Pith. sign in

REVIEW 3 major objections 5 minor 29 references

Bridging the Dynamic Perception Gap: Training-Free Draft Chain-of-Thought for Dynamic Multimodal Spatial Reasoning

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper claims that adding visual drafts of the model's current state to textual chains of thought improves dynamic spatial reasoning in multimodal models without fine-tuning.

desk verdict Useful benchmark and a plausible training-free method, but the 'consistent gains' claim overshoots the data and the missing text-only iterative baseline leaves the key mechanism unproven. read the letter →

arxiv 2505.16579 v1 pith:M7YL5Y54 submitted 2025-05-22 cs.AI cs.CV

Siqu Ou , Hongcheng Liu , Pingjie Wang , Yusheng Liao , Chuan Xuan , Yanfeng Wang , Yu Wang This is my paper
classification cs.AIcs.CV
keywords dynamicspatialreasoningchain-of-thoughtmultimodallargelanguagemodelsvisualdraftstraining-freeframeworkmazenavigationvideo
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper sets out to show that multimodal language models reason about changing scenes far better when the textual chain of thought is paired with visual drafts drawn directly on the video frames, marking where the agent is and what the next move will hit. It introduces GRASSLAND, a grid-world maze benchmark with moving lava and stationary water, and two tasks: judging whether a given action sequence ends safely, and planning a safe route to a goal. Across three MLLMs, the paper reports that this Draft CoT approach beats text-only chain-of-thought and static visual prompting, and that D2R, a training-free framework that uses external tools to produce the drafts, reproduces most of the gain without fine-tuning. The result matters because it suggests a cheap, model-agnostic way to externalize a model's intermediate spatial state rather than forcing the model to keep it all in text.

What carries the argument

The load-bearing object is the Draft Chain-of-Thought produced by D2R's iterative loop: a plan, a sequence of augmented perceptual thoughts in which each textual step is fused with a visual draft drawn onto the input image, and a final answer read out after the loop stops. The mechanism that carries the argument is the external toolchain, specifically the Dynamic-Information-Extract tool for reading the agent and environment state from the frames and the Position-Draw tool for overlaying it, together with the scheduling hub that decides when to stop and ask for the final answer.

What would settle it

Run D2R on the hard maze-judgment set while logging each drawn overlay against the true agent and lava coordinates; if accuracy does not rise when all overlays are correct, or if correct overlays still fail as often as incorrect ones, the claim that the draft itself carries the benefit would be falsified.

Watch

Extended reading notes

Core claim

The paper's central claim is that the limiting factor in dynamic spatial reasoning is not the model's textual logic but its access to its own evolving spatial state, and that drawing that state on the input frames fixes the gap. Concretely, for each step of a maze task, D2R has an external scheduling hub decide which tool to call, asks the MLLM for its next action, then uses a Dynamic-Information-Extract tool and a Position-Draw tool to overlay a marker of the current position and the consequence of the action on the current video frame; the augmented frame and text are fed back into the MLLM. On GRASSLAND, this bimodal chain of thought outperforms text-only CoT, one-shot CoT, and static visual prompting, and the ablation shows that removing the visual draft costs more than removing the textual thought.

Load-bearing premise

Everything rests on the external tools correctly extracting the agent's location and the moving hazards from the video and drawing them accurately on the frame; the paper does not report the error rate of that extraction step.

Editorial extensions

If this is right

  • Text-only chain-of-thought is not enough for dynamic spatial reasoning; the model's current state needs to be made visible on the frame.
  • A training-free augmentation can narrow most of the gap to a version that uses ground-truth drafts, at least on the hard maze-judgment task.
  • The benefit is larger for stronger models, so D2R externalizes reasoning but does not create reasoning capacity where none exists.
  • The method transfers across MLLMs with different baselines and only requires image-and-text input, not image-generation ability.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A natural untested extension is to apply the same draft-augmented loop to other state-tracking domains, such as traffic scenes, sports plays, or robot navigation, where position can be overlaid on video; nothing in the method is maze-specific.
  • The ablation suggests the visual draft, not the textual rationale, carries most of the gain; this implies that a version with oracle overlays and no text might perform nearly as well, which the paper does not run.
  • Because D2R does not train the model, its ceiling is set by the toolchain's extraction accuracy; measuring that error and its propagation through the loop would predict where the method breaks.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper introduces GRASSLAND, a dynamic maze benchmark with two tasks (Maze Judgment and Maze Navigation), and proposes D2R, a training-free framework that combines textual chain-of-thought with visual drafts—markers and routes overlaid on the input video frames—generated by external tools and orchestrated by an LLM scheduler. The authors report experiments on three MLLMs, claiming that D2R significantly and consistently outperforms direct prompting, text-only CoT, 1-shot CoT, and static visual prompting (VAP), with ablations suggesting that both the textual thoughts and the visual drafts contribute to the gains. The central claim is that dynamic visual drafts, rather than text alone, enable better dynamic spatial reasoning in MLLMs.

Significance. If the causal role of the visual draft is substantiated, D2R would be a simple, training-free prompt augmentation with broad applicability to multimodal spatial reasoning. The paper's strengths include a new benchmark for dynamic spatial reasoning, an open-source implementation, a multi-model evaluation, and a training-free method with no parameters fitted to the benchmark. The reported ablation (Table 4) showing a large drop when drafts are removed is evidence that the overlay matters. However, the significance is currently limited by two gaps: the claimed 'consistent' and 'significant' improvements are not backed by statistical testing, and the ablations do not isolate the visual draft from the iterative position-feedback mechanism. If these are addressed, the paper could make a useful contribution.

major comments (3)
  1. [Section 5.2, Table 3] The claim of consistent and significant improvement is not supported by the reported numbers. In Table 3, Qwen2.5VL-7B on hard Maze Judgment scores 28.0 with D2R versus 28.5 with Direct prompting, and Qwen2.5VL-72B on easy Maze Judgment scores 67.0 with D2R versus 71.0 with 1-shot CoT, so D2R does not consistently win. No error bars, confidence intervals, or significance tests are reported anywhere, and the number of evaluation items per cell is not provided, making it impossible to judge whether differences such as 28.0 vs 28.5 are meaningful. The abstract's 'significantly outperforms' and 'consistently enhances' are therefore not backed by the evidence as presented. Please report item counts, variance across seeds or resamples, and significance tests, and revise the wording to match the actual support.
  2. [Section 4.2, Table 4] The causal attribution of the gain to the visual draft is confounded. D2R differs from all baselines in at least three simultaneous dimensions: (i) it iterates with multiple MLLM calls rather than a single pass, (ii) it receives current-position information extracted and drawn by external tools (Position-Draw and Dynamic-Information-Extract), and (iii) it sees the visual draft overlay. The 'w/o Drafts' row in Table 4 removes the overlay but retains iterative decomposition and whatever positional information survives in the textual chain; it is not a matched text-only iterative baseline because it does not feed the same positional state as explicit text. Consequently, the performance drop in that row could be due to the removal of explicit position feedback rather than to the loss of the visual modality. The Draft CoT(GT) comparison in Table 5 is shown only for the hard Maze Judgment condition and uses ground-truth paths, so it does not isolate the visual contribution either. Please add a control that supplies the current position as text (e.g., 'Your position is (x,y)') in the same iterative protocol, without any overlay, to separate the visual-draft effect from the iterative position-feedback effect.
  3. [Section 4.1, Algorithm 1] The reliability of the toolchain is unquantified. D2R's correctness depends on Dynamic-Information-Extract and Position-Draw accurately identifying the agent's position and the dynamic environment state at every iteration, and on the MLLM correctly interpreting the black-square marker as the current position. The paper never reports the error rate of these extraction and drawing tools, nor any analysis of cases where a wrong overlay might mislead the MLLM. If these tools are error-prone, the method could produce systematic errors that are not visible in aggregate accuracy. At minimum, please report tool success rates on GRASSLAND and, ideally, an ablation comparing D2R with ground-truth overlays against D2R with predicted overlays to separate tool errors from reasoning errors.
minor comments (5)
  1. [Table 7] In the normal-level row for Qwen2.5VL-7B, the cell '325.3' appears to be a typo; percentage accuracies should lie between 0 and 100.
  2. [Figure 4] The legend and axis labels in Figure 4 are corrupted with '/uni000000...' sequences and are not legible; please replace with readable text.
  3. [Section 5.3] The heading 'Draft DoT(GT)' should read 'Draft CoT(GT)'.
  4. [Equations (6)-(7)] Equation (7) uses 'C_all' but only 'C_{<n}' is defined in Equation (6); please clarify the notation.
  5. [Section 5.1] The paper does not report the number of evaluation instances per difficulty level for GRASSLAND; please add this information to support the percentage accuracies.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: D2R is a training-free heuristic evaluated on a new benchmark with explicit ablations; oracle conditions are labeled as ground truth and no fitted parameter is renamed as a prediction.

full rationale

GRASSLAND is a newly generated synthetic benchmark, and D2R is a training-free prompting/tool pipeline; no model parameter is fitted to the benchmark and no reported accuracy is obtained by optimizing on test labels. The paper's own ablations (Table 4) remove textual thought and drafts separately, providing independent evidence for the visual-draft contribution. The closest potentially circular element is Draft CoT(GT), where ground-truth routes are used to draw overlays; however, the paper explicitly labels this condition as GT and uses it as an oracle/sanity check rather than as the claimed training-free method, and the main D2R results do not consume test labels. Self-citations in the related-work section (Liu et al. 2024a,b) are illustrative and not load-bearing. The absence of a matched iterative text-only baseline and the privileged Position-Draw tool are experimental-validity concerns (potential confounds), not definitional reductions: the final answer is not identical by construction to the tool output, and the paper does not rename a fitted parameter as a prediction. Therefore no circular step meeting the required evidentiary standard is identifiable.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

No free parameters are fitted in D2R; the method uses an external LLM scheduler and drawing tools, with hyperparameters (temperature 0.1, max tokens) set conventionally. The central assumptions are about the reliability of the visual-draft toolchain and the validity of the synthetic benchmark as a measure of dynamic spatial reasoning. No new physical entities are introduced.

assumptions (4)
  • domain assumption MLLMs can reliably interpret overlaid visual markers (e.g., a black square) as the agent's current position.
    The iteration prompt (Table 13) instructs the model to treat the black square as the current position; if the model misreads it, D2R's gains would vanish.
  • domain assumption The Dynamic-Information-Extract tool can accurately determine the agent's position and dynamic environment state from video frames at each time step.
    Section 4.1 introduces tools for visual editing and position drawing; Algorithm 1 relies on them to produce drafts. If extraction is noisy, the drafts will mislead the model.
  • domain assumption GRASSLAND is a valid proxy for dynamic spatial reasoning in real-world tasks.
    The benchmark is a synthetic pixel grid world; its relationship to real dynamic spatial reasoning is assumed, not demonstrated.
  • domain assumption The 'player moves first' rule for simultaneous lava collisions is clear to the MLLM from the prompt.
    The prompt in Table 9 defines this rule; the benchmark's state transitions (Eq. 1-2) depend on it.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Bridging the Dynamic Perception Gap: Training-Free Draft Chain-of-Thought for Dynamic Multimodal Spatial Reasoning." pith.science (2026). https://pith.science/paper/M7YL5Y54

@misc{pith2026250516579,
  author       = {Pith},
  title        = {Pith review of: Bridging the Dynamic Perception Gap: Training-Free Draft Chain-of-Thought for Dynamic Multimodal Spatial Reasoning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/M7YL5Y54}},
  note         = {Machine review of arXiv:2505.16579}
}
read the original abstract

While chains-of-thought (CoT) have advanced complex reasoning in multimodal large language models (MLLMs), existing methods remain confined to text or static visual domains, often faltering in dynamic spatial reasoning tasks. To bridge this gap, we present GRASSLAND, a novel maze navigation benchmark designed to evaluate dynamic spatial reasoning. Our experiments show that augmenting textual reasoning chains with dynamic visual drafts, overlaid on input images, significantly outperforms conventional approaches, offering new insights into spatial reasoning in evolving environments. To generalize this capability, we propose D2R (Dynamic Draft-Augmented Reasoning), a training-free framework that seamlessly integrates textual CoT with corresponding visual drafts into MLLMs. Extensive evaluations demonstrate that D2R consistently enhances performance across diverse tasks, establishing a robust baseline for dynamic spatial reasoning without requiring model fine-tuning. Project is open at https://github.com/Cratileo/D2R.

Figures

Figures reproduced from arXiv: 2505.16579 by the authors.

Figure 1
Figure 1. The demonstration of the Draft CoT with D2R. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Illustration of the difference between our method and others. Direct prompting and language-centric CoT [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Example of dynamic scenario sequence in GRASSLAND. The left part is the illustration of the dynamic [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (5 more)
Figure 5
Figure 5. Figure 5: Average accuracy of models for each choice [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 4
Figure 4. Figure 4: Accuracy with different models and methods [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 6
Figure 6. Figure 6: Illustration of D2R reasoning process. After the schedule hub initialization, the process consists of [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: An example of the thought process for D2R in maze judgment task. [PITH_FULL_IMAGE:figures/full_fig_p019_7.png]
Figure 8
Figure 8. Figure 8: An example of the thought process for D2R in maze navigation task. [PITH_FULL_IMAGE:figures/full_fig_p019_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

29 extracted references · 3 canonical work pages

  1. [1]

    Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. 2023. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966

    arXiv 2023

  2. [2]

    Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, Humen Zhong, Yuanzhi Zhu, Mingkun Yang, Zhaohai Li, Jianqiang Wan, Pengfei Wang, Wei Ding, Zheren Fu, Yiheng Xu, and 8 others. 2025. Qwen2.5-vl technical report. arXiv preprint arXiv:2502.13923

    arXiv 2025

  3. [3]

    Qiguang Chen, Libo Qin, Jin Zhang, Zhi Chen, Xiao Xu, and Wanxiang Che. 2024 a . https://arxiv.org/abs/2405.16473 M ^3 cot: A novel benchmark for multi-domain multi-step multi-modal chain-of-thought . Preprint, arXiv:2405.16473

    arXiv 2024

  4. [4]

    Zhe Chen, Weiyun Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Erfei Cui, Jinguo Zhu, Shenglong Ye, Hao Tian, Zhaoyang Liu, and 1 others. 2024 b . Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling. arXiv preprint arXiv:2412.05271

    arXiv 2024

  5. [5]

    Jun Gao, Yongqi Li, Ziqiang Cao, and Wenjie Li. 2025. https://arxiv.org/abs/2411.19488 Interleaved-modal chain-of-thought . Preprint, arXiv:2411.19488

    arXiv 2025

  6. [6]

    The Abduction of Sherlock Holmes: A Dataset for Visual Abductive Reasoning

    Jack Hessel, Jena D. Hwang, Jae Sung Park, Rowan Zellers, Chandra Bhagavatula, Anna Rohrbach, Kate Saenko, and Yejin Choi. 2022. https://arxiv.org/abs/2202.04800 The abduction of sherlock holmes: A dataset for visual abductive reasoning . Preprint, arXiv:2202.04800

    work page Pith review arXiv 2022

  7. [7]

    Dongzhi Jiang, Renrui Zhang, Ziyu Guo, Yanwei Li, Yu Qi, Xinyan Chen, Liuhui Wang, Jianhan Jin, Claire Guo, Shen Yan, Bo Zhang, Chaoyou Fu, Peng Gao, and Hongsheng Li. 2025. https://arxiv.org/abs/2502.09621 Mme-cot: Benchmarking chain-of-thought in large multimodal models for reasoning quality, robustness, and efficiency . Preprint, arXiv:2502.09621

    arXiv 2025

  8. [8]

    Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. 2024. https://arxiv.org/abs/2408.03326 Llava-onevision: Easy visual task transfer . Preprint, arXiv:2408.03326

    arXiv 2024

Show all 29 references
  1. [9]

    Chengzu Li, Wenshan Wu, Huanyu Zhang, Yan Xia, Shaoguang Mao, Li Dong, Ivan Vulić, and Furu Wei. 2025. https://arxiv.org/abs/2501.07542 Imagine while reasoning in space: Multimodal visualization-of-thought . Preprint, arXiv:2501.07542

    2025 arXiv

  2. [10]

    Hongcheng Liu, Zhe Chen, Hui Li, Pingjie Wang, Yanfeng Wang, and Yu Wang. 2024 a . https://doi.org/10.1109/ICASSP48485.2024.10447469 Msg-bart: Multi-granularity scene graph-enhanced encoder-decoder language model for video-grounded dialogue generation . In ICASSP 2024 - 2024 I...

    2024

  3. [11]

    Hongcheng Liu, Yusheng Liao, Siqv Ou, Yuhao Wang, Heyang Liu, Yanfeng Wang, and Yu Wang. 2024 b . https://api.semanticscholar.org/CorpusID:271891942 Med-pmc: Medical personalized multi-modal consultation with a proactive ask-first-observe-next paradigm . ArXiv, abs/2408.08693

    2024 arXiv

  4. [12]

    Zhijian Liu, Ligeng Zhu, Baifeng Shi, Zhuoyang Zhang, Yuming Lou, Shang Yang, Haocheng Xi, Shiyi Cao, Yuxian Gu, Dacheng Li, Xiuyu Li, Yunhao Fang, Yukang Chen, Cheng-Yu Hsieh, De-An Huang, An-Chieh Cheng, Vishwesh Nath, Jinyi Hu, Sifei Liu, and 8 others. 2025. https://arxiv.o...

    2025 arXiv

  5. [13]

    Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. 2024. https://arxiv.org/abs/2310.02255 Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts . Preprint, a...

    2024 arXiv

  6. [14]

    Chancharik Mitra, Brandon Huang, Trevor Darrell, and Roei Herzig. 2024. https://arxiv.org/abs/2311.17076 Compositional chain-of-thought prompting for large multimodal models . Preprint, arXiv:2311.17076

    2024 arXiv

  7. [15]

    OpenAI. 2025. https://openai.com/index/introducing-o3-and-o4-mini/ Introducing openai o3 and o4-mini

    2025

  8. [16]

    Zhiyu Wu, Xiaokang Chen, Zizheng Pan, Xingchao Liu, Wen Liu, Damai Dai, Huazuo Gao, Yiyang Ma, Chengyue Wu, Bingxuan Wang, Zhenda Xie, Yu Wu, Kai Hu, Jiawei Wang, Yaofeng Sun, Yukun Li, Yishi Piao, Kang Guan, Aixin Liu, and 8 others. 2024. https://arxiv.org/abs/2412.10302 Deep...

    2024 arXiv

  9. [17]

    Ziyang Xiao, Dongxiang Zhang, Xiongwei Han, Xiaojin Fu, Yin Yu, Tao Zhong, Sai Wu, Yuan Wang, Jianwei Yin, and Gang Chen. 2024. https://proceedings.neurips.cc/paper_files/paper/2024/file/328c922d068dd4ccb23cec5c64e6c7fc-Paper-Conference.pdf Enhancing llm reasoning via vision-a...

    2024

  10. [18]

    Guowei Xu, Peng Jin, Hao Li, Yibing Song, Lichao Sun, and Li Yuan. 2025 a . https://arxiv.org/abs/2411.10440 Llava-cot: Let vision language models reason step-by-step . Preprint, arXiv:2411.10440

    2025 arXiv

  11. [19]

    Jin Xu, Zhifang Guo, Jinzheng He, Hangrui Hu, Ting He, Shuai Bai, Keqin Chen, Jialin Wang, Yang Fan, Kai Dang, Bin Zhang, Xiong Wang, Yunfei Chu, and Junyang Lin. 2025 b . https://arxiv.org/abs/2503.20215 Qwen2.5-omni technical report . Preprint, arXiv:2503.20215

    2025 arXiv

  12. [20]

    Zhengyuan Yang, Linjie Li, Kevin Lin, Jianfeng Wang, Chung-Ching Lin, Zicheng Liu, and Lijuan Wang. 2023 a . https://arxiv.org/abs/2309.17421 The dawn of lmms: Preliminary explorations with gpt-4v(ision) . Preprint, arXiv:2309.17421

    2023 arXiv

  13. [21]

    Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Ehsan Azarnasab, Faisal Ahmed, Zicheng Liu, Ce Liu, Michael Zeng, and Lijuan Wang. 2023 b . https://arxiv.org/abs/2303.11381 Mm-react: Prompting chatgpt for multimodal reasoning and action . Preprint, arXiv:2303.11381

    2023 arXiv

  14. [22]

    Griffiths, Yuan Cao, and Karthik Narasimhan

    Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Thomas L. Griffiths, Yuan Cao, and Karthik Narasimhan. 2023. https://arxiv.org/abs/2305.10601 Tree of thoughts: Deliberate problem solving with large language models . Preprint, arXiv:2305.10601

    2023 arXiv

  15. [23]

    Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, Qianyu Chen, Huarong Zhou, Zhensheng Zou, Haoye Zhang, Shengding Hu, Zhi Zheng, Jie Zhou, Jie Cai, Xu Han, and 4 others. 2024. https://arxiv.org/abs/2408.01800 Mi...

    2024 arXiv

  16. [24]

    Shukang Yin, Chaoyou Fu, Sirui Zhao, Ke Li, Xing Sun, Tong Xu, and Enhong Chen. 2024. https://doi.org/10.1093/nsr/nwae403 A survey on multimodal large language models . National Science Review, 11(12)

    2024 doi

  17. [25]

    Boqiang Zhang, Kehan Li, Zesen Cheng, Zhiqiang Hu, Yuqian Yuan, Guanzheng Chen, Sicong Leng, Yuming Jiang, Hang Zhang, Xin Li, Peng Jin, Wenqi Zhang, Fan Wang, Li Bing, and Deli Zhao. 2025. https://api.semanticscholar.org/CorpusID:275789161 Videollama 3: Frontier multimodal fo...

    2025 arXiv

  18. [26]

    Ge Zheng, Bin Yang, Jiajin Tang, Hong-Yu Zhou, and Sibei Yang. 2023. https://arxiv.org/abs/2310.16436 Ddcot: Duty-distinct chain-of-thought prompting for multimodal reasoning in language models . Preprint, arXiv:2310.16436

    2023 arXiv

  19. [27]

    Jinguo Zhu, Weiyun Wang, Zhe Chen, Zhaoyang Liu, Shenglong Ye, Lixin Gu, Hao Tian, Yuchen Duan, Weijie Su, Jie Shao, Zhangwei Gao, Erfei Cui, Xuehui Wang, Yue Cao, Yangzhou Liu, Xingguang Wei, Hongjie Zhang, Haomin Wang, Weiye Xu, and 32 others. 2025. https://arxiv.org/abs/250...

    2025 arXiv

  20. [28]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list...

  21. [29]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.