Pith. sign in

REVIEW 4 major objections 4 minor 27 references

LenGuard-GPC claims that adding a guided-prompt token-wise KL consistency reward and a staged length bonus to group-relative policy optimization improves multi-view spatial reasoning accuracy while reducing response length.

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 · deepseek-v4-flash

2026-08-01 18:35 UTC pith:3ABSEYMD

load-bearing objection The paper has a plausible reward-shaping idea and directionally positive results, but its only ablation doesn't test the KL reward, and the ablation text contradicts its own table. the 4 major comments →

arxiv 2607.17243 v1 pith:3ABSEYMD submitted 2026-07-19 cs.AI

LenGuard-GPC: Length Guarding with Guided-Prompt Consistency for Spatial Reasoning Reinforce Learning

Xingjian Tao , Yiwei Wang , Yujun Cai , Jing Tang This is my paper
classification cs.AI
keywords multi-view spatial reasoningreinforcement learningdense rewardKL divergenceguided promptschain-of-thought lengthlength guardingvision-language models
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.

Multi-view spatial reasoning — deciding spatial relations from several images of the same scene — is a setting where chain-of-thought reasoning tends to grow verbose without becoming more accurate. The paper argues that outcome-only reinforcement learning gives a model too little information about where a reasoning trace goes wrong, so it adds a dense signal: for each sampled response, the policy's token predictions under the original prompt are compared with its predictions under a prompt that includes a one-sentence reasoning hint, and the summed KL divergence between the two distributions enters the reward. A staged length bonus — zero below 160 tokens, rising to a plateau between 360 and 768 tokens — counters the tendency of a token-sum penalty to drive responses to near-zero length. On six benchmarks the method reports 57.3% average accuracy versus 55.3% for the plain group-relative baseline, with larger gains on in-domain multi-view benchmarks and shorter average responses. If the claim is right, dense distributional consistency is a practical process-level reward that needs no extra reward model and no extra inference-time computation.

Core claim

The central claim is that group-relative policy optimization for multi-view spatial reasoning fails not for lack of exploration but for lack of process-level feedback: a final-answer reward cannot say which step in a long trace mislocates an object or breaks cross-view correspondence. The proposed remedy is to compute, for every sampled trajectory, the token-wise KL divergence between the policy's teacher-forced predictive distributions under the standard prompt and under a guided prompt, and to penalize the accumulated sum. The guidance is a single sentence generated per training instance by an external chat model; it is used only at training time. Because the KL sum grows with length, the

What carries the argument

The load-bearing object is the guided-prompt consistency reward. For a sampled trajectory y, the model computes two teacher-forced next-token distributions at each position t: p_std_t = πθ(· | x, y<t) under the original prompt and p_gp_t = πθ(· | x_gp, y<t) under the guided prompt, where x_gp appends a one-sentence reasoning hint h to the question. Teacher-forced here means each next-token distribution is conditioned on the already-sampled prefix, not on the model's own continuation. The per-token KL divergence D_t = KL(p_std_t || p_gp_t) is summed over the trajectory and enters the reward as r_KL = -λ_KL Σ_t D_t. Because this sum grows with token count, the second component, the staged leng

Load-bearing premise

The reward is only as informative as the one-sentence guidance used to build the guided prompt: if that guidance is not consistently more helpful than the standard prompt, or if the token-wise KL distance mostly reflects phrasing and style rather than reasoning quality, then the consistency term is not a dense signal for spatial reasoning and the reported accuracy gains could come from the length bonus and the final-answer reward alone.

What would settle it

Train the same pipeline with the consistency term removed (keeping the length bonus and task reward) and compare average accuracy: if it stays near 57.3%, the guided-prompt mechanism is not the source of the gain. A second check is to replace the per-instance guidance sentence with a deliberately irrelevant or misleading one; if accuracy does not fall, the consistency term is not carrying reasoning-specific information.

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

If this is right

  • Group-relative policy optimization can carry dense process feedback without a separate reward model or auxiliary supervised loss, because the consistency score reuses the policy itself under a second prompt condition.
  • Average accuracy on the six reported benchmarks improves by 2.0 points over the plain group-relative baseline and by 5.9 points over the original instruction-tuned model; the largest gains are in cross-view positional reasoning, suggesting the reward targets the main failure mode.
  • Token cost drops by about 14.9% on MMSI and 38.1% on ViewSpatial, so reasoning becomes more compact while accuracy rises; the method does not simply truncate answers.
  • The reward does not overfit to the training domain: out-of-domain spatial benchmarks show small positive gains over the baseline.
  • Training overhead is roughly 1.6x the plain pipeline for the extra teacher-forced pass, and the shorter generated trajectories offset part of that cost.

Where Pith is reading between the lines

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

  • If the KL distance is mostly measuring how confidently the model shifts under any extra instruction, a fixed generic 'think step by step' suffix might reproduce the gain; that would make the method cheaper and remove the per-instance external guidance.
  • A natural extension is to use the per-token KL values as logged process supervision: high-KL positions may mark the reasoning steps where the model most disagrees with guidance, which could identify hard views or erroneous correspondences in the training data.
  • The staged length guard effectively defines a fixed length budget; one could make the boundaries instance-dependent by predicting task difficulty from the number of views, so easier cases get a shorter band and harder cases a longer one.
  • The same reward shape applies to any multimodal task with long contexts and sparse answers, such as video question answering or multi-document visual grounding, but the guidance content would need to be re-derived per task, and the paper does not demonstrate those settings.

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

4 major / 4 minor

Summary. The paper proposes LenGuard-GPC, an extension of GRPO for multi-view spatial reasoning. For each sampled trajectory it computes a token-wise KL divergence between the model's predictive distributions under a standard prompt and a per-instance guided prompt, and uses the negative accumulated KL as a dense reward. It also adds a staged length bonus to prevent the model from collapsing to overly short responses. Experiments on six benchmarks report that LenGuard-GPC improves average accuracy to 57.3% versus 55.3% for vanilla GRPO and 51.4% for the base Qwen3-VL-8B-Instruct, while reducing response length on MMSI and ViewSpatial. The central claim is that the guided-prompt KL reward plus the staged length guard jointly improve reasoning accuracy and token efficiency.

Significance. If validated, the method is a useful contribution: it provides a dense, process-level reward for GRPO without an auxiliary supervised loss, adds no inference-time overhead, and is motivated for long-context multimodal reasoning. The paper includes several strengths: evaluation on six benchmarks, including out-of-domain sets; a fine-grained MMSI breakdown; an explicit statement of training-time overhead (1.6x); and a code repository link. However, the main ablation is internally inconsistent and does not isolate the guided-prompt KL reward, several reward coefficients are never reported, and the guided prompt itself is a black box. These issues leave the central causal claim unsupported in the current manuscript.

major comments (4)
  1. [Section 4.3 / Table 3] The ablation study is internally inconsistent. The text describes ablations named 'Free format Reasoning + RL', 'w/o GRPO', and 'w/o Format Reward' with values 35.4, 33.4, 32.4, and 35.0, and refers to a 'structured two-stage output' and 'spatial pre-thinking stage' that are not defined in Section 3. Table 3, however, lists only 'Vanilla GRPO', 'w/o Staged Length Reward', and 'Full Method'. The only component ablation present in Table 3 removes the staged length reward, not the guided-prompt KL reward. Therefore the paper's main contribution—the KL-based guided-prompt consistency reward—is never causally isolated. The reported gain of 36.1 versus 32.2 could in principle come from the task reward and length bonus alone. This needs to be fixed with a proper ablation that removes r_KL and/or replaces h with a control prompt.
  2. [Eq. (5) / Eq. (9) / Section 4.1] The reward coefficients are incompletely specified. Eq. (9) defines r(y) = λ_task r_task(y) − λ_KL S_KL(y) + λ_len r_len(T), and Eq. (5) uses the same coefficients, but λ_task and λ_len are never reported. Section 4.1 gives λ_KL = 5e-3 and the length-guard parameters, but not λ_task or λ_len. Likewise, Eq. (4) and Eq. (7) introduce λ_acc and λ_format without reporting their values, and the phrase 'we use a format reward of 0.2' leaves unclear whether this is λ_format times r_format or a separate absolute reward. Without these coefficients the reward scale and the relative contribution of each term cannot be checked, and the experiments cannot be reproduced.
  3. [Section 3.2] The guided prompt h is a black box. The paper states that h is 'a one-sentence guidance generated by GPT-5.3-Chat for each training instance', but provides no examples, no generation template, and no discussion of what kinds of guidance were produced. Since r_KL measures consistency between the standard and guided conditions, the informativeness of the reward depends entirely on whether h is actually more helpful than the standard prompt. Without any detail or control, a reader cannot tell whether the KL term reflects reasoning quality, prompt style, or even accidental answer leakage. Please provide the prompt template, representative h instances, and an analysis of what h contains.
  4. [Section 4.2 / Table 1] The empirical comparison reports no error bars, standard deviations, or number of seeds. Several of the claimed improvements over vanilla GRPO are small: +0.7 on 3DSR, +1.1 on BLINK, +0.3 on CV-Bench, and the average gain is +2.0 points. Without variance estimates or multiple seeds, these differences—especially the out-of-domain ones—cannot be distinguished from noise. At minimum, report mean and variance over at least three seeds for the main accuracy and token-length results, and ideally a paired test across benchmarks.
minor comments (4)
  1. [Figure 2] The figure labels appear corrupted in the text (strings such as '/uni00000039 /uni00000044/uni00000051...' are visible in the manuscript). Please replace with a clean rendered figure.
  2. [Section 4.2, final paragraph] There is a stray text fragment 'uality here is therefore not a function of length; what matters is whether the trace resolves cross-view correspondence at all.' that appears to have been inserted in the middle of the training-time discussion. Please remove or relocate it.
  3. [Eq. (4) and Eq. (7)] Equations (4) and (7) are duplicates. Define the task-reward coefficients once, at first use, and give their values in the implementation section.
  4. [Section 3.4.1, Eq. (6)] The advantage normalization formula is standard, but it would be clearer to state whether the same epsilon appears in the denominator and whether rewards are normalized per group of G trajectories sampled for each instance; currently the notation is implicit.

Circularity Check

0 steps flagged

No circular derivation: LenGuard-GPC's claims are supported by external benchmark measurements, not by construction or self-citation.

full rationale

The paper's central claims are empirical: accuracy is measured on six external benchmarks (MMSI-Bench, ViewSpatial, MindCube, 3DSR, BLINK, CV-Bench) and compared against the base model and vanilla GRPO. The reward design r(y) = λ_task r_task - λ_KL S_KL + λ_len r_len(T) does not define accuracy in terms of the KL term; r_task is computed from external answer correctness, and the KL/length terms are auxiliary rewards. The 'guided-prompt consistency' KL term compares the same policy under standard and guided prompts; this is a self-distillation-style regularizer, and while it is self-referential in the sense of comparing the policy to itself, it is not circular validation because the reported gains are not derived from the KL term but measured on held-out benchmarks. The length reduction is partly a designed consequence of the token-sum KL penalty and staged length bonus, but the paper openly states this design intention and reports the magnitude empirically; it does not present the length reduction as an independent prediction. Self-citations (Tao et al. 2025, 2026a, 2026b) appear only in related-work context and are not load-bearing; the core method cites external work (Self On-Policy Distillation, Hübotter et al. 2026) for inspiration. The ablation inconsistency in §4.3 (text describes a 'structured two-stage output' with numbers 35.4/33.4/32.4/35.0, while Table 3 lists different rows) is a serious reporting and causal-identification problem for the claim that the KL reward contributes, but it is a correctness and reproducibility issue, not a circularity step: no equation is shown to reduce to its input, and no fitted parameter is renamed as a prediction.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

No new physical or model-level entities are introduced; the guided prompt and staged length bonus are training constructs, not independent falsifiable entities.

free parameters (5)
  • λ_KL (guided-prompt KL penalty coefficient) = 5e-3
    Chosen by hand; controls the strength of the consistency reward (Section 4.1).
  • λ_task, λ_len (reward coefficients) = not reported
    Equation (9) includes these weights but the paper never states their values.
  • l1=160, l2=360, l3=768 (length guard boundaries) = 160/360/768 tokens
    Hand-set staging boundaries.
  • r1=0.25, r2=0.30 (length bonus levels) = 0.25/0.30
    Hand-set bonus values; slope 2.5e-4 per token derived from these.
  • r_format = 0.2 (format reward) = 0.2
    Fixed format reward weight (Section 4.1).
axioms (4)
  • domain assumption Guided prompt h (one-sentence GPT-5.3-Chat guidance) is consistently informative for spatial reasoning
    Section 3.2 builds the entire dense reward on this h, but no independent validation is shown.
  • domain assumption Teacher-forcing KL between standard and guided prompts measures reasoning quality
    Equation (1) assumes lower KL to the guided condition is better; no evidence connects it to correctness beyond the final reward.
  • domain assumption GRPO group-relative advantage normalization is a valid optimizer for this composite reward
    Section 3.4.1 applies standard GRPO unchanged.
  • domain assumption The 13k-sample VST-500K subset is representative and does not overlap test benchmarks
    Section 4.1 describes sampling but no filtering or contamination check.

pith-pipeline@v1.3.0-alltime-deepseek · 15667 in / 11860 out tokens · 107434 ms · 2026-08-01T18:35:33.965333+00:00 · methodology

0 comments
read the original abstract

Multi-view spatial reasoning requires vision-language models to compare visual evidence across images, align object correspondences, and infer spatial relations over long visual contexts, a setting where chain-of-thought reasoning tends to grow verbose without becoming more accurate. Reinforcement learning with verifiable rewards is a natural fit for this task, but standard GRPO reward relies on sparse outcome-level feedback and gives no signal about where a reasoning trajectory goes wrong, nor any control over its length. We propose LenGuard-GPC, a dense reward framework that addresses both problems together. For each sampled trajectory, it compares the token-wise predictive distributions under a standard prompt and a guided prompt, and uses the resulting token-sum KL divergence as a dense reward signal. Since this KL penalty accumulates over tokens and would otherwise reward shorter responses regardless of their quality, we introduce a staged length bonus that keeps reasoning length within a controlled range without simply encouraging brevity. On six multi-view spatial reasoning benchmarks, LenGuard-GPC improves accuracy over vanilla GRPO while reducing average response length.

Figures

Figures reproduced from arXiv: 2607.17243 by Jing Tang, Xingjian Tao, Yiwei Wang, Yujun Cai.

Figure 1
Figure 1. Figure 1: Overview of LenGuard-GPC for multi-view spatial reasoning. Within the staged interval [l1, l2], LenGuard￾GPC adaptively balances the negative KL-based consistency reward and the length bonus. When the mean per-token KL is high, the consistency reward dominates and encourages more concise reasoning. When the mean per-token KL is low, the length bonus becomes relatively stronger, preventing the model from co… view at source ↗
Figure 2
Figure 2. Figure 2 [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗

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

27 extracted references · 19 linked inside Pith

  1. [4]

    Video-r1: Reinforcing video reasoning in mllms.arXiv preprint arXiv:2503.21776,

    Kaituo Feng, Kaixiong Gong, Bohao Li, Zonghao Guo, Yibing Wang, Tianshuo Peng, Junfei Wu, Xiaoying Zhang, Benyou Wang, and Xiangyu Yue. Video-r1: Reinforcing video reasoning in mllms.arXiv preprint arXiv:2503.21776,

    Pith/arXiv arXiv

  2. [6]

    Zirun Guo, Minjie Hong, and Tao Jin

    URL https://openreview.net/forum?id= kxnoqaisCT. Zirun Guo, Minjie Hong, and Tao Jin. Observe-r1: Unlocking reasoning abilities of mllms with dynamic progressive reinforcement learning.arXiv preprint arXiv:2505.12432,

    Pith/arXiv arXiv

  3. [7]

    Don’t overthink it

    Michael Hassid, Gabriel Synnaeve, Yossi Adi, and Roy Schwartz. Don’t overthink it. preferring shorter thinking chains for improved llm reasoning.arXiv preprint arXiv:2505.17813,

    arXiv

  4. [8]

    Vision-r1: Incentivizing reasoning capability in multimodal large language models.arXiv preprint arXiv:2503.06749,

    Preprint : Under Review Wenxuan Huang, Bohan Jia, Zijie Zhai, Shaosheng Cao, Zheyu Ye, Fei Zhao, Zhe Xu, Yao Hu, and Shaohui Lin. Vision-r1: Incentivizing reasoning capability in multimodal large language models.arXiv preprint arXiv:2503.06749,

    Pith/arXiv arXiv

  5. [9]

    Reinforcement learning via self-distillation

    Jonas Hübotter, Frederike Lübeck, Lejs Behric, Anton Baumann, Marco Bagatella, Daniel Marta, Ido Hakimi, Idan Shenfeld, Thomas Kleine Buening, Carlos Guestrin, et al. Reinforcement learning via self-distillation. arXiv preprint arXiv:2601.20802,

    Pith/arXiv arXiv

  6. [11]

    Spatialmosaic: A multiview vlm dataset for partial visibility.arXiv preprint arXiv:2512.23365,

    Kanghee Lee, Injae Lee, Minseok Kwak, Kwonyoung Ryu, Jungi Hong, and Jaesik Park. Spatialmosaic: A multiview vlm dataset for partial visibility.arXiv preprint arXiv:2512.23365,

    Pith/arXiv arXiv

  7. [12]

    Topviewrs: Vision- language models as top-view spatial reasoners

    Chengzu Li, Caiqi Zhang, Han Zhou, Nigel Collier, Anna Korhonen, and Ivan Vulić. Topviewrs: Vision- language models as top-view spatial reasoners. InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pp. 1786–1807, 2024a. Dingming Li, Hongxing Li, Zixuan Wang, Yuchen Yan, Hang Zhang, Siqi Chen, Guiyang Hou, Shengpei Ji...

    arXiv 2024

  8. [13]

    Concise thoughts: Impact of output length on llm reasoning and cost.arXiv preprint arXiv:2407.19825,

    Sania Nayab, Giulio Rossolini, Marco Simoni, Andrea Saracino, Giorgio Buttazzo, Nicolamaria Manes, and Fabrizio Giacomelli. Concise thoughts: Impact of output length on llm reasoning and cost.arXiv preprint arXiv:2407.19825,

    Pith/arXiv arXiv

  9. [14]

    Spacer: Reinforcing mllms in video spatial reasoning.arXiv preprint arXiv:2504.01805,

    Kun Ouyang, Yuanxin Liu, Haoning Wu, Yi Liu, Hao Zhou, Jie Zhou, Fandong Meng, and Xu Sun. Spacer: Reinforcing mllms in video spatial reasoning.arXiv preprint arXiv:2504.01805,

    Pith/arXiv arXiv

  10. [15]

    Image captioning evaluation in the age of multimodal llms: Challenges and future perspectives.arXiv preprint arXiv:2503.14604,

    Sara Sarto, Marcella Cornia, and Rita Cucchiara. Image captioning evaluation in the age of multimodal llms: Challenges and future perspectives.arXiv preprint arXiv:2503.14604,

    Pith/arXiv arXiv

  11. [16]

    Aligning large multimodal models with factually augmented rlhf

    Zhiqing Sun, Sheng Shen, Shengcao Cao, Haotian Liu, Chunyuan Li, Yikang Shen, Chuang Gan, Liangyan Gui, Yu-Xiong Wang, Yiming Yang, et al. Aligning large multimodal models with factually augmented rlhf. InFindings of the Association for Computational Linguistics: ACL 2024, pp. 13088–13110,

    2024

  12. [17]

    Understanding gui agent localization biases through logit sharpness

    Preprint : Under Review Xingjian Tao, Yiwei Wang, Yujun Cai, Zhicheng Yang, and Jing Tang. Understanding gui agent localization biases through logit sharpness. InFindings of the Association for Computational Linguistics: EMNLP 2025, pp. 23361–23374,

    2025

  13. [18]

    Mitigating coordinate prediction bias from positional encoding failures

    Xingjian Tao, Yiwei Wang, Yujun Cai, Yihong Luo, Kai Han, and Jing Tang. Mitigating coordinate prediction bias from positional encoding failures. InFindings of the Association for Computational Linguistics: ACL 2026, pp. 20635–20650, 2026a. Xingjian Tao, Yiwei Wang, Yujun Cai, Yifan Song, and Jing Tang. Viewfusion: Structured spatial thinking chains for m...

    arXiv 2026

  14. [19]

    Visioncube: 3d-aware vision-language model for multi-step spatial reasoning

    Feiyang Wang, Nan Luo, and Wangyu Wu. Visioncube: 3d-aware vision-language model for multi-step spatial reasoning. InProceedings of the Computer Vision and Pattern Recognition Conference, pp. 3270–3279, 2025a. Jiacong Wang, Zijian Kang, Haochen Wang, Haiyong Jiang, Jiawen Li, Bohong Wu, Ya Wang, Jiao Ran, Xiao Liang, Chao Feng, et al. Vgr: Visual grounded...

    Pith/arXiv arXiv

  15. [20]

    Video-str: Reinforcing mllms in video spatio-temporal reasoning with relation graph.arXiv preprint arXiv:2510.10976, 2025c

    Wentao Wang, Heqing Zou, Tianze Luo, Rui Huang, Yutian Zhao, Zhuochen Wang, Hansheng Zhang, Chengwei Qin, Yan Wang, Lin Zhao, et al. Video-str: Reinforcing mllms in video spatio-temporal reasoning with relation graph.arXiv preprint arXiv:2510.10976, 2025c. Ziyang Wang, Jaehong Yoon, Shoubin Yu, Md Mohaiminul Islam, Gedas Bertasius, and Mohit Bansal. Video...

    arXiv 2025

  16. [21]

    Diankun Wu, Fangfu Liu, Yi-Hsin Hung, and Yueqi Duan

    URLhttps://openreview.net/forum?id=_VjQlMeSB_J. Diankun Wu, Fangfu Liu, Yi-Hsin Hung, and Yueqi Duan. Spatial-mllm: Boosting mllm capabilities in visual-based spatial intelligence.arXiv preprint arXiv:2505.23747, 2025a. Haoning Wu, Xiao Huang, Yaohui Chen, Ya Zhang, Yanfeng Wang, and Weidi Xie. Spatialscore: Towards comprehensive evaluation for spatial in...

    Pith/arXiv arXiv

  17. [22]

    V-dpo: Mitigating hallucination in large vision language models via vision-guided direct preference optimization.arXiv preprint arXiv:2411.02712,

    Preprint : Under Review Yuxi Xie, Guanzhen Li, Xiao Xu, and Min-Yen Kan. V-dpo: Mitigating hallucination in large vision language models via vision-guided direct preference optimization.arXiv preprint arXiv:2411.02712,

    Pith/arXiv arXiv

  18. [23]

    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,

    Pith/arXiv arXiv

  19. [24]

    Humanomniv2: From understanding to omni-modal reasoning with context.arXiv preprint arXiv:2506.21277, 2025a

    Qize Yang, Shimin Yao, Weixuan Chen, Shenghao Fu, Detao Bai, Jiaxing Zhao, Boyuan Sun, Bowen Yin, Xihan Wei, and Jingren Zhou. Humanomniv2: From understanding to omni-modal reasoning with context.arXiv preprint arXiv:2506.21277, 2025a. Rui Yang, Ziyu Zhu, Yanwei Li, Jingjia Huang, Shen Yan, Siyuan Zhou, Zhe Liu, Xiangtai Li, Shuangye Li, Wenqian Wang, et ...

    Pith/arXiv arXiv

  20. [25]

    Rlhf-v: Towards trustworthy mllms via behavior alignment from fine-grained correctional human feedback

    Tianyu Yu, Yuan Yao, Haoye Zhang, Taiwen He, Yifeng Han, Ganqu Cui, Jinyi Hu, Zhiyuan Liu, Hai-Tao Zheng, Maosong Sun, et al. Rlhf-v: Towards trustworthy mllms via behavior alignment from fine-grained correctional human feedback. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 13807–13816, 2024a. Tianyu Yu, Haoye Z...

    Pith/arXiv arXiv

  21. [26]

    Spacemind: Camera-guided modality fusion for spatial reasoning in vision-language models.arXiv preprint arXiv:2511.23075,

    Ruosen Zhao, Zhikang Zhang, Jialei Xu, Jiahao Chang, Dong Chen, Lingyun Li, Weijian Sun, and Zizhuang Wei. Spacemind: Camera-guided modality fusion for spatial reasoning in vision-language models.arXiv preprint arXiv:2511.23075,

    arXiv

  22. [27]

    Internvl3: Exploring advanced training and test-time recipes for open-source multimodal models.arXiv preprint arXiv:2504.10479, 2025

    Jinguo Zhu, Weiyun Wang, Zhe Chen, Zhaoyang Liu, Shenglong Ye, Lixin Gu, Hao Tian, Yuchen Duan, Weijie Su, Jie Shao, et al. Internvl3: Exploring advanced training and test-time recipes for open-source multimodal models.arXiv preprint arXiv:2504.10479, 2025

    Pith/arXiv arXiv 2025

  23. [2022]

    Qwen technical report.arXiv preprint arXiv:2309.16609,

    Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report.arXiv preprint arXiv:2309.16609,

    Pith/arXiv arXiv

  24. [2023]

    Qwen3-vl technical report, 2025a

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

    Pith/arXiv arXiv

  25. [2024]

    Spatial reasoning with vision-language models in ego-centric multi-view scenes.arXiv preprint arXiv:2509.06266,

    Mohsen Gholami, Ahmad Rezaei, Zhou Weimin, Sitong Mao, Shunbo Zhou, Yong Zhang, and Mohammad Akbari. Spatial reasoning with vision-language models in ego-centric multi-view scenes.arXiv preprint arXiv:2509.06266,

    arXiv

  26. [2025]

    Vlm-3r: Vision-language models augmented with instruction-aligned 3d reconstruction.arXiv preprint arXiv:2505.20279,

    Zhiwen Fan, Jian Zhang, Renjie Li, Junge Zhang, Runjin Chen, Hezhen Hu, Kevin Wang, Huaizhi Qu, Dilin Wang, Zhicheng Yan, et al. Vlm-3r: Vision-language models augmented with instruction-aligned 3d reconstruction.arXiv preprint arXiv:2505.20279,

    Pith/arXiv arXiv

  27. [2026]

    Taco: Think-answer consistency for optimized long-chain reasoning and efficient data learning via reinforcement learning in lvlms.arXiv preprint arXiv:2505.20777,

    Zhehan Kan, Yanlin Liu, Kun Yin, Xinghua Jiang, Xin Li, Haoyu Cao, Yinsong Liu, Deqiang Jiang, Xing Sun, Qingmin Liao, et al. Taco: Think-answer consistency for optimized long-chain reasoning and efficient data learning via reinforcement learning in lvlms.arXiv preprint arXiv:2505.20777,

    Pith/arXiv arXiv