflect genuine improvements in temporal reasoning rather than post-hoc selection artifacts. revision: partial

Circularity Check

0 steps flagged

No significant circularity; method and claims are empirically grounded

full rationale

The paper introduces RCoT-Seg as a factorization of VRS into TVR (agentic keyframe selection via GRPO on a curated corpus) and KTP (SAM2-based segmentation). Performance is asserted via experimental comparison to SOTA methods rather than any closed-form derivation or prediction that reduces to the inputs by construction. No equations are presented that equate outputs to fitted parameters or self-evaluations; the GRPO rewards are described as task-aligned but the validation remains external empirical testing. No self-citation chains, uniqueness theorems, or ansatz smuggling appear in the abstract or described pipeline. The derivation chain is self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The approach rests on the empirical effectiveness of GRPO under task-specific rewards and the assumption that a curated CoT-start corpus supplies a sufficiently strong initialization; no new physical or mathematical entities are introduced.

free parameters (1)

task-aligned reward weights
The rewards guiding GRPO are defined relative to final segmentation quality and therefore constitute fitted or hand-tuned parameters whose exact form is not specified in the abstract.

axioms (2)

domain assumption A curated CoT-start corpus provides a good initialization for the agentic keyframe selector.
The abstract states the module is 'initialized with a curated CoT-start corpus' without independent verification that this corpus is representative or bias-free.
domain assumption SAM2-based mask propagation preserves inter-frame consistency after high-resolution segmentation on the selected keyframe.
The method delegates temporal consistency to SAM2 without additional justification or ablation in the provided abstract.

pith-pipeline@v0.9.0 · 5563 in / 1583 out tokens · 41559 ms · 2026-05-11T01:12:47.288501+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Foundation/RealityFromDistinction.lean reality_from_one_distinction unclear
RCoT-Seg factorizes VRS into temporal video reasoning (TVR) and keyframe target perception (KTP) ... agentic keyframe selection module, initialized with a curated CoT-start corpus and refined by GRPO under task-aligned rewards
IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear
We design a Hungarian algorithm-based reward to robustly handle multi-object scenarios

Reference graph

Works this paper leans on

35 extracted references · 35 canonical work pages · 11 internal anchors

[1]

Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2. 5-vl technical report.arXiv preprint arXiv:2502.13923, 2025

work page internal anchor Pith review Pith/arXiv arXiv 2025
[2]

Univg-r1: Reasoning guided universal visual grounding with reinforcement learning.arXiv preprint arXiv:2505.14231, 2025

Sule Bai, Mingxing Li, Yong Liu, Jing Tang, Haoji Zhang, Lei Sun, Xiangxiang Chu, and Yansong Tang. Univg-r1: Reasoning guided universal visual grounding with reinforcement learning.arXiv preprint arXiv:2505.14231, 2025

work page arXiv 2025
[3]

One token to seg them all: Language instructed reasoning seg- mentation in videos.Advances in Neural Information Processing Systems, 37:6833–6859, 2024

Zechen Bai, Tong He, Haiyang Mei, Pichao Wang, Ziteng Gao, Joya Chen, Lei Liu, Zheng Zhang, and Mike Z Shou. One token to seg them all: Language instructed reasoning seg- mentation in videos.Advances in Neural Information Processing Systems, 37:6833–6859, 2024

work page 2024
[4]

Mevis: A large-scale benchmark for video segmentation with motion expressions

Henghui Ding, Chang Liu, Shuting He, Xudong Jiang, and Chen Change Loy. Mevis: A large-scale benchmark for video segmentation with motion expressions. InProceedings of the IEEE/CVF international conference on computer vision, pages 2694–2703, 2023

work page 2023
[5]

Got: Unleashing reasoning capability of multimodal large language model for visual generation and editing.arXiv preprint arXiv:2503.10639, 2025a

Rongyao Fang, Chengqi Duan, Kun Wang, Linjiang Huang, Hao Li, Shilin Yan, Hao Tian, Xingyu Zeng, Rui Zhao, Jifeng Dai, et al. Got: Unleashing reasoning capability of multimodal large language model for visual generation and editing.arXiv preprint arXiv:2503.10639, 2025

work page arXiv 2025
[6]

Video-R1: Reinforcing Video Reasoning in MLLMs

Kaituo Feng, Kaixiong Gong, Bohao Li, Zonghao Guo, Yibing Wang, Tianshuo Peng, Junfei Wu, Xiaoying Zhang, Benyou Wang, and Xiangyu Yue. Video-r1: Reinforcing video reasoning in mllms.arXiv preprint arXiv:2503.21776, 2025

work page internal anchor Pith review arXiv 2025
[7]

Reinforcing video reasoning segmentation to think before it segments.arXiv preprint arXiv:2508.11538, 2025

Sitong Gong, Lu Zhang, Yunzhi Zhuge, Xu Jia, Pingping Zhang, and Huchuan Lu. Reinforcing video reasoning segmentation to think before it segments.arXiv preprint arXiv:2508.11538, 2025

work page arXiv 2025
[8]

The devil is in temporal token: High quality video reasoning segmentation

Sitong Gong, Yunzhi Zhuge, Lu Zhang, Zongxin Yang, Pingping Zhang, and Huchuan Lu. The devil is in temporal token: High quality video reasoning segmentation. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 29183–29192, 2025

work page 2025
[9]

DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning

Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948, 2025

work page internal anchor Pith review Pith/arXiv arXiv 2025
[10]

Lora: Low-rank adaptation of large language models.ICLR, 1 (2):3, 2022

Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models.ICLR, 1 (2):3, 2022

work page 2022
[11]

OpenAI o1 System Card

Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richardson, Ahmed El-Kishky, Aiden Low, Alec Helyar, Aleksander Madry, Alex Beutel, Alex Carney, et al. Openai o1 system card.arXiv preprint arXiv:2412.16720, 2024. 10

work page internal anchor Pith review Pith/arXiv arXiv 2024
[12]

Videochat-r1: Enhancing spatio-temporal perception via reinforcement fine-tuning,

Xinhao Li, Ziang Yan, Desen Meng, Lu Dong, Xiangyu Zeng, Yinan He, Yali Wang, Yu Qiao, Yi Wang, and Limin Wang. Videochat-r1: Enhancing spatio-temporal perception via reinforce- ment fine-tuning.arXiv preprint arXiv:2504.06958, 2025

work page arXiv 2025
[13]

Glus: Global-local reasoning unified into a single large language model for video segmentation

Lang Lin, Xueyang Yu, Ziqi Pang, and Yu-Xiong Wang. Glus: Global-local reasoning unified into a single large language model for video segmentation. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 8658–8667, 2025

work page 2025
[14]

Seg-zero: Reasoning-chain guided segmentation via cognitive reinforcement.arXiv preprint arXiv:2503.06520, 2025

Yuqi Liu, Bohao Peng, Zhisheng Zhong, Zihao Yue, Fanbin Lu, Bei Yu, and Jiaya Jia. Seg- zero: Reasoning-chain guided segmentation via cognitive reinforcement.arXiv preprint arXiv:2503.06520, 2025

work page arXiv 2025
[15]

Visionreasoner: Unified visual perception and reasoning via reinforcement learning.arXiv preprint arXiv:2505.12081, 2025

Yuqi Liu, Tianyuan Qu, Zhisheng Zhong, Bohao Peng, Shu Liu, Bei Yu, and Jiaya Jia. Vision- reasoner: Unified visual perception and reasoning via reinforcement learning.arXiv preprint arXiv:2505.12081, 2025

work page arXiv 2025
[16]

Visual-RFT: Visual Reinforcement Fine-Tuning

Ziyu Liu, Zeyi Sun, Yuhang Zang, Xiaoyi Dong, Yuhang Cao, Haodong Duan, Dahua Lin, and Jiaqi Wang. Visual-rft: Visual reinforcement fine-tuning.arXiv preprint arXiv:2503.01785, 2025

work page internal anchor Pith review arXiv 2025
[17]

The 2017 DAVIS Challenge on Video Object Segmentation

Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation.arXiv preprint arXiv:1704.00675, 2017

work page internal anchor Pith review arXiv 2017
[18]

SAM 2: Segment Anything in Images and Videos

Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos.arXiv preprint arXiv:2408.00714, 2024

work page internal anchor Pith review Pith/arXiv arXiv 2024
[19]

Urvos: Unified referring video object segmentation network with a large-scale benchmark

Seonguk Seo, Joon-Young Lee, and Bohyung Han. Urvos: Unified referring video object segmentation network with a large-scale benchmark. InComputer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XV 16, pages 208–223. Springer, 2020

work page 2020
[20]

DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models

Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models.arXiv preprint arXiv:2402.03300, 2024

work page internal anchor Pith review Pith/arXiv arXiv 2024
[21]

Hybridflow: A flexible and efficient rlhf framework

Guangming Sheng, Chi Zhang, Zilingfeng Ye, Xibin Wu, Wang Zhang, Ru Zhang, Yanghua Peng, Haibin Lin, and Chuan Wu. Hybridflow: A flexible and efficient rlhf framework. In Proceedings of the Twentieth European Conference on Computer Systems, pages 1279–1297, 2025

work page 2025
[22]

MIT press Cambridge, 1998

Richard S Sutton, Andrew G Barto, et al.Reinforcement learning: An introduction, volume 1. MIT press Cambridge, 1998

work page 1998
[23]

Pixelthink: Towards efficient chain-of-pixel reasoning.arXiv preprint arXiv:2505.23727, 2025

Song Wang, Gongfan Fang, Lingdong Kong, Xiangtai Li, Jianyun Xu, Sheng Yang, Qiang Li, Jianke Zhu, and Xinchao Wang. Pixelthink: Towards efficient chain-of-pixel reasoning.arXiv preprint arXiv:2505.23727, 2025

work page arXiv 2025
[24]

Time-r1: Post-training large vision language model for temporal video grounding,

Ye Wang, Ziheng Wang, Boshen Xu, Yang Du, Kejun Lin, Zihan Xiao, Zihao Yue, Jianzhong Ju, Liang Zhang, Dingyi Yang, et al. Time-r1: Post-training large vision language model for temporal video grounding.arXiv preprint arXiv:2503.13377, 2025

work page arXiv 2025
[25]

Hyperseg: Towards universal visual segmentation with large language model.arXiv preprint arXiv:2411.17606, 2024

Cong Wei, Yujie Zhong, Haoxian Tan, Yong Liu, Zheng Zhao, Jie Hu, and Yujiu Yang. Hy- perseg: Towards universal visual segmentation with large language model.arXiv preprint arXiv:2411.17606, 2024

work page arXiv 2024
[26]

Instructseg: Unifying instructed visual segmentation with multi-modal large language models

Cong Wei, Yujie Zhong, Haoxian Tan, Yingsen Zeng, Yong Liu, Hongfa Wang, and Yujiu Yang. Instructseg: Unifying instructed visual segmentation with multi-modal large language models. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 20193–20203, 2025. 11

work page 2025
[27]

Gsva: Generalized segmentation via multimodal large language models

Zhuofan Xia, Dongchen Han, Yizeng Han, Xuran Pan, Shiji Song, and Gao Huang. Gsva: Generalized segmentation via multimodal large language models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3858–3869, 2024

work page 2024
[28]

Mindomni: Unleashing reasoning generation in vision language models with rgpo.arXiv preprint arXiv:2505.13031, 2025

Yicheng Xiao, Lin Song, Yukang Chen, Yingmin Luo, Yuxin Chen, Yukang Gan, Wei Huang, Xiu Li, Xiaojuan Qi, and Ying Shan. Mindomni: Unleashing reasoning generation in vision language models with rgpo.arXiv preprint arXiv:2505.13031, 2025

work page arXiv 2025
[29]

DanceGRPO: Unleashing GRPO on Visual Generation

Zeyue Xue, Jie Wu, Yu Gao, Fangyuan Kong, Lingting Zhu, Mengzhao Chen, Zhiheng Liu, Wei Liu, Qiushan Guo, Weilin Huang, et al. Dancegrpo: Unleashing grpo on visual generation. arXiv preprint arXiv:2505.07818, 2025

work page internal anchor Pith review arXiv 2025
[30]

Visa: Reasoning video object segmentation via large language models

Cilin Yan, Haochen Wang, Shilin Yan, Xiaolong Jiang, Yao Hu, Guoliang Kang, Weidi Xie, and Efstratios Gavves. Visa: Reasoning video object segmentation via large language models. In European Conference on Computer Vision, pages 98–115. Springer, 2024

work page 2024
[31]

Improving the Reasoning of Multi-Image Grounding in MLLMs via Reinforcement Learning

Bob Zhang, Haoran Li, Tao Zhang, Cilin Yan, Jiayin Cai, and Yanbin Hao. Improving the reasoning of multi-image grounding in mllms via reinforcement learning.arXiv preprint arXiv:2507.00748, 2025

work page internal anchor Pith review Pith/arXiv arXiv 2025
[32]

Villa: Video reasoning segmentation with large language model.arXiv preprint arXiv:2407.14500, 2024

Rongkun Zheng, Lu Qi, Xi Chen, Yi Wang, Kun Wang, Yu Qiao, and Hengshuang Zhao. Villa: Video reasoning segmentation with large language model.arXiv preprint arXiv:2407.14500, 2024

work page arXiv 2024
[33]

LlamaFactory: Unified Efficient Fine-Tuning of 100+ Language Models

Yaowei Zheng, Richong Zhang, Junhao Zhang, Yanhan Ye, Zheyan Luo, Zhangchi Feng, and Yongqiang Ma. Llamafactory: Unified efficient fine-tuning of 100+ language models.arXiv preprint arXiv:2403.13372, 2024

work page internal anchor Pith review arXiv 2024
[34]

Omni-r1: Reinforcement learning for omnimodal reasoning via two-system collaboration.arXiv preprint arXiv:2505.20256,

Hao Zhong, Muzhi Zhu, Zongze Du, Zheng Huang, Canyu Zhao, Mingyu Liu, Wen Wang, Hao Chen, and Chunhua Shen. Omni-r1: Reinforcement learning for omnimodal reasoning via two-system collaboration.arXiv preprint arXiv:2505.20256, 2025

work page arXiv 2025
[35]

arXiv preprint arXiv:2312.17448 (2023)

Jiawen Zhu, Zhi-Qi Cheng, Jun-Yan He, Chenyang Li, Bin Luo, Huchuan Lu, Yifeng Geng, and Xuansong Xie. Tracking with human-intent reasoning.arXiv preprint arXiv:2312.17448, 2023. 12 <image> The above is a frame from a video, and the text description of the video is: "{Video_Description}". One or more objects are marked with red rectangles in each frame.Th...

work page arXiv 2023