Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

PixelThink: Towards Efficient Chain-of-Pixel Reasoning

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read PixelThink claims that multiplying the Seg-Zero reward by a soft, difficulty- and uncertainty-aware token penalty teaches multimodal LLMs to reason concisely and segment more accurately.

desk verdict Useful efficiency recipe and benchmark for reasoning segmentation, but the difficulty-adaptive mechanism isn't yet supported by the experiments. read the letter →

arxiv 2505.23727 v1 pith:DXBGCOOX submitted 2025-05-29 cs.CV cs.MM

classification cs.CVcs.MM
keywords reasoningsegmentationefficientchain-of-thoughttokenbudgetrewardGRPOmultimodallargelanguagemodeldifficulty-adaptiveoverthinkingmitigationreferringexpression
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

PixelThink aims to fix overthinking in reasoning segmentation, where reinforcement-trained multimodal models emit long reasoning chains even for easy queries. The paper proposes a reward-level control: multiply the existing Seg-Zero reward by a soft penalty that shrinks whenever the generated chain exceeds a token budget, with the budget itself set by the sample's externally scored difficulty and the model's internally measured uncertainty. Under group-relative policy optimization this teaches the model to spend reasoning tokens where they help and to stop early where they do not. On the paper's new ReasonSeg-DIFF benchmark the method halves average reasoning length while improving mask Intersection-over-Union, and the same trained model transfers zero-shot to established referring-segmentation benchmarks.

What carries the argument

The load-bearing mechanism is the soft length penalty $s(L_{\text{used}}, L_{\text{budget}})$ (Eq. 4), a multiplicative reward modifier that returns $1$ for chains inside their budget and decays linearly once the chain exceeds it. The budget $L_{\text{budget}}$ is a piecewise function (Eq. 3) of $D$, an external difficulty score in $[1,10]$ averaged from scene complexity, segmentation challenge, and linguistic ambiguity, and of $U$, an internal uncertainty defined as $1$ minus the average top-2 token-probability margin. The final reward $R_{\text{final}} = R_{\text{original}} \cdot s(L_{\text{used}}, L_{\text{budget}})$ is optimized with GRPO, giving the model a smooth incentive to stay concise while leaving medium-difficulty samples unconstrained.

What would settle it

A shuffle test would settle it: retrain PixelThink with the difficulty labels randomly permuted across samples, preserving the same budget distribution. If token counts and gIoU stay essentially the same, the gains come from the soft penalty's overall pressure rather than from difficulty-adaptive allocation; the paper does not report this control.

Watch

Extended reading notes

Core claim

The paper's central claim is that a reward signal which gently taxes token overuse can teach a reasoning-segmentation model when to stop thinking. PixelThink multiplies the Seg-Zero reward $R_{\text{original}}$ by a soft penalty $s(L_{\text{used}}, L_{\text{budget}})$: the penalty is $1$ when the chain stays within budget and $1 - \beta(L_{\text{used}} - L_{\text{budget}})$ when it exceeds it, where the budget itself is $L_{\text{base}} + \alpha U$ for hard examples, $L_{\text{low}}$ for easy examples, and unconstrained for medium ones. Because $L_{\text{budget}}$ depends on an externally estimated difficulty $D$ and an internally measured uncertainty $U$, the model learns to allocate reasoning effort to cases that need it. On the paper's ReasonSeg-DIFF test benchmark this yields 47.66 tokens per chain versus Seg-Zero's 90.58, with gIoU rising from 58.20 to 60.17.

Load-bearing premise

The mechanism depends on the outside difficulty rating actually tracking how much reasoning each image needs; if that rating is noisy or biased, the token budgets and the difficulty-stratified benchmark both rest on an unreliable ruler.

Editorial extensions

If this is right

  • On ReasonSeg-DIFF test, average reasoning tokens drop from 90.58 to 47.66 while gIoU rises from 58.20 to 60.17, so the same architecture becomes cheaper and more accurate at inference.
  • Trained only on RefCOCOg, PixelThink improves over Seg-Zero on ReasonSeg (val gIoU 63.8 vs 61.6; test 60.2 vs 58.2) and remains competitive on RefCOCO, RefCOCO+, and RefCOCOg, indicating the efficiency gain is not bought by losing generalization.
  • The 3B variant of PixelThink performs on par with prior 7B counterparts on ReasonSeg, suggesting length control can substitute for some model-capacity scaling in this task.
  • No-thinking variants that remove reasoning entirely score lower on segmentation, supporting the paper's conclusion that concise reasoning, not reasoning omission, carries the benefit.
  • The new metrics SAT, RST, and URSS make reasoning quality and segmentation accuracy comparable per token, giving future work a common yardstick for efficiency-aware evaluation.

Reading between the lines

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

  • Because the uncertainty term $U$ is computed from token-level margins at no extra annotation cost, one could deploy the same budget rule at inference time without the external difficulty scorer, asking the model to stop when its own confidence is high; the paper does not test this adaptive-compute extension.
  • The benchmark's RScore rewards completeness regardless of length, so PixelThink's modest RScore drop (6.92 vs 7.67) may partly reflect the reference being verbose; a length-controlled judge or length-matched references would isolate genuine reasoning-quality loss from brevity.
  • The same multiplicative soft-penalty recipe could be transplanted to other GRPO-trained multimodal tasks such as visual question answering or grounding wherever overthinking is diagnosed, though the paper only demonstrates it for segmentation.
  • If the difficulty score $D$ were replaced by a cheap proxy such as expression length or scene-object count, the reward's benefit might be reproducible without a large 72B annotator; the paper's difficulty-granularity ablation suggests sensitivity to how $D$ is defined.
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

4 major / 5 minor

Summary. PixelThink proposes an RL fine-tuning scheme for reasoning segmentation that modulates the Seg-Zero GRPO reward with a soft length penalty s(L_used, L_budget), where L_budget depends on an external difficulty score D (from Qwen2.5-VL-72B) and an internal uncertainty estimate U (based on top-2 token probabilities). Easy samples are assigned a small budget, hard samples a larger budget scaled by U, and medium samples are left unconstrained. The paper also introduces ReasonSeg-DIFF, a benchmark with difficulty scores and short/long reference reasoning chains, plus metrics RScore/RST, SAT, and URSS. On ReasonSeg-DIFF, PixelThink reports reduced reasoning token counts (e.g., 47.66 vs 90.58 test tokens) with improved gIoU (60.17 vs 58.20) over Seg-Zero, and presents ablations of the D/U components, difficulty splits, and budget settings.

Significance. If the difficulty-adaptive compression claim is established, the reward design would be a simple and useful recipe for controlling reasoning length in multimodal LLMs, and ReasonSeg-DIFF would be a valuable resource for jointly evaluating reasoning quality, segmentation accuracy, and efficiency. The paper is clearly written, includes extensive ablations and qualitative analyses, and the planned release of code and benchmark is welcome. However, the central mechanism is under-supported: the missing constant-budget control and the circular use of D in training and evaluation leave room for alternative explanations of the reported token reduction, so the main claims require additional evidence.

major comments (4)
  1. [§5.2, Table 4; Eq. (3)–(5)] The ablation in Table 4 compares no length penalty to penalties modulated by D and/or U, but it never trains a control with a fixed, non-adaptive L_budget. Without this control, the token reduction from 86.92 to 39.95–46.98 could be caused by the soft penalty alone, with the D/U modulation in Eq. (3) making little contribution. This concern is supported by the paper's own observations: Appendix D.2 states final reasoning lengths converge across difficulty levels, and Table 2 shows easy/medium/hard tokens of 44.73/47.00/51.79 despite budgets of 96/unconstrained/approximately 256–281, a gradient no larger than Seg-Zero's own difficulty gradient (84.97/90.73/95.37). Please add a training run with a constant budget (e.g., 96 or 128 tokens for all samples) under the identical pipeline and report per-difficulty token counts and gIoU; without it, the central claim that the model 'learns to compress reasoning length in accordance with scene complexity' is not demonstrated. The non-monotonicity in Table 4 (39.95 tokens with D only, 42.41 with U only, 46.98 with both) also requires explanation.
  2. [§4.1, §4.2; Eq. (3)] The difficulty score D is computed by Qwen2.5-VL-72B (Section 4.1) and used both to set L_budget in training (Eq. 3) and to stratify the evaluation and select short versus long reference chains for RScore (Section 4.2); the RScore judge is also a Qwen2.5 model. This creates a closed loop between the training signal and the evaluation signal. The only guard is the statement that D is 'cross-validated against human annotations' (Section 4.1), but no quantitative agreement is reported. Please report human-model agreement statistics (e.g., correlation or Cohen's kappa on a held-out set), and evaluate with an independent difficulty estimator or with D held fixed across all methods to show that the difficulty-aware results are not an artifact of aligning to the same estimator.
  3. [§5.1, Tables 1–3] No seed or error-bar information is reported for any experiment. The headline claims on the ReasonSeg-DIFF test set (60.17 vs 58.20 gIoU; 55.77 vs 52.37 cIoU) are small enough that RL training noise and LLM-judge variance could alter the conclusions, and RScore is an LLM-based scalar. Please report means and standard deviations over at least three runs (or confidence intervals) for the main tables, including RScore, so the reader can assess whether the reported improvements are statistically meaningful.
  4. [§5.1 vs Appendix D.3] The text states that PixelThink 'preserves reasoning quality' (Section 5.1) and 'maintaining reasoning quality' (Introduction), but Table 1 shows RScore dropping from 7.67 to 6.92 on both validation and test splits. Appendix D.3 acknowledges this drop and attributes it to lower completeness in concise chains. As written, the paper's own metric contradicts the 'maintains reasoning quality' claim. Please either soften the claim or provide statistical evidence that the RScore drop is within noise; otherwise the reader cannot conclude that reasoning quality is maintained.
minor comments (5)
  1. [Eq. (3) and §5 Implementation Details] The thresholds τ1 and τ2 are first used in Eq. (3) but defined only later in Section 5; please define them at first use and state whether the easy condition is D < τ2 or D ≤ τ2.
  2. [Table 1 and Appendix A.1] The L1-Exact rows report 65.66 and 65.24 tokens despite a '64-token upper limit' in Appendix A.1; please clarify whether L1-Exact is a hard cap or an exact-length target and explain the discrepancy.
  3. [§4.2] The SAT and RST formulas use arbitrary constants (100 and 10) and P in billions; please state whether the resulting dimensionless values are comparable across model sizes and provide a sensitivity analysis or justification for the default γ=0.7 in URSS.
  4. [References and figures] There are minor typos: 'Empirical Eethods' in reference [30] should be 'Empirical Methods', and the token counts in Figure 1 (44 vs 47.7) should be reconciled with Table 1.
  5. [Appendix D.1] The explanation that the soft penalty 'implicitly encourages' shorter outputs is informal; a sentence describing how the multiplicative penalty interacts with GRPO group-relative normalization would help readers understand the mechanism.

Circularity Check

0 steps flagged · score 2.0 of 10

No material circularity: the difficulty-estimator training/evaluation overlap is a benchmark-validity concern, not an equation-level reduction, and token counts and gIoU/cIoU are externally measured.

full rationale

PixelThink's core mechanism (Eqs. 3-5) is an explicit reward modulation, R_final = R_original * s(L_used, L_budget), with L_budget set by an external difficulty estimate D and the model's own uncertainty U. The reported efficiency and segmentation numbers are measured from actual model outputs (token counts, gIoU, cIoU) and are not equal to any input by construction. In fact, Appendix D.2 concedes that final reasoning lengths converge across difficulty levels, and Table 2 shows the model does not reproduce L_budget (easy: 44.73 tokens vs L_low=96; hard: 51.79 vs ~256+25U), so the difficulty-aware part of the claim is empirically weak but not definitionally forced. The same Qwen2.5-VL-72B estimate D is used in Eq. 3 to set the training budget and in Section 4.2 to select short/long reference chains for RScore; this creates an evaluation-dependence loop for the 'difficulty-aware' RScore comparison, but RScore content and segmentation accuracy are not equal to the reward input, so this is a validation-validity concern rather than a circular derivation. The paper's self-citations (e.g., the authors' own survey [26], CoT-Valve [56], TokenPacker [63]) appear only in related work and are not load-bearing for the method. Section E.1 explicitly acknowledges reliance on coarse-grained difficulty scores and manual budget rules, which limits the strength of the central claim but does not make it circular. Overall, no 'prediction' reduces by construction; the main weaknesses are missing controls (no constant-budget ablation) and unquantified human cross-validation of D, which are empirical omissions, not circularity.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The central claim rests on two external estimators, an LLM difficulty scorer and a token-margin uncertainty measure, plus several hand-set reward constants. The uncertainty measure is computed from the model being trained, creating a self-referential reward signal. The difficulty scores are reused for both training and evaluation. No new physical or mathematical entities are introduced.

free parameters (7)
  • Lbase = 256
    Base token budget for hard tasks in Eq. 3, chosen by hand and ablated in Table 6.
  • Llow = 96
    Minimal token budget for easy tasks in Eq. 3, ablated in Table 6.
  • alpha = 25
    Uncertainty gain for hard tasks in Eq. 3, ablated in Table B.
  • beta = 2e-3
    Penalty factor in Eq. 4, set by hand with no sensitivity analysis reported.
  • tau1 and tau2 = 5.0 and 3.5
    Difficulty thresholds separating hard, medium, and easy tasks, ablated in Table 5.
  • gamma = 0.7
    Weight in the URSS unified metric, chosen by hand to emphasize segmentation accuracy.
  • RScore reference selection rule = short for easy/medium, long for hard
    Evaluation design choice in Section 4.2 that directly affects reported RST and URSS values.
assumptions (5)
  • domain assumption GRPO is a stable and valid optimizer for the non-differentiable reward R_final.
    Invoked in Section 3.4 without convergence guarantees; standard in the field.
  • domain assumption The average top-2 token-probability margin in Eq. 1 is a valid proxy for model uncertainty in reward shaping.
    Section 3.2; no calibration or correlation analysis with actual errors is provided.
  • domain assumption Qwen2.5-VL-72B difficulty scores are accurate and human-aligned.
    Section 4.1 claims cross-validation against human annotations but gives no quantitative agreement.
  • domain assumption LLM-based RScore with Qwen-generated short/long references is a valid measure of reasoning quality.
    Section 4.2; the judge is from the same model family as the generator, and inter-annotator agreement is not reported.
  • ad hoc to paper The re-implemented L1 and prompt baselines are faithful adaptations of L1 to segmentation.
    Appendix A.1; L1 was originally trained on DeepScaleR with explicit token-count prompts, and the adaptation is not validated against original code performance.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PixelThink: Towards Efficient Chain-of-Pixel Reasoning." pith.science (2026). https://pith.science/paper/DXBGCOOX

@misc{pith2026250523727,
  author       = {Pith},
  title        = {Pith review of: PixelThink: Towards Efficient Chain-of-Pixel Reasoning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DXBGCOOX}},
  note         = {Machine review of arXiv:2505.23727}
}
read the original abstract

Existing reasoning segmentation approaches typically fine-tune multimodal large language models (MLLMs) using image-text pairs and corresponding mask labels. However, they exhibit limited generalization to out-of-distribution scenarios without an explicit reasoning process. Although recent efforts leverage reinforcement learning through group-relative policy optimization (GRPO) to enhance reasoning ability, they often suffer from overthinking - producing uniformly verbose reasoning chains irrespective of task complexity. This results in elevated computational costs and limited control over reasoning quality. To address this problem, we propose PixelThink, a simple yet effective scheme that integrates externally estimated task difficulty and internally measured model uncertainty to regulate reasoning generation within a reinforcement learning paradigm. The model learns to compress reasoning length in accordance with scene complexity and predictive confidence. To support comprehensive evaluation, we introduce ReasonSeg-Diff, an extended benchmark with annotated reasoning references and difficulty scores, along with a suite of metrics designed to assess segmentation accuracy, reasoning quality, and efficiency jointly. Experimental results demonstrate that the proposed approach improves both reasoning efficiency and overall segmentation performance. Our work contributes novel perspectives towards efficient and interpretable multimodal understanding. The code and model will be publicly available.

Figures

Figures reproduced from arXiv: 2505.23727 by the authors.

Figure 1
Figure 1. Motivation of Efficient Chain-of-Pixel Reasoning (PIXELTHINK). We propose a novel scheme for reasoning segmentation that effectively regulates reasoning length based on task difficulty and uncertainty. Our method improves segmentation quality while significantly reducing token usage. A suite of metrics is introduced for holistic evaluations of reasoning quality, segmentation accuracy, and computational efficiency. A… view at source ↗
Figure 2
Figure 2. Overview of PIXELTHINK. (a) Workflow of the reasoning segmentation framework. Given an input image and query, the model generates a reasoning chain and segmentation reference that guides the segmentation outcome. (b) The group-relative policy optimization (GRPO) procedure employed during reinforcement fine-tuning. Our new reward design incorporates both task difficulty and model uncertainty, enabling the model to le… view at source ↗
Figure 3
Figure 3. The construction of ReasonSeg-DIFF. (a) Design on the Difficulty Scoring scheme. (b) Generation of the Short Thinking process. (c) Generation of Long Thinking process. Difficulty Scoring. To quantify the intrinsic reasoning challenge of each sample, we propose a structured scoring framework based on three interpretable factors: (1) Scene Complexity, measuring the number and similarity of distractor objects; (2) Segm… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Qualitative comparisons between Seg-Zero [18] and the proposed PIXELTHINK. Repre￾sentative samples across different difficulty levels are selected to highlight differences in the reasoning process and segmentation performance [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  1. DGSeg: Dynamic Gating of Semantic-Spatial Guided Predictions for Reasoning Segmentation

    cs.CV 2026-07 accept novelty 6.0 of 10

    Separate semantic-spatial MLLM cues plus a quality-supervised dynamic gate yield 69.6/67.3 gIoU on ReasonSeg val/test, beating strong MLLM+SAM baselines.

Reference graph

Works this paper leans on

81 extracted references · 26 canonical work pages · cited by 1 Pith paper

  1. [1]

    Modeling context in referring expressions

    Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. InEuropean Conference on Computer Vision, pages 69–85. Springer, 2016

  2. [2]

    Lisa: Reasoning segmentation via large language model

    Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9579–9589, 2024

  3. [3]

    POPEN: Preference-Based Optimization and Ensemble for LVLM-Based Reasoning Segmentation

    Lanyun Zhu, Tianrun Chen, Qianxiong Xu, Xuanyi Liu, Deyi Ji, Haiyang Wu, De Wen Soh, and Jun Liu. Popen: Preference-based optimization and ensemble for lvlm-based reasoning segmentation.arXiv preprint arXiv:2504.00640, 2025

  4. [4]

    Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs

    Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE Transactions on Pattern Analysis and Machine Intelligence, 40(4):834–848, 2017

  5. [5]

    Masked- attention mask transformer for universal image segmentation

    Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked- attention mask transformer for universal image segmentation. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1290–1299, 2022

  6. [6]

    Mask r-cnn

    Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask r-cnn. InIEEE International Conference on Computer Vision, pages 2961–2969, 2017

  7. [7]

    Lamm: Language-assisted multi-modal instruction-tuning dataset, framework, and benchmark.Advances in Neural Information Processing Systems, 36:26650–26685, 2023

    Zhenfei Yin, Jiong Wang, Jianjian Cao, Zhelun Shi, Dingning Liu, Mukai Li, Xiaoshui Huang, Zhiyong Wang, Lu Sheng, Lei Bai, et al. Lamm: Language-assisted multi-modal instruction-tuning dataset, framework, and benchmark.Advances in Neural Information Processing Systems, 36:26650–26685, 2023

  8. [8]

    Inst3d-lmm: Instance-aware 3d scene understanding with multi-modal instruction tuning.arXiv preprint arXiv:2503.00513, 2025

    Hanxun Yu, Wentong Li, Song Wang, Junbo Chen, and Jianke Zhu. Inst3d-lmm: Instance-aware 3d scene understanding with multi-modal instruction tuning.arXiv preprint arXiv:2503.00513, 2025

Show all 81 references
  1. [9]

    Drivevlm: The convergence of autonomous driving and large vision-language models.arXiv preprint arXiv:2402.12289, 2024

    Xiaoyu Tian, Junru Gu, Bailin Li, Yicheng Liu, Yang Wang, Zhiyong Zhao, Kun Zhan, Peng Jia, Xianpeng Lang, and Hang Zhao. Drivevlm: The convergence of autonomous driving and large vision-language models.arXiv preprint arXiv:2402.12289, 2024

  2. [10]

    Are vlms ready for autonomous driving? an empirical study from the reliability, data, and metric perspectives.arXiv preprint arXiv:2501.04003, 2025

    Shaoyuan Xie, Lingdong Kong, Yuhao Dong, Chonghao Sima, Wenwei Zhang, Qi Alfred Chen, Ziwei Liu, and Liang Pan. Are vlms ready for autonomous driving? an empirical study from the reliability, data, and metric perspectives.arXiv preprint arXiv:2501.04003, 2025

  3. [11]

    Visual instruction tuning.Advances in Neural Information Processing Systems, 36:34892–34916, 2023

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning.Advances in Neural Information Processing Systems, 36:34892–34916, 2023

  4. [12]

    Improved baselines with visual instruction tuning

    Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296–26306, 2024

  5. [13]

    Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution.arXiv preprint arXiv:2409.12191, 2024

    Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution.arXiv preprint arXiv:2409.12191, 2024

  6. [14]

    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

  7. [15]

    Pixellm: Pixel reasoning with large multimodal model

    Zhongwei Ren, Zhicheng Huang, Yunchao Wei, Yao Zhao, Dongmei Fu, Jiashi Feng, and Xiaojie Jin. Pixellm: Pixel reasoning with large multimodal model. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26374–26383, 2024

  8. [16]

    Omg-llava: Bridging image-level, object-level, pixel-level reasoning and understanding

    Tao Zhang, Xiangtai Li, Hao Fei, Haobo Yuan, Shengqiong Wu, Shunping Ji, Chen Change Loy, and Shuicheng Yan. Omg-llava: Bridging image-level, object-level, pixel-level reasoning and understanding. InAdvances in Neural Information Processing Systems, volume 37, pages 71737–71767, 2024

  9. [17]

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

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

  10. [18]

    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

  11. [19]

    Vlm-r1: A stable and generalizable r1-style large vision-language model.arXiv preprint arXiv:2504.07615, 2025

    Haozhan Shen, Peng Liu, Jingcheng Li, Chunxin Fang, Yibo Ma, Jiajia Liao, Qiaoli Shen, Zilun Zhang, Kangjia Zhao, Qianqian Zhang, et al. Vlm-r1: A stable and generalizable r1-style large vision-language model.arXiv preprint arXiv:2504.07615, 2025

  12. [20]

    Towards reasoning era: A survey of long chain-of-thought for reasoning large language models.arXiv preprint arXiv:2503.09567, 2025

    Qiguang Chen, Libo Qin, Jinhao Liu, Dengyun Peng, Jiannan Guan, Peng Wang, Mengkang Hu, Yuhang Zhou, Te Gao, and Wanxiang Che. Towards reasoning era: A survey of long chain-of-thought for reasoning large language models.arXiv preprint arXiv:2503.09567, 2025

  13. [21]

    From system 1 to system 2: A survey of reasoning large language models.arXiv preprint arXiv:2502.17419, 2025

    Zhong-Zhi Li, Duzhen Zhang, Ming-Liang Zhang, Jiaxin Zhang, Zengyan Liu, Yuxuan Yao, Haotian Xu, Junhao Zheng, Pei-Jie Wang, Xiuyi Chen, et al. From system 1 to system 2: A survey of reasoning large language models.arXiv preprint arXiv:2502.17419, 2025. 20

  14. [22]

    Deepseekmath: Pushing the limits of mathematical reasoning in open language models.arXiv preprint arXiv:2402.03300, 2024

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

  15. [23]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948, 2025

    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

  16. [24]

    R1-V.https://github.com/Deep-Agent/R1-V?tab=readme-ov-file, 2025

    R1-V Team. R1-V.https://github.com/Deep-Agent/R1-V?tab=readme-ov-file, 2025

  17. [25]

    Visual-rft: Visual reinforcement fine-tuning.arXiv preprint arXiv:2503.01785, 2025

    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

  18. [26]

    Efficient reasoning models: A survey.arXiv preprint arXiv:2504.10903, 2025

    Sicheng Feng, Gongfan Fang, Xinyin Ma, and Xinchao Wang. Efficient reasoning models: A survey.arXiv preprint arXiv:2504.10903, 2025

  19. [27]

    Stop overthinking: A survey on efficient reasoning for large language models.arXiv preprint arXiv:2503.16419, 2025

    Yang Sui, Yu-Neng Chuang, Guanchu Wang, Jiamu Zhang, Tianyi Zhang, Jiayi Yuan, Hongyi Liu, Andrew Wen, Hanjie Chen, Xia Hu, et al. Stop overthinking: A survey on efficient reasoning for large language models.arXiv preprint arXiv:2503.16419, 2025

  20. [28]

    Sam4mllm: Enhance multi-modal large language model for referring expression segmentation

    Yi-Chia Chen, Wei-Hua Li, Cheng Sun, Yu-Chiang Frank Wang, and Chu-Song Chen. Sam4mllm: Enhance multi-modal large language model for referring expression segmentation. InEuropean Conference on Computer Vision, pages 323–340. Springer, 2024

  21. [29]

    L1: Controlling how long a reasoning model thinks with reinforcement learning.arXiv preprint arXiv:2503.04697, 2025

    Pranjal Aggarwal and Sean Welleck. L1: Controlling how long a reasoning model thinks with reinforcement learning.arXiv preprint arXiv:2503.04697, 2025

  22. [30]

    Referitgame: Referring to objects in photographs of natural scenes

    Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. InConference on Empirical Eethods in Natural Language Processing, pages 787–798, 2014

  23. [31]

    Towards robust referring image segmentation.IEEE Transactions on Image Processing, 2024

    Jianzong Wu, Xiangtai Li, Xia Li, Henghui Ding, Yunhai Tong, and Dacheng Tao. Towards robust referring image segmentation.IEEE Transactions on Image Processing, 2024

  24. [32]

    Remamber: Referring image segmentation with mamba twister

    Yuhuan Yang, Chaofan Ma, Jiangchao Yao, Zhun Zhong, Ya Zhang, and Yanfeng Wang. Remamber: Referring image segmentation with mamba twister. InEuropean Conference on Computer Vision, pages 108–126. Springer, 2024

  25. [33]

    Mask grounding for referring image segmentation

    Yong Xien Chng, Henry Zheng, Yizeng Han, Xuchong Qiu, and Gao Huang. Mask grounding for referring image segmentation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26573–26583, 2024

  26. [34]

    A mutual supervision framework for referring expression segmentation and generation.International Journal of Computer Vision, pages 1–16, 2025

    Shijia Huang, Feng Li, Hao Zhang, Shilong Liu, Lei Zhang, and Liwei Wang. A mutual supervision framework for referring expression segmentation and generation.International Journal of Computer Vision, pages 1–16, 2025

  27. [35]

    Pixel-sail: Single transformer for pixel-grounded understanding.arXiv, 2025

    Tao Zhang, Xiangtai Li, Zilong Huang, Yanwei Li, Weixian Lei, Xueqing Deng, Shihao Chen, Shunping Ji, , and Jiashi Feng. Pixel-sail: Single transformer for pixel-grounded understanding.arXiv, 2025

  28. [36]

    U-net: Convolutional networks for biomedical image segmentation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. InInternational Conference on Medical Image Computing and Computer-Assisted Intervention, pages 234–241. Springer, 2015

  29. [37]

    Segment anything

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. InIEEE/CVF International Conference on Computer Vision, pages 4015–4026, 2023

  30. [38]

    Lisa++: An improved baseline for reasoning segmentation with large language model.arXiv preprint arXiv:2312.17240, 2023

    Senqiao Yang, Tianyuan Qu, Xin Lai, Zhuotao Tian, Bohao Peng, Shu Liu, and Jiaya Jia. Lisa++: An improved baseline for reasoning segmentation with large language model.arXiv preprint arXiv:2312.17240, 2023

  31. [39]

    Sa2va: Marrying sam2 with llava for dense grounded understanding of images and videos.arXiv, 2025

    Haobo Yuan, Xiangtai Li, Tao Zhang, Zilong Huang, Shilin Xu, Shunping Ji, Yunhai Tong, Lu Qi, Jiashi Feng, and Ming-Hsuan Yang. Sa2va: Marrying sam2 with llava for dense grounded understanding of images and videos.arXiv, 2025

  32. [40]

    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

  33. [41]

    Automatic chain of thought prompting in large language models.arXiv preprint arXiv:2210.03493, 2022

    Zhuosheng Zhang, Aston Zhang, Mu Li, and Alex Smola. Automatic chain of thought prompting in large language models.arXiv preprint arXiv:2210.03493, 2022

  34. [42]

    Towards better chain-of-thought prompting strategies: A survey.arXiv preprint arXiv:2310.04959, 2023

    Zihan Yu, Liang He, Zhen Wu, Xinyu Dai, and Jiajun Chen. Towards better chain-of-thought prompting strategies: A survey.arXiv preprint arXiv:2310.04959, 2023

  35. [43]

    Math-shepherd: Verify and reinforce llms step-by-step without human annotations.arXiv preprint arXiv:2312.08935, 2023

    Peiyi Wang, Lei Li, Zhihong Shao, RX Xu, Damai Dai, Yifei Li, Deli Chen, Yu Wu, and Zhifang Sui. Math-shepherd: Verify and reinforce llms step-by-step without human annotations.arXiv preprint arXiv:2312.08935, 2023. 21

  36. [44]

    Visualprm: An effective process reward model for multimodal reasoning.arXiv preprint arXiv:2503.10291, 2025

    Weiyun Wang, Zhangwei Gao, Lianjie Chen, Zhe Chen, Jinguo Zhu, Xiangyu Zhao, Yangzhou Liu, Yue Cao, Shenglong Ye, Xizhou Zhu, et al. Visualprm: An effective process reward model for multimodal reasoning.arXiv preprint arXiv:2503.10291, 2025

  37. [45]

    Prmbench: A fine-grained and challenging benchmark for process-level reward models.arXiv preprint arXiv:2501.03124, 2025

    Mingyang Song, Zhaochen Su, Xiaoye Qu, Jiawei Zhou, and Yu Cheng. Prmbench: A fine-grained and challenging benchmark for process-level reward models.arXiv preprint arXiv:2501.03124, 2025

  38. [46]

    OpenAI o1.https://openai.com/o1/, 2024

    OpenAI. OpenAI o1.https://openai.com/o1/, 2024

  39. [47]

    Kimi Team, Angang Du, Bofei Gao, Bowei Xing, Changjiu Jiang, Cheng Chen, Cheng Li, Chenjun Xiao, Chenzhuang Du, Chonghua Liao, et al. Kimi k1. 5: Scaling reinforcement learning with llms.arXiv preprint arXiv:2501.12599, 2025

  40. [48]

    s1: Simple test-time scaling

    Niklas Muennighoff, Zitong Yang, Weijia Shi, Xiang Lisa Li, Li Fei-Fei, Hannaneh Hajishirzi, Luke Zettlemoyer, Percy Liang, Emmanuel Candès, and Tatsunori Hashimoto. s1: Simple test-time scaling. arXiv preprint arXiv:2501.19393, 2025

  41. [49]

    Can 1b llm surpass 405b llm? rethinking compute-optimal test-time scaling.arXiv preprint arXiv:2502.06703, 2025

    Runze Liu, Junqi Gao, Jian Zhao, Kaiyan Zhang, Xiu Li, Biqing Qi, Wanli Ouyang, and Bowen Zhou. Can 1b llm surpass 405b llm? rethinking compute-optimal test-time scaling.arXiv preprint arXiv:2502.06703, 2025

  42. [50]

    Ttrl: Test-time reinforcement learning.arXiv preprint arXiv:2504.16084, 2025

    Yuxin Zuo, Kaiyan Zhang, Shang Qu, Li Sheng, Xuekai Zhu, Biqing Qi, Youbang Sun, Ganqu Cui, Ning Ding, and Bowen Zhou. Ttrl: Test-time reinforcement learning.arXiv preprint arXiv:2504.16084, 2025

  43. [51]

    Open R1 Multimodal

    EvolvingLMMs Lab. Open R1 Multimodal. https://github.com/EvolvingLMMs-Lab/ open-r1-multimodal, 2025

  44. [52]

    Reason-rft: Reinforcement fine-tuning for visual reasoning.arXiv preprint arXiv:2503.20752, 2025

    Huajie Tan, Yuheng Ji, Xiaoshuai Hao, Minglan Lin, Pengwei Wang, Zhongyuan Wang, and Shanghang Zhang. Reason-rft: Reinforcement fine-tuning for visual reasoning.arXiv preprint arXiv:2503.20752, 2025

  45. [53]

    Efficient inference for large reasoning models: A survey.arXiv preprint arXiv:2503.23077, 2025

    Yue Liu, Jiaying Wu, Yufei He, Hongcheng Gao, Hongyu Chen, Baolong Bi, Jiaheng Zhang, Zhiqi Huang, and Bryan Hooi. Efficient inference for large reasoning models: A survey.arXiv preprint arXiv:2503.23077, 2025

  46. [54]

    A survey of efficient reasoning for large reasoning models: Language, multimodality, and beyond.arXiv preprint arXiv:2503.21614, 2025

    Xiaoye Qu, Yafu Li, Zhaochen Su, Weigao Sun, Jianhao Yan, Dongrui Liu, Ganqu Cui, Daizong Liu, Shuxian Liang, Junxian He, et al. A survey of efficient reasoning for large reasoning models: Language, multimodality, and beyond.arXiv preprint arXiv:2503.21614, 2025

  47. [55]

    Token-budget- aware llm reasoning.arXiv preprint arXiv:2412.18547, 2024

    Tingxu Han, Zhenting Wang, Chunrong Fang, Shiyu Zhao, Shiqing Ma, and Zhenyu Chen. Token-budget- aware llm reasoning.arXiv preprint arXiv:2412.18547, 2024

  48. [56]

    Cot-valve: Length- compressible chain-of-thought tuning.arXiv preprint arXiv:2502.09601, 2025

    Xinyin Ma, Guangnian Wan, Runpeng Yu, Gongfan Fang, and Xinchao Wang. Cot-valve: Length- compressible chain-of-thought tuning.arXiv preprint arXiv:2502.09601, 2025

  49. [57]

    O1-pruner: Length-harmonizing fine-tuning for o1-like reasoning pruning.arXiv preprint arXiv:2501.12570, 2025

    Haotian Luo, Li Shen, Haiying He, Yibo Wang, Shiwei Liu, Wei Li, Naiqiang Tan, Xiaochun Cao, and Dacheng Tao. O1-pruner: Length-harmonizing fine-tuning for o1-like reasoning pruning.arXiv preprint arXiv:2501.12570, 2025

  50. [58]

    Chain of draft: Thinking faster by writing less

    Silei Xu, Wenhao Xie, Lingxiao Zhao, and Pengcheng He. Chain of draft: Thinking faster by writing less. arXiv preprint arXiv:2502.18600, 2025

  51. [59]

    Tokenskip: Controllable chain-of-thought compression in llms.arXiv preprint arXiv:2502.12067, 2025

    Heming Xia, Yongqi Li, Chak Tou Leong, Wenjie Wang, and Wenjie Li. Tokenskip: Controllable chain-of-thought compression in llms.arXiv preprint arXiv:2502.12067, 2025

  52. [60]

    Thinkprune: Pruning long chain-of-thought of llms via reinforcement learning.arXiv preprint arXiv:2504.01296, 2025

    Bairu Hou, Yang Zhang, Jiabao Ji, Yujian Liu, Kaizhi Qian, Jacob Andreas, and Shiyu Chang. Thinkprune: Pruning long chain-of-thought of llms via reinforcement learning.arXiv preprint arXiv:2504.01296, 2025

  53. [61]

    Self-training elicits concise reasoning in large language models.arXiv preprint arXiv:2502.20122, 2025

    Tergel Munkhbat, Namgyu Ho, Seo Hyun Kim, Yongjin Yang, Yujin Kim, and Se-Young Yun. Self-training elicits concise reasoning in large language models.arXiv preprint arXiv:2502.20122, 2025

  54. [62]

    Llava-prumerge: Adaptive token reduction for efficient large multimodal models.arXiv preprint arXiv:2403.15388, 2024

    Yuzhang Shang, Mu Cai, Bingxin Xu, Yong Jae Lee, and Yan Yan. Llava-prumerge: Adaptive token reduction for efficient large multimodal models.arXiv preprint arXiv:2403.15388, 2024

  55. [63]

    Tokenpacker: Efficient visual projector for multimodal llm.arXiv preprint arXiv:2407.02392, 2024

    Wentong Li, Yuqian Yuan, Jian Liu, Dongqi Tang, Song Wang, Jie Qin, Jianke Zhu, and Lei Zhang. Tokenpacker: Efficient visual projector for multimodal llm.arXiv preprint arXiv:2407.02392, 2024

  56. [64]

    An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models

    Liang Chen, Haozhe Zhao, Tianyu Liu, Shuai Bai, Junyang Lin, Chang Zhou, and Baobao Chang. An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models. InEuropean Conference on Computer Vision, pages 19–35. Springer, 2024

  57. [65]

    Visionzip: Longer is better but not necessary in vision language models.arXiv preprint arXiv:2412.04467, 2024

    Senqiao Yang, Yukang Chen, Zhuotao Tian, Chengyao Wang, Jingyao Li, Bei Yu, and Jiaya Jia. Visionzip: Longer is better but not necessary in vision language models.arXiv preprint arXiv:2412.04467, 2024

  58. [66]

    Learning to inference adaptively for multimodal large language models.arXiv preprint arXiv:2503.10905, 2025

    Zhuoyan Xu, Khoi Duc Nguyen, Preeti Mukherjee, Saurabh Bagchi, Somali Chaterji, Yingyu Liang, and Yin Li. Learning to inference adaptively for multimodal large language models.arXiv preprint arXiv:2503.10905, 2025. 22

  59. [67]

    Sam 2: Segment anything in images and videos.arXiv preprint arXiv:2408.00714, 2024

    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

  60. [68]

    Minimum-margin active learning.arXiv preprint arXiv:1906.00025, 2019

    Heinrich Jiang and Maya Gupta. Minimum-margin active learning.arXiv preprint arXiv:1906.00025, 2019

  61. [69]

    Chain-of-thought reasoning without prompting.arXiv preprint arXiv:2402.10200, 2024

    Xuezhi Wang and Denny Zhou. Chain-of-thought reasoning without prompting.arXiv preprint arXiv:2402.10200, 2024

  62. [70]

    An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al. Qwen2. 5 technical report.arXiv preprint arXiv:2412.15115, 2024

  63. [71]

    Deepspeed: System optimizations enable training deep learning models with over 100 billion parameters

    Jeff Rasley, Samyam Rajbhandari, Olatunji Ruwase, and Yuxiong He. Deepspeed: System optimizations enable training deep learning models with over 100 billion parameters. InACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 3505–3506, 2020

  64. [72]

    Open-vocabulary semantic segmentation with mask-adapted clip

    Feng Liang, Bichen Wu, Xiaoliang Dai, Kunpeng Li, Yinan Zhao, Hang Zhang, Peizhao Zhang, Peter Vajda, and Diana Marculescu. Open-vocabulary semantic segmentation with mask-adapted clip. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7061–7070, 2023

  65. [73]

    Gres: Generalized referring expression segmentation

    Chang Liu, Henghui Ding, and Xudong Jiang. Gres: Generalized referring expression segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23592–23601, 2023

  66. [74]

    Grounded sam: Assembling open-world models for diverse visual tasks.arXiv preprint arXiv:2401.14159, 2024

    Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, et al. Grounded sam: Assembling open-world models for diverse visual tasks.arXiv preprint arXiv:2401.14159, 2024

  67. [75]

    Lavt: Language- aware vision transformer for referring image segmentation

    Zhao Yang, Jiaqi Wang, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip HS Torr. Lavt: Language- aware vision transformer for referring image segmentation. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18155–18165, 2022

  68. [76]

    Perceptiongpt: Effectively fusing visual perception into llm

    Renjie Pi, Lewei Yao, Jiahui Gao, Jipeng Zhang, and Tong Zhang. Perceptiongpt: Effectively fusing visual perception into llm. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27124–27133, 2024

  69. [77]

    Think or not think: A study of explicit thinking in rule-based visual reinforcement fine-tuning.arXiv preprint arXiv:2503.16188, 2025

    Ming Li, Shitian Zhao, Jike Zhong, Yuxiang Lai, and Kaipeng Zhang. Think or not think: A study of explicit thinking in rule-based visual reinforcement fine-tuning.arXiv preprint arXiv:2503.16188, 2025

  70. [78]

    Reasoning models can be effective without thinking.arXiv preprint arXiv:2504.09858, 2025

    Wenjie Ma, Jingxuan He, Charlie Snell, Tyler Griggs, Sewon Min, and Matei Zaharia. Reasoning models can be effective without thinking.arXiv preprint arXiv:2504.09858, 2025

  71. [79]

    Deepscaler: Surpassing o1-preview with a 1.5 b model by scaling rl.Notion Blog, 2025

    Michael Luo, Sijun Tan, Justin Wong, Xiaoxiang Shi, William Y Tang, Manan Roongta, Colin Cai, Jeffrey Luo, Tianjun Zhang, Li Erran Li, et al. Deepscaler: Surpassing o1-preview with a 1.5 b model by scaling rl.Notion Blog, 2025

  72. [80]

    Microsoft coco: Common objects in context

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. InEuropean Conference on Computer Vision, pages 740–755, 2014

  73. [81]

    Hybridflow: A flexible and efficient rlhf framework.arXiv preprint arXiv: 2409.19256, 2024

    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.arXiv preprint arXiv: 2409.19256, 2024. 23

Pith tools

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