Pith. sign in

REVIEW 3 major objections 5 minor 33 references

Supervising latent states as reasoning processes, not visual feature copies, lets multimodal models think without external CoT or help images.

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

2026-07-31 16:10 UTC pith:GSFKSCQR

load-bearing objection Solid empirical recipe: on-policy top-k KL plus latent cosine from a privileged multimodal-CoT teacher beats several visual-latent baselines on a 7B backbone; the “process-level internalization” story is stronger than the isolation experiments fully support. the 3 major comments →

arxiv 2607.28154 v1 pith:GSFKSCQR submitted 2026-07-30 cs.CV cs.AI

OPLD: On-Policy Latent Distillation for Multimodal Reasoning

Shoutai Zhu , Tianyang Xu , Bin Sun , Mingyuan Xu , Yu Liu , Qinzhen Guo This is my paper
classification cs.CV cs.AI
keywords multimodal reasoninglatent reasoningchain-of-thoughtknowledge distillationvision-language modelson-policy distillationvisual chain-of-thought
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.

Interleaved multimodal chain-of-thought helps models reason about images by inserting auxiliary visual evidence into intermediate steps, but those traces stay external and rigid at inference. Prior latent methods mostly force continuous states to match compressed visual features, so they store evidence without internalizing the abstract process CoT induces. OPLD trains a teacher that sees privileged multimodal CoT and a student that sees only the original image and question. After separate warmups, the student rolls out answers under its own policy; the frozen teacher then aligns the student’s latent trajectory and supplies token-level distribution feedback on that same trajectory. The student thereby absorbs CoT-induced reasoning into a fixed number of recursive latent steps and, at inference, answers from the raw input alone. Across perception-heavy and general multimodal benchmarks the method outperforms other visual-latent approaches and reaches top reported scores on several tasks, arguing that process-level latent supervision is more effective than feature-level alignment.

Core claim

The paper claims that on-policy distillation of a privileged multimodal-CoT teacher’s token distributions and latent trajectories transfers abstract reasoning into a CoT-free student’s continuous latent slots, so the student can reason without auxiliary images or explicit rationales and outperforms feature-alignment latent methods on multiple multimodal benchmarks.

What carries the argument

On-Policy Latent Distillation (OPLD): after teacher and student warmups, the student generates an on-policy answer and latent trajectory; the frozen CoT-augmented teacher supplies top-k forward KL on that trajectory plus slot-wise cosine alignment of the K recursive latent codes.

Load-bearing premise

Matching latent slots by recursive depth with cosine loss is assumed to transfer the teacher’s abstract reasoning process, even though equal depth need not mean the same semantic stage and the two models warm up independently.

What would settle it

If ablating the latent-cosine term (keeping only token-level KL) or intervening at test time by zeroing, noising, or shuffling the student’s latent trajectory left benchmark gains essentially unchanged, the claim that process-level latent supervision internalizes multimodal CoT would fail.

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

If this is right

  • Multimodal models can drop explicit visual tools, crops, and text CoT at inference while retaining much of the accuracy those traces provide in training.
  • Latent reasoning design should prioritize process-level teacher signals over forcing latents to reconstruct auxiliary visual embeddings.
  • On-policy student rollouts plus teacher scoring on the student’s own trajectory give stabler transfer than offline or larger-teacher token distillation alone.
  • A compact encoder–decoder latent bottleneck is a practical interface for calibrating student imagination to a privileged teacher space.

Where Pith is reading between the lines

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

  • If slot-wise depth alignment is only a weak temporal bias, learned mappers or optimal-transport matching between teacher and student latents could raise how much of the privileged gap is recovered.
  • The same on-policy latent-distillation pattern may transfer to pure-language continuous-thought models whenever a privileged discrete CoT teacher is available.
  • Gains concentrated on fine-grained and high-resolution perception suggest the latent chain mainly organizes evidence integration rather than long symbolic deduction.

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 / 5 minor

Summary. The paper proposes OPLD, a three-stage teacher–student framework that transfers reasoning induced by privileged interleaved multimodal CoT into a CoT-free student’s recursive latent slots. A CoT-augmented teacher and a CoT-free student are warmed up separately; then the frozen teacher supervises student on-policy rollouts via top-k forward KL on answer tokens (Eqs. 16–19) and slot-wise cosine alignment of latent trajectories (Eq. 15), combined as L_OPLD (Eq. 20). At inference the student uses only the original image–question input and K recursive Enc–Dec latent steps. On Qwen2.5-VL-7B with a cleaned Zebra-CoT∪Visual-CoT set (217K), OPLD reports large gains over the backbone and SFT on V* and HRBench, outperforms several published visual-latent methods on most reported metrics (Table 1), and is supported by training-strategy, latent-design, and test-time latent-intervention ablations (Fig. 4, Tables 2, 5–7).

Significance. If the results hold under fairer isolation of the claimed mechanism, the work is a solid systems contribution to multimodal latent reasoning: it shows that privileged multimodal CoT can be internalized into fixed-step continuous latents so that inference needs neither textual CoT nor auxiliary images/tools, while remaining competitive with thinking-with-images agents on perception-heavy benchmarks. Strengths include a clear three-stage recipe, careful leakage filtering of teacher help images, systematic ablations (strategy, K, Enc–Dec vs raw hidden, loss terms), and latent interventions (noise/zero/shuffle) that functionally link the trajectory to answers. The framing that process-level latent supervision is a more effective paradigm than feature-level visual alignment would be high-impact if isolated; even as an empirical distillation recipe the gains on V* Attribute and HRBench FSP are practically relevant.

major comments (3)
  1. [Abstract; Method Stage 3, Eqs. 15–20; Table 2; Supplementary limitations] Abstract/Introduction/Conclusion claim that OPLD supervises latents at the “reasoning-process level” and is more effective than “conventional feature-level alignment.” Stage 3 defines zT_1:K = f_lat_ϕ(xT) with xT = (I, Q, R, Ihelp) (Eqs. 2, 14–15), so L_lat pulls student latents toward states computed under privileged evidence the student never sees. The supplement (“Interpretation and limitation of slot-wise alignment”) already states that equal recursive depth is only a temporal inductive bias, not semantic stage correspondence, and that independent full-parameter warmups can drift coordinates. Table 2 shows cosine is necessary (top-k KL only collapses on HRBench), but there is no same-backbone, same-data control that aligns the same K Enc–Dec slots to compressed Ihelp / key visual features—the paradigm OPLD claims to beat. Without that isolation (or substantially softer claims), Table
  2. [Table 1; Abstract; Main Results] SOTA wording is stronger than Table 1 supports. Among visual-latent methods OPLD is often best, but HyLaR exceeds OPLD on HR-4K overall (75.00 vs 73.75), HR-4K FSP (93.75 vs 90.00), HR-8K FSP (88.25 vs 87.75), and V* Spatial (85.53 vs 81.58); several thinking-with-images agents also lead on individual cells. Please restrict “state-of-the-art on multiple benchmarks” to the metrics/settings where OPLD is actually first, and mark visual-latent vs agent comparisons consistently.
  3. [Fig. 4; Table 5; §Effect of Training Strategy] Fig. 4 / Table 5: standard OPD from a Qwen2.5-VL-32B teacher with the same forward-KL objective yields a large average drop (−3.30), including severe MME-RW-Lite degradation. This is used to argue that token-level distillation without latent alignment is insufficient, but the 32B teacher is a different model family scale and is not described as using the same latent protocol or multimodal-CoT teacher view as OPLD’s 7B teacher. Either match the teacher setup (same 7B CoT teacher, text-only OPD) or qualify that the failure mode may mix capacity/distribution shift with missing L_lat, so it does not cleanly isolate the necessity of latent trajectory alignment.
minor comments (5)
  1. [Method, Eqs. 11–17] Notation drift: student/teacher parameters are θ/ϕ in Eqs. 11–17 but the text sometimes refers to “frozen teacher” without restating that Enc/Dec are included; a single parameter table in the main text (beyond Supp. Table 4) would help.
  2. [Experimental Setup; Table 2] K=8, D=2048, λ_lat=0.05, top-k=128 are fixed with limited sensitivity beyond K∈{0,4,8} and a few loss variants; a short sweep or stability note for λ_lat and k would strengthen reproducibility claims.
  3. [Supplementary Case Studies] Case studies (Figs. 6–12) are useful, but several show the warmed-up teacher wrong and OPLD student correct; the prose explanation (soft KL + scoring on student prefixes) belongs in a short main-text paragraph so readers do not infer hard-label imitation.
  4. [Title block; Abstract; Table 1] Minor writing issues: spacing in “Interleavedmultimodal”, “On-PolicyLatent”, author line “Xumingyuan1”, and inconsistent benchmark names (SeedBench2-Plus vs Seed2+).
  5. [Related Work] Related Work is thorough; briefly situate OPLD against pure text latent CoT (e.g., Coconut-style) versus visual-latent feature alignment in one comparative sentence for non-MLLM readers.

Circularity Check

0 steps flagged

No significant circularity: empirical distillation evaluated on external benchmarks, not a by-construction derivation.

full rationale

OPLD is a systems/training paper. Its load-bearing claim is empirical: a CoT-free student trained with on-policy top-k forward KL plus slot-wise latent cosine to a frozen privileged multimodal-CoT teacher (Eqs. 13–20, Stage 3) improves external multimodal benchmarks (Table 1: V*, HRBench, MMStar, etc.) over baselines and prior visual-latent methods. That chain does not reduce the reported scores to the training objective by definition. Privileged teacher inputs xT=(I,Q,C) and student inputs xS=(I,Q) are a standard asymmetric distillation setup; L_lat and L_text are optimization losses, not the evaluation metrics. Warmup CE on y* and held-out LMMS-Eval greedy scores are independent of claiming that cosine equality is itself the scientific result. The supplement’s caveat that equal recursive depth is only a temporal inductive bias weakens the semantic interpretation of “process-level” transfer, but that is an assumption/correctness issue, not circularity (no fitted parameter renamed as prediction; no self-citation uniqueness theorem forcing the method; no X defined as Y). No steps match the enumerated circular patterns.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 2 invented entities

Load-bearing commitments are methodological: privileged multimodal CoT induces transferable abstract latent trajectories; student on-policy rollouts plus frozen-teacher scoring are valid supervision; fixed-slot cosine alignment is a sufficient process prior; leakage-filtered Zebra-CoT∪Visual-CoT is a fair training distribution; Qwen2.5-VL-7B with K-step Enc–Dec latents is an adequate testbed. Free parameters are standard ML hyperparameters chosen for the reported run rather than physical constants.

free parameters (5)
  • K (latent slots) = 8
    Number of recursive latent steps; ablated at 0/4/8, default K=8 chosen for best average metrics.
  • D (latent dimension) = 2048
    Projected latent width for Enc–Dec bottleneck; fixed without broad sweep in main text.
  • λ_lat (latent cosine weight) = 0.05
    Balances L_text vs L_lat in the distillation objective; small weight 0.05 selected for the main run.
  • top-k for forward KL = 128
    Restricts token distillation support to teacher top-k mass; k=128 in main config.
  • distillation learning rate / schedule = 1e-6, 1 epoch
    Student update rate 1e-6, one epoch, batch settings on 8×B200; warmup uses different LRs—choices that affect final numbers.
axioms (5)
  • domain assumption Privileged interleaved multimodal CoT induces latent trajectories that encode transferable abstract reasoning rather than only shortcut visual features (after leakage filtering).
    Core motivation in Introduction/Method; without it, latent cosine targets are just another feature alignment.
  • ad hoc to paper Aligning student latents slot-wise by recursive depth to teacher latents transfers process-level reasoning despite possible semantic misalignment and warmup drift.
    Stage-3 L_lat and supplement limitation discussion; authors treat ordered cosine as inductive bias, not proven semantic correspondence.
  • domain assumption On-policy student rollouts scored by a frozen teacher under mixed views (teacher sees x_T but conditions on student z^S and ŷ) yield stable, non-degenerate distillation targets.
    Distillation stage design (Eqs. for p^T_t, p^S_t, L_text); standard in OPD-style work but load-bearing here.
  • ad hoc to paper GPT-5.4-based removal of answer-leaking auxiliary images yields a cleaner reasoning supervision set without systematically discarding hard genuine reasoning cases.
    Supplementary filtering pipeline removes ~18.4% of merged data; results depend on this curated 217K set.
  • domain assumption Answer-only CE warmups establish compatible latent protocols so subsequent trajectory alignment is meaningful.
    Stages 1–2 justified as necessary for stable distillation; ablation shows weak results without full pipeline.
invented entities (2)
  • OPLD on-policy latent distillation objective (L_text + λ_lat L_lat) with teacher scoring on student latents no independent evidence
    purpose: Name and operationalize process-level transfer from multimodal CoT into student latent space without inference-time CoT.
    New training framework entity relative to pure visual-feature latent alignment methods; defined entirely by the paper’s procedure.
  • Enc–Dec projected latent reasoning slots with recursive continuous feedback in the MLLM chat template no independent evidence
    purpose: Provide a compact latent interface shared by teacher and student for trajectory alignment.
    Architectural construct building on prior latent-reasoning work; paper’s specific Enc/Dec MLP form and slot placement.

pith-pipeline@v1.2.0-daily-grok45 · 27585 in / 3997 out tokens · 86329 ms · 2026-07-31T16:10:54.845989+00:00 · methodology

0 comments
read the original abstract

Interleaved multimodal Chain-of-Thought (CoT) improves visual reasoning by incorporating auxiliary visual evidence into intermediate reasoning. However, existing approaches remain constrained by externally defined reasoning traces and visual operations, limiting their ability to develop flexible and abstract visual thinking. Reasoning with latent has recently offered a promising direction by internalizing intermediate computation into continuous representations. Nevertheless, existing visual-latent methods mainly supervise latent states through alignment with compressed auxiliary visual features, treating them as proxies for visual observations rather than active reasoning states. Consequently, they capture the provided evidence but fail to fully internalize the abstract reasoning process induced by multimodal CoT. In this paper, we propose OPLD (On-Policy Latent Distillation), a simple framework that transfers the reasoning capability induced by privileged multimodal CoT into latent reasoning representations. Extensive experiments on diverse multimodal benchmarks demonstrate that OPLD consistently outperforms existing latent reasoning methods and achieves state-of-the-art performance on multiple benchmarks. The results suggest that supervising latent representations at the reasoning-process level provides a more effective paradigm for multimodal latent reasoning than conventional feature-level alignment.

Figures

Figures reproduced from arXiv: 2607.28154 by Bin Sun, Mingyuan Xu, Qinzhen Guo, Shoutai Zhu, Tianyang Xu, Yu Liu.

Figure 1
Figure 1. Figure 1: Comparison of multimodal reasoning paradigms. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of OPLD. The left shows inference, where the model answers from the original image and question through [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: OPLD distillation stage. The student rolls out response [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Training strategy ablation with the same backbone [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: In-domain accuracy on 5K held-out samples. OPLD [PITH_FULL_IMAGE:figures/full_fig_p014_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Inference Result [PITH_FULL_IMAGE:figures/full_fig_p016_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Inference Result [PITH_FULL_IMAGE:figures/full_fig_p017_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Inference Result [PITH_FULL_IMAGE:figures/full_fig_p018_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Inference Result [PITH_FULL_IMAGE:figures/full_fig_p019_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Inference Result [PITH_FULL_IMAGE:figures/full_fig_p020_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Inference Result [PITH_FULL_IMAGE:figures/full_fig_p021_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Inference Result [PITH_FULL_IMAGE:figures/full_fig_p022_12.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

33 extracted references · 19 linked inside Pith

  1. [1]

    Advances in Neural Information Processing Systems , volume =

    Chain-of-Thought Prompting Elicits Reasoning in Large Language Models , author =. Advances in Neural Information Processing Systems , volume =

  2. [2]

    Advances in Neural Information Processing Systems , volume =

    Large Language Models are Zero-Shot Reasoners , author =. Advances in Neural Information Processing Systems , volume =

  3. [3]

    arXiv preprint arXiv:2302.00923 , year =

    Multimodal Chain-of-Thought Reasoning in Language Models , author =. arXiv preprint arXiv:2302.00923 , year =. 2302.00923 , archivePrefix =

    Pith/arXiv arXiv

  4. [4]

    arXiv preprint arXiv:2505.14362 , year =

    DeepEyes: Incentivizing ``Thinking with Images'' via Reinforcement Learning , author =. arXiv preprint arXiv:2505.14362 , year =. 2505.14362 , archivePrefix =

    Pith/arXiv arXiv

  5. [5]

    arXiv preprint arXiv:2505.15436 , year =

    Chain-of-Focus: Adaptive Visual Search and Zooming for Multimodal Reasoning via RL , author =. arXiv preprint arXiv:2505.15436 , year =. 2505.15436 , archivePrefix =

    Pith/arXiv arXiv

  6. [6]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , year =

    Bootstrapping Grounded Chain-of-Thought in Multimodal LLMs for Data-Efficient Model Adaptation , author =. Proceedings of the IEEE/CVF International Conference on Computer Vision , year =

  7. [8]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , year =

    Monet: Reasoning in Latent Visual Space Beyond Images and Language , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , year =. 2511.21395 , archivePrefix =

    arXiv

  8. [9]

    arXiv preprint arXiv:2412.06769 , year =

    Training Large Language Models to Reason in a Continuous Latent Space , author =. arXiv preprint arXiv:2412.06769 , year =. 2412.06769 , archivePrefix =

    Pith/arXiv arXiv

  9. [10]

    arXiv preprint arXiv:2602.04476 , year =

    Vision-aligned Latent Reasoning for Multi-modal Large Language Model , author =. arXiv preprint arXiv:2602.04476 , year =. 2602.04476 , archivePrefix =

    Pith/arXiv arXiv

  10. [11]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume =

    Visual Chain-of-Thought Prompting for Knowledge-Based Visual Reasoning , author =. Proceedings of the AAAI Conference on Artificial Intelligence , volume =. 2024 , doi =

    2024

  11. [12]

    arXiv preprint arXiv:2403.16999 , year =

    Visual CoT: Unleashing Chain-of-Thought Reasoning in Multi-Modal Language Models , author =. arXiv preprint arXiv:2403.16999 , year =. 2403.16999 , archivePrefix =

    Pith/arXiv arXiv

  12. [13]

    arXiv preprint arXiv:2405.13872 , year =

    Image-of-Thought Prompting for Visual Reasoning Refinement in Multimodal Large Language Models , author =. arXiv preprint arXiv:2405.13872 , year =. 2405.13872 , archivePrefix =

    Pith/arXiv arXiv

  13. [14]

    arXiv preprint arXiv:2503.12799 , year =

    Grounded Chain-of-Thought for Multimodal Large Language Models , author =. arXiv preprint arXiv:2503.12799 , year =. 2503.12799 , archivePrefix =

    Pith/arXiv arXiv

  14. [15]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , year =

    Argus: Vision-Centric Reasoning with Grounded Chain-of-Thought , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , year =. 2505.23766 , archivePrefix =

    Pith/arXiv arXiv

  15. [16]

    arXiv preprint arXiv:2506.11991 , year =

    VGR: Visual Grounded Reasoning , author =. arXiv preprint arXiv:2506.11991 , year =. 2506.11991 , archivePrefix =

    Pith/arXiv arXiv

  16. [17]

    Advances in Neural Information Processing Systems , year =

    VLM-R ^3 : Region Recognition, Reasoning, and Refinement for Enhanced Multimodal Chain-of-Thought , author =. Advances in Neural Information Processing Systems , year =. 2505.16192 , archivePrefix =

    Pith/arXiv arXiv

  17. [18]

    arXiv preprint arXiv:2606.08464 , year =

    TVI-CoT: Text-Visual Interleaved Chain-of-Thought Reasoning for Multimodal Understanding , author =. arXiv preprint arXiv:2606.08464 , year =. 2606.08464 , archivePrefix =

    Pith/arXiv arXiv

  18. [19]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume =

    Let's Think with Images Efficiently! An Interleaved-Modal Chain-of-Thought Reasoning Framework with Dynamic and Precise Visual Thoughts , author =. Proceedings of the AAAI Conference on Artificial Intelligence , volume =. 2026 , doi =

    2026

  19. [20]

    Proceedings of the 42nd International Conference on Machine Learning , series =

    Imagine While Reasoning in Space: Multimodal Visualization-of-Thought , author =. Proceedings of the 42nd International Conference on Machine Learning , series =

  20. [21]

    arXiv preprint arXiv:2505.22525 , year =

    Thinking with Generated Images , author =. arXiv preprint arXiv:2505.22525 , year =. 2505.22525 , archivePrefix =

    Pith/arXiv arXiv

  21. [22]

    arXiv preprint arXiv:2606.16783 , year =

    Gen-VCoT: Generative Visual Chain-of-Thought Reasoning via Diffusion-Based RGB Intermediate Representations , author =. arXiv preprint arXiv:2606.16783 , year =. 2606.16783 , archivePrefix =

    arXiv

  22. [23]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    CoT-VLA: Visual Chain-of-Thought Reasoning for Vision-Language-Action Models , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

  23. [24]

    arXiv preprint arXiv:2603.01893 , year =

    Generative Visual Chain-of-Thought for Image Editing , author =. arXiv preprint arXiv:2603.01893 , year =. 2603.01893 , archivePrefix =

    arXiv

  24. [25]

    arXiv preprint arXiv:2509.24251 , year=

    Latent visual reasoning , author=. arXiv preprint arXiv:2509.24251 , year=

    Pith/arXiv arXiv

  25. [26]

    arXiv preprint arXiv:2506.17218 , year =

    Machine Mental Imagery: Empower Multimodal Reasoning with Latent Visual Tokens , author =. arXiv preprint arXiv:2506.17218 , year =. 2506.17218 , archivePrefix =

    Pith/arXiv arXiv

  26. [27]

    arXiv preprint arXiv:2512.16584 , year =

    Sketch-in-Latents: Eliciting Unified Reasoning in MLLMs , author =. arXiv preprint arXiv:2512.16584 , year =. 2512.16584 , archivePrefix =

    arXiv

  27. [28]

    arXiv preprint arXiv:2601.06803 , year =

    Forest Before Trees: Latent Superposition for Efficient Visual Reasoning , author =. arXiv preprint arXiv:2601.06803 , year =. 2601.06803 , archivePrefix =

    Pith/arXiv arXiv

  28. [29]

    Findings of the Association for Computational Linguistics: NAACL 2025 , pages=

    Lmms-eval: Reality check on the evaluation of large multimodal models , author=. Findings of the Association for Computational Linguistics: NAACL 2025 , pages=

    2025

  29. [30]

    arXiv preprint arXiv:2507.16746 , year=

    Zebra-cot: A dataset for interleaved vision language reasoning , author=. arXiv preprint arXiv:2507.16746 , year=

    arXiv

  30. [31]

    Advances in Neural Information Processing Systems , volume=

    Visual cot: Advancing multi-modal language models with a comprehensive dataset and benchmark for chain-of-thought reasoning , author=. Advances in Neural Information Processing Systems , volume=

  31. [32]

    arXiv preprint arXiv:2604.20328 , year =

    Hybrid Latent Reasoning with Decoupled Policy Optimization , author =. arXiv preprint arXiv:2604.20328 , year =. 2604.20328 , archivePrefix =

    Pith/arXiv arXiv

  32. [33]

    arXiv preprint arXiv:2602.22766 , year =

    Imagination Helps Visual Reasoning, But Not Yet in Latent Space , author =. arXiv preprint arXiv:2602.22766 , year =. 2602.22766 , archivePrefix =

    Pith/arXiv arXiv

  33. [34]

    arXiv preprint arXiv:2605.02735 , year =

    Visual Latents Know More Than They Say: Unsilencing Latent Reasoning in MLLMs , author =. arXiv preprint arXiv:2605.02735 , year =. 2605.02735 , archivePrefix =

    Pith/arXiv arXiv