Pith. sign in

REVIEW 4 major objections 3 minor 2 cited by

LanteRn claims that a multimodal model can reason more accurately by interleaving text with compact latent visual embeddings, and that reinforcement learning on top of supervised grounding is what makes those latent thoughts task-useful.

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 ↗

A 3B vision-language model trained to emit latent visual thought tokens interleaved with text, then refined by reinforcement learning, outperforms a matched text-only baseline on several visual reasoning benchmarks.

T0 review reviewed 2026-08-02 challenge →

load-bearing objection LanteRn interleaves text and latent visual blocks in a sensible two-stage SFT+RL recipe, and its RL numbers are consistently positive—but the RL objective as written severs the gradient to the latent states it claims to train, and the eval has train/eval overlap issues. the 4 major comments →

arxiv 2603.25629 v2 pith:LTHAQG6Y submitted 2026-03-26 cs.CV cs.LG

LanteRn: Latent Visual Structured Reasoning

Andr\'e G. Viveiros , Nuno Gon\c{c}alves , Matthias Lindemann , Andr\'e Martins This is my paper
classification cs.CV cs.LG
keywords latent visual reasoningvisual thought embeddingsmultimodal reasoningvision-language modelssupervised fine-tuningreinforcement learninginterleaved reasoningperception benchmarks
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 reading

The paper tries to establish that a multimodal model can reason about images by interleaving ordinary text with compact continuous 'visual thought' vectors, rather than translating everything into words or generating images. It augments a vision-language transformer so that, at inference, the model can decide to emit a block of latent embeddings derived from its own visual feature space, then continue in text. The authors train this capability in two stages: supervised fine-tuning that pulls latent tokens toward pooled vision-encoder features for the image region a reasoning step refers to, and reinforcement learning that rewards final answer correctness. They report that the RL stage yields consistent gains over both the fine-tuned base model and a text-only control on perception-heavy benchmarks, with the largest improvements on out-of-distribution spatial and relational tasks. A sympathetic reader would care because it suggests a route to more efficient perceptual reasoning than scaling models or generating intermediate images.

Core claim

LanteRn's central claim is that visual reasoning does not have to be verbalized: a model can be taught to think in truncated latent states that encode visual content, and those states can be optimized for task success rather than visual fidelity. The method adds three control tokens to a standard vision-language transformer. When the model outputs the start token, the language modeling head is bypassed for K steps and the transformer's hidden states are emitted directly as latent tokens; the model later returns to text. Supervised training uses the frozen vision encoder's pooled patch embeddings over a referenced bounding box as the regression target for those latent tokens. Reinforcement le

What carries the argument

The load-bearing mechanism is the latent visual thought block: K continuous vectors emitted by the final transformer layer while the language-modeling head is bypassed, bracketed by <|lvr_start|> and <|lvr_end|>. The block is trained by MSE regression toward average-pooled patch embeddings from the frozen vision encoder over the bounding box associated with each reasoning step, and then refined by GRPO-style policy optimization in which latent vectors are replayed as fixed conditioning context while only text likelihoods are scored. This lets the model generate, attend to, and later condition on its own high-dimensional visual states without ever rendering pixels or words for them.

Load-bearing premise

The method assumes that the pooled vision-encoder features for a referenced bounding box are a faithful and useful target for a 'visual thought'—that by minimizing MSE toward those features, the model learns to reason about the region rather than merely copy input features it already has in earlier layers.

What would settle it

Take the trained LantErn-RL-8 model and replace the latent vectors at inference with random vectors, or with pooled features from an unrelated region, while keeping the same token schedule; if benchmark scores do not drop, the specific latent content is not what carries the improvement. Alternatively, linearly probe the latent vectors for the box's location or object identity: if probing accuracy is no better than chance after RL, the latent states are not storing the visual information the paper attributes to them.

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

If this is right

  • If the central claim holds, perception-heavy reasoning tasks do not require the model to verbalize everything it sees; a compact latent channel can carry the visual information that text cannot express.
  • The two-stage recipe—grounding latent states with a frozen vision encoder, then aligning them with outcome rewards—gives a transferable training template for other backbones and modalities.
  • The observed gains on relative-position and object-localization subsets suggest that RL reshapes latent states from appearance reconstruction toward task-relevant spatial features, which could generalize to other tasks with sparse rewards.
  • The result that 3B-scale latent reasoning matches or exceeds text-only controls implies a compute-efficiency story: latent thoughts add fewer tokens than verbose descriptions or generated images, so better perception may come without proportional inference cost.
  • Because latent blocks can be generated and attended to directly, future systems can interleave multiple such blocks per trace, allowing iterative visual refinement inside a single answer.

Where Pith is reading between the lines

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

  • A natural extension the paper leaves implicit is dynamic latent block sizing: since performance does not rise monotonically with K, a gating mechanism that emits variable-length latent blocks per reasoning step could beat any fixed K.
  • One could test whether improvements come from the latent content itself or merely from added computation and token structure by ablating latent vectors to noise at inference; if scores hold, the latent channel is not yet doing the hypothesized work.
  • The reliance on pooled vision-encoder features as targets suggests a ceiling tied to the encoder's representational quality; a jointly trained or adversarially refined latent target might lead to more abstract visual thoughts than the frozen encoder provides.
  • Because the SFT targets come from bounding boxes in synthetic traces, the approach could extend to unlabeled images by using the model's own attention maps or object proposals as pseudo-boxes, removing the need for annotated reasoning traces.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 3 minor

Summary. The paper proposes LanteRn, a framework that augments Qwen2.5-VL with the ability to emit and attend to continuous latent visual tokens, interleaved with text, so that visual reasoning can occur in latent space. Training is in two stages: SFT regresses latent states to pooled ROI features from the model's frozen vision encoder, and RL (GRPO) is then applied over text tokens while treating latent states as conditioning variables, with 'latent state replay' intended to stabilize the policy update. The method is evaluated on subsets of Visual-CoT, V*, and Blink, reporting consistent improvements from the RL stage over a text-only NTP-RL baseline. The paper claims that RL makes latent states transition from perceptually faithful reconstructions to task-driven internal representations, with parity to a 7B model on several benchmarks.

Significance. The idea of interleaving compact latent visual representations with text during reasoning is timely and potentially impactful for perception-heavy multimodal tasks. The controlled comparison against a next-token-prediction baseline (LantErn-NTP) is a genuine strength, as is the explicit intention to use the model's own vision encoder as a teacher. However, the central RL mechanism, as written, is internally inconsistent and cannot support the paper's main empirical and conceptual claims. The evaluation also has protocol weaknesses (no significance tests, possible train/eval overlap, benchmark selection). If the RL formulation is corrected and the experiments are properly validated, the approach could be interesting; as it stands, the evidence for the central claims is not convincing.

major comments (4)
  1. [§3.3.1, Eq. (5)] The paper makes incompatible claims. Eq. (5) defines a GRPO objective over text tokens with π_θ(y_{i,t} | q, I, h_latent_i, y_{i,<t}), and the 'Latent State Replay' paragraph says the model is 'forced to condition on the exact latent vectors H_rollout generated during sampling.' If H_rollout is a fixed constant during the update, the gradient of the objective with respect to the parameters that generated those latent vectors is identically zero, so no RL signal reaches the latent-generating parameters. If H_rollout is re-generated under the current θ, then the latent trace is not fixed, the importance ratios no longer correspond to the rollout distribution, and the stated replay stabilization is void. The paper cannot have it both ways. Consequently, the claim that RL makes latent states 'transition from perceptually faithful reconstruction to task-driven internal visual representations'
  2. [§4.2] Evaluation protocol concerns undermine the empirical claims. (a) The SFT dataset is constructed from Visual-CoT, and Table 1 evaluates on 'a subset of VisCoT' without stating that the evaluation split is disjoint from the training split, so train/eval overlap is possible. (b) The latent size K=8 is selected from SFT ablations on the same evaluation benchmarks (Section 4.4: 'as using 8 latent tokens seems to yield the best overall performance, as indicated in Section 4.3'), so the reported RL gains on those benchmarks are not independent selections. (c) Table 2 reports no error bars, confidence intervals, or significance tests; differences such as 0.83 vs 0.82 on VisCoT are within typical noise for these benchmarks. Please report standard errors across multiple seeds and test for significance.
  3. [§5] The conclusion states that LanteRn achieves 'parity with a 7B model on several benchmarks,' but no 7B baseline is included in any table, and Section 4.1 explicitly declines to compare with 7B-scale latent reasoning methods. This claim is unsupported and should be removed or substantiated with concrete results.
  4. [§3.2.1] The SFT regression targets are computed with the model's own frozen vision encoder, and the model already receives image patch embeddings as input. This raises a circularity concern: the MSE loss may simply teach the LLM to copy visual features already present in its input sequence, rather than learning to generate abstract 'visual thoughts.' The paper states that bounding boxes are not exposed to the model, but the ROI-pooled features are derived from the same encoder whose outputs are inputs to the LLM. An ablation using, e.g., random targets, a different vision encoder, or a zero-shot probe of the learned latents would help establish that the latent states are not merely memorized input features.
minor comments (3)
  1. [§7] The reproducibility statement refers to an appendix, but the manuscript does not include an appendix. Please either add the appendix or remove the reference.
  2. [Throughout] The method name is inconsistent: 'LanteRn' in the title and abstract, 'LantErn' in tables and figures. Please unify the naming.
  3. [§4.2] The benchmark names are used inconsistently ('VisCoT' vs 'Visual-CoT'; 'V*' vs 'VStar'). Also, the V* reference should cite the original benchmark paper more precisely.

Circularity Check

2 steps flagged

Validation leakage (K chosen on eval set, Visual-CoT train/eval overlap) and a latent-replay gradient gap weaken the central claims, but there is no definitional circularity or load-bearing self-citation.

specific steps
  1. other [§4.1 Dataset Construction; §4.2 Evaluation Benchmarks; §4.4 Hyperparameters]
    "we construct a synthetic dataset derived from Visual-CoT (Shao et al., 2024a) … To evaluate LanteRn’s performance, we used a subset of Visual-CoT … latent size k=8, as using 8 latent tokens seems to yield the best overall performance, as indicated in Section 4.3)."

    The SFT supervision data is derived from the same Visual-CoT corpus later used for evaluation, and no disjoint train/eval split is reported, so the §4.3 claim of 'generalization beyond the supervision data' is not established. In addition, K=8 is selected by maximizing Table 1 performance on the very benchmarks that Table 2 uses for the final RL claims, so the RL comparisons are hyperparameter-tuned on the test set rather than independent predictions.

  2. other [§3.3.1 Latent State Replay (Eq. 5)]
    "during policy updates, the model is forced to condition on the exact latent vectors H_rollout generated during sampling. This ensures that probability ratios reflect changes in the text policy under a fixed internal reasoning trace, while still allowing gradients to flow back to the parameters responsible for producing latent states."

    If H_rollout is a fixed replay constant in Eq. (5), the gradient of J(θ) with respect to the latent-generating parameters is identically zero, so RL cannot by construction refine latent states; any observed gains would arise only from text-policy updates on SFT-initialized latents. If H_rollout is re-generated under the current θ, the 'exact latent vectors' are no longer fixed and the replay stabilization is void. The paper asserts both, so the conclusion that 'RL is the stage at which latent states transition…' is not entailed by the stated objective.

full rationale

The SFT stage is not circular: Z_target is computed from the frozen vision encoder on image regions (Eqs. 1–2), and training the LLM hidden states to match it is a distillation/supervision objective rather than a result derived from itself. There are no load-bearing self-citations or imported uniqueness theorems. The main issues are empirical circularity and a broken optimization path: (1) K is chosen from the same evaluation benchmarks on which final results are reported, and Visual-CoT-derived training data is evaluated on Visual-CoT without a stated held-out split; (2) the latent-state replay described in §3.3.1 fixes H_rollout, which severs the gradient to latent-generating parameters, so the paper's key claim that RL reshapes latent representations is not supported by the optimization as written. These are validation and correctness gaps rather than definitional equivalence, hence a moderate score rather than a high one.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The paper introduces no new physical entities; its load-bearing assumptions are domain assumptions about whether hidden states can be supervised and optimized as 'visual thoughts'. The main manual/fitted choices are latent size, loss weight, KL coefficient, and reward weights; latent size K is selected on the evaluation benchmarks.

free parameters (5)
  • latent_block_size_K = 8 for RL (ablations K=4,8,16,32)
    K controls latent capacity and was selected because '8 latent tokens seems to yield best overall performance' (§4.4), i.e., tuned on the same evaluation benchmarks.
  • latent_loss_weight_gamma = 0.1
    Hand-set weighting between text and latent MSE loss (§3.2.2, §4.1).
  • KL_beta = 0.1
    Hand-set KL regularization coefficient in the GRPO objective to limit policy drift (§3.3.1, §4.4).
  • reward_weights = accuracy 1.0, format 1.0
    Hand-chosen weights for the sparse outcome reward and the format reward (§3.3.2, §4.4).
  • rollout_sampling_hyperparameters = T=0.6, top-p=0.85, G=4
    Exploration hyperparameters chosen without reported sweeps (§4.4).
axioms (5)
  • domain assumption Pooled ROI features from the frozen vision encoder are an appropriate regression target for latent visual thoughts.
    Central SFT premise: Eqs. (1)-(2) construct Z_target and Eq. (4) trains hidden states to match it (§3.2.1-3.2.2).
  • domain assumption Latent generation can be treated as an intermediate computation, so optimizing only text-token likelihood with gradients flowing through latent states is a valid policy update.
    Explicitly stated in §3.3 because no explicit probability density over continuous latent vectors is defined; Eq. (5) only scores text tokens.
  • domain assumption Qwen3-VL-235B-Thinking synthetic reasoning traces and bounding boxes are accurate enough to supervise latent states.
    Dataset construction in §4.1 relies on the reference model's output and 'matching ground-truth boxes'; no human agreement is reported.
  • domain assumption Sparse accuracy plus format reward suffices to induce task-utility-aligned latent representations.
    Reward design in §3.3.2 has no intermediate latent-quality metric.
  • domain assumption Qwen2.5-VL-3B can learn to emit meaningful hidden states for newly added control tokens.
    Implied by the architecture modification in §3.1; no proof or preliminary analysis is given.

reviewed 2026-08-02 · how reviews work

0 comments
Cite this review

Pith. "Pith review of LanteRn: Latent Visual Structured Reasoning." pith.science (2026). https://pith.science/paper/LTHAQG6Y

@misc{pith2026260325629,
  author       = {Pith},
  title        = {Pith review of: LanteRn: Latent Visual Structured Reasoning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LTHAQG6Y}},
  note         = {Machine review of arXiv:2603.25629}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

While language reasoning models excel in many tasks, visual reasoning remains challenging for current large multimodal models (LMMs). As a result, most LMMs default to verbalizing perceptual content into text, a strong limitation for tasks requiring fine-grained spatial and visual understanding. While recent approaches take steps toward thinking with images by invoking tools or generating intermediate images, they either rely on external modules, or incur unnecessary computation by reasoning directly in pixel space. In this paper, we introduce LanteRn, a framework that enables LMMs to interleave language with compact latent visual representations, allowing visual reasoning to occur directly in latent space. LanteRn augments a vision-language transformer with the ability to generate and attend to continuous visual thought embeddings during inference. We train the model in two stages: supervised fine-tuning to ground visual features in latent states, followed by reinforcement learning to align latent reasoning with task-level utility. We evaluate LanteRn on three perception-centric benchmarks (VisCoT, V*, and Blink), observing consistent improvements in visual grounding and fine-grained reasoning. These results suggest that internal latent representations provide a promising direction for more efficient multimodal reasoning.

Figures

Figures reproduced from arXiv: 2603.25629 by Andr\'e G. Viveiros, Andr\'e Martins, Matthias Lindemann, Nuno Gon\c{c}alves.

Figure 1
Figure 1. Figure 1: The LanteRn framework enables interleaved reasoning between text and latent representations that encode visual “thoughts”. During inference, LanteRn can automati￾cally decide when to start latent reasoning by outputting a special token. ∗Corresponding author: andre.viveiros@tecnico.ulisboa.pt. 1 arXiv:2603.25629v1 [cs.CV] 26 Mar 2026 [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗

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. DeepLatent: Think with Images via Parallel Latent Visual Reasoning

    cs.CV 2026-05 unverdicted novelty 7.0

    DeepLatent introduces a parallel latent visual reasoning framework with learnable 2D tokens and continuous RL, trained via distillation then RL, plus a new 180K dataset, claiming SOTA benchmark results.

  2. What's Holding Back Latent Visual Reasoning?

    cs.CV 2026-05 unverdicted novelty 5.0

    Latent visual reasoning fails in current models because standard datasets make oracle latents uninformative and inference-time latents collapse away from useful representations.

Reference graph

Works this paper leans on

16 extracted references · 13 linked inside Pith · cited by 2 Pith papers

  1. [3]

    Zixu Cheng, Jian Hu, Ziquan Liu, Chenyang Si, Wei Li, and Shaogang Gong

    URL https://arxiv.org/abs/2405.09818. Zixu Cheng, Jian Hu, Ziquan Liu, Chenyang Si, Wei Li, and Shaogang Gong. V- star: Benchmarking video-llms on video spatio-temporal reasoning,

    Pith/arXiv arXiv

  2. [4]

    Chaorui Deng, Deyao Zhu, Kunchang Li, Chenhui Gou, Feng Li, Zeyu Wang, Shu Zhong, Weihao Yu, Xiaonan Nie, Ziang Song, Guang Shi, and Haoqi Fan

    URLhttps: //arxiv.org/abs/2503.11495. Chaorui Deng, Deyao Zhu, Kunchang Li, Chenhui Gou, Feng Li, Zeyu Wang, Shu Zhong, Weihao Yu, Xiaonan Nie, Ziang Song, Guang Shi, and Haoqi Fan. Emerging properties in unified multimodal pretraining,

    Pith/arXiv arXiv

  3. [5]

    Xingyu Fu, Yushi Hu, Bangzheng Li, Yu Feng, Haoyu Wang, Xudong Lin, Dan Roth, Noah A Smith, Wei-Chiu Ma, and Ranjay Krishna

    URLhttps://arxiv.org/abs/2505.14683. Xingyu Fu, Yushi Hu, Bangzheng Li, Yu Feng, Haoyu Wang, Xudong Lin, Dan Roth, Noah A Smith, Wei-Chiu Ma, and Ranjay Krishna. Blink: Multimodal large language models can see but not perceive.arXiv preprint arXiv:2404.12390,

    Pith/arXiv arXiv

  4. [6]

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee

    URL https://arxiv.org/abs/2509.24251. Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Advances in Neural Information Processing Systems,

    Pith/arXiv arXiv

  5. [8]

    URL https://arxiv.org/abs/2512.13961. Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schul- man, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul F Christiano, Jan Leike, and Ryan Lowe. Training lan- guage models...

    Pith/arXiv arXiv

  6. [9]

    John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov

    URLhttps://proceedings.neurips.cc/paper_files/paper/2022/file/ b1efde53be364a73914f58805a001731-Paper-Conference.pdf. John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms,

    2022

  7. [11]

    Leandro von Werra, Younes Belkada, Victor Sanh, et al

    doi: 10.1109/ICCV51070.2023.01092. Leandro von Werra, Younes Belkada, Victor Sanh, et al. Trl: Transformer reinforcement learning.https://github.com/huggingface/trl,

    arXiv 2023

  8. [12]

    Vl- rethinker: Incentivizing self-reflection of vision-language models with reinforcement learn- ing.arXiv preprint arXiv:2504.08837,

    Haozhe Wang, Chao Qu, Zuming Huang, Wei Chu, Fangzhen Lin, and Wenhu Chen. Vl- rethinker: Incentivizing self-reflection of vision-language models with reinforcement learn- ing.arXiv preprint arXiv:2504.08837,

    Pith/arXiv arXiv

  9. [13]

    Yijia Xiao, Edward Sun, Tianyu Liu, and Wei Wang

    URLhttps://arxiv.org/abs/2312.14135. Yijia Xiao, Edward Sun, Tianyu Liu, and Wei Wang. Logicvista: Multimodal llm logical rea- soning benchmark in visual contexts,

    Pith/arXiv arXiv

  10. [15]

    Zhuosheng Zhang, Aston Zhang, Mu Li, Hai Zhao, George Karypis, and Alex Smola

    URLhttps://arxiv.org/abs/2303.11381. Zhuosheng Zhang, Aston Zhang, Mu Li, Hai Zhao, George Karypis, and Alex Smola. Multi- modal chain-of-thought reasoning in language models,

    Pith/arXiv arXiv

  11. [16]

    URLhttps://arxiv.org/ abs/2302.00923. 11

    Pith/arXiv arXiv

  12. [2017]

    10 Published at ICLR 2026 Workshop on Multimodal Intelligence HaoShao, ShengjuQian, HanXiao, GuangluSong, ZhuofanZong, LetianWang, YuLiu, and Hongsheng Li

    URLhttps://arxiv.org/abs/1707.06347. 10 Published at ICLR 2026 Workshop on Multimodal Intelligence HaoShao, ShengjuQian, HanXiao, GuangluSong, ZhuofanZong, LetianWang, YuLiu, and Hongsheng Li. Visual cot: Advancing multi-modal language models with a comprehensive dataset and benchmark for chain-of-thought reasoning. InThe Thirty-eight Conference on Neural...

    Pith/arXiv arXiv 2026

  13. [2022]

    Qwen2.5-vl technical report,

    9 Published at ICLR 2026 Workshop on Multimodal Intelligence Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, Humen Zhong, Yuanzhi Zhu, Mingkun Yang, Zhaohai Li, Jianqiang Wan, Pengfei Wang, Wei Ding, Zheren Fu, Yiheng Xu, Jiabo Ye, Xi Zhang, Tianbao Xie, Zesen Cheng, Hang Zhang, Zhibo Yang...

    2026

  14. [2023]

    URLhttps://arxiv.org/ abs/2304.08485. Team Olmo, :, Allyson Ettinger, Amanda Bertsch, Bailey Kuehl, David Graham, David Heineman, Dirk Groeneveld, Faeze Brahman, Finbarr Timbers, Hamish Ivison, Jacob Morrison, Jake Poznanski, Kyle Lo, Luca Soldaini, Matt Jordan, Mayee Chen, Michael Noukhovitch, Nathan Lambert, Pete Walsh, Pradeep Dasigi, Robert Berry, Sau...

    Pith/arXiv arXiv

  15. [2024]

    URLhttps://arxiv.org/abs/2407.04973. An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, Chujie Zheng, Dayiheng Liu, Fan Zhou, Fei Huang, Feng Hu, Hao Ge, Haoran Wei, Huan Lin, Jialong Tang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jing Zhou, Jingren Zhou, Junyang L...

    Pith/arXiv arXiv

  16. [2025]

    Chameleon Team

    URLhttps://arxiv.org/abs/2502.13923. Chameleon Team. Chameleon: Mixed-modal early-fusion foundation models,

    Pith/arXiv arXiv

This paper was first reviewed by deepseek-v4-flash on August 2, 2026.