Pith. sign in

REVIEW 3 major objections 3 minor 34 references

Reported gains from iterative visual self-correction in VLMs are a measurement mirage that disappears under any shippable stopping rule.

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

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

T0 review · grok-4.5

2026-07-12 14:10 UTC pith:6AVSL3D6

load-bearing objection The +2.4pp iterative gain is an oracle mirage; under every shippable stopping rule the loop does not beat step 0, and the honest-trajectory protocol is the part worth keeping. the 3 major comments →

arxiv 2606.13156 v2 pith:6AVSL3D6 submitted 2026-06-11 cs.CV cs.AI

Iterative Visual Thinking and the Self-Correction Mirage in VLM Grounding

Animesh Tripathy , Aswanth Krishnan This is my paper
classification cs.CV cs.AI
keywords vision-language modelsvisual groundingself-correctiontest-time scalingbounding-box predictionevaluation protocolverification failure
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper builds a clean instance of visual self-correction for spatial grounding: a vision-language model predicts a bounding box, sees the box drawn on the image, and revises over multiple steps. After a two-stage training recipe (teacher-generated correction traces followed by simple IoU reward), the system appears to beat its single-shot base by 2.4 points. The authors show that the number is produced by an oracle that, for every sample, keeps the trajectory step closest to the true box—an answer the system cannot know at deployment. Under every fixed, label-free stopping policy the gain vanishes; the best practical choice is simply not to iterate. The underlying failure is verification, not generation: better boxes appear inside trajectories, yet the model cannot reliably select them. The work therefore replaces the usual reporting practice with an honest-trajectory protocol that reports accuracy under deployable rules and makes the oracle-to-shippable gap explicit.

Core claim

The +2.4 pp Acc@0.5 improvement claimed for Iterative Visual Thinking over a single-shot base is an artifact of an oracle that retains, per sample, the trajectory step nearest the ground-truth box. Once accuracy is re-measured under any fixed label-free stopping rule, the improvement disappears and stopping at the first prediction matches the base while beating every shippable alternative. The model can generate a better box somewhere in its trajectory but cannot identify it; self-verification confidence correlates only weakly with correctness and the loop responds to the mere presence of a rendered box rather than its quality.

What carries the argument

Honest-trajectory evaluation: accuracy measured only under fixed, label-free stopping policies, together with an explicit oracle-shippable gap that quantifies how much of the reported gain depends on knowing the answer in advance.

Load-bearing premise

The weak correlation between self-verification scores and correctness, plus one counterfactual overlay experiment, are taken as sufficient proof that the failure is pure verification rather than limited generation ability or training artifacts.

What would settle it

Train or prompt a model that, on the same trajectories, can select the highest-IoU step with label-free confidence scores whose correlation with true IoU exceeds roughly 0.6; if that selection recovers most of the oracle gain under a deployable rule, the verification-failure claim is falsified.

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

If this is right

  • Any iterative visual grounding result that reports only the best-in-trajectory box must be re-scored under fixed stopping rules before it can be treated as a real improvement.
  • Test-time scaling for spatial tasks cannot be assumed to work by simple self-correction loops; the missing piece is a reliable verifier, not more generation steps.
  • Training recipes that use teacher-generated correction traces and IoU rewards can teach the model to produce better intermediate boxes without teaching it to stop on them.
  • Published numbers that silently keep the oracle step systematically overstate the practical value of multi-step visual reasoning.

Where Pith is reading between the lines

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

  • The same oracle-versus-deployable gap likely contaminates other multi-step VLM pipelines that claim test-time gains from self-reflection or tool use.
  • A cheap external verifier (even a small specialist model) could convert the latent better boxes into real shippable gains without changing the generator.
  • If verification remains hard, the practical research priority for grounding may shift from longer thinking to better single-shot localization or retrieval of external spatial knowledge.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 3 minor

Summary. The paper constructs Iterative Visual Thinking (IVT), a multi-step visual self-correction procedure for VLM spatial grounding: the model predicts a box, sees it rendered, and refines. Training uses a two-phase recipe (supervised warm-up in which a teacher VLM converts the base model’s own errors into corrective reasoning traces, followed by GRPO with a simple IoU reward). Under the conventional practice of retaining, per sample, the trajectory step closest to ground truth, IVT yields +2.4 pp Acc@0.5 over the single-shot base. The authors show this gain is an oracle mirage: under every fixed label-free stopping rule the improvement vanishes and stopping at step 0 matches the base and dominates all shippable policies. They diagnose a verification failure (self-verification confidence correlates only weakly with correctness, r≈0.22; a counterfactual overlay reacts to box presence rather than correctness) and propose an honest-trajectory evaluation protocol that reports accuracy under label-free policies together with the explicit oracle–shippable gap.

Significance. If the measurement result holds, the work supplies a timely and practically important corrective to evaluation practice in iterative / test-time-compute VLMs. It cleanly separates generation capacity from selection capacity and demonstrates that a common (often implicit) oracle selection systematically inflates reported gains. The honest-trajectory protocol is a concrete, adoptable recommendation. The annotation-free training recipe is a useful engineering contribution. The causal claim that the failure mode is verification rather than generation is secondary; even if that attribution is only partially supported, the primary empirical finding (oracle gain is non-deployable) remains valuable for the field.

major comments (3)
  1. [Experiments / re-scoring results] The central empirical claim (that the +2.4 pp Acc@0.5 disappears under every deployable label-free rule and that step-0 is optimal) requires a fully explicit table of all policies tested (fixed step k, confidence threshold, entropy, majority vote over trajectory, etc.), with Acc@0.5, confidence intervals or multi-seed statistics, and the precise definition of each rule. Without that table the assertion that “every shippable rule” is beaten by step 0 cannot be audited.
  2. [Analysis / verification failure] The interpretive claim that “the model can generate a better box somewhere but cannot identify it” is load-bearing for the verification-failure diagnosis. The single Pearson r≈0.22 is a weak association; the manuscript should add rank correlation, reliability diagrams, and a controlled ablation that replaces the model’s own verifier with an external oracle (or stronger VLM) verifier to quantify how much of the oracle–shippable gap is closed by better verification alone.
  3. [Introduction / Related Work] The claim that oracle best-step selection is representative of “how such systems are usually reported” is an axiom of the paper. It needs concrete citations to prior iterative grounding or visual self-correction works that actually used (or can be shown to have used) this selection; otherwise the “mirage” framing risks over-generalizing from a single evaluation convention.
minor comments (3)
  1. [entire manuscript] The full manuscript text supplied for review is heavily corrupted by encoding artifacts, rendering many section headings, equations and table captions unreadable. A clean PDF is required for final production.
  2. [Method / GRPO stage] Notation for the IoU reward scaling and the precise GRPO hyperparameters should be collected in a single table or appendix for reproducibility.
  3. [Discussion / Protocol] A short discussion of how the honest-trajectory protocol would apply to non-grounding iterative VLM tasks (e.g., visual chain-of-thought) would broaden impact without lengthening the main argument.

Circularity Check

0 steps flagged

No significant circularity: the paper exposes an oracle-selection mirage in prior evaluation practice and re-scores under label-free policies that do not reduce to ground-truth by construction.

full rationale

This is an empirical ML paper whose central claim is that a common reporting practice (keeping, per sample, the trajectory step closest to the ground-truth box) is non-deployable and produces a measurement mirage. The paper's own positive numbers are obtained under fixed label-free stopping rules (including step-0), and the oracle-shippable gap is reported explicitly rather than hidden. There is no derivation in which a fitted parameter, self-defined quantity, or self-citation uniqueness theorem is renamed as a prediction of the same quantity. The training recipe (base-model errors turned into teacher traces, then GRPO with IoU) and the diagnostic correlations (r≈0.22, counterfactual overlay) are ordinary experimental inputs; they do not force the headline result by construction. Self-citation, if any, is not load-bearing for the measurement claim. The derivation chain is therefore self-contained against external benchmarks and exhibits no circular reduction of the kind the analyzer flags. Score 0 is the correct honest finding.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 2 invented entities

The central claim rests on standard VLM grounding assumptions (IoU/Acc@0.5 as the right metric, GRPO as a valid RL fine-tuning method, teacher-generated corrective traces as valid supervision) plus the experimental design choices that define the oracle vs. shippable comparison. No new physical entities are postulated; the invented constructs are methodological (IVT loop, honest-trajectory protocol). Free parameters are the usual training and evaluation knobs (iteration budget, reward threshold, stopping rules) rather than quantities fitted to force the mirage conclusion.

free parameters (4)
  • iteration budget / max trajectory length
    Number of refinement steps allowed; determines the size of the set from which the oracle can pick and over which label-free policies are evaluated.
  • Acc@0.5 / IoU success threshold
    Binary success criterion used for the headline +2.4pp number; conventional but still a free evaluation choice.
  • GRPO / RL hyperparameters and IoU reward scaling
    Training knobs that shape whether the model learns generation, verification, or both; not derived from first principles.
  • self-verification confidence threshold for early stopping
    Any confidence-based shippable rule requires a cutoff; the paper reports weak correlation (r≈0.22) but the threshold itself is free.
axioms (4)
  • domain assumption IoU (and Acc@0.5) is an adequate scalar reward and success metric for spatial grounding quality.
    Standard in detection/grounding literature; used both as GRPO reward and as the reported accuracy metric.
  • domain assumption A teacher VLM can turn base-model prediction errors into useful corrective reasoning traces without human annotation.
    Load-bearing for the supervised warm-up phase that produces training data.
  • ad hoc to paper Oracle selection of the trajectory step with highest IoU to ground truth is representative of how iterative systems are "usually reported."
    The mirage claim depends on this characterization of prior evaluation practice; if prior work already used fixed stopping rules, the critique narrows.
  • domain assumption Label-free stopping rules (fixed step, model confidence, etc.) exhaust the set of deployable policies worth comparing.
    Needed to conclude that "the best policy is not to iterate at all."
invented entities (2)
  • Iterative Visual Thinking (IVT) no independent evidence
    purpose: Concrete two-phase training recipe (supervised warm-up on base-model errors + GRPO with IoU) that realizes visual self-correction for bounding-box grounding.
    Methodological construct introduced to study the self-correction loop; not claimed as a new physical or cognitive entity.
  • Honest-trajectory evaluation protocol no independent evidence
    purpose: Require reporting accuracy under fixed label-free stopping policies plus the explicit gap to the oracle best-step score.
    Proposed measurement standard that separates generation capacity from selection/verification capacity.

pith-pipeline@v1.1.0-grok45 · 7793 in / 3180 out tokens · 37180 ms · 2026-07-12T14:10:59.400853+00:00 · methodology

0 comments
read the original abstract

Letting a vision-language model (VLM) think longer at test time has driven much recent progress. A natural way to bring this to spatial grounding is visual self-correction: the model predicts a bounding box, sees it rendered on the image, and refines it over several steps. We build a faithful instance of this idea, Iterative Visual Thinking (IVT), with a two-phase recipe: a supervised warm-up in which the base model's own predictions serve as realistic errors that a teacher VLM turns into corrective reasoning traces (yielding training data without human annotation), followed by GRPO with a simple IoU reward. Measured the way such systems are usually reported, it works: the trained model surpasses the single-shot base by +2.4pp Acc@0.5. We show this gain is a measurement mirage. The reported number silently keeps, per sample, the trajectory step closest to the ground-truth box: an oracle that needs the very answer it predicts. Re-scored under deployable, label-free stopping rules the improvement vanishes, and the best policy is not to iterate at all: stopping at step 0 matches the base and beats every shippable rule. The cause is a verification failure, since the model can generate a better box somewhere in its trajectory but cannot identify it. Self-verification confidence correlates only weakly with correctness (r about 0.22), and a counterfactual overlay shows the loop reacts to the presence of a rendered box rather than its correctness. We distill the lesson into an honest-trajectory evaluation protocol: accuracy under fixed label-free policies plus an explicit oracle-shippable gap.

Figures

Figures reproduced from arXiv: 2606.13156 by Animesh Tripathy, Aswanth Krishnan.

Figure 1
Figure 1. Figure 1: The spatial self-correction gap. Acc@0.5 across training phases. Naively applying iterative visual thinking catastrophically degrades performance (−31pp). SFT warm-up recovers and surpasses the base model across all metrics. GRPO contributes refinement stabil￾ity, reducing per-step IoU degradation by 5×. 2. GRPO Fine-tuning. Starting from the SFT￾initialized policy, we apply Group Relative Policy Optimizat… view at source ↗
Figure 2
Figure 2. Figure 2: Training recipe and inference procedure for IVT. (Top) SFT warm-up: student predictions are interpolated toward the ground truth to form correction trajectories; a teacher VLM generates step-conditioned reasoning traces, which are used to train the student via cross-entropy loss. (Middle) GRPO: N=6 trajectories are sampled per prompt, scored by IoU-based reward (Eq. 3), and used for policy gradient updates… view at source ↗
Figure 3
Figure 3. Figure 3: IVT inference pipeline. Given an image and referring expression, the model predicts a bounding box (Step 0, IoU 0.320), observes its rendered prediction as a red overlay, and iteratively refines through reasoning traces. Over two refinement steps the prediction progressively corrects toward the target (IoU 0.532, ∆IoU = +0.212). GRPO formulation. For each training sample, we generate N=6 complete trajector… view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative examples. (Top) Hard case: the model corrects from the wrong bronze sculpture to the correct one across refinement steps. (Bottom) Fine refinement: an already accurate prediction on a blue vase is tightened through visual feedback [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗

Review history (2 revisions) →

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

34 extracted references · 9 linked inside Pith

  1. [1]

    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 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    Qwen3-VL technical report

    Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, et al. Qwen3-VL technical report. arXiv preprint arXiv:2511.21631, 2025 a

    Pith/arXiv arXiv 2025

  3. [3]

    UniVG-R1 : Reasoning guided universal visual grounding with reinforcement learning

    Sule Bai, Mingxing Li, Yong Liu, Jing Tang, Haoji Zhang, Lei Sun, Xiangxiang Chu, and Yansong Tang. UniVG-R1 : Reasoning guided universal visual grounding with reinforcement learning. arXiv preprint arXiv:2505.14231, 2025 b

    Pith/arXiv arXiv 2025

  4. [4]

    Ground-R1 : Incentivizing grounded visual reasoning via reinforcement learning

    Meng Cao, Haoze Zhao, Can Zhang, Xiaojun Chang, Ian Reid, and Xiaodan Liang. Ground-R1 : Incentivizing grounded visual reasoning via reinforcement learning. arXiv preprint arXiv:2505.20272, 2025

    arXiv 2025

  5. [5]

    Gary Chan, and Hongyang Zhang

    Jierun Chen, Fangyun Wei, Jinjing Zhao, Sizhe Song, Bohuai Wu, Zhuoxuan Peng, S.-H. Gary Chan, and Hongyang Zhang. Revisiting Referring Expression Comprehension Evaluation in the Era of Large Multimodal Models . In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 513--524, 2025 a

    2025

  6. [6]

    Shikra : Unleashing multimodal LLM 's referential dialogue magic

    Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra : Unleashing multimodal LLM 's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023

    Pith/arXiv arXiv 2023

  7. [7]

    Learning only with images: Visual reinforcement learning with reasoning, rendering, and visual feedback

    Yang Chen, Yufan Shen, Wenxuan Huang, Sheng Zhou, Qunshu Lin, Xinyu Cai, Zhi Yu, Jiajun Bu, Botian Shi, and Yu Qiao. Learning only with images: Visual reinforcement learning with reasoning, rendering, and visual feedback. arXiv preprint arXiv:2507.20766, 2025 b

    Pith/arXiv arXiv 2025

  8. [8]

    QLoRA : Efficient finetuning of quantized language models

    Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. QLoRA : Efficient finetuning of quantized language models. Advances in Neural Information Processing Systems, 36, 2023

    2023

  9. [9]

    Ref-adv: Exploring MLLM visual reasoning in referring expression tasks

    Qihua Dong, Kuo Yang, Lin Ju, Handong Zhao, Yitian Zhang, Yizhou Wang, Huimin Zeng, Jianglin Lu, and Yun Fu. Ref-adv: Exploring MLLM visual reasoning in referring expression tasks. In The Fourteenth International Conference on Learning Representations, 2026

    2026

  10. [10]

    GRIT : Teaching MLLMs to think with images

    Yue Fan, Xuehai He, Diji Yang, Kaizhi Zheng, Ching-Chen Kuo, Yuting Zheng, Sravana Jyothi Narayanaraju, Xinze Guan, and Xin Eric Wang. GRIT : Teaching MLLMs to think with images. In NeurIPS, 2025

    2025

  11. [11]

    DeepSeek-R1 : Incentivizing reasoning capability in LLMs via reinforcement learning

    Daya Guo, Dejian Yang, He 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

    Pith/arXiv arXiv 2025

  12. [12]

    LoRA : Low-rank adaptation of large language models

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Sheng Wang, Lu Wang, and Weizhu Chen. LoRA : Low-rank adaptation of large language models. In ICLR, 2022

    2022

  13. [13]

    Large language models cannot self-correct reasoning yet

    Jie Huang, Xinyun Chen, Swaroop Mishra, Huaixiu Steven Zheng, Adams Wei Yu, Xinying Song, and Denny Zhou. Large language models cannot self-correct reasoning yet. In ICLR, 2024

    2024

  14. [14]

    Vision- R1 : Incentivizing reasoning capability in multimodal large language models

    Wenxuan Huang, Bohan Jia, Zijie Zhai, Shaosheng Cao, Zheyu Ye, Fei Zhao, Zhe Xu, Xu Tang, Yao Hu, and Shaohui Lin. Vision- R1 : Incentivizing reasoning capability in multimodal large language models. In ICLR, 2026

    2026

  15. [15]

    MDETR -- modulated detection for end-to-end multi-modal understanding

    Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. MDETR -- modulated detection for end-to-end multi-modal understanding. In ICCV, 2021

    2021

  16. [16]

    Can large vision-language models correct semantic grounding errors by themselves? In CVPR, 2025

    Yuan-Hong Liao, Rafid Mahmood, Sanja Fidler, and David Acuna. Can large vision-language models correct semantic grounding errors by themselves? In CVPR, 2025

    2025

  17. [17]

    Microsoft COCO : Common objects in context

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Doll \'a r, and C Lawrence Zitnick. Microsoft COCO : Common objects in context. In ECCV, 2014

    2014

  18. [18]

    Grounding DINO : Marrying DINO with grounded pre-training for open-set object detection

    Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding DINO : Marrying DINO with grounded pre-training for open-set object detection. In ECCV, 2024

    2024

  19. [19]

    Self-refine: Iterative refinement with self-feedback

    Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, et al. Self-refine: Iterative refinement with self-feedback. Advances in Neural Information Processing Systems, 36, 2023

    2023

  20. [20]

    Generation and comprehension of unambiguous object descriptions

    Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In CVPR, 2016

    2016

  21. [21]

    Training language models to follow instructions with human feedback

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35: 0 27730--27744, 2022

    2022

  22. [22]

    Grounding multimodal large language models to the world

    Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, Qixiang Ye, and Furu Wei. Grounding multimodal large language models to the world. In ICLR, 2024

    2024

  23. [23]

    CogCoM : A visual language model with chain-of-manipulations reasoning

    Ji Qi, Ming Ding, Weihan Wang, Yushi Bai, Qingsong Lv, Wenyi Hong, Bin Xu, Lei Hou, Juanzi Li, Yuxiao Dong, and Jie Tang. CogCoM : A visual language model with chain-of-manipulations reasoning. In ICLR, 2025

    2025

  24. [24]

    Tarr, Aviral Kumar, and Katerina Fragkiadaki

    Gabriel Sarch, Snigdha Saha, Naitik Khandelwal, Ayush Jain, Michael J. Tarr, Aviral Kumar, and Katerina Fragkiadaki. Grounded reinforcement learning for visual reasoning. arXiv preprint arXiv:2505.23678, 2025

    Pith/arXiv arXiv 2025

  25. [25]

    Objects365 : A large-scale, high-quality dataset for object detection

    Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365 : A large-scale, high-quality dataset for object detection. In ICCV, 2019

    2019

  26. [26]

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, Y.K. Li, Y. Wu, and Daya Guo. DeepSeekMath : Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300, 2024

    Pith/arXiv arXiv 2024

  27. [27]

    VLM-R1 : A stable and generalizable R1 -style large vision-language model

    Haozhan Shen, Peng Liu, Jingcheng Li, Chunxin Fang, Yibo Ma, Jiajia Liao, Qiaoli Shen, Zilun Zhang, Kangjia Zhao, Qianqian Zhang, Ruochen Xu, and Tiancheng Zhao. VLM-R1 : A stable and generalizable R1 -style large vision-language model. arXiv preprint arXiv:2504.07615, 2025

    Pith/arXiv arXiv 2025

  28. [28]

    Reflexion: Language agents with verbal reinforcement learning

    Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. Reflexion: Language agents with verbal reinforcement learning. Advances in Neural Information Processing Systems, 36, 2023

    2023

  29. [29]

    Qwen2-VL : Enhancing vision-language model's perception of the world at any resolution

    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

    Pith/arXiv arXiv 2024

  30. [30]

    Chain-of-thought prompting elicits reasoning in large language models

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc V Le, and Denny Zhou. Chain-of-thought prompting elicits reasoning in large language models. In NeurIPS, 2022

    2022

  31. [31]

    Visual planning: Let's think only with images

    Yi Xu, Chengzu Li, Han Zhou, Xingchen Wan, Caiqi Zhang, Anna Korhonen, and Ivan Vuli\' c . Visual planning: Let's think only with images. In ICLR, 2026

    2026

  32. [32]

    Ferret: Refer and ground anything anywhere at any granularity

    Haoxuan You, Haotian Zhang, Zhe Gan, Xianzhi Du, Bowen Zhang, Zirui Wang, Liangliang Cao, Shih-Fu Chang, and Yinfei Yang. Ferret: Refer and ground anything anywhere at any granularity. In ICLR, 2024

    2024

  33. [33]

    Critic- V : VLM critics help catch VLM errors in multimodal reasoning

    Di Zhang, Jingdi Lei, Junxian Li, Xunzhi Wang, Yujie Liu, Zonglin Yang, Jiatong Li, Weida Wang, Suorong Yang, Jianbo Wu, Peng Ye, Wanli Ouyang, and Dongzhan Zhou. Critic- V : VLM critics help catch VLM errors in multimodal reasoning. In CVPR, 2025 a

    2025

  34. [34]

    R1-VL : Learning to reason with multimodal large language models via step-wise group relative policy optimization

    Jingyi Zhang, Jiaxing Huang, Huanjin Yao, Shunyu Liu, Xikun Zhang, Shijian Lu, and Dacheng Tao. R1-VL : Learning to reason with multimodal large language models via step-wise group relative policy optimization. In ICCV, 2025 b

    2025