Pith. sign in

REVIEW 3 major objections 3 cited by

Diffusion models can run genuine chain-of-thought reasoning by iteratively refining latent thought states inside the multimodal encoder, not by a single static embedding.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-14 22:23 UTC pith:ZC2JPTMD

load-bearing objection Practical iterative latent CoT for diffusion that works well on hard synthetic spatial tasks, but the GT intermediate images leave open multi-stage imitation rather than pure endogenous logic. the 3 major comments →

arxiv 2603.12252 v4 pith:ZC2JPTMD submitted 2026-03-12 cs.CV cs.CL

EndoCoT: Scaling Endogenous Chain-of-Thought Reasoning in Diffusion Models

Xuanlang Dai , Yujie Zhou , Long Xing , Jiazi Bu , Xilin Wei , Yuhong Liu , Beichen Zhang , Kai Chen
show 1 more author
Yuhang Zang
This is my paper
classification cs.CV cs.CL
keywords endogenous chain-of-thoughtdiffusion transformermultimodal large language modellatent thought statesvisual reasoningmaze solvingSudokuflow matching
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Current diffusion models that bolt multimodal language models on as text encoders fail at hard spatial and logical tasks because the encoder thinks only once and then freezes its guidance while the image is denoised. EndoCoT fixes both problems: it lets the encoder repeatedly update a continuous latent thought state, and it ties each of those states to a full denoising pass so the diffusion transformer can act on progressive plans. A final alignment loss keeps the last thought state faithful to the true answer. On maze solving, traveling-salesman tours, Sudoku, and visual spatial planning the method reaches 92 percent average accuracy and still works when maze size or font style changes. The result matters because it shows that multi-step reasoning can live inside the generation process itself rather than being pre-computed and merely rendered.

Core claim

The authors show that genuine endogenous chain-of-thought reasoning arises inside a diffusion pipeline when an MLLM is allowed to iteratively refine continuous latent thought states and each state is used to condition a complete denoising trajectory of a DiT; a terminal L2 alignment to ground-truth text embeddings then keeps the trajectory from drifting. This joint iterative mechanism, not single-pass encoding or static guidance, is what lets the model solve complex multi-step visual reasoning tasks step by step and generalize to harder instances.

What carries the argument

Iterative Thought Guidance: the recursive update h_τ = extract(MLLM([prefix; h_τ−1])) that produces a fresh conditioning vector for a full DiT flow-matching pass at every reasoning step, paired with Terminal Thought Grounding that L2-aligns only the final h_𝒯 to ground-truth answer embeddings.

Load-bearing premise

That intermediate visual targets obtained by cutting ground-truth solutions into partial paths or progressive fills, plus one final text-embedding alignment, are enough to force real multi-step logic rather than multi-stage visual imitation of the supplied trajectory.

What would settle it

Train and evaluate an identical architecture on the same hard Maze-32 and Sudoku-35 splits but withhold all intermediate visual targets, supervising only the final image and the terminal thought alignment; if accuracy collapses to the level of static-guidance baselines, the claim that iterative latent refinement induces genuine reasoning is falsified.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Accuracy on long-horizon visual puzzles scales smoothly with the number of latent reasoning steps allocated at inference time.
  • A single model trained jointly on Maze, TSP, Sudoku and VSP retains competitive performance, indicating transferable latent reasoning skills.
  • Progressive image editing becomes controllable simply by choosing how many reasoning steps to run, without separate editing modules.
  • The same latent-refinement loop can be attached to other DiT backbones that already use MLLM text encoders.
  • High-resolution generation becomes relatively cheaper because the expensive DiT denoising is not repeated for every reasoning step.

Where Pith is reading between the lines

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

  • If the intermediate-supervision assumption holds only for algorithmically generated paths, the method may not transfer to open-ended visual planning where no clean partial solutions exist.
  • The continuous latent tokens could in principle support tree-structured search or backtracking inside the MLLM, turning the current linear chain into a latent planner.
  • The observed early commitment of vanilla diffusion models suggests that any static-conditioning pipeline, even with stronger MLLMs, will hit the same reasoning-depth wall on combinatorial tasks.
  • Adaptive stopping of the latent loop, rather than a fixed step budget, would remove the manual hyper-parameter the authors currently leave to the user.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 0 minor

Summary. The paper proposes EndoCoT, a framework that couples an MLLM text encoder with a Diffusion Transformer so that latent thought states are refined iteratively (Iterative Thought Guidance; Eq. 4–6) and the terminal latent is L2-aligned to ground-truth text embeddings (Terminal Thought Grounding; Eqs. 7–8). A two-stage progressive training schedule first supervises all intermediate visual targets obtained by sequential ground-truth decomposition, then consolidates the final output. Built on Qwen-Image-Edit-2511 with LoRA, the method is evaluated on Maze, TSP, Sudoku, and VSP (task-specific and unified settings). Task-specific training reports 92.1% average accuracy, +8.3 pp over DiffThinker, with large gains on hard instances (Maze-32 90% vs 65%; Sudoku-35 95% vs 55%). Supporting analyses include layer-wise sensitivity, attention entropy, ablations of semantic loss and implicit vs explicit tokens, MLLM/DiT-only baselines, and inference-time scaling of the reasoning budget τ.

Significance. If the iterative latent mechanism truly induces multi-step logical planning rather than multi-stage visual imitation of provided trajectories, the work would be a meaningful step toward endogenous reasoning inside diffusion generators, with practical value for spatial planning and controllable progressive editing. Strengths that should be credited regardless of interpretation: multi-task tables under both task-specific and unified regimes, systematic ablations (Tabs. 2–4, 6), inference-time CoT scaling curves with latency trade-offs, qualitative intermediate trajectories, layer-wise sensitivity motivating joint MLLM–DiT adaptation, and public code plus a large constructed dataset (182.4K). These make the empirical contribution solid even if the strongest ‘genuine CoT’ framing requires tighter evidence.

major comments (3)
  1. Sec. 4.2.1 and Eqs. (4)–(6): each reasoning step is supervised by sequential ground-truth visual intermediates I*_τ (partial maze paths, progressive Sudoku fills, etc.). Terminal text alignment (Eqs. 7–8) is applied only at τ=𝒯. The central claim of ‘genuine endogenous CoT via iterative latent refinement’ (Abstract, Sec. 1, contribution 1) is therefore not isolated from multi-stage visual trajectory imitation. Existing ablations (Tab. 2 semantic loss; Tab. 3 MLLM/DiT-only; Tab. 4 τ scaling) do not remove intermediate visual targets while retaining terminal grounding. A load-bearing experiment is needed: train with only final-image (and/or terminal text) supervision, or probe whether h_τ encode constraint-satisfying logic when the visual GT trajectory is withheld or scrambled.
  2. Table 1, unified-training block: accuracy drops sharply relative to task-specific training (e.g., Maze-32 90%→52%, TSP-18 73%→46%, Sudoku-35 95%→80%). The paper still frames EndoCoT as learning transferable reasoning (Sec. 5.1). These drops are large enough that claims of abstract, cross-task endogenous CoT need either stronger analysis of what transfers or explicit tempering of the abstraction narrative; otherwise the gains look closer to task-specialized multi-stage imitation.
  3. Sec. 1 / conclusion: the manuscript asserts the ‘first’ diffusion framework for genuine CoT and contrasts with DiffThinker as ‘superficial alignment.’ Given that both methods use powerful MLLM+DiT backbones and heavy task-specific supervision, and that EndoCoT’s intermediates are algorithmically generated GT decompositions (Appendix B), the novelty claim should be restated operationally (what is endogenous, what is measured) and supported by the isolation experiment above. Without that, the contribution remains a strong empirical system with progressive latent conditioning, not yet a demonstrated cognitive CoT mechanism.

Circularity Check

0 steps flagged

No derivation-level circularity; supervised empirical method matches external algorithmic GT intermediates and reports accuracy, with only mild interpretive overclaim that trajectory matching equals endogenous CoT.

full rationale

EndoCoT is a standard supervised ML framework paper, not a first-principles derivation. Intermediate targets I*_τ are obtained from external algorithmic ground truth (DFS/BFS for mazes, Held-Karp for TSP, backtracking for Sudoku, Dijkstra for VSP; Appendix B) and used as visual supervision in the Conditional Flow Matching loss (Eq. 6) plus terminal L2 alignment of h_𝒯 to GT text embeddings (Eqs. 7–9). Progressive two-stage training (Eqs. 10–11) and inference-time τ scaling (Tab. 4) are ordinary training/inference choices. Accuracy numbers (92.1% avg, Tab. 1) are measured against held-out instances of the same external solvers; ablations (Tabs. 2–3) remove components and show degradation. There is no equation that reduces a claimed prediction to a fitted identity by construction, no uniqueness theorem imported via self-citation, and no ansatz smuggled from prior author work. The only mild issue is the interpretive framing (Abstract, Sec. 1, Sec. 4.2, Conclusion) that recursive latent updates under visual GT supervision constitute “genuine endogenous CoT” rather than multi-stage trajectory imitation; the paper itself flags reliance on high-quality intermediate supervision. This is overclaim risk, not circularity of the derivation chain. Score 1 reflects that single interpretive step without any load-bearing reduction.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 3 invented entities

The central accuracy claims rest on standard flow-matching and LoRA machinery plus several paper-specific design choices (continuous latent thoughts, intermediate visual GT supervision, terminal L2 alignment weight, progressive stages, hand-set 𝒯). No new physical entities; invented modules are architectural. Free parameters are ordinary ML hyperparameters and the alignment weight set to 1.

free parameters (5)
  • lambda_align = 1
    Weight on terminal semantic L2 loss; set empirically to 1 and activated only at final reasoning step (Eq. 9).
  • reasoning_steps_T = task-dependent (e.g. up to 50 at inference)
    Number of iterative thought updates at train/inference; manually chosen and shown to scale accuracy (Tab. 4); conclusion admits manual tuning.
  • LoRA_rank = 32
    Rank of LoRA adapters on MLLM and DiT; fixed at 32 for all targeted modules.
  • learning_rate = 1e-4
    Training LR for LoRA fine-tuning.
  • stage2_iterations = limited (not uniquely fixed in main text)
    Short-cycle fine-tuning length for terminal consolidation; appendix shows excessive Stage-2 steps sparsify intermediate behavior.
axioms (5)
  • standard math Flow matching linear path and Conditional Flow Matching objective correctly train the DiT vector field (Sec. 4.1).
    Standard generative-modeling assumption imported from Lipman et al.
  • domain assumption A single forward pass of the MLLM cannot encode all logical constraints for complex spatial tasks; multi-round latent refinement is required (Sec. 3).
    Empirical diagnosis from maze attention/path failures; motivates iterative thought guidance.
  • domain assumption Static text embeddings cannot maintain DiT–MLLM coupling on dense topologies; conditioning must be updated each reasoning step (Sec. 3).
    Supported by attention-entropy maps; underpins dynamic bridging of h_τ to DiT.
  • ad hoc to paper Sequential ground-truth decomposition yields valid intermediate visual targets I*_τ that teach incremental reasoning (Sec. 4.2.1).
    Dataset construction choice; if intermediates only teach visual imitation, the CoT interpretation weakens.
  • ad hoc to paper Aligning final latent state h_𝒯 to embeddings of ground-truth reasoning text prevents cumulative drift (Eqs. 7–8).
    Design hypothesis validated by ablation but not derived from first principles.
invented entities (3)
  • latent thought state h_τ (endogenous CoT token) no independent evidence
    purpose: Continuous hidden vector iteratively refined by the MLLM and used as dynamic DiT condition.
    Core invented representation; independent evidence is only internal ablations and path metrics, not external theory.
  • Iterative Thought Guidance module no independent evidence
    purpose: Recursively update h_τ and run a full flow-matching generation per reasoning step.
    Architectural module defined in Sec. 4.2.1; purpose is to activate multi-step reasoning.
  • Terminal Thought Grounding module no independent evidence
    purpose: L2-align final thought to GT text embeddings to keep trajectory grounded.
    Auxiliary objective (Eq. 8); ablation shows large Maze-32 drop without it.

pith-pipeline@v1.1.0-grok45 · 23232 in / 3640 out tokens · 34803 ms · 2026-07-14T22:23:51.467255+00:00 · methodology

0 comments
read the original abstract

Recently, Multimodal Large Language Models (MLLMs) have been widely integrated into diffusion frameworks primarily as text encoders to tackle complex tasks such as spatial reasoning. However, this paradigm suffers from two critical limitations: (i) MLLMs text encoder exhibits insufficient reasoning depth. Single-step encoding fails to activate the Chain-of-Thought process, which is essential for MLLMs to provide accurate guidance for complex tasks. (ii) The guidance remains invariant during the decoding process. Invariant guidance during decoding prevents DiT from progressively decomposing complex instructions into actionable denoising steps, even with correct MLLM encodings. To this end, we propose Endogenous Chain-of-Thought (EndoCoT), a novel framework that first activates MLLMs' reasoning potential by iteratively refining latent thought states through an iterative thought guidance module, and then bridges these states to the DiT's denoising process. Second, a terminal thought grounding module is applied to ensure the reasoning trajectory remains grounded in textual supervision by aligning the final state with ground-truth answers. With these two components, the MLLM text encoder delivers meticulously reasoned guidance, enabling the DiT to execute it progressively and ultimately solve complex tasks in a step-by-step manner. Extensive evaluations across diverse benchmarks (e.g., Maze, TSP, VSP, and Sudoku) achieve an average accuracy of 92.1%, outperforming the strongest baseline by 8.3 percentage points. The code and dataset are publicly available at https://internlm.github.io/EndoCoT/.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Beyond the Current Observation: Evaluating Multimodal Large Language Models in Controllable Non-Markov Games

    cs.CV 2026-06 unverdicted novelty 7.0

    RNG-Bench evaluates MLLMs on hidden-observation reconstruction in non-Markov games, finds forgetting as the dominant error source, and shows fine-tuning on optimal rollouts improves performance with transfer to other ...

  2. ChartSync: A Benchmark for Visuo-Logical Cascading Chart Editing

    cs.CV 2026-07 conditional novelty 6.0

    ChartSync formalizes visuo-logical cascading chart editing and finds only two frontier proprietary models show strong text-to-geometry synchronization; open-source models largely fail.

  3. DDA-Thinker: Decoupled Dual-Atomic Reinforcement Learning for Reasoning-Driven Image Editing

    cs.CV 2026-04 unverdicted novelty 6.0

    DDA-Thinker decouples planning from generation and applies dual-atomic RL with checklist-based rewards to boost reasoning in image editing, yielding competitive results on RISE-Bench and KRIS-Bench.

Reference graph

Works this paper leans on

43 extracted references · 16 linked inside Pith · cited by 3 Pith papers

  1. [1]

    The art of scaling test-time compute for large language models.arXiv preprint arXiv:2512.02008, 2025

    Aradhye Agarwal, Ayan Sengupta, and Tanmoy Chakraborty. The art of scaling test-time compute for large language models.arXiv preprint arXiv:2512.02008, 2025. 2

    arXiv 2025

  2. [2]

    Claude opus 4.6

    Anthropic. Claude opus 4.6. Anthropic Official Website, February 2026. Release; Claude Opus 4.6 achieves new state-of-the-art performance in complex reasoning, coding, and long-context understanding. 1

    2026

  3. [3]

    Qwen3-vl technical report, 2025

    Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, Wenbin Ge, Zhifang Guo, Qidong Huang, Jie Huang, Fei Huang, Binyuan Hui, Shutong Jiang, Zhaohai Li, Mingsheng Li, Mei Li, Kaixin Li, Zicheng Lin, Junyang Lin, Xuejing Liu, Jiawei Liu, Chenglong Liu, Yang Liu, Dayiheng Liu, Shixuan ...

    2025

  4. [4]

    Graph of thoughts: Solving elaborate problems with large language models

    MaciejBesta, NilsBlach, AlesKubicek, RobertGerstenberger, MichalPodstawski, LukasGianinazzi, Joanna Gajda, Tomasz Lehmann, Hubert Niewiadomski, Piotr Nyczyk, et al. Graph of thoughts: Solving elaborate problems with large language models. InProceedings of the AAAI conference on artificial intelligence, volume 38, pages 17682–17690, 2024. 2

    2024

  5. [5]

    Flux.2: Frontier visual intelligence

    Black Forest Labs. Flux.2: Frontier visual intelligence. Black Forest Labs Official Website. Release; FLUX.2 provides significant improvements in visual quality, prompt adherence, and architectural efficiency. 1, 4.2

  6. [6]

    Video- of-thought: Step-by-step video reasoning from perception to cognition.arXiv preprint arXiv:2501.03230,

    Hao Fei, Shengqiong Wu, Wei Ji, Hanwang Zhang, Meishan Zhang, Mong-Li Lee, and Wynne Hsu. Video- of-thought: Step-by-step video reasoning from perception to cognition.arXiv preprint arXiv:2501.03230,

    Pith/arXiv arXiv

  7. [7]

    D-ar: Diffusion via autoregressive models.arXiv preprint arXiv:2505.23660, 2025

    Ziteng Gao and Mike Zheng Shou. D-ar: Diffusion via autoregressive models.arXiv preprint arXiv:2505.23660, 2025. 2

    Pith/arXiv arXiv 2025

  8. [8]

    Gemini 3

    Google DeepMind. Gemini 3. Google DeepMind Official Website, February 2026. Release; Gemini 3 is Google’s most capable model, featuring breakthrough capabilities in long-context understanding, multimodal reasoning, and native agentic performance. F

    2026

  9. [9]

    Dart: Denoising autoregressive transformer for scalable text-to-image generation

    Jiatao Gu, Yuyang Wang, Yizhe Zhang, Qihang Zhang, Dinghuai Zhang, Navdeep Jaitly, Josh Susskind, and Shuangfei Zhai. Dart: Denoising autoregressive transformer for scalable text-to-image generation. arXiv preprint arXiv:2410.08159, 2024. 2

    Pith/arXiv arXiv 2024

  10. [10]

    Training large language models to reason in a continuous latent space.arXiv preprint arXiv:2412.06769,

    Shibo Hao, Sainbayar Sukhbaatar, DiJia Su, Xian Li, Zhiting Hu, Jason Weston, and Yuandong Tian. Training large language models to reason in a continuous latent space.arXiv preprint arXiv:2412.06769,

    Pith/arXiv arXiv

  11. [11]

    Diffthinker: Towards generative multimodal reasoning with diffusion models.arXiv preprint arXiv:2512.24165, 2025

    Zefeng He, Xiaoye Qu, Yafu Li, Tong Zhu, Siyuan Huang, and Yu Cheng. Diffthinker: Towards generative multimodal reasoning with diffusion models.arXiv preprint arXiv:2512.24165, 2025. 1, 2, 5.1, 1, 5.1, 4

    arXiv 2025

  12. [12]

    Denoising diffusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020. 1

    2020

  13. [13]

    LoRA: Low-rank adaptation of large language models

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Liang Wang, Weizhu Chen, et al. LoRA: Low-rank adaptation of large language models. InICLR, 2022. 5

    2022

  14. [14]

    Real-time intermediate flow estimation for video frame interpolation

    Zhewei Huang, Tianyuan Zhang, Wen Heng, Boxin Shi, and Shuchang Zhou. Real-time intermediate flow estimation for video frame interpolation. InEuropean conference on computer vision, pages 624–642. Springer, 2022. F

    2022

  15. [15]

    Thinkgen: Generalized thinking for visual generation.arXiv preprint arXiv:2512.23568, 2025

    Siyu Jiao, Yiheng Lin, Yujie Zhong, Qi She, Wei Zhou, Xiaohan Lan, Zilong Huang, Fei Yu, Yingchen Yu, Yunqing Zhao, et al. Thinkgen: Generalized thinking for visual generation.arXiv preprint arXiv:2512.23568, 2025. 2, 1, 5.1, C.2 12 EndoCoT : Scaling Endogenous Chain-of-Thought Reasoning in Diffusion Models

    arXiv 2025

  16. [16]

    Think-then-generate: Reasoning-aware text-to-image diffusion with llm encoders.arXiv preprint arXiv:2601.10332, 2026

    Siqi Kou, Jiachun Jin, Zetong Zhou, Ye Ma, Yugang Wang, Quan Chen, Peng Jiang, Xiao Yang, Jun Zhu, Kai Yu, et al. Think-then-generate: Reasoning-aware text-to-image diffusion with llm encoders.arXiv preprint arXiv:2601.10332, 2026. 2

    arXiv 2026

  17. [17]

    Flow matching for generative modeling.arXiv preprint arXiv:2210.02747, 2022

    Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling.arXiv preprint arXiv:2210.02747, 2022. 1, 4.1

    Pith/arXiv arXiv 2022

  18. [18]

    Ministral 3.arXiv preprint arXiv:2601.08584, 2026

    Alexander H Liu, Kartik Khandelwal, Sandeep Subramanian, Victor Jouault, Abhinav Rastogi, Adrien Sadé, Alan Jeffares, Albert Jiang, Alexandre Cahill, Alexandre Gavaudan, et al. Ministral 3.arXiv preprint arXiv:2601.08584, 2026. 1

    Pith/arXiv arXiv 2026

  19. [19]

    Thinking with images

    OpenAI. Thinking with images. OpenAI Official Website, April 2025. Release; OpenAI o3 and o4-mini visual reasoning models can think with images in chain-of-thought. 2

    2025

  20. [20]

    Chain-of-visual-thought: Teaching vlms to see and think better with continuous visual tokens

    Yiming Qin, Bomin Wei, Jiaxin Ge, Konstantinos Kallidromitis, Stephanie Fu, Trevor Darrell, and XuDong Wang. Chain-of-visual-thought: Teaching vlms to see and think better with continuous visual tokens. arXiv preprint arXiv:2511.19418, 2025. 2

    arXiv 2025

  21. [21]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. InInternational conference on machine learning, pages 8748–8763. PmLR,

  22. [22]

    Exploring the limits of transfer learning with a unified text-to-text transformer

    Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1–67, 2020. 2

    2020

  23. [23]

    Pathways on the image manifold: Image editing via video generation

    Noam Rotstein, Gal Yona, Daniel Silver, Roy Velich, David Bensaïd, and Ron Kimmel. Pathways on the image manifold: Image editing via video generation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7857–7866, 2025. 2

    2025

  24. [24]

    Codi: Compressing chain-of-thought into continuous space via self-distillation.arXiv preprint arXiv:2502.21074, 2025

    Zhenyi Shen, Hanqi Yan, Linhai Zhang, Zhanghao Hu, Yali Du, and Yulan He. Codi: Compressing chain-of-thought into continuous space via self-distillation.arXiv preprint arXiv:2502.21074, 2025. 2, 5.2

    Pith/arXiv arXiv 2025

  25. [25]

    Denoising diffusion implicit models.arXiv preprint arXiv:2010.02502, 2020

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models.arXiv preprint arXiv:2010.02502, 2020. 1

    Pith/arXiv arXiv 2010

  26. [26]

    Score-based generative modeling through stochastic differential equations.arXiv preprint arXiv:2011.13456, 2020

    Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations.arXiv preprint arXiv:2011.13456, 2020. 1

    Pith/arXiv arXiv 2011

  27. [27]

    Think silently, think fast: Dynamic latent compression of llm reasoning chains.arXiv preprint arXiv:2505.16552, 2025

    Wenhui Tan, Jiaze Li, Jianzhong Ju, Zhenbo Luo, Jian Luan, and Ruihua Song. Think silently, think fast: Dynamic latent compression of llm reasoning chains.arXiv preprint arXiv:2505.16552, 2025. 2

    arXiv 2025

  28. [28]

    Thinking with video: Video generation as a promising multimodal reasoning paradigm.arXiv preprint arXiv:2511.04570, 2025

    Jingqi Tong, Yurong Mou, Hangcheng Li, Mingzhe Li, Yongzhuo Yang, Ming Zhang, Qiguang Chen, Tianyi Liang, Xiaomeng Hu, Yining Zheng, et al. Thinking with video: Video generation as a promising multimodal reasoning paradigm.arXiv preprint arXiv:2511.04570, 2025. 2

    Pith/arXiv arXiv 2025

  29. [29]

    A very big video reasoning suite, 2026

    Maijunxian Wang, Ruisi Wang, Juyi Lin, Ran Ji, Thaddäus Wiedemer, Qingying Gao, Dezhi Luo, Yaoyao Qian, Lianyu Huang, Zelong Hong, Jiahui Ge, Qianli Ma, Hang He, Yifan Zhou, Lingzi Guo, Lantao Mei, Jiachen Li, Hanwen Xing, Tianqi Zhao, Fengyuan Yu, Weihang Xiao, Yizheng Jiao, Jianheng Hou, Danyang Zhang, Pengcheng Xu, Boyang Zhong, Zehong Zhao, Gaoyun Fan...

    2026

  30. [30]

    thinking with videos

    Shijian Wang, Jiarui Jin, Xingjian Wang, Linxin Song, Runhao Fu, Hecheng Wang, Zongyuan Ge, Yuan Lu, and Xuelian Cheng. Video-thinker: Sparking" thinking with videos" via reinforcement learning.arXiv preprint arXiv:2510.23473, 2025. 2 13 EndoCoT : Scaling Endogenous Chain-of-Thought Reasoning in Diffusion Models

    arXiv 2025

  31. [31]

    Chain-of-thought prompting elicits reasoning in large language models.Advances in neural information processing systems, 35:24824–24837, 2022

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models.Advances in neural information processing systems, 35:24824–24837, 2022. 1

    2022

  32. [32]

    Qwen-image technical report.arXiv preprint arXiv:2508.02324, 2025

    Chenfei Wu, Jiahao Li, Jingren Zhou, Junyang Lin, Kaiyuan Gao, Kun Yan, Sheng-ming Yin, Shuai Bai, Xiao Xu, Yilei Chen, et al. Qwen-image technical report.arXiv preprint arXiv:2508.02324, 2025. 1, 2, 3, 4.2, 5, 1, F

    Pith/arXiv arXiv 2025

  33. [33]

    Chronoedit: Towards temporal reasoning for image editing and world simulation.arXiv preprint arXiv:2510.04290, 2025

    Jay Zhangjie Wu, Xuanchi Ren, Tianchang Shen, Tianshi Cao, Kai He, Yifan Lu, Ruiyuan Gao, Enze Xie, Shiyi Lan, Jose M Alvarez, et al. Chronoedit: Towards temporal reasoning for image editing and world simulation.arXiv preprint arXiv:2510.04290, 2025. 2, 1, 5.1, C.2

    arXiv 2025

  34. [34]

    Videochat-r1

    Ziang Yan, Xinhao Li, Yinan He, Zhengrong Yue, Xiangyu Zeng, Yali Wang, Yu Qiao, Limin Wang, and Yi Wang. Videochat-r1. 5: Visual test-time scaling to reinforce multimodal reasoning by iterative perception.arXiv preprint arXiv:2509.21100, 2025. 2

    arXiv 2025

  35. [35]

    Qwen3 technical report.arXiv preprint arXiv:2505.09388, 2025

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. Qwen3 technical report.arXiv preprint arXiv:2505.09388, 2025. 1, 1

    Pith/arXiv arXiv 2025

  36. [36]

    Tree of thoughts: Deliberate problem solving with large language models.Advances in neural information processing systems, 36:11809–11822, 2023

    Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan. Tree of thoughts: Deliberate problem solving with large language models.Advances in neural information processing systems, 36:11809–11822, 2023. 2

    2023

  37. [37]

    Thinking with videos: Multimodal tool-augmented reinforcement learning for long video reasoning.arXiv preprint arXiv:2508.04416, 2025

    Haoji Zhang, Xin Gu, Jiawen Li, Chixiang Ma, Sule Bai, Chubin Zhang, Bowen Zhang, Zhichao Zhou, Dongliang He, and Yansong Tang. Thinking with videos: Multimodal tool-augmented reinforcement learning for long video reasoning.arXiv preprint arXiv:2508.04416, 2025. 2

    Pith/arXiv arXiv 2025

  38. [38]

    Latent sketchpad: Sketching visual thoughts to elicit multimodal reasoning in mllms.arXiv preprint arXiv:2510.24514, 2025

    Huanyu Zhang, Wenshan Wu, Chengzu Li, Ning Shang, Yan Xia, Yangyu Huang, Yifan Zhang, Li Dong, Zhang Zhang, Liang Wang, et al. Latent sketchpad: Sketching visual thoughts to elicit multimodal reasoning in mllms.arXiv preprint arXiv:2510.24514, 2025. 2

    arXiv 2025

  39. [39]

    A survey on test-time scaling in large language models: What, how, where, and how well?arXiv preprint arXiv:2503.24235, 2025

    Qiyuan Zhang, Fuyuan Lyu, Zexu Sun, Lei Wang, Weixu Zhang, Wenyue Hua, Haolun Wu, Zhihan Guo, Yufei Wang, Niklas Muennighoff, et al. A survey on test-time scaling in large language models: What, how, where, and how well?arXiv preprint arXiv:2503.24235, 2025. 2

    Pith/arXiv arXiv 2025

  40. [40]

    Are image-to-video models good zero-shot image editors?arXiv preprint arXiv:2511.19435, 2025

    Zechuan Zhang, Zhenyuan Chen, Zongxin Yang, and Yi Yang. Are image-to-video models good zero-shot image editors?arXiv preprint arXiv:2511.19435, 2025. 2

    arXiv 2025

  41. [41]

    Soft thinking: Unlocking the reasoning potential of llms in continuous concept space

    Zhen Zhang, Xuehai He, Weixiang Yan, Ao Shen, Chenyang Zhao, Shuohang Wang, Yelong Shen, and Xin Eric Wang. Soft thinking: Unlocking the reasoning potential of llms in continuous concept space. arXiv preprint arXiv:2505.15778, 2025. 2

    Pith/arXiv arXiv 2025

  42. [42]

    Beyond imitation: Reinforcement learning for active latent planning, 2026

    Zhi Zheng and Wee Sun Lee. Beyond imitation: Reinforcement learning for active latent planning, 2026. 2

    2026

  43. [43]

    Monte carlo tree search for comprehensive exploration in llm-based automatic heuristic design.arXiv preprint arXiv:2501.08603, 2025

    Zhi Zheng, Zhuoliang Xie, Zhenkun Wang, and Bryan Hooi. Monte carlo tree search for comprehensive exploration in llm-based automatic heuristic design.arXiv preprint arXiv:2501.08603, 2025. 2 14 EndoCoT : Scaling Endogenous Chain-of-Thought Reasoning in Diffusion Models Appendix A. Overview In the supplementary material, Sec. B details the dataset statisti...

    Pith/arXiv arXiv 2025