Pith. sign in

REVIEW 3 major objections 8 minor 2 cited by

Mixing text tokens with continuous visual latents stabilizes RL reasoning

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 · glm-5.2

2026-07-05 04:19 UTC pith:E264CY2M

load-bearing objection DePO is a genuinely novel RL algorithm for hybrid discrete-continuous MLLM reasoning; the vMF theory-practice gap is real but does not sink the empirical results the 3 major comments →

arxiv 2604.20328 v2 pith:E264CY2M submitted 2026-04-22 cs.CV

HyLaR: Hybrid Latent Reasoning with Decoupled Policy Optimization

Tao Cheng , Shi-Zhe Chen , Hao Zhang , Yixin Qin , Jinwen Luo , Zheng Wei This is my paper
classification cs.CV
keywords latentreasoninghybridhylarpolicydecoupleddepofine-grained
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 HyLaR, a framework that lets a multimodal large language model interleave discrete text generation with continuous visual latent representations during reasoning, avoiding the information loss that occurs when visual signals are prematurely discretized into text tokens. The central technical contribution is DePO (Decoupled Policy Optimization), a reinforcement learning algorithm designed for this hybrid discrete-continuous action space. DePO addresses two mismatches that arise when standard RL is applied to hybrid trajectories: first, the importance-sampling ratios for discrete tokens and continuous latent vectors exhibit vastly different variance, so DePO applies separate trust-region clipping ranges to each; second, standard Gaussian assumptions imply a flat Euclidean geometry that does not match the hyperspherical structure of layer-normalized LLM hidden states, so DePO models the continuous latent policy using the von Mises-Fisher (vMF) distribution on the unit hypersphere, which yields a closed-form KL divergence reducible to a scaled cosine distance. The paper claims that this combination yields stable optimization and concrete accuracy gains on fine-grained visual perception benchmarks, with improvements of +7.33% on V* and +7.00% on HRBench-4K over the Qwen2.5-VL-7B baseline.

Core claim

The paper identifies that optimizing a hybrid reasoning trajectory—where discrete text tokens and continuous visual latent vectors coexist in the same sequence—requires treating the two action types differently in the RL objective. The discovery is that (1) continuous latent actions need a much tighter clipping range (0.05 vs. 0.2 for text) because small directional shifts in high-dimensional hidden states cause disproportionately large swings in importance ratios, and (2) modeling the latent policy with a vMF distribution rather than a Gaussian aligns with the hyperspherical geometry of normalized LLM representations and converts the KL regularizer from a high-variance sample-based estimate

What carries the argument

The mechanism is a two-stage pipeline. Stage 1 (SFT) teaches the model to alternate between text tokens and continuous canvas tokens—hidden states recurrently fed back as input embeddings without passing through the discrete vocabulary—using an MSE loss to align predicted hidden states with compressed ground-truth visual embeddings. Stage 2 (DePO) refines the hybrid trajectory with RL: the policy log-probability is a categorical distribution for text positions and a vMF density for latent positions, the surrogate loss applies independent dual-clip PPO objectives with separate clipping ranges to text and latent positions, and the KL penalty for latent positions is computed in closed form as a

Load-bearing premise

The vMF distribution is justified by the claim that layer-normalized LLM representations reside on a hypersphere, but the actual implementation relaxes strict unit-norm normalization and uses unnormalized inner products, meaning the formal vMF density-ratio and KL guarantees no longer strictly hold—the objective is a directional surrogate rather than an exact probabilistic quantity.

What would settle it

Replace the vMF-based latent policy and closed-form cosine KL with a standard Gaussian policy and sample-based KL; if performance is comparable, the hyperspherical modeling is not the load-bearing ingredient. The paper reports this ablation and finds vMF wins, but the gap is modest (e.g., 83.77 vs. 83.25 on V*), so the falsifier would be a stronger Gaussian baseline with tuned variance that closes or reverses the gap.

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

If this is right

  • If decoupled clipping is necessary, other hybrid action spaces—such as models that interleave text with tool calls, structured outputs, or continuous embeddings in other modalities—may also require per-action-type trust regions rather than uniform clipping.
  • The vMF-to-cosine-distance reduction suggests that any RL objective over layer-normalized hidden states could benefit from hyperspherical modeling, potentially extending to latent-space planning in language-only models.
  • The finding that RL enables length extrapolation (stable performance when inference latent steps exceed training steps) while SFT degrades could inform training-budget allocation for other latent reasoning architectures.
  • If the unnormalized surrogate consistently outperforms strict vMF, the practical lesson is that geometrically motivated objectives can work even when their formal probabilistic assumptions are relaxed—directional preference may matter more than exact distributional fidelity.

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 / 8 minor

Summary. This paper introduces HyLaR, a hybrid latent reasoning framework for multimodal large language models (MLLMs) that interleaves discrete text generation with continuous visual latent representations. The method uses a two-stage training pipeline: (1) a cold-start SFT phase with canvas compression and alignment, and (2) a reinforcement learning phase using DePO (Decoupled Policy Optimization), which applies separate trust-region clipping to text and latent actions and models the latent policy with a von Mises-Fisher (vMF) distribution. Experiments on high-resolution perception benchmarks (V*, HRBench) and general VQA benchmarks show improvements over the Qwen2.5-VL-7B baseline and competing latent reasoning methods.

Significance. The paper addresses a real problem: optimizing RL in hybrid discrete-continuous action spaces for MLLMs. The decoupled clipping idea is well-motivated by the variance mismatch between discrete and continuous importance ratios (Fig. 3). The vMF KL derivation (Supplement §1.2) is a clean, parameter-free mathematical result under stated assumptions. The ablation suite is reasonably thorough, covering latent step counts, RL algorithms, hyperparameters, and distribution assumptions. Code is publicly available. The improvements on V* (+7.33%) and HRBench-4K (+7.00%) over the base model are notable.

major comments (3)
  1. §3.3, Abstract, and contributions list: The paper repeatedly claims 'exact closed-form vMF KL' and 'geometrically rigorous' optimization as a primary selling point. However, Supplement §1.3 explicitly states that the implementation uses unnormalized inner products where 'the exact probabilistic interpretation as a vMF density ratio or KL divergence no longer strictly holds.' Table S1 confirms the unnormalized surrogate is what produces the reported results. This is a framing inconsistency: the theoretical contribution is presented as exact and rigorous, but the actual implementation is a surrogate that only 'inherits directional preference.' The main text should transparently state this relaxation rather than presenting exactness as a property of the deployed system. As it stands, a reader of the main text alone would conclude the implementation enjoys exact vMF guarantees, which it does
  2. Tables 1, 2, 3, 6: No error bars, confidence intervals, or significance tests are reported for any experiment. Several key comparisons rest on differences of 0.5–3% on benchmarks with a few hundred samples (e.g., V* has ~121 samples based on 0.83% granularity in Table 1). The DePO-vs-GRPO ablation (Table 3) shows 83.77 vs 78.53 on V*—a 5.24-point gap on a small benchmark. The vMF-vs-Gaussian ablation (Table 6) shows differences of 0.52–3.04% across benchmarks. Without any variance estimates, it is difficult to distinguish genuine improvements from seed-level noise, particularly for the smaller deltas in Tables 4, 5, and 6. At minimum, the paper should report results over multiple seeds or provide bootstrap confidence intervals for the headline numbers and the key ablation comparisons.
  3. Table 1, V* column: The paper reports HyLaR-7B at 83.77% but the reproduced SkiLa* at 78.53%, while the original SkiLa reports 84.3%. The paper claims HyLaR outperforms SkiLa by 5.24%, but this comparison is against the reproduced number, not the original. The original SkiLa (84.3%) actually slightly outperforms HyLaR (83.77%) on V*. The paper should clarify whether the comparison is fair (same evaluation pipeline) and acknowledge that the original SkiLa number is higher. The same issue applies to Monet (original 83.25% vs reproduced 80.1%). The current framing in §4.2 ('outperforms SkiLa (78.53%)... by margins of 5.24%') is misleading without noting the original reported score.
minor comments (8)
  1. §3.3, Eq. (4): The notation uses both z̃_t and μ_θ^t. It would help to clarify upfront that z̃_t = μ_θ_old^t (the old policy mode), as this identification is only made later in Supplement §1.1.
  2. Table 1: The '∆(vs Qwen2.5-VL 7B)' row shows HRBench-8K FCP as '↓1.00', indicating a regression. This is not discussed in the text. A brief acknowledgment would be appropriate.
  3. §4.1: The reward function uses 'GPT-5' as the evaluation model for accuracy reward. This seems unusual for a 2026 paper; please verify this is correct or clarify if a different model was used.
  4. Supplement §2.2, Table S4: The 'valid group accuracy [0.1, 0.9]' filtering mechanism is described, but it is unclear what fraction of rollout groups are discarded under this criterion. Reporting this statistic would help readers assess training efficiency.
  5. Fig. 3: The x-axis label 'policy perturbation magnitude' is defined in the caption but the y-axis is unlabeled. Adding a y-axis label (e.g., 'importance ratio r_t') would improve clarity.
  6. §3.2, Eq. (1): The MSE loss L_Canvas uses h_t to predict e_{t+1}, but the subscript relationship between canvas position t and target embedding t+1 could be stated more explicitly in the text.
  7. Table 6 caption: The vMF row specifies κ=0.01, but the main text §3.3 does not mention this value. It only appears in Supplement Table S4. Including κ in the main text would make the experiments self-contained.
  8. References: Several arXiv preprints are cited with 2025–2026 dates. Please verify these are accessible and correctly attributed.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for a careful and constructive review. All three major comments are well-taken. We will (1) transparently state the unnormalized surrogate relaxation in the main text, (2) add multi-seed variance estimates for headline results and key ablations, and (3) correct the SkiLa/Monet comparison framing to acknowledge original reported scores. We respectfully note one point of partial disagreement regarding the characterization of the vMF contribution as 'misleading.'

read point-by-point responses

  1. Referee: vMF exactness framing inconsistency: main text claims 'exact closed-form vMF KL' and 'geometrically rigorous' but Supplement §1.3 admits the implementation uses unnormalized inner products where the exact vMF interpretation 'no longer strictly holds.'

    Authors: The referee is correct that the main text does not transparently convey the relaxation described in Supplement §1.3, and we will revise the manuscript accordingly. Specifically, we will: (a) add a clear statement in §3.3 that the deployed implementation uses unnormalized inner products rather than strict unit-norm cosine distances, and that the exact vMF KL derivation serves as the theoretical foundation while the practical surrogate 'inherits directional preference' but does not strictly satisfy the vMF density ratio / KL interpretation; (b) soften the abstract and contributions list to say 'a vMF-derived closed-form KL regularizer' rather than 'exact closed-form vMF KL'; (c) add a forward reference to Supplement §1.3 at the point where Eq. (7) is introduced, so readers of the main text are immediately aware of the relaxation. We will also ensure Table S1 (normalized vs. unnormalized ablation) is referenced in the main text. We respectfully note one nuance: the vMF KL derivation itself (Supplement §1.2) is mathematically exact under its stated assumptions—this is not in question. What the referee correctly identifies is that the *deployed system* relaxes those assumptions, and the main text should not present exactness as a property of the implementation. We agree this distinction must be made transparent. revision: yes

  2. Referee: No error bars, confidence intervals, or significance tests reported for any experiment. Several key comparisons rest on small deltas on small benchmarks (V* ~121 samples).

    Authors: This is a fair and important criticism. We will add multi-seed results (minimum 3 seeds) for the headline comparisons in Table 1 (at least V* and HRBench-4K) and for the key ablation comparisons in Tables 3 and 6 (DePO vs. GRPO, vMF vs. Gaussian). We will report standard deviations and, where feasible, bootstrap 95% confidence intervals. We acknowledge that on benchmarks with ~121 samples (V*), per-seed variance can be substantial, and reporting this honestly will strengthen the paper. For the broader suite of ablations in Tables 4 and 5, we will add at least single-seed bootstrap CIs. We note that some of the larger deltas—e.g., DePO vs. GRPO on V* (83.77 vs. 78.53, a 5.24-point gap) and the HRBench improvements over the Qwen2.5-VL baseline (+7.00% on HRBench-4K, +6.75% on HRBench-8K)—are unlikely to be attributable to seed noise, but we agree the burden of proof is on us to demonstrate this with variance estimates. revision: yes

  3. Referee: Table 1 V* comparison is misleading: HyLaR (83.77%) is compared against reproduced SkiLa (78.53%) but original SkiLa reports 84.3%, which is higher than HyLaR. Same issue for Monet (original 83.25% vs reproduced 80.1%). The §4.2 framing 'outperforms SkiLa (78.53%) by 5.24%' is misleading without noting the original score.

    Authors: The referee is correct on the factual point: the original SkiLa reports 84.3% on V*, which is higher than our 83.77%, and the original Monet reports 83.25%, which is also higher than our number. Our current framing in §4.2, which compares against the reproduced numbers without adequately flagging the originals, is misleading. We will revise the text to explicitly acknowledge both the original and reproduced scores for SkiLa and Monet on V*, and we will not claim superiority over SkiLa on V*. We will clarify that the reproduced numbers are presented to control for evaluation pipeline differences (prompt format, answer extraction, judge model), and that the original numbers may reflect different evaluation protocols. We will also note that on HRBench-4K and HRBench-8K, HyLaR's improvements over both SkiLa and Monet hold regardless of whether original or reproduced numbers are used (e.g., HyLaR 75.00% vs. original SkiLa 72.0% on HRBench-4K; HyLaR 70.50% vs. original Monet 68.0% on HRBench-8K). The comparison on V* specifically will be presented as competitive rather than superior. revision: yes

Circularity Check

0 steps flagged

No significant circularity; the vMF KL derivation is standard mathematics with stated assumptions, and empirical claims are externally benchmarked. The theory-practice relaxation gap is a framing/correctness concern, not circularity.

full rationale

The paper's derivation chain is largely self-contained. The vMF KL derivation (Supplement §1.2, Eqs. S7–S14) is a standard mathematical result: given two vMF distributions with shared concentration κ on the unit hypersphere, the KL divergence reduces to κ·A_D(κ)·(1 − cos(μ_new, μ_old)). This derivation is parameter-free with explicitly stated assumptions (unit norm, shared κ) and does not reduce to its own outputs by definition. The vMF distribution is cited to external works [3, 6, 33] — none by the present authors. The practical relaxation to unnormalized inner products (§1.3) is openly acknowledged: the paper states 'the exact probabilistic interpretation as a vMF density ratio or KL divergence no longer strictly holds.' This is a theory-practice gap (overclaiming the theoretical guarantee), not circularity — the implementation does not become equivalent to its inputs by construction. The empirical claims (Tables 1–6) are benchmarked against external baselines (GPT-4o, Gemini, Qwen2.5-VL, Monet, SkiLa, etc.) and the ablations compare DePO against GRPO, DAPO, and VLPO. No self-citation chain is load-bearing for the central claims. The vMF-vs-Gaussian ablation (Table 6) and normalized-vs-unnormalized ablation (Table S1) are genuine empirical comparisons, not fitted-input-renamed-as-prediction. The minor score of 2 reflects the framing mismatch between the 'exact closed-form vMF KL' language in the main text and the surrogate actually implemented, which could be seen as a mild overclaim but does not constitute circular reasoning.

Axiom & Free-Parameter Ledger

8 free parameters · 4 axioms · 2 invented entities

The free parameters are mostly tuned via ablation studies, which is standard practice but means the reported numbers reflect optimized hyperparameters. The key domain assumption about hyperspherical geometry is partially contradicted by the practical relaxation. No invented entities are unsupported—canvas tokens are functionally validated and the compressor is a training-only tool.

free parameters (8)
  • alpha (latent loss weight) = 0.5
    Tuned via ablation in Table 4; balances text and latent loss terms in L_PPO = L_tok + alpha * L_lat
  • epsilon_lat (latent clipping range) = 0.05
    Tuned via ablation in Table 5; symmetric clipping for latent positions, much tighter than text clipping of 0.2
  • epsilon_tok_l, epsilon_tok_h (text clipping) = 0.2, 0.28
    Asymmetric text clipping thresholds, stated as experimental settings
  • beta_tok (text KL weight) = 0.01
    Controls KL regularization strength for text tokens
  • beta_lat (latent KL weight) = 0.005
    Controls KL regularization strength for latent tokens
  • kappa (vMF concentration) = 0.01
    Global base scale for vMF distribution; stated as preventing premature entropy collapse
  • K_train (training latent steps) = 8
    Number of latent steps during SFT training; ablated in Fig. 4
  • lambda (SFT canvas loss weight) = not specified
    Balances L_CE and L_Canvas in L_SFT = L_CE + lambda * L_Canvas; value not given in the paper
axioms (4)
  • domain assumption Layer-normalized LLM representations reside on a high-dimensional hyperspherical manifold
    Stated in §3.3 to justify vMF over Gaussian; this is the geometric foundation of the entire DePO formulation. The supplementary (§1.3) partially walks this back by noting norms are typically ~sqrt(D), not 1.
  • domain assumption Importance sampling ratios for discrete and continuous actions exhibit fundamentally different variance characteristics
    Stated in §3.3 and supported by Fig. 3; motivates decoupled clipping. The figure is synthetic/illustrative rather than from actual training.
  • domain assumption Semantic information in high-dimensional LLM hidden states is primarily encoded in vector direction rather than magnitude
    Invoked in §4.3 ablation discussion to explain vMF superiority over Gaussian; cited to prior representation learning studies [6, 33].
  • standard math GRPO-style group normalization is a valid basis for advantage estimation in hybrid action spaces
    DePO builds on GRPO (Eq. S16); the group-relative advantage computation is inherited without modification.
invented entities (2)
  • Canvas tokens (<|canvas_start|>, <|canvas_end|>) independent evidence
    purpose: Delimit continuous latent reasoning steps within the autoregressive sequence
    These are functional architectural additions with clear behavioral semantics; their effectiveness is tested via ablation (K_test=0 vs K_test>0 in Fig. 4). The model learns to trigger them via SFT+RL.
  • Cross-attention canvas compressor no independent evidence
    purpose: Aggregates 729 SigLIP2 patch tokens into 16 compact canvas embeddings for SFT supervision
    Auxiliary module used only during SFT and discarded post-training. Its architecture (L=2 layers, N=16 query tokens) is specified but no independent validation of this specific design choice is provided beyond end-to-end results.

pith-pipeline@v1.1.0-glm · 23643 in / 2930 out tokens · 197752 ms · 2026-07-05T04:19:12.439761+00:00 · methodology

0 comments
read the original abstract

Chain-of-Thought (CoT) reasoning significantly elevates the complex problem-solving capabilities of multimodal large language models (MLLMs). However, adapting CoT to vision typically discretizes signals to fit LLM inputs, causing early semantic collapse and discarding fine-grained details. While external tools can mitigate this, they introduce a rigid bottleneck, confining reasoning to predefined operations. Although recent latent reasoning paradigms internalize visual states to overcome these limitations, optimizing the resulting hybrid discrete-continuous action space remains challenging. In this work, we propose HyLaR (Hybrid Latent Reasoning), a framework that seamlessly interleaves discrete text generation with continuous visual latent representations. Specifically, following an initial cold-start supervised fine-tuning (SFT), we introduce DePO (Decoupled Policy Optimization) to enable effective reinforcement learning within this hybrid space. DePO decomposes the policy gradient objective, applying independent trust-region constraints to the textual and latent components, alongside an exact closed-form von Mises-Fisher (vMF) KL regularizer. Extensive experiments demonstrate that HyLaR outperforms standard MLLMs and state-of-the-art latent reasoning approaches across fine-grained perception and general multimodal understanding benchmarks. Code is available at https://github.com/EthenCheng/HyLaR.

Figures

Figures reproduced from arXiv: 2604.20328 by Hao Zhang, Jinwen Luo, Shi-Zhe Chen, Tao Cheng, Yixin Qin, Zheng Wei.

Figure 1
Figure 1. Figure 1: Comparison between HyLaR and two reasoning paradigms. (A) Text-only CoT: relies solely on explicit CoT, often causing visual grounding errors and redundant steps. (B) Think-with-Image reasoning: depends on external perception tools, leading to unstable invocations and extra latency. (C) HyLaR (ours): refines latent think tokens directly within the latent space to preserve fine-grained visual evidence. disc… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the HyLaR two-stage framework. Stage-I (SFT): Jointly opti￾mizes discrete text via cross-entropy (LCE) and aligns continuous hidden states with compressed ground-truth canvases via an MSE loss (LCanvas). Stage-II (DePO): Re￾fines the hybrid trajectory using RL. Text tokens are updated via standard probability ratios, while latent vectors are optimized on a hypersphere by maximizing vMF-based co… view at source ↗
Figure 3
Figure 3. Figure 3: Importance ratio rt under increas￾ing policy perturbation magnitude for dis￾crete token actions vs. continuous latent ac￾tions. The x-axis, policy perturbation mag￾nitude, is defined as the relative ℓ2 distance between the parameters of the perturbed policy θ and the reference policy θold, i.e., ∥θ − θold∥2/∥θold∥2. This metric quantifies the degree of policy update in the parame￾ter space. (2) Geometric m… view at source ↗
Figure 4
Figure 4. Figure 4: Ablation on inference latent steps (Ktest). We evaluate SFT and RL models trained with varying horizons (Ktrain) on V* and HRBench-8K. The horizontal dashed line represents the baseline of Qwen2.5-VL-7B. Results show that while SFT models suffer from “over-thinking” degradation when Ktest ≫ Ktrain, RL optimization robustly mitigates this drift and extrapolates effectively to extended reasoning budgets [PI… 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.

Forward citations

Cited by 2 Pith papers

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

  1. Imagine Before You Predict: Interleaved Latent Visual Reasoning for Video Event Prediction

    cs.CV 2026-06 unverdicted novelty 7.0

    Future-L1 interleaves latent visual spans with text in MLLM decoding, trained on a custom Future-L1-50K dataset via LA-DAPO RL, and reports SOTA gains on FutureBench (61.0 to 85.4) and TwiFF-Bench (2.44 to 3.04).

  2. OPLD: On-Policy Latent Distillation for Multimodal Reasoning

    cs.CV 2026-07 conditional novelty 6.0

    On-policy distillation of a CoT-privileged teacher’s token preferences and latent trajectories yields a CoT-free student that beats prior visual-latent methods on several multimodal benchmarks.

Reference graph

Works this paper leans on

49 extracted references · 49 canonical work pages · cited by 2 Pith papers · 19 internal anchors

  1. [1]

    Qwen3-VL Technical Report

    Bai, S., Cai, Y., Chen, R., Chen, K., Chen, X., Cheng, Z., Deng, L., Ding, W., Gao, C., Ge, C., et al.: Qwen3-vl technical report. arXiv preprint arXiv:2511.21631 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  2. [2]

    Cheng et al

    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., Lin, J.: Qwen2.5-vl technical report (2025) 16 T. Cheng et al

    work page 2025

  3. [3]

    In: Forty-second International Conference on Machine Learning (2025)

    Bendada, W., Salha-Galvan, G., Hennequin, R., Bontempelli, T., Bouabça, T., Cazenave, T.: Exploring large action sets with hyperspherical embeddings using von mises-fisher sampling. In: Forty-second International Conference on Machine Learning (2025)

    work page 2025

  4. [4]

    Chen, L., Li, J., Dong, X., Zhang, P., Zang, Y., Chen, Z., Duan, H., Wang, J., Qiao, Y., Lin, D., et al.: Are we on the right way for evaluating large vision- language models? Advances in Neural Information Processing Systems37, 27056– 27087 (2024)

    work page 2024

  5. [5]

    Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities

    Comanici, G., Bieber, E., Schaekermann, M., Pasupat, I., Sachdeva, N., Dhillon, I., Blistein, M., Ram, O., Zhang, D., Rosen, E., et al.: Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities. arXiv preprint arXiv:2507.06261 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  6. [6]

    In: Conference on Uncertainty in Artificial Intelligence (UAI) (2018)

    Davidson, T.R., Falorsi, L., De Cao, N., Kipf, T., Tomczak, J.M.: Hyperspherical variational auto-encoders. In: Conference on Uncertainty in Artificial Intelligence (UAI) (2018)

    work page 2018

  7. [7]

    In: European Conference on Computer Vision

    Fu, X., Hu, Y., Li, B., Feng, Y., Wang, H., Lin, X., Roth, D., Smith, N.A., Ma, W.C., Krishna, R.: Blink: Multimodal large language models can see but not per- ceive. In: European Conference on Computer Vision. pp. 148–166. Springer (2024)

    work page 2024

  8. [8]

    Google DeepMind: Gemini 3 flash model card. Tech. rep., Google (2025),https: //storage.googleapis.com/deepmind- media/Model- Cards/Gemini- 3- Flash- Model-Card.pdf

    work page 2025

  9. [9]

    In: Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Guan, T., Liu, F., Wu, X., Xian, R., Li, Z., Liu, X., Wang, X., Chen, L., Huang, F., Yacoob, Y., et al.: Hallusionbench: an advanced diagnostic suite for entangled language hallucination and visual illusion in large vision-language models. In: Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14375–14385 (2024)

    work page 2024

  10. [10]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Gupta, T., Kembhavi, A.: Visual programming: Compositional visual reasoning without training. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14953–14962 (2023)

    work page 2023

  11. [11]

    Training Large Language Models to Reason in a Continuous Latent Space

    Hao, S., Sukhbaatar, S., Su, D., Li, X., Hu, Z., Weston, J., Tian, Y.: Train- ing large language models to reason in a continuous latent space. arXiv preprint arXiv:2412.06769 (2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  12. [12]

    In: The Fourteenth International Conference on Learning Representations (2026)

    Hong, J., Zhao, C., Lu, W., Zhu, C., Xu, G., XingYu: Deepeyesv2: Toward agen- tic multimodal model. In: The Fourteenth International Conference on Learning Representations (2026)

    work page 2026

  13. [13]

    GLM-4.5V and GLM-4.1V-Thinking: Towards Versatile Multimodal Reasoning with Scalable Reinforcement Learning

    Hong, W., Yu, W., Gu, X., Wang, G., Gan, G., Tang, H., Cheng, J., Qi, J., Ji, J., Pan, L., et al.: Glm-4.5 v and glm-4.1 v-thinking: Towards versatile multimodal reasoning with scalable reinforcement learning. arXiv preprint arXiv:2507.01006 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  14. [14]

    Vision-R1: Incentivizing Reasoning Capability in Multimodal Large Language Models

    Huang, W., Jia, B., Zhai, Z., Cao, S., Ye, Z., Zhao, F., Xu, Z., Hu, Y., Lin, S.: Vision-r1: Incentivizing reasoning capability in multimodal large language models. arXiv preprint arXiv:2503.06749 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  15. [15]

    GPT-4o System Card

    Hurst, A., Lerer, A., Goucher, A.P., Perelman, A., Ramesh, A., Clark, A., Os- trow, A., Welihinda, A., Hayes, A., Radford, A., et al.: Gpt-4o system card. arXiv preprint arXiv:2410.21276 (2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  16. [16]

    In: The Fourteenth International Con- ference on Learning Representations (2026)

    Li, A., Wang, C.L., Fu, D., Yue, K., Cai, Z., Zhu, W.B., Liu, O., Guo, P., Neiswanger, W., Huang, F., Goldstein, T., Goldblum, M.: Zebra-cot: A dataset for interleaved vision-language reasoning. In: The Fourteenth International Con- ference on Learning Representations (2026)

    work page 2026

  17. [17]

    Latent Visual Reasoning

    Li, B., Sun, X., Liu, J., Wang, Z., Wu, J., Yu, X., Chen, H., Barsoum, E., Chen, M., Liu, Z.: Latent visual reasoning. arXiv preprint arXiv:2509.24251 (2025) Hybrid Latent Reasoning with Decoupled Policy Optimization 17

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  18. [18]

    LLaVA-OneVision: Easy Visual Task Transfer

    Li, B., Zhang, Y., Guo, D., Zhang, R., Li, F., Zhang, H., Zhang, K., Zhang, P., Li, Y., Liu, Z., et al.: Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326 (2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  19. [19]

    arXiv preprint arXiv:2404.16790 (2024)

    Li, B., Ge, Y., Chen, Y., Ge, Y., Zhang, R., Shan, Y.: Seed-bench-2-plus: Bench- marking multimodal large language models with text-rich visual comprehension. arXiv preprint arXiv:2404.16790 (2024)

    work page arXiv 2024

  20. [20]

    In: The Thirteenth International Conference on Learning Representations (2025)

    Luo, R., Li, Y., Chen, L., He, W., Lin, T.E., Liu, Z., Zhang, L., Song, Z., Alinejad- Rokny, H., Xia, X., Liu, T., Hui, B., Yang, M.: DEEM: Diffusion models serve as the eyes of large language models for image perception. In: The Thirteenth International Conference on Learning Representations (2025)

    work page 2025

  21. [21]

    In: Forty-second International Conference on Machine Learning (2025)

    Mi, Z., Wang, K.C., Qian, G., Ye, H., Liu, R., Tulyakov, S., Aberman, K., Xu, D.: I think, therefore i diffuse: Enabling multimodal in-context reasoning in diffusion models. In: Forty-second International Conference on Machine Learning (2025)

    work page 2025

  22. [22]

    Mull-Tokens: Modality-Agnostic Latent Thinking

    Ray,A.,Abdelkader,A.,Mao,C.,Plummer,B.A.,Saenko,K.,Krishna,R.,Guibas, L., Chu, W.S.: Mull-tokens: Modality-agnostic latent thinking. arXiv preprint arXiv:2512.10941 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  23. [23]

    Proximal Policy Optimization Algorithms

    Schulman, J., Wolski, F., Dhariwal, P., Radford, A., Klimov, O.: Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347 (2017)

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  24. [24]

    DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models

    Shao, Z., Wang, P., Zhu, Q., Xu, R., Song, J., Bi, X., Zhang, H., Zhang, M., Li, Y., Wu, Y., et al.: Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300 (2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  25. [25]

    In: Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing

    Shen, H., Zhao, K., Zhao, T., Xu, R., Zhang, Z., Zhu, M., Yin, J.: Zoomeye: Enhancing multimodal llms with human-like zooming capabilities through tree- based image exploration. In: Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. pp. 6613–6629 (2025)

    work page 2025

  26. [26]

    In: Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing

    Shen, Z., Yan, H., Zhang, L., Hu, Z., Du, Y., He, Y.: CODI: Compressing chain- of-thought into continuous space via self-distillation. In: Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. pp. 677–693 (2025)

    work page 2025

  27. [27]

    arXiv preprint arXiv:2512.17312 (2025)

    Song, Q., Li, H., Yu, Y., Zhou, H., Yang, L., Bai, S., She, Q., Huang, Z., Zhao, Y.: Codedance: A dynamic tool-integrated mllm for executable visual reasoning. arXiv preprint arXiv:2512.17312 (2025)

    work page arXiv 2025

  28. [28]

    arXiv preprint arXiv:2512.16584 (2025)

    Tong, J., Gu, J., Lou, Y., Fan, L., Zou, Y., Wu, Y., Ye, J., Li, R.: Sketch-in-latents: Eliciting unified reasoning in mllms. arXiv preprint arXiv:2512.16584 (2025)

    work page arXiv 2025

  29. [29]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Tong, S., Liu, Z., Zhai, Y., Ma, Y., LeCun, Y., Xie, S.: Eyes wide shut? exploring the visual shortcomings of multimodal llms. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9568–9578 (2024)

    work page 2024

  30. [30]

    SigLIP 2: Multilingual Vision-Language Encoders with Improved Semantic Understanding, Localization, and Dense Features

    Tschannen, M., Gritsenko, A., Wang, X., Naeem, M.F., Alabdulmohsin, I., Parthasarathy, N., Evans, T., Beyer, L., Xia, Y., Mustafa, B., et al.: Siglip 2: Multilingual vision-language encoders with improved semantic understanding, lo- calization, and dense features. arXiv preprint arXiv:2502.14786 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  31. [31]

    VL-Rethinker: Incentivizing Self-Reflection of Vision-Language Models with Reinforcement Learning

    Wang, H., Qu, C., Huang, Z., Chu, W., Lin, F., Chen, W.: Vl-rethinker: Incen- tivizing self-reflection of vision-language models with reinforcement learning. arXiv preprint arXiv:2504.08837 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  32. [32]

    arXiv preprint arXiv:2511.21395 (2025)

    Wang, Q., Shi, Y., Wang, Y., Zhang, Y., Wan, P., Gai, K., Ying, X., Wang, Y.: Monet: Reasoning in latent visual space beyond images and language. arXiv preprint arXiv:2511.21395 (2025)

    work page arXiv 2025

  33. [33]

    In: International conference on ma- chine learning

    Wang, T., Isola, P.: Understanding contrastive representation learning through alignment and uniformity on the hypersphere. In: International conference on ma- chine learning. pp. 9929–9939. PMLR (2020) 18 T. Cheng et al

    work page 2020

  34. [34]

    InternVL3.5: Advancing Open-Source Multimodal Models in Versatility, Reasoning, and Efficiency

    Wang, W., Gao, Z., Gu, L., Pu, H., Cui, L., Wei, X., Liu, Z., Jing, L., Ye, S., Shao, J., et al.: Internvl3. 5: Advancing open-source multimodal models in versatility, reasoning, and efficiency. arXiv preprint arXiv:2508.18265 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  35. [35]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Wang, W., Ding, L., Zeng, M., Zhou, X., Shen, L., Luo, Y., Yu, W., Tao, D.: Divide, conquer and combine: A training-free framework for high-resolution image perception in multimodal large language models. In: Proceedings of the AAAI Conference on Artificial Intelligence. pp. 7907–7915 (2025)

    work page 2025

  36. [36]

    Forest Before Trees: Latent Superposition for Efficient Visual Reasoning

    Wang, Y., Zhang, J., Wu, Y., Lin, Y., Lukas, N., Liu, Y.: Forest before trees: Latent superposition for efficient visual reasoning. arXiv preprint arXiv:2601.06803 (2026)

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  37. [37]

    In: The Fourteenth International Conference on Learning Representations (2026)

    Wei, X., Liu, X., Zang, Y., Dong, X., Cao, Y., Wang, J., Qiu, X., Lin, D.: SIM-cot: Supervised implicit chain-of-thought. In: The Fourteenth International Conference on Learning Representations (2026)

    work page 2026

  38. [38]

    https://github.com/huggingface/trl(2020)

    von Werra, L., Belkada, Y., Tunstall, L., Beeching, E., Thrush, T., Lambert, N., Huang, S., Rasul, K., Gallouédec, Q.: TRL: Transformer Reinforcement Learning. https://github.com/huggingface/trl(2020)

    work page 2020

  39. [39]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Wu, P., Xie, S.: V*: Guided visual search as a core mechanism in multimodal llms. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13084–13094 (2024)

    work page 2024

  40. [40]

    In: The Thirteenth International Conference on Learning Representations (2025)

    Xie, J., Mao, W., Bai, Z., Zhang, D.J., Wang, W., Lin, K.Q., Gu, Y., Chen, Z., Yang, Z., Shou, M.Z.: Show-o: One single transformer to unify multimodal under- standing and generation. In: The Thirteenth International Conference on Learning Representations (2025)

    work page 2025

  41. [41]

    Machine Mental Imagery: Empower Multimodal Reasoning with Latent Visual Tokens

    Yang, Z., Yu, X., Chen, D., Shen, M., Gan, C.: Machine mental imagery: Empower multimodal reasoning with latent visual tokens. arXiv preprint arXiv:2506.17218 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  42. [42]

    In: The Thirty-ninth Annual Conference on Neural Information Processing Systems (2025)

    Yu, Q., Zhang, Z., Zhu, R., Yuan, Y., Zuo, X., YuYue, Dai, W., Fan, T., Liu, G., Liu, J., Liu, L., Liu, X., Lin, H., Lin, Z., Ma, B., Sheng, G., Tong, Y., Zhang, C., Zhang, M., Zhang, R., Zhang, W., Zhu, H., Zhu, J., Chen, J., Chen, J., Wang, C., Yu, H., Song, Y., Wei, X., Zhou, H., Liu, J., Ma, W.Y., Zhang, Y.Q., Yan, L., Wu, Y., Wang, M.: DAPO: An open-...

    work page 2025

  43. [43]

    Advances in neural information processing systems32(2019)

    Zhang, B., Sennrich, R.: Root mean square layer normalization. Advances in neural information processing systems32(2019)

    work page 2019

  44. [44]

    Thyme: Think Beyond Images

    Zhang, Y.F., Lu, X., Yin, S., Fu, C., Chen, W., Hu, X., Wen, B., Jiang, K., Liu, C., Zhang, T., et al.: Thyme: Think beyond images. arXiv preprint arXiv:2508.11630 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  45. [45]

    In: The Thirty- ninth Annual Conference on Neural Information Processing Systems (2025)

    Zhang, Z., He, X., Yan, W., Shen, A., Zhao, C., Wang, X.E.: Soft thinking: Unlock- ing the reasoning potential of LLMs in continuous concept space. In: The Thirty- ninth Annual Conference on Neural Information Processing Systems (2025)

    work page 2025

  46. [46]

    Multimodal Chain-of-Thought Reasoning in Language Models

    Zhang, Z., Zhang, A., Li, M., Zhao, H., Karypis, G., Smola, A.: Multimodal chain- of-thought reasoning in language models. arXiv preprint arXiv:2302.00923 (2023)

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  47. [47]

    arXiv preprint arXiv:2501.12345 (2025)

    Zheng, Y., Lu, J., Wang, S., Feng, Z., Kuang, D., Xiong, Y.: Easyr1: An efficient, scalable, multi-modality rl training framework. arXiv preprint arXiv:2501.12345 (2025)

    work page arXiv 2025

  48. [48]

    thinking with images

    Zheng, Z., Yang, M., Hong, J., Zhao, C., Xu, G., Yang, L., Shen, C., XingYu: Deepeyes: Incentivizing “thinking with images” via reinforcement learning. In: The Fourteenth International Conference on Learning Representations (2026)

    work page 2026

  49. [49]

    InternVL3: Exploring Advanced Training and Test-Time Recipes for Open-Source Multimodal Models

    Zhu, J., Wang, W., Chen, Z., Liu, Z., Ye, S., Gu, L., Tian, H., Duan, Y., Su, W., Shao, J., et al.: Internvl3: Exploring advanced training and test-time recipes for open-source multimodal models. arXiv preprint arXiv:2504.10479 (2025) HyLaR: Hybrid Latent Reasoning with Decoupled Policy Optimization (Supplementary Material) Tao Cheng1, Shi-Zhe Chen1, Hao ...

    work page internal anchor Pith review Pith/arXiv arXiv 2025