pith. sign in

arxiv: 2606.02842 · v1 · pith:OMKL5NVSnew · submitted 2026-06-01 · 💻 cs.LG

Spectral-Progressive Thought Flow for Lightweight Multimodal Reasoning

Yixian Shen , Zhiheng Yang , Qi Bi , Changshuo Wang , Shuai Wang , Jia-Hong Huang , George Floros , Prayag Tiwari
show 1 more author
Anuj Pathania
This is my paper

Pith reviewed 2026-06-28 15:38 UTC · model grok-4.3

classification 💻 cs.LG
keywords multimodal reasoningdiscrete cosine transformclassifier-free guidancespatial reasoninglightweight inferencevisual workspacethought flowspectral representation
0
0 comments X

The pith

SpecFlow encodes visual thoughts in fixed-size discrete cosine space so multimodal spatial reasoning stays bounded in memory and latency no matter the chain length.

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

The paper proposes SpecFlow to cut the growing cost of long chains of visual and textual thoughts in multimodal spatial reasoning. It stores each intermediate visual state in a compact frequency representation that keeps global layout and relations intact through energy compaction. Classifier-free guidance lets the running textual trace steer updates to this fixed workspace without ever expanding the context window. Because the visual state size never grows, both compute and memory stay constant across arbitrary reasoning depth. Experiments report matching or higher accuracy on spatial tasks at up to 2.1 times lower KV cache and computation cost.

Core claim

SpecFlow maintains a bounded visual workspace whose updates depend only on the current visual state and accumulated textual trace, enabling long-horizon inference with stable latency and memory usage independent of reasoning depth.

What carries the argument

Fixed-size discrete cosine transform representation of visual thoughts, steered by classifier-free guidance from textual thoughts.

If this is right

  • Reasoning depth can increase without quadratic growth in attention or cache costs.
  • High-frequency visual details are introduced only when spatial precision requires them.
  • Textual generation can direct visual state evolution without enlarging shared context.
  • Overall computation and KV cache usage drop by up to 2.1 times while accuracy remains competitive.

Where Pith is reading between the lines

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

  • The same fixed spectral buffer could be tested on sequential video or audio reasoning chains.
  • A direct comparison on puzzles whose solution depth exceeds the current benchmark maxima would expose any hidden limits of the fixed representation.
  • Deployed agents running extended multimodal dialogues would see reduced per-step energy draw.

Load-bearing premise

A fixed-size frequency encoding plus guidance is sufficient to preserve all layout and relational structure needed for the target tasks without losing critical details.

What would settle it

Compare SpecFlow against expanding-context baselines on a spatial reasoning benchmark whose hardest items require more than twenty chained visual-text steps; if accuracy stays competitive while measured KV cache and latency remain flat, the central claim holds.

Figures

Figures reproduced from arXiv: 2606.02842 by Anuj Pathania, Changshuo Wang, George Floros, Jia-Hong Huang, Prayag Tiwari, Qi Bi, Shuai Wang, Yixian Shen, Zhiheng Yang.

Figure 1
Figure 1. Figure 1: Comparison of multimodal spatial reasoning paradigms. (a) Text-only CoT lacks spatial grounding; (b) Image–Text Co￾Thought improves grounding but causes context and KV-cache growth; (c) SpecFlow updates a compact spectral visual state, enabling efficient multi-hop reasoning with stable memory. sual tokens, which allows the visual tokens ground linguistic reasoning in spatial structure and provides essentia… view at source ↗
Figure 2
Figure 2. Figure 2: Spectral-Progressive Thought Flow (SpecFlow) alternates autoregressive text thoughts with text-conditioned, flow-based updates of a continuous visual state. Visual states are overwritten at each hop and represented in the cosine domain with progressively activated frequency bands, enabling efficient multimodal spatial reasoning without accumulating visual tokens or growing the context length. are structure… view at source ↗
Figure 3
Figure 3. Figure 3: Spectral-progressive frequency allocation with block cosine projection. (a) An intermediate visual state. (b) Partition into b × b blocks and employ block cosine projection. (c) Average coefficient energy concentrates in low frequencies. (d) Reconstruction using only the low-frequency bands preserves global layout. (e) Retaining a small subset of coefficients yields significant visual-token reduction per h… view at source ↗
Figure 4
Figure 4. Figure 4: KV-cache memory across benchmarks. SpecFlow achieves 1.6×–2.1× KV-cache reduction over baselines. 0 2 4 8 10 CFG scale w 0.00 0.25 0.50 0.75 1.00 Accuracy (a) CFG scale sweep on Maze Maze-4 Maze-8 Maze-12 Maze-16 0 2 4 8 10 CFG scale w 0.00 0.25 0.50 0.75 1.00 Accuracy (b) CFG scale sweep on MiniBehavior MB-8 MB-12 MB-16 MB-20 [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Effect of CFG guidance scale on reasoning accuracy, Performance improves with increasing guidance strength and peaks at a moderate scale (w = 4), while excessively large guidance yields diminishing returns due to over-deterministic conditioning. Baselines. We compare SpecFlow against baselines span￾ning prompt-based, heuristic, and latent-space token￾compression paradigms, including SoT (Aytes et al., 2025… view at source ↗
Figure 6
Figure 6. Figure 6: Effect of the number of ODE inference steps T on rea￾soning accuracy. Marker size and color encode inference latency. Accuracy improves with increasing T and saturates at moderate step counts, while larger T leads to diminishing returns and higher computational cost [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Effect of spectral block size b on reasoning accuracy and KV-cache usage. We vary the block size used for block-wise cosine projection and report task accuracy (left y-axis) together with the total KV cache footprint (right y-axis) on (a) Winoground and (b) FrozenLake. Smaller blocks preserve finer local structure and yield higher accuracy, but increase the KV cache due to higher token granularity. Larger … view at source ↗
Figure 8
Figure 8. Figure 8: Success case of multi-hop Maze planning with SpecFlow. Each panel shows the intermediate visual workspace at one hop, overlaid with the current planned trajectory in red. The route is progressively extended while preserving global Maze geometry, and the final hop yields a coherent collision-free path that reaches the goal [PITH_FULL_IMAGE:figures/full_fig_p022_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Failure case under insufficient text alignment (under-aligned CFG). In SpecFlow, the CFG scale w controls how strongly textual guidance constrains the flow update of the visual workspace. When the model is under-aligned, the workspace update is not sufficiently anchored by the intended route and constraints, and the generator may drift toward visually plausible but incorrect structures, producing spurious … view at source ↗
Figure 10
Figure 10. Figure 10: Qualitative MiniBehavior example: fetch then place. We visualize multi-hop visual thoughts for a two-stage manipulation task. The red rectangle denotes the current subgoal region encoded in the workspace, which first targets the printer for pickup and then switches to the table for placement. The sequence illustrates how SpecFlow overwrites a compact workspace to track subgoals and progress, enabling mult… view at source ↗
Figure 11
Figure 11. Figure 11: FrozenLake qualitative example with multi-hop visual thoughts. Top-left shows the initial environment. The remaining panels show SpecFlow’s decoded visual workspaces across hops under the hop-wise controller prompt. The sequence progressively refines a collision-free route that avoids holes and reaches the goal, demonstrating bounded-workspace long-horizon reasoning. bounded workspace, rather than accumul… view at source ↗
read the original abstract

Multimodal spatial reasoning often relies on long chains of intermediate textual and visual thoughts, where accumulating visual tokens and dense cross-modal attention incur substantial computation and memory overhead. To address this challenge, we propose Spectral-Progressive Thought Flow (SpecFlow), a novel lightweight multimodal spatial reasoning framework that represents intermediate visual thoughts in a fixed-size discrete cosine space. By exploiting strong energy compaction, SpecFlow preserves global layout and relational structure while introducing high-frequency details only when increased spatial precision is required. To align visual state evolution with linguistic intent, classifier-free guidance enables autoregressive textual thoughts to steer flow-based updates of the visual workspace/state without expanding the context. As a result, SpecFlow maintains a bounded visual workspace whose updates depend only on the current visual state and accumulated textual trace, enabling long-horizon inference with stable latency and memory usage independent of reasoning depth. Empirical results show that SpecFlow achieves competitive or superior reasoning performance while reducing computation and KV cache costs by up to 2.1 times.

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.

Referee Report

1 major / 0 minor

Summary. The paper proposes Spectral-Progressive Thought Flow (SpecFlow), a lightweight multimodal spatial reasoning framework that encodes intermediate visual thoughts in a fixed-size discrete cosine transform (DCT) space to preserve global layout while adding high-frequency details progressively. Classifier-free guidance aligns autoregressive textual thoughts with visual state updates without context expansion. The central claim is that this yields a bounded visual workspace whose updates depend only on the current visual state and accumulated textual trace, enabling long-horizon inference with latency and memory usage independent of reasoning depth, plus up to 2.1× reductions in computation and KV-cache costs while maintaining competitive reasoning performance.

Significance. If the bounded-workspace and depth-independent memory claims hold with the stated mechanisms, the work would offer a practical route to scalable long-horizon multimodal reasoning by avoiding the quadratic costs of accumulating visual tokens and dense cross-modal attention. The use of energy-compaction properties of DCT and classifier-free guidance for steering without expansion are potentially reusable ideas, but the absence of any derivation or experimental verification of the independence property limits the assessed impact at present.

major comments (1)
  1. [Abstract] Abstract: the claim that 'SpecFlow maintains a bounded visual workspace whose updates depend only on the current visual state and accumulated textual trace, enabling long-horizon inference with stable latency and memory usage independent of reasoning depth' is not supported by the described components. The visual workspace is stated to be fixed-size via DCT, but the textual trace is described as accumulated across autoregressive steps; standard transformer KV-cache implementations cause linear growth in context length (and thus memory) with reasoning depth. No mechanism for bounding, summarizing, windowing, or compressing the textual trace is referenced, so the independence property does not follow from the stated dependencies.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive feedback on the bounded-workspace claim. We address the concern point by point below and will revise the manuscript accordingly.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the claim that 'SpecFlow maintains a bounded visual workspace whose updates depend only on the current visual state and accumulated textual trace, enabling long-horizon inference with stable latency and memory usage independent of reasoning depth' is not supported by the described components. The visual workspace is stated to be fixed-size via DCT, but the textual trace is described as accumulated across autoregressive steps; standard transformer KV-cache implementations cause linear growth in context length (and thus memory) with reasoning depth. No mechanism for bounding, summarizing, windowing, or compressing the textual trace is referenced, so the independence property does not follow from the stated dependencies.

    Authors: We agree that the current manuscript text does not explicitly describe a mechanism (such as summarization, fixed-window caching, or compression) that would bound the textual trace's KV cache growth under standard autoregressive transformer implementations. The abstract's phrasing that classifier-free guidance enables steering 'without expanding the context' was intended to indicate that visual-state updates do not require cross-modal attention over the full textual history, but this does not automatically bound the text model's own context. We will revise the abstract and add a dedicated paragraph in Section 3 (or a new subsection) clarifying the precise implementation of classifier-free guidance, whether the textual trace is maintained only as a compact conditioning vector for the guidance scale, and any practical limits on textual context length. This will either substantiate the independence claim with additional implementation details or qualify the claim to apply primarily to the visual workspace and cross-modal costs. revision: yes

Circularity Check

0 steps flagged

No circularity; derivation self-contained

full rationale

The paper presents SpecFlow via new mechanisms (fixed-size DCT visual workspace, classifier-free guidance for textual steering) whose claimed bounded memory and depth-independent latency are asserted to follow from those design choices. No equations, fitted parameters, or self-citations are shown that would make the independence property reduce by construction to the inputs or to prior author work. The textual trace is described as an input to the update rule rather than being redefined as the output; the skeptic concern addresses empirical validity of the claim, not circularity in the derivation chain itself.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The approach rests on the standard mathematical property that DCT provides strong energy compaction for natural images; no free parameters, ad-hoc axioms, or invented entities are named in the abstract.

axioms (1)
  • domain assumption DCT exhibits strong energy compaction that preserves global layout and relational structure in visual data
    Invoked to justify fixed-size representation without loss of necessary information for reasoning.

pith-pipeline@v0.9.1-grok · 5726 in / 1146 out tokens · 20236 ms · 2026-06-28T15:38:14.629536+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

30 extracted references · 12 linked inside Pith

  1. [1]

    Albergo, M. S. and Vanden-Eijnden, E. Building normal- izing flows with stochastic interpolants.arXiv preprint arXiv:2209.15571,

    Pith/arXiv arXiv

  2. [2]

    A., Baek, J., and Hwang, S

    Aytes, S. A., Baek, J., and Hwang, S. J. Sketch-of-thought: Efficient LLM reasoning with adaptive cognitive-inspired sketching.arXiv preprint arXiv:2503.05179,

    arXiv

  3. [3]

    Qwen technical report.arXiv preprint arXiv:2309.16609,

    Bai, J., Bai, S., Chu, Y ., Cui, Z., Dang, K., Deng, X., Fan, Y ., Ge, W., Han, Y ., Huang, F., et al. Qwen technical report.arXiv preprint arXiv:2309.16609,

    Pith/arXiv arXiv

  4. [4]

    SpatialRGPT: Grounded spatial reasoning in vision language models.arXiv preprint arXiv:2406.01584,

    Cheng, A.-C., Yin, H., Fu, Y ., Guo, Q., Yang, R., Kautz, J., Wang, X., and Liu, S. SpatialRGPT: Grounded spatial reasoning in vision language models.arXiv preprint arXiv:2406.01584,

    arXiv

  5. [5]

    and Van Durme, B

    Cheng, J. and Van Durme, B. Compressed chain of thought: Efficient reasoning through dense representations.arXiv preprint arXiv:2412.13171,

    Pith/arXiv arXiv

  6. [6]

    From explicit CoT to implicit CoT: Learning to internalize CoT step by step

    Deng, Y ., Choi, Y ., and Shieber, S. From explicit CoT to implicit CoT: Learning to internalize CoT step by step. arXiv preprint arXiv:2405.14838,

    Pith/arXiv arXiv

  7. [7]

    Embspatial- bench: Benchmarking spatial understanding for embod- ied tasks with large vision-language models.arXiv preprint arXiv:2406.05756,

    Du, M., Wu, B., Li, Z., Huang, X., and Wei, Z. Embspatial- bench: Benchmarking spatial understanding for embod- ied tasks with large vision-language models.arXiv preprint arXiv:2406.05756,

    arXiv

  8. [8]

    Loca: Location-aware cosine adapta- tion for parameter-efficient fine-tuning.arXiv preprint arXiv:2502.06820,

    Du, Z., Min, Y ., Li, J., Lu, K., Zou, C., Peng, L., Chu, T., and Gong, M. Loca: Location-aware cosine adapta- tion for parameter-efficient fine-tuning.arXiv preprint arXiv:2502.06820,

    arXiv

  9. [9]

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

    Hao, S., Sukhbaatar, S., Su, D., Li, X., Hu, Z., Weston, J., and Tian, Y . Training large language models to reason in a continuous latent space.arXiv preprint arXiv:2412.06769,

    Pith/arXiv arXiv

  10. [10]

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

    He, Z., Qu, X., Li, Y ., Zhu, T., Huang, S., and Cheng, Y . DiffThinker: Towards generative multimodal reasoning with diffusion models.arXiv preprint arXiv:2512.24165,

    arXiv

  11. [11]

    I., Shah, R., Spies, A

    Ivanitskiy, M. I., Shah, R., Spies, A. F., R¨auker, T., Valen- tine, D., Rager, C., Quirke, L., Mathwin, C., Corlouer, G., Behn, C. D., et al. A configurable library for gen- erating and manipulating maze datasets.arXiv preprint arXiv:2309.10498,

    arXiv

  12. [12]

    Mini-BEHA VIOR: A procedurally generated benchmark for long-horizon decision-making in embodied AI.arXiv preprint arXiv:2310.01824,

    Jin, E., Hu, J., Huang, Z., Zhang, R., Wu, J., Fei-Fei, L., and Mart´ın-Mart´ın, R. Mini-BEHA VIOR: A procedurally generated benchmark for long-horizon decision-making in embodied AI.arXiv preprint arXiv:2310.01824,

    arXiv

  13. [13]

    Gradient weight-normalized low-rank projection for efficient llm training

    Kanoulas, E., HUANG, J.-H., Shen, Y ., Zhu, H., and Rud- inac, S. Gradient weight-normalized low-rank projection for efficient llm training. InGreeks in AI Symposium 2025,

    2025

  14. [14]

    Imagine while reasoning in space: Multimodal visualization-of-thought.Forty-Second Inter- national Conference on Machine Learning, 2025a

    Li, C., Wu, W., Zhang, H., Xia, Y ., Mao, S., Dong, L., Vuli´c, I., and Wei, F. Imagine while reasoning in space: Multimodal visualization-of-thought.Forty-Second Inter- national Conference on Machine Learning, 2025a. Li, J., Zheng, S., Shen, Y ., Huang, J.-H., Lu, X., Ni, M., and Guan, Y . Keeping the evidence chain: Semantic evi- dence allocation for tr...

    Pith/arXiv arXiv

  15. [15]

    Enhancing advanced visual reasoning ability of large language models.Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing,

    Li, Z., Liu, D., Zhang, C., Wang, H., Xue, T., and Cai, W. Enhancing advanced visual reasoning ability of large language models.Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing,

    2024

  16. [16]

    VoCoT: Unleashing visually grounded multi-step reasoning in large multi-modal models

    Li, Z., Luo, R., Zhang, J., Qiu, M., Huang, X., and Wei, Z. VoCoT: Unleashing visually grounded multi-step reasoning in large multi-modal models. InProceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), 2025b. Lipman, Y ., Chen, R. T.,...

    Pith/arXiv arXiv 2025

  17. [17]

    Flow straight and fast: Learning to generate and transfer data with rectified flow

    Liu, X., Gong, C., and Liu, Q. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003,

    Pith/arXiv arXiv

  18. [18]

    C., and Santoro, A

    Raposo, D., Ritter, S., Richards, B., Lillicrap, T., Humphreys, P. C., and Santoro, A. Mixture-of-depths: Dynamically allocating compute in transformer-based lan- guage models.arXiv preprint arXiv:2404.02258,

    Pith/arXiv arXiv

  19. [19]

    Efficient reasoning with hidden thinking.arXiv preprint arXiv:2501.19201, 2025a

    10 SpecFlow for Multimodal Reasoning Shen, X., Wang, Y ., Shi, X., Wang, Y ., Zhao, P., and Gu, J. Efficient reasoning with hidden thinking.arXiv preprint arXiv:2501.19201, 2025a. Shen, Y ., Bi, Q., Huang, J.-H., Zhu, H., Pimentel, A. D., and Pathania, A. Macp: Minimal yet mighty adaptation via hierarchical cosine projection. InProceedings of the 63rd Ann...

    Pith/arXiv arXiv 2025

  20. [20]

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

    Shen, Z., Yan, H., Zhang, L., Hu, Z., Du, Y ., and He, Y . Codi: Compressing chain-of-thought into continuous space via self-distillation.arXiv preprint arXiv:2502.21074, 2025d. Sheng, G., Zhang, C., Ye, Z., Wu, X., Zhang, W., Zhang, R., Peng, Y ., Lin, H., and Wu, C. HybridFlow: A flexible and efficient rlhf framework. InProceedings of the Twentieth Euro...

    Pith/arXiv arXiv

  21. [21]

    Token assorted: Mixing latent and text tokens for improved language model reasoning.arXiv preprint arXiv:2502.03275,

    Su, D., Zhu, H., Xu, Y ., Jiao, J., Tian, Y ., and Zheng, Q. Token assorted: Mixing latent and text tokens for improved language model reasoning.arXiv preprint arXiv:2502.03275,

    arXiv

  22. [22]

    TokenCarve: Information-preserving visual token compression in multimodal large language models.arXiv preprint arXiv:2503.10501,

    Tan, X., Ye, P., Tu, C., Cao, J., Yang, Y ., Zhang, L., Zhou, D., and Chen, T. TokenCarve: Information-preserving visual token compression in multimodal large language models.arXiv preprint arXiv:2503.10501,

    arXiv

  23. [23]

    FlashSloth: Lightning multimodal large language models via embedded visual compression

    Tong, B., Lai, B., Zhou, Y ., Luo, G., Shen, Y ., Li, K., Sun, X., and Ji, R. FlashSloth: Lightning multimodal large language models via embedded visual compression. arXiv preprint arXiv:2412.04317,

    arXiv

  24. [24]

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

    Wu, C., Li, J., Zhou, J., Lin, J., Gao, K., Yan, K., Yin, S.-m., Bai, S., Xu, X., Chen, Y ., et al. Qwen-image technical report.arXiv preprint arXiv:2508.02324, 2025a. Wu, H., Teng, Z., and Tu, K. Parallel continuous chain-of-thought with jacobi iteration.arXiv preprint arXiv:2506.18582, 2025b. Wu, P. and Xie, S. V?: Guided visual search as a core mechani...

    Pith/arXiv arXiv

  25. [25]

    Y ., Zhang, Y ., and Chang, S

    Wu, Q., Zhao, H., Saxon, M., Bui, T., Wang, W. Y ., Zhang, Y ., and Chang, S. VSP: Assessing the dual challenges of perception and reasoning in spatial planning tasks for VLMs.arXiv preprint arXiv:2407.01863, 2024a. Wu, W., Mao, S., Zhang, Y ., Xia, Y ., Dong, L., Cui, L., and Wei, F. Mind’s Eye of LLMs: Visualization-of-thought elicits spatial reasoning ...

    arXiv

  26. [26]

    Eligen: Entity-level controlled image generation with regional attention

    11 SpecFlow for Multimodal Reasoning Zhang, H., Duan, Z., Wang, X., Chen, Y ., and Zhang, Y . Eligen: Entity-level controlled image generation with regional attention. InProceedings of the 7th ACM In- ternational Conference on Multimedia in Asia, pp. 1–7, 2025a. Zhang, J., Zhu, Y ., Sun, M., Luo, Y ., Qiao, S., Du, L., Zheng, D., Chen, H., and Zhang, N. L...

    arXiv

  27. [27]

    Image-of-thought prompting for visual reasoning re- finement in multimodal large language models.arXiv preprint arXiv:2405.13872,

    Zhou, Q., Zhou, R., Hu, Z., Lu, P., Gao, S., and Zhang, Y . Image-of-thought prompting for visual reasoning re- finement in multimodal large language models.arXiv preprint arXiv:2405.13872,

    arXiv

  28. [28]

    Dynamic Spatial Reasoning Tasks

    Visual search – 238 No extra prompting; use benchmark as-is A.2. Dynamic Spatial Reasoning Tasks. We evaluateSpecFlowon three dynamic spatial reasoning benchmarks, Maze (Ivanitskiy et al., 2023), MiniBehavior (Jin et al., 2023), and FrozenLake (Wu et al., 2024a), all of which require multi-step reasoning over simulated grid environments. These tasks stres...

    2023

  29. [29]

    This formulation avoids stochastic sampling and enables stable, low-variance generation suitable for multi-step reasoning

    Epochs 5 5 1 Learning rate1×10 −4 1×10 −4 1×10 −6 LoRA rank 32 32 – Batch size 4 16 64 (8B) / 32 (32B) Rollout size (n) – – 4 KL coefficient – –1×10 −2 is performed by solving the resulting ordinary differential equation with a fixed-step solver, yielding deterministic visual- thought trajectories with predictable computational cost. This formulation avoi...

    2025

  30. [30]

    Backbone initialization Qwen-Image-Edit-2509 (Esser et al., 2024; Wu et al., 2025a) & Qwen3-VL-8B (Bai et al.,

    2024