Pith. sign in

REVIEW 3 major objections 4 minor 17 references

Latent visual reasoning can be trained from answer labels alone, without bounding boxes or sketches, and still beat methods that use such extra supervision.

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

2026-08-05 00:22 UTC pith:MGNCP2ZM

load-bearing objection A serious empirical pipeline for VQA-only latent visual reasoning, but the utility-scoring signal is weaker than the story suggests and the paper needs replication details before I'd bet on it. the 3 major comments →

arxiv 2608.00743 v1 pith:MGNCP2ZM submitted 2026-08-01 cs.CV

LUT: Latent Utility Training for Visual Reasoning

Jiaxuan Kang , Siyu Chen , Mingda Li , Mingjie Liu , Tianyue Wang , Zhaoyang Wei , Yongheng Zhang , Yanchao Hao
show 1 more author
Zheng Wei
This is my paper
classification cs.CV
keywords latent visual reasoninglatent utilityutility-aware distillationlatent curriculum learninglatent policy optimizationattention attributionVQA trainingmultimodal LLM
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.

The paper proposes LUT, a two-stage training framework that teaches a multimodal LLM to reason in latent space using only ordinary image-question-answer triples. The key claim is that answer usefulness, measured at both trajectory and step levels, can substitute for the expensive intermediate supervision (bounding boxes, sketches, rationales) used by prior latent reasoning methods. If true, this makes latent visual reasoning scalable to standard VQA data and cheaper to deploy, since inference uses a fixed small set of latent steps instead of long text chains. Experiments report LUT outperforming latent reasoning baselines on perception-intensive benchmarks and coming within 1.7 points of a latent-text interleaved method while emitting far fewer tokens.

Core claim

The central discovery is that latent reasoning can be driven by answer-utility signals rather than externally shaped latent targets. LUT first trains a teacher under a visual bottleneck mask, forcing answer-relevant visual information to pass through latent states, then scores each teacher trajectory by its information gain over a random latent control using a frozen judge (Δfair). Only trajectories above threshold are distilled into a student, ordered in a low-to-high utility curriculum with cosine alignment to teacher hidden states. A second stage, LAPO, treats latent states as continuous actions, estimates per-step answer relevance from attention attribution, and reweights the policy adva

What carries the argument

Two coupled mechanisms carry the argument. (1) Utility-Aware Latent Distillation SFT: a visual bottleneck mask forces the teacher's answer to depend on latent states, a decorrelation regularizer keeps latent steps diverse, and a frozen judge computes Δfair = S(Z_T) − S(Z_R), the log-probability gap between teacher latent and random latent, to filter and rank trajectories for curriculum distillation. (2) LAPO: latent states are treated as Gaussian-distributed continuous actions (following the VLPO formulation), and an answer-to-latent attention attribution score per step is mapped to gates g_i in [1.0, 2.5] that reweight the sequence advantage, so steps most associated with the correct answer

Load-bearing premise

The whole pipeline depends on the frozen base model's Δfair score being a reliable measure of which teacher latent trajectories will actually help the student, even though that judge has never seen the bottleneck-masked teacher and the student is a fine-tuned version of the same judge.

What would settle it

Run the SFT stage on the same 60K VQA samples but replace Δfair-based trajectory selection with an equal-size random retention of teacher trajectories (keeping the same curriculum split by random order). If the average score across VStar, HRBench4K, HRBench8K, and MME-RealWorld-Lite does not drop by the roughly 0.66 points the paper attributes to filtering (68.45 vs 67.79), then the information-gain criterion is not the active mechanism.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • If LUT's result holds, latent reasoning methods no longer need bounding-box, sketch, or auxiliary-image annotations, removing the main scaling bottleneck for this line of work.
  • A fixed latent budget of 8 steps suffices across all four benchmarks after LAPO, suggesting that answer-utility optimization can make latent reasoning compact and predictable at inference time.
  • The Δfair selection filter, which removes only 6.5% of trajectories, implies that a small amount of poorly scoring supervision can mislead latent distillation; trajectory-level quality checks may become standard practice.
  • The attribution-gated advantage reweighting improves over both GRPO and VLPO, indicating that fine-grained, per-latent credit assignment is a productive direction for latent policy optimization.
  • Latent-only reasoning can compete with latent-text interleaved methods (within 1.7 points of SCOLAR) while using roughly a tenth of the output tokens, shifting the cost–accuracy tradeoff in favor of latent-space computation.

Where Pith is reading between the lines

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

  • The judge for Δfair is a frozen base model derived from the same family as the student; a natural extension would be to test whether a judge from a different model family or a committee of judges yields a more transferable trajectory ranking, since the paper does not explore judge–student mismatch.
  • The curriculum orders samples by judged utility, but the paper does not analyze whether the benefit comes from the ordering per se or from the implicit reweighting of data; training on the higher-utility group with frequency proportional to Δfair might achieve similar gains with a simpler procedure.
  • The RAR analysis shows that high-attribution latent steps concentrate attention on task-relevant regions (RAR 11.25 vs 0.81 for low-attribution steps), suggesting that latent steps specialize; an untested extension is to adapt the latent budget per sample based on attribution confidence, rather than using a fixed 8 steps.
  • Because the method relies on answer cross-entropy and binary accuracy rewards, it may inherit biases in the VQA labels; training on deliberately misleading or ambiguous VQA pairs could reveal whether Δfair and attribution gates robustly detect genuinely answer-useful latents or merely fit label idiosyncrasies.

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 / 4 minor

Summary. The paper proposes LUT, a two-stage training framework for latent visual reasoning in multimodal LLMs, trained only on standard VQA pairs. In the first stage, a teacher model explores latent trajectories under a visual bottleneck attention mask; trajectories are scored by an information-gain utility metric Δfair, filtered, ranked, and distilled into a student model in a low-to-high curriculum. In the second stage, LAPO applies answer-to-latent attention attribution to reweight per-step advantages during RL. Experiments on VStar, HRBench4K, HRBench8K, and MME-RealWorld-Lite report LUT-7B averaging 69.96, outperforming existing latent-only methods and remaining competitive with latent-text interleaved methods such as SCOLAR-7B while using fewer inference tokens. The paper also reports ablations of each design choice and a qualitative Region Attention Ratio analysis.

Significance. If the central claim holds, LUT would be a practically valuable contribution: it shows that answer-utility-guided latent training can substitute for expensive intermediate supervision (bounding boxes, sketches, auxiliary images) while retaining the inference-time efficiency of latent-only reasoning. The paper has genuine strengths: it includes a controlled reproduction of LVR on matched 60K data, systematic ablations over selection/curriculum and over teacher/student design choices, an intervention-style control (LAPO reverse), and post-hoc grounding analysis with held-out annotations. However, the specific utility-selection mechanism is not yet convincingly isolated, and several key performance differences are small relative to the absence of reported variance estimates.

major comments (3)
  1. [§3.2, Latent Trajectory Selection] The Δfair definition, S(Z_T) − S(Z_R), uses the frozen base Qwen2.5-VL as judge. This judge has never seen teacher trajectories produced under the visual bottleneck mask, and Z_T is injected under ordinary causal attention, so the scoring is performed on out-of-distribution inputs. The random control Z_R is only norm-matched and is otherwise unstructured, so a higher S(Z_T) could reflect distributional similarity rather than utility for the student. No evidence is provided that the Δfair ranking predicts the student's actual per-sample learning gain. The reported effects are also small: the threshold removes only 6.5% of samples, and removing the curriculum costs 0.44 average points (Table 4). To make the central 'utility' claim load-bearing, please add a validation: compare Δfair-based selection with random selection at the same retention rate, or report the correlation between per-samp
  2. [§4.3/§4.4, Tables 4 and 5] All reported numbers appear to be from single training runs, with no error bars, multiple seeds, or significance tests. Several load-bearing comparisons are small: removing the curriculum costs 0.44 points (68.45 vs 68.01), LAPO vs VLPO costs 0.61 points (69.96 vs 69.35), and the advantage of LUT-7B over UniVLR is 1.08 points. These gaps could easily fall within run-to-run seed variance. Please report means and standard deviations over at least three seeds for the main ablations and key baselines, or provide per-sample bootstrap confidence intervals. This is necessary to support the paper's mechanistic claims that filtering, curriculum order, and attribution-guided reweighting are responsible for the reported gains.
  3. [§4.2, Table 1 and LVR comparison] The controlled comparison with LVR is a strength, but only LUT-7B-SFT is training-matched to LVR† (both use 60K Visual-CoT samples). LUT-7B then adds 24K RL samples from ZwZ-RL-VQA and ViRL, so the headline LUT-7B number is not matched to LVR†. Similarly, other baselines (SkiLa-V, UniVLR, SCOLAR) are taken from public reports with different training data. Please explicitly mark which comparisons are strictly controlled and which are not, and consider adding a matched reproduction of at least one additional latent-reasoning baseline under the same data setting as LUT-7B-SFT.
minor comments (4)
  1. [§3.2, Eq. (1)] The formula for L_dec is typeset incorrectly: it should read L_dec = (1/(N(N−1))) Σ_{i≠j} (ẑ_i^T ẑ_j)^2. The current inline rendering is difficult to parse.
  2. [§3.2] The phrase 'frozen original student model' is confusing because the student is later initialized from the same base model. Call the judge the 'base MLLM' or 'base Qwen2.5-VL' and use a distinct symbol from the trained student to avoid apparent circularity.
  3. [§4.1, training setup] The threshold Δfair > −0.5 is introduced as 'preset' with no justification or sensitivity analysis beyond the one stricter threshold in Table 4. Since selection is a core component, please report the fraction retained and performance for at least one more threshold value, or argue why this particular value is stable.
  4. [§4.5, Figure 5] The RAR analysis is a nice addition, but the sample size (6K held-out Visual-CoT samples) and the use of original annotations only for analysis should be stated more prominently in the caption. Also clarify whether RAR values are averages over samples or pooled attention.

Circularity Check

0 steps flagged

No significant circularity: the reported benchmark results are externally measured and the utility score is a data-filtering heuristic, not a fitted predictor of the reported accuracies.

full rationale

The paper's central claims are empirical benchmark scores (Tables 1–3) on external benchmarks (VStar, HRBench4K, HRBench8K, MME-RealWorld-Lite, MMStar, BLINK, VisualPuzzles). These scores are measured, not derived from the utility score Δfair by construction. Δfair = S(Z_T) − S(Z_R) (Section 3.2) is used only to filter and rank teacher trajectories before distillation; it is a data-selection criterion, not a term in the evaluation metric. The student is initialized from the same base model used as judge, which is self-referential, but this does not make the reported improvements circular: the student is trained with answer cross-entropy and hidden-state alignment, and final performance is measured on held-out benchmarks with the student's own latent generation. The teacher is trained from answer supervision and a decorrelation regularizer; the student alignment loss is a standard distillation objective. No fitted parameter is renamed as a prediction. Self-citations appear only in related-work and data-source context, not as load-bearing uniqueness theorems or ansatz justifications. The ablation shows small effects (e.g., removing curriculum costs 0.44 average points), but this is an empirical finding, not an indication that the benchmark outcome is encoded in the method's inputs. No equation or construction in the paper equates the reported performance with the utility score or with any self-cited result. Therefore no circular step is present.

Axiom & Free-Parameter Ledger

7 free parameters · 4 axioms · 0 invented entities

The method relies on several free hyperparameters (threshold, curriculum split, α, τ, budget, λs, σ) and on domain assumptions about the bottleneck mask, the frozen-judge utility measure, attention attribution as credit, and Gaussian latent policies. There are no wholly invented physical or conceptual entities beyond the training framework itself.

free parameters (7)
  • Δfair threshold = -0.5 (retains 93.5%)
    Sets the cutoff for retaining teacher trajectories; chosen heuristically in Section 3.2 and Section 4.1, and its value affects the distillation quality.
  • Curriculum median split = median Δfair
    Divides retained trajectories into lower and higher utility groups; the order (lower-then-higher) is a design choice shown to matter in Table 4.
  • LAPO gate parameters α, τ = α=1.5, τ=0.5
    Control the strength and sharpness of the attribution gate in Section 3.3; chosen by the authors, with gates in [1.0, 2.5].
  • Latent budget N = 8 default (also 10,12,16)
    Number of latent steps; a hyperparameter that is swept in experiments, and the default is used for all reported LUT-7B results.
  • Decorrelational weight λ_dec = not reported
    Weight for the latent decorrelation regularizer in teacher training; no value is given in the paper.
  • Alignment weight λ_align = not reported
    Weight for the all-layer cosine alignment loss in student distillation; no value is given in the paper.
  • Gaussian policy scale σ = not specified
    Scale in the latent policy log-probability; cancels in the importance ratio but affects the effective update magnitude in LAPO.
axioms (4)
  • ad hoc to paper Visual bottleneck mask forces answer-relevant visual information to flow through the latent trajectory
    The teacher prevents answer and question tokens from attending to image tokens, so any visual evidence must be encoded in the latent states. If this assumption fails, e.g., if the model can solve the answer from question-only shortcuts, the teacher latents would not be visually grounded.
  • domain assumption Δfair computed with a frozen base model is a valid estimate of latent trajectory utility for the student
    The paper uses the log-probability gain of the frozen Qwen2.5-VL judge to rank trajectories. This assumes the judge's preferences are transferable to the trained student, which is plausible but not demonstrated against alternative judges.
  • ad hoc to paper Answer-to-latent attention attribution reflects the causal contribution of each latent step
    LAPO uses averaged attention from answer tokens to latent steps as a credit signal. The reversal control experiment supports the direction, but attention is only a proxy for causal effect and could be confounded by other factors.
  • domain assumption Continuous latent actions can be modeled with a Gaussian policy
    The paper adopts VLPO's Gaussian form for log πθ(zt|st). Latent states may not be Gaussian, and the covariance is assumed fixed and shared, which is a simplifying modeling choice.

pith-pipeline@v1.3.0-alltime-deepseek · 12927 in / 12791 out tokens · 136438 ms · 2026-08-05T00:22:05.990005+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of LUT: Latent Utility Training for Visual Reasoning." pith.science (2026). https://pith.science/paper/MGNCP2ZM

@misc{pith2026260800743,
  author       = {Pith},
  title        = {Pith review of: LUT: Latent Utility Training for Visual Reasoning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MGNCP2ZM}},
  note         = {Machine review of arXiv:2608.00743}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Multimodal large language models have advanced visual understanding, yet perception-intensive reasoning remains challenging. Recent latent visual reasoning methods introduce hidden-space computation before answering, but they often rely on costly intermediate supervision, such as bounding boxes, sketches, or interleaved rationales. These strategies focus on how latent states should be shaped, but do not explicitly assess whether the latent is useful for the final answer. We propose LUT, a latent reasoning framework trained with only standard VQA pairs. LUT centers training on Latent Utility at two levels. At the trajectory level, we propose Utility-Aware Latent Distillation SFT, which explores answer-relevant latent trajectories, selects qualified trajectories by their information gain, and distills more reliable and learnable supervision through curriculum learning. At the step level, we propose Latent Attribution Policy Optimization, which uses answer-to-latent attribution to differentially optimize latent steps during reinforcement learning. Experiments on perception-intensive visual reasoning benchmarks show that LUT outperforms previous latent reasoning methods and remains competitive with latent-text interleaved methods with lower annotation cost.

Figures

Figures reproduced from arXiv: 2608.00743 by Jiaxuan Kang, Mingda Li, Mingjie Liu, Siyu Chen, Tianyue Wang, YanChao Hao, Yongheng Zhang, Zhaoyang Wei, Zheng Wei.

Figure 1
Figure 1. Figure 1: Overview of LUT compared with prior latent visual [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of LUT. The left part illustrates Utility-Aware Latent Distillation SFT, which explores teacher latent [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Effect of latent budgets during training and inference. We compare LUT-7B-SFT models trained with different latent [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative visualization of latent-to-image atten [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Average RAR of latent-to-image attention on held [PITH_FULL_IMAGE:figures/full_fig_p007_5.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

17 extracted references · 1 canonical work pages · 1 internal anchor

  1. [3]

    Chen, L.; Li, J.; Dong, X.; Zhang, P.; Zang, Y.; Chen, Z.; Duan, H.; Wang, J.; Qiao, Y.; Lin, D.; et al

    Reasoning in the dark: Interleaved vision-text reason- ing in latent space.arXiv preprint arXiv:2510.12603. Chen, L.; Li, J.; Dong, X.; Zhang, P.; Zang, Y.; Chen, Z.; Duan, H.; Wang, J.; Qiao, Y.; Lin, D.; et al

    arXiv

  2. [5]

    Fu,X.;Hu,Y.;Li,B.;Feng,Y.;Wang,H.;Lin,X.;Roth,D.; Smith,N.A.;Ma,W.-C.;andKrishna,R.2024.Blink:Multi- modallargelanguagemodelscanseebutnotperceive

    Gemini 2.5: Pushing the fron- tier with advanced reasoning, multimodality, long context, and next generation agentic capabilities.arXiv preprint arXiv:2507.06261. Fu,X.;Hu,Y.;Li,B.;Feng,Y.;Wang,H.;Lin,X.;Roth,D.; Smith,N.A.;Ma,W.-C.;andKrishna,R.2024.Blink:Multi- modallargelanguagemodelscanseebutnotperceive. InEu- ropeanConferenceonComputerVision,148–166....

    Pith/arXiv arXiv 2024

  3. [6]

    Huang,W.;Jia,B.;Zhai,Z.;Cao,S.;Ye,Z.;Zhao,F.;Xu,Z.; Tang, X.; Hu, Y.; and Lin, S

    Glm-4.5 v and glm-4.1 v-thinking: Towards versatile multimodal rea- soning with scalable reinforcement learning.arXiv preprint arXiv:2507.01006. Huang,W.;Jia,B.;Zhai,Z.;Cao,S.;Ye,Z.;Zhao,F.;Xu,Z.; Tang, X.; Hu, Y.; and Lin, S

    Pith/arXiv arXiv

  4. [7]

    arXiv preprint arXiv:2503.06749

    Vision-r1: Incentivizing reasoning capability in multimodal large language models. arXiv preprint arXiv:2503.06749. Hurst, A.; Lerer, A.; Goucher, A. P.; Perelman, A.; Ramesh, A.; Clark, A.; Ostrow, A.; Welihinda, A.; Hayes, A.; Rad- ford, A.; et al

    Pith/arXiv arXiv

  5. [8]

    Jeon, B.; Jeong, Y.; Lee, H.; Cho, M.; and Shin, J

    Gpt-4o system card.arXiv preprint arXiv:2410.21276. Jeon, B.; Jeong, Y.; Lee, H.; Cho, M.; and Shin, J

    Pith/arXiv arXiv

  6. [9]

    Jiang, C.; Heng, Y.; Ye, W.; Xu, H.; Yan, M.; Zhang, J.; Huang, F.; and Zhang, S

    Vision-alignedLatentReasoningforMulti-modalLargeLan- guage Model.arXiv preprint arXiv:2602.04476. Jiang, C.; Heng, Y.; Ye, W.; Xu, H.; Yan, M.; Zhang, J.; Huang, F.; and Zhang, S. 2026a. Vlm-r3: Region recog- nition, reasoning, and refinement for enhanced multimodal chain-of-thought.Advances in Neural Information Process- ing Systems, 38: 63841–63869. Jia...

    Pith/arXiv arXiv

  7. [10]

    Li, K.; Shang, C.; Karlinsky, L.; Feris, R.; Darrell, T.; and Herzig, R

    Llava-onevision: Easy visual task transfer.arXiv preprint arXiv:2408.03326. Li, K.; Shang, C.; Karlinsky, L.; Feris, R.; Darrell, T.; and Herzig, R. 2025b. Latent Implicit Visual Reasoning.arXiv preprint arXiv:2512.21218. Li, P.; Wang, Y.; Zhang, X.; Hu, Z.; Li, S.; Wei, Z.; and Tang, Y. 2026a. ProLaViT: Learning Progressive Latent Visual Thoughts in Stru...

    Pith/arXiv arXiv

  8. [11]

    Fill the GAP: A Granular Alignment Paradigm for Visual Reasoning in Multimodal Large Language Models

    Fill the GAP: A Granular Alignment Paradigm for Visual Reason- ing in Multimodal Large Language Models.arXiv preprint arXiv:2605.12374. Pham,T.-H.;andNgo,C.2025. Multimodalchainofcontin- uous thought for latent-space reasoning in vision-language models.arXiv preprint arXiv:2508.12587. Qiang,C.;Wei,Z.;Han,X.;Wang,Z.;Li,S.;Lan,X.;Jiao,J.; andHan,Z.2025. VER...

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  9. [12]

    Tong,J.;Gu,J.;Lou,Y.;Fan,L.;Zou,Y.;Wu,Y.;Ye,J.;and Li,R.2025

    Visualpuzzles: Decoupling multimodal reason- ing evaluation from domain knowledge.arXiv preprint arXiv:2504.10342. Tong,J.;Gu,J.;Lou,Y.;Fan,L.;Zou,Y.;Wu,Y.;Ye,J.;and Li,R.2025. Sketch-in-latents:Elicitingunifiedreasoningin mllms.arXiv preprint arXiv:2512.16584. Wang,C.;He,W.;Zhu,X.;Zhou,C.;Li,Q.;Yan,S.;Zheng, Y.;Yu,C.;Lu,F.;Zhai,W.;etal.2026a. Self-Consis...

    Pith/arXiv arXiv 2025

  10. [13]

    Wu, P.; and Xie, S

    LaViT: Aligning Latent Visual Thoughts for Multi-modal Reasoning.arXiv preprint arXiv:2601.10129. Wu, P.; and Xie, S

    arXiv

  11. [14]

    Yang, Z.; Yu, X.; Chen, D.; Shen, M.; and Gan, C

    Visual Planning: Let’s Think Only with Images.arXiv preprint arXiv:2505.11409. Yang, Z.; Yu, X.; Chen, D.; Shen, M.; and Gan, C

    arXiv

  12. [15]

    Zhang, Y.; Liu, X.; Tao, R.; Chen, Q.; Fei, H.; Che, W.; and Qin, L

    Machine mental imagery: Empower multimodal reasoning with latent visual tokens.arXiv preprint arXiv:2506.17218. Zhang, Y.; Liu, X.; Tao, R.; Chen, Q.; Fei, H.; Che, W.; and Qin, L. 2025a. Vitcot: Video-text interleaved chain-of- thought for boosting video understanding in large language models. InProceedingsofthe33rdACMInternationalCon- ference on Multime...

    Pith/arXiv arXiv 2026

  13. [17]

    Thinking with Images

    DeepEyes: Incentivizing “Thinking with Images” via Reinforcement Learning.arXiv preprint arXiv:2505.14362. Zhu, J.; Wang, W.; Chen, Z.; Liu, Z.; Ye, S.; Gu, L.; Tian, H.;Duan,Y.;Su,W.;Shao,J.;etal.2025. Internvl3:Explor- ing advanced training and test-time recipes for open-source multimodal models.arXiv preprint arXiv:2504.10479

    Pith/arXiv arXiv 2025

  14. [2023]

    Zheng, Z.; Yang, M.; Hong, J.; Zhao, C.; Xu, G.; Yang, L.; Shen, C.; and Yu, X

    Multimodal chain-of-thought reasoning in language models.arXiv preprint arXiv:2302.00923. Zheng, Z.; Yang, M.; Hong, J.; Zhao, C.; Xu, G.; Yang, L.; Shen, C.; and Yu, X

    Pith/arXiv arXiv

  15. [2024]

    Cheng,T.;Chen,S.-Z.;Zhang,H.;Qin,Y.;Luo,J.;andWei, Z

    Are we on the right way for evaluating large vision-language mod- els?Advances in Neural Information Processing Systems, 37: 27056–27087. Cheng,T.;Chen,S.-Z.;Zhang,H.;Qin,Y.;Luo,J.;andWei, Z. 2026a. Hybrid Latent Reasoning with Decoupled Policy Optimization.arXiv preprint arXiv:2604.20328. Cheng, Z.; Chen, Q.; Xu, X.; Wang, J.; Wang, W.; Fei, H.; Wang,Y.;...

    Pith/arXiv arXiv 2025

  16. [2025]

    Chen,C.;Ma,Z.;Li,Y.;Hu,Y.;Wei,Y.;Li,W.;andNie,L

    Qwen2.5-VL Technical Report.arXiv preprint arXiv:2502.13923. Chen,C.;Ma,Z.;Li,Y.;Hu,Y.;Wei,Y.;Li,W.;andNie,L

    Pith/arXiv arXiv

  17. [2026]

    Multimodal La- tent Reasoning via Predictive Embeddings.arXiv preprint arXiv:2604.08065. Bai,S.;Chen,K.;Liu,X.;Wang,J.;Ge,W.;Song,S.;Dang, K.; Wang, P.; Wang, S.; Tang, J.; Zhong, H.; Zhu, Y.; Yang, M.; Li, Z.; Wan, J.; Wang, P.; Ding, W.; Fu, Z.; Xu, Y.; Ye, J.; Zhang, X.; Xie, T.; Cheng, Z.; Zhang, H.; Yang, Z.; Xu, H.; and Lin, J

    Pith/arXiv arXiv