pith. sign in

arxiv: 2606.05769 · v1 · pith:CXAUIZ3Onew · submitted 2026-06-04 · 💻 cs.CV

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

Tianxiang Jiang , Linquan Wu , Sheng Xia , Songze Li , Ziang Yan , Haoyu Yang , Yu Qiao , Yi Wang This is my paper

Pith reviewed 2026-06-28 02:42 UTC · model grok-4.3

classification 💻 cs.CV
keywords video event predictionlatent visual reasoninginterleaved decodingmultimodal large language modelsfuture predictionvisual semantics preservation
0
0 comments X

The pith

Interleaved latent visual reasoning improves video event prediction by keeping intermediate steps in visual latent space rather than text.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Video event prediction requires inferring future states from partial video, but existing multimodal models often convert all reasoning steps into text and thereby lose fine-grained motion, geometry, and interaction details. The paper introduces an approach that lets the model alternate between ordinary language tokens and continuous latent visual spans while generating its output. Training uses a curated set of examples chosen because future visual hints improve accuracy, aligns the latent states to actual future-frame embeddings, and applies a latent-aware reinforcement-learning objective that rewards outcome contrast and temporal diversity. Large gains on two benchmarks follow from this design, supporting the view that preserving visual semantics in latent form reduces visually ungrounded predictions.

Core claim

Future-L1 enables an MLLM to perform autoregressive decoding that interleaves language tokens with continuous latent visual spans, trained on Future-L1-50K examples selected for helpful future visual hints, with latent states aligned to future-frame embeddings and further optimized by LA-DAPO using outcome-contrastive and temporal-diversity rewards, producing state-of-the-art scores of 85.4 on FutureBench and 3.04 on TwiFF-Bench.

What carries the argument

Interleaved latent visual reasoning that alternates language tokens with continuous latent visual spans to retain intermediate visual semantics during prediction.

If this is right

  • Future-L1 raises the base model's FutureBench score from 61.0 to 85.4.
  • Future-L1 surpasses the prior best method by 10.4 points on FutureBench.
  • Future-L1 raises the average TwiFF-Bench score from 2.44 to 3.04.
  • Future-oriented video reasoning improves when intermediate visual semantics stay in latent space instead of being translated to text.

Where Pith is reading between the lines

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

  • The interleaving technique may transfer to other predictive multimodal tasks such as robotic action planning where visual continuity matters.
  • Similar latent preservation could reduce hallucinations in longer-horizon video tasks if the alignment procedure scales without additional supervision.
  • The approach suggests a broader design pattern: keep visual information in continuous latent form until the final output step is required.

Load-bearing premise

Selecting examples where future visual hints help prediction and aligning latent states to future-frame embeddings produces a generalizable signal rather than overfitting to the chosen subset or alignment artifacts.

What would settle it

An ablation that trains the same base model on the identical data but removes the latent-state alignment step or the selection criterion, then measures whether the large benchmark gains disappear.

Figures

Figures reproduced from arXiv: 2606.05769 by Haoyu Yang, Linquan Wu, Sheng Xia, Songze Li, Tianxiang Jiang, Yi Wang, Yu Qiao, Ziang Yan.

Figure 1
Figure 1. Figure 1: Motivation of interleaved latent visual reasoning. Text-CoT can be verbose and visually lossy, while pixel-space future simulation is computationally heavy. FUTURE-L1 instead inserts compact latent visual spans that preserve dynamic future semantics without generating full frames. trajectories with outcome-contrastive and temporal￾diversity rewards, encouraging successful latent fu￾tures while discouraging… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of FUTURE-L1. (Left) FUTURE-L1-50K is built by ranking TwiFF candidates by visual gain pv − pt. (Center) SFT trains interleaved text–latent trajectories, aligning latent spans with future visual states. (Right) LA-DAPO further optimizes sampled trajectories with outcome-contrastive and temporal-diversity rewards. FUTURE-L1-50K Training Example <reason> [Textual CoT 0] </reason> <|latent_start|> [L… view at source ↗
Figure 3
Figure 3. Figure 3: FUTURE-L1-50K training format: textual reasoning interleaved with bounded latent visual spans supervised by future-frame embeddings. carry. We therefore filter examples by the marginal utility of their intermediate reasoning frames. For each candidate, we evaluate Qwen3-VL-8B￾Instruct under two conditions: (1) a text-only input with the observed video prefix and question; and (2) a hinted input that additi… view at source ↗
Figure 4
Figure 4. Figure 4: Latent-span usage by reasoning depth. Donuts show span-count distributions; values report mean spans over six RL settings. 5K 10K 20K 2.7 2.8 2.9 3.0 3.1 Acc COT 2.86 2.95 3.11 5K 10K 20K ANS 2.70 2.83 2.97 5K 10K 20K AVG 2.78 2.89 3.04 RL Data Size [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: RL data scaling on TwiFF-Bench. Scores improve as LA-DAPO uses 5K, 10K, and 20K retained visual-gain samples. 68.4, showing that interleaved demonstrations help, but it remains 4.8 points below our visual-gain se￾lected set (73.2). The gap persists on the harder splits, including 3-Hop (70.1 vs. 77.6) and Interp. (67.7 vs. 72.2). Thus FUTURE-L1-50K improves transfer not only by exposing the model to TwiFF￾… view at source ↗
Figure 6
Figure 6. Figure 6: Statistics of FUTURE-L1-50K. Category, visual-gain, reasoning-frame count, and word-count dis￾tributions [PITH_FULL_IMAGE:figures/full_fig_p015_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Word frequency in FUTURE-L1-50K. Text Vision 1st Latent Span 2nd Latent Span 3rd Latent Span 4th Latent Span [PITH_FULL_IMAGE:figures/full_fig_p015_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Stage-wise latent representation. t-SNE of FUTURE-L1-RL embeddings on FutureBench; sequen￾tial latent spans form distinct clusters. hypothesis before the final prediction. Reward Dynamics [PITH_FULL_IMAGE:figures/full_fig_p015_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Reward dynamics during RL. FUTURE-L1 shows higher and more stable rewards than DAPO. Future-L1 System Prompt You are a multimodal reasoning assistant capable of thinking in textual and visual modes. Use the following tags to switch your thinking mode: 1. Textual Mode: <reason>Your textual reasoning process</reason> For logical analysis, planning, and verbal thought. 2. Visual Mode: <|latent_start|>Your vis… view at source ↗
Figure 10
Figure 10. Figure 10: Future-L1 system prompt. TwiFF-Bench User Prompt Template You are an AI assistant capable of reasoning with visual imagery. You should conduct a detailed analysis of the question. Consider different angles, potential solutions, and reason through the problem step-by-step with image. After fully reasoning through the problem–potentially using image-based thinking–provide only a clear, concise, and direct a… view at source ↗
Figure 11
Figure 11. Figure 11: TwiFF-Bench user prompt template. TwiFF-Bench Judge System Prompt You are a strict evaluator. You will have to evaluate the model response reasoning chain and answer based on the reference reasoning chain and ground truth answer. Given: Question: The original forecasting question with image originates from the first video frame. Reference Reasoning Chain: What actually happened, as a reference for the rat… view at source ↗
Figure 12
Figure 12. Figure 12: TwiFF-Bench judge system prompt. 16 [PITH_FULL_IMAGE:figures/full_fig_p016_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: TwiFF-Bench judge user payload template. Accuracy Judge System Prompt You are a strict and objective answer judge. Your sole task is to determine if the model’s predicted answer matches the ground-truth answer based on the question provided. Important Rules: 1. Absolute Truth: The ground truth is the ONLY standard. Even if you think it is factually incorrect, judge based on it. 2. Multiple Choice: Accept … view at source ↗
Figure 14
Figure 14. Figure 14: Accuracy judge system prompt. 17 [PITH_FULL_IMAGE:figures/full_fig_p017_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Successful case: grooming routine. From an observed bedroom scene, FUTURE-L1 predicts the missing sequence of beard trimming, mirror inspection, and returning to bed. The latent spans are inserted around scene and action transitions, while the text keeps the forecast interpretable. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Successful case: product demonstration. FUTURE-L1 tracks the SHOVEL HELPER demonstration from table setup to attachment, outdoor use, and endorsement. The interleaved trajectory separates physical manipulation from later usage scenes. 19 [PITH_FULL_IMAGE:figures/full_fig_p019_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Successful case: staged action sequence. FUTURE-L1 follows a martial-arts montage through performance, balance practice, challenge preparation, and the final meditation scene. The latent spans help bridge visually distinct future stages before the final answer. 20 [PITH_FULL_IMAGE:figures/full_fig_p020_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Failure case: event-specific detail loss. FUTURE-L1 recognizes the baseball-dog setting but predicts a generic continuation rather than the ground-truth sequence with the carpet, refrigerator, and dugout events. The example shows that latent invocation must still preserve fine-grained visual event identity. 21 [PITH_FULL_IMAGE:figures/full_fig_p021_18.png] view at source ↗
read the original abstract

Video event prediction (VEP) requires models to infer unobserved future states from partial video evidence. Existing video MLLMs usually verbalize intermediate future reasoning in text space: once visual evidence is verbalized, fine-grained motion, geometry, and interaction cues can be lost, leading to plausible but visually ungrounded hallucinations. We introduce Future-L1, an interleaved latent visual reasoning framework that lets an MLLM alternate between language tokens and continuous latent visual spans during autoregressive decoding. To train this capability, we construct Future-L1-50K by selecting examples where future visual hints help prediction and align latent states to future-frame embeddings, then further optimize sampled latent trajectories with LA-DAPO, a latent-aware RL objective with outcome-contrastive and temporal-diversity rewards. Future-L1 achieves new state-of-the-art results on both benchmarks: on FutureBench, it improves Qwen3-VL-8B from 61.0 to 85.4 and exceeds the previous best Video-CoE by 10.4 points; on TwiFF-Bench, it improves the average score from 2.44 to 3.04. These results suggest that future-oriented video reasoning benefits from preserving intermediate visual semantics in latent space rather than translating every reasoning step into text.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 1 minor

Summary. The manuscript introduces Future-L1, an interleaved latent visual reasoning framework for video event prediction in multimodal large language models. It constructs the Future-L1-50K dataset by filtering for examples where future visual hints aid prediction and aligning latent states to future-frame embeddings, then optimizes using the LA-DAPO objective with outcome-contrastive and temporal-diversity rewards. The method claims to achieve state-of-the-art performance, improving Qwen3-VL-8B from 61.0 to 85.4 on FutureBench and from 2.44 to 3.04 on TwiFF-Bench, attributing the gains to preserving intermediate visual semantics in latent space rather than verbalizing all reasoning steps.

Significance. If the central performance improvements can be shown to arise specifically from the interleaved latent reasoning mechanism rather than from the dataset curation and alignment procedures, the work would represent a meaningful advance in video event prediction by highlighting the benefits of maintaining visual information in continuous latent form during reasoning. The reported gains are large, but their attribution requires further validation.

major comments (3)
  1. [Abstract / Dataset construction] The Future-L1-50K dataset is constructed by selecting only those examples where future visual hints help prediction. This filtering step, combined with alignment of latent states to future-frame embeddings, may preferentially include cases amenable to visual matching or introduce leakage; without an ablation on the unfiltered dataset or without the alignment loss, the claim that gains (e.g., +24.4 points on FutureBench) result from interleaved latent reasoning is not isolated from these training choices.
  2. [Abstract / Training procedure] The optimization uses LA-DAPO with custom outcome-contrastive and temporal-diversity rewards after alignment. These steps explicitly tie predictions to quantities derived from the training data's future frames. The manuscript provides no controls or ablations demonstrating that the same architecture without selection filter or alignment would yield comparable results, undermining the mechanistic attribution in the abstract.
  3. [Abstract / Experimental results] The abstract reports large benchmark gains but supplies no experimental details, error bars, ablation studies, or controls for the dataset selection step. This prevents verification of whether the central performance claim is supported by the methods.
minor comments (1)
  1. The abstract mentions 'Future-L1-50K' and 'LA-DAPO' without defining them in the provided text; ensure full definitions and motivations are clear in the introduction.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback emphasizing the need to isolate the contribution of interleaved latent visual reasoning from dataset construction and optimization choices. We address each major comment below.

read point-by-point responses

  1. Referee: [Abstract / Dataset construction] The Future-L1-50K dataset is constructed by selecting only those examples where future visual hints help prediction. This filtering step, combined with alignment of latent states to future-frame embeddings, may preferentially include cases amenable to visual matching or introduce leakage; without an ablation on the unfiltered dataset or without the alignment loss, the claim that gains (e.g., +24.4 points on FutureBench) result from interleaved latent reasoning is not isolated from these training choices.

    Authors: We agree that the filtering and alignment steps require explicit controls to strengthen attribution. In the revision we will add ablations on the unfiltered dataset and without the alignment loss, reporting the resulting performance to demonstrate the incremental benefit of the interleaved latent mechanism. revision: yes

  2. Referee: [Abstract / Training procedure] The optimization uses LA-DAPO with custom outcome-contrastive and temporal-diversity rewards after alignment. These steps explicitly tie predictions to quantities derived from the training data's future frames. The manuscript provides no controls or ablations demonstrating that the same architecture without selection filter or alignment would yield comparable results, undermining the mechanistic attribution in the abstract.

    Authors: The LA-DAPO rewards are intended to promote latent-space future reasoning. We will include the requested controls (architecture without the selection filter and without alignment) in the revised manuscript to clarify the source of the observed gains. revision: yes

  3. Referee: [Abstract / Experimental results] The abstract reports large benchmark gains but supplies no experimental details, error bars, ablation studies, or controls for the dataset selection step. This prevents verification of whether the central performance claim is supported by the methods.

    Authors: Abstract length limits preclude full experimental detail. The main text and supplement already contain the core experimental protocol; we will add error bars, the new ablation results, and explicit controls for the selection step in the revision to enable direct verification. revision: yes

Circularity Check

0 steps flagged

No significant circularity; empirical method with external benchmarks

full rationale

The paper presents an empirical training procedure (data curation, latent alignment, and custom RL) and reports numerical improvements on named external benchmarks (FutureBench, TwiFF-Bench). No equations, uniqueness theorems, or self-citations are invoked to derive the central claim; the performance numbers are presented as measured outcomes rather than identities or forced statistical consequences of the training inputs. The method is therefore self-contained against its stated benchmarks.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 3 invented entities

The central claim rests on a newly constructed dataset filtered by future-hint utility, a new RL objective with two reward terms, and the untested premise that latent-to-future-frame alignment preserves predictive signal without distortion.

free parameters (1)
  • LA-DAPO reward weights
    Outcome-contrastive and temporal-diversity reward coefficients are introduced to optimize sampled latent trajectories and are therefore fitted or chosen hyperparameters.
axioms (1)
  • domain assumption Aligning model latent states to future-frame embeddings preserves useful visual semantics for downstream prediction
    Invoked when constructing the training signal for Future-L1.
invented entities (3)
  • Future-L1-50K dataset no independent evidence
    purpose: Training corpus filtered for examples where future visual hints aid prediction
    Custom selection criterion introduced by the authors.
  • LA-DAPO objective no independent evidence
    purpose: Latent-aware RL loss combining outcome-contrastive and temporal-diversity rewards
    New optimization method presented in the work.
  • latent visual spans no independent evidence
    purpose: Continuous visual representations interleaved with text tokens during decoding
    Core representational innovation of the framework.

pith-pipeline@v0.9.1-grok · 5778 in / 1597 out tokens · 76068 ms · 2026-06-28T02:42:19.437671+00:00 · methodology

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

91 extracted references · 52 canonical work pages · 29 internal anchors

  1. [1]

    ECCV , pages =

    Yi Wang and Kunchang Li and Xinhao Li and Jiashuo Yu and Yinan He and Guo Chen and Baoqi Pei and Rongkun Zheng and Zun Wang and Yansong Shi and Tianxiang Jiang and Songze Li and Jilan Xu and Hongjie Zhang and Yifei Huang and Yu Qiao and Yali Wang and Limin Wang , title =. ECCV , pages =

  2. [2]

    Hello GPT-4o , author=

  3. [3]

    Janus-Pro: Unified Multimodal Understanding and Generation with Data and Model Scaling

    Janus-pro: Unified multimodal understanding and generation with data and model scaling , author=. arXiv preprint arXiv:2501.17811 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  4. [4]

    VideoChat: Chat-Centric Video Understanding

    Videochat: Chat-centric video understanding , author=. arXiv preprint arXiv:2305.06355 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  5. [5]

    LLaVA-OneVision: Easy Visual Task Transfer

    Llava-onevision: Easy visual task transfer , author=. arXiv preprint arXiv:2408.03326 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  6. [6]

    2024 , journal =

    VideoChat-Flash: Hierarchical Compression for Long-Context Video Modeling , author =. 2024 , journal =

    2024

  7. [7]

    2024 , journal =

    TimeSuite: Improving MLLMs for Long Video Understanding via Grounded Tuning , author =. 2024 , journal =

    2024

  8. [8]

    2025 , journal =

    Task Preference Optimization: Improving Multimodal Large Language Models with Vision Task Alignment , author =. 2025 , journal =

    2025

  9. [9]

    2025 , journal =

    InternVideo2.5: Empowering Video MLLMs with Long and Rich Context Modeling , author =. 2025 , journal =

    2025

  10. [10]

    2025 , journal =

    Emerging Properties in Unified Multimodal Pretraining , author =. 2025 , journal =

    2025

  11. [11]

    Advances in Neural Information Processing Systems , volume=

    Video-r1: Reinforcing video reasoning in mllms , author=. Advances in Neural Information Processing Systems , volume=

  12. [12]

    VideoChat-R1: Enhancing Spatio-Temporal Perception via Reinforcement Fine-Tuning

    Videochat-r1: Enhancing spatio-temporal perception via reinforcement fine-tuning , author=. arXiv preprint arXiv:2504.06958 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  13. [13]

    2025 , journal =

    Qwen2.5-VL Technical Report , author =. 2025 , journal =

    2025

  14. [14]

    Video-MME: The First-Ever Comprehensive Evaluation Benchmark of Multi-modal LLMs in Video Analysis

    Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis , author=. arXiv preprint arXiv:2405.21075 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  15. [15]

    MLVU: Benchmarking Multi-task Long Video Understanding

    Mlvu: A comprehensive benchmark for multi-task long video understanding , author=. arXiv preprint arXiv:2406.04264 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  16. [16]

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

    Mvbench: A comprehensive multi-modal video understanding benchmark , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  17. [17]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Sigmoid loss for language image pre-training , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  18. [18]

    YOLOv11: An Overview of the Key Architectural Enhancements

    Yolov11: An overview of the key architectural enhancements , author=. arXiv preprint arXiv:2410.17725 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  19. [19]

    VideoLLaMA 3: Frontier Multimodal Foundation Models for Image and Video Understanding

    VideoLLaMA 3: Frontier Multimodal Foundation Models for Image and Video Understanding , author=. arXiv preprint arXiv:2501.13106 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  20. [20]

    VideoLLaMA 2: Advancing Spatial-Temporal Modeling and Audio Understanding in Video-LLMs

    Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms , author=. arXiv preprint arXiv:2406.07476 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  21. [21]

    LLaVA-Video: Video Instruction Tuning With Synthetic Data

    Llava-video: Video instruction tuning with synthetic data , author=. arXiv preprint arXiv:2410.02713 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  22. [22]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    Thinking in space: How multimodal large language models see, remember, and recall spaces , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  23. [23]

    LLaVA-NeXT: A Strong Zero-shot Video Understanding Model , url=

    Zhang, Yuanhan and Li, Bo and Liu, haotian and Lee, Yong jae and Gui, Liangke and Fu, Di and Feng, Jiashi and Liu, Ziwei and Li, Chunyuan , month=. LLaVA-NeXT: A Strong Zero-shot Video Understanding Model , url=

  24. [24]

    2025 , eprint=

    MiMo-VL Technical Report , author=. 2025 , eprint=

    2025

  25. [25]

    2025 , journal =

    InternVL3: Exploring Advanced Training and Test-Time Recipes for Open-Source Multimodal Models , author =. 2025 , journal =

    2025

  26. [26]

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

    Internvl3. 5: Advancing open-source multimodal models in versatility, reasoning, and efficiency , author=. arXiv preprint arXiv:2508.18265 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  27. [27]

    Claude 3.5 Sonnet , year =

  28. [28]

    Findings of the Association for Computational Linguistics: ACL 2024 , pages=

    Tempcompass: Do video llms really understand videos? , author=. Findings of the Association for Computational Linguistics: ACL 2024 , pages=

    2024

  29. [29]

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

    Make your training flexible: Towards deployment-efficient video models , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  30. [30]

    arXiv preprint arXiv:2511.20272 , year=

    VKnowU: Evaluating Visual Knowledge Understanding in Multimodal LLMs , author=. arXiv preprint arXiv:2511.20272 , year=

    work page arXiv

  31. [31]

    arXiv preprint arXiv:2410.12381 , year=

    HumanEval-V: Benchmarking High-Level Visual Reasoning with Complex Diagrams in Coding Tasks , author=. arXiv preprint arXiv:2410.12381 , year=

    work page arXiv

  32. [32]

    Multimodal Chain-of-Thought Reasoning in Language Models

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

    work page internal anchor Pith review Pith/arXiv arXiv

  33. [33]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=

    Anticipating Human Activities Using Object Affordances for Reactive Robotic Response , author=. IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=

  34. [34]

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

    Anticipating Visual Representations from Unlabeled Video , author=. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , pages=

  35. [35]

    arXiv preprint arXiv:2403.13315 , year=

    PuzzleVQA: Diagnosing Multimodal Reasoning Challenges of Language Models with Abstract Visual Patterns , author=. arXiv preprint arXiv:2403.13315 , year=

    work page arXiv

  36. [36]

    arXiv preprint arXiv:2404.03622 , year=

    Mind's Eye of LLMs: Visualization-of-Thought Elicits Spatial Reasoning in Large Language Models , author=. arXiv preprint arXiv:2404.03622 , year=

    work page arXiv

  37. [37]

    arXiv preprint arXiv:2406.09403 , year=

    Visual Sketchpad: Sketching as a Visual Chain of Thought for Multimodal Language Models , author=. arXiv preprint arXiv:2406.09403 , year=

    work page arXiv

  38. [38]

    Imagine while Reasoning in Space: Multimodal Visualization-of-Thought

    Imagine while Reasoning in Space: Multimodal Visualization-of-Thought , author=. arXiv preprint arXiv:2501.07542 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  39. [39]

    Multimodal chain of continuous thought for latent-space reasoning in vision-language models.arXiv preprint arXiv:2508.12587, 2025

    Multimodal Chain of Continuous Thought for Latent-Space Reasoning in Vision-Language Models , author=. arXiv preprint arXiv:2508.12587 , year=

    work page arXiv

  40. [40]

    arXiv preprint arXiv:2511.19418 , year=

    Chain-of-Visual-Thought: Teaching VLMs to See and Think Better with Continuous Visual Tokens , author=. arXiv preprint arXiv:2511.19418 , year=

    work page arXiv

  41. [41]

    arXiv preprint arXiv:2510.11606 , year=

    ExpVid: A Benchmark for Experiment Video Understanding & Reasoning , author=. arXiv preprint arXiv:2510.11606 , year=

    work page arXiv

  42. [42]

    European conference on computer vision , pages=

    Internvideo2: Scaling foundation models for multimodal video understanding , author=. European conference on computer vision , pages=. 2024 , organization=

    2024

  43. [43]

    International Conference on Learning Representations , volume=

    Timesuite: Improving mllms for long video understanding via grounded tuning , author=. International Conference on Learning Representations , volume=

  44. [44]

    arXiv preprint arXiv:2603.03985 , year=

    RIVER: A Real-Time Interaction Benchmark for Video LLMs , author=. arXiv preprint arXiv:2603.03985 , year=

    work page arXiv

  45. [45]

    Video-o3: Native Interleaved Clue Seeking for Long Video Multi-Hop Reasoning

    Video-o3: Native Interleaved Clue Seeking for Long Video Multi-Hop Reasoning , author=. arXiv preprint arXiv:2601.23224 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  46. [46]

    Qwen3-VL Technical Report

    Qwen3-vl technical report , author=. arXiv preprint arXiv:2511.21631 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  47. [47]

    Kimi K2.5: Visual Agentic Intelligence

    Kimi K2. 5: Visual Agentic Intelligence , author=. arXiv preprint arXiv:2602.02276 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  48. [48]

    Video-MME-v2: Towards the Next Stage in Benchmarks for Comprehensive Video Understanding

    Video-MME-v2: Towards the Next Stage in Benchmarks for Comprehensive Video Understanding , author=. arXiv preprint arXiv:2604.05015 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  49. [49]

    2026 , eprint=

    LLaVA-OneVision-2: Towards Next-Generation Perceptual Intelligence , author=. 2026 , eprint=

    2026

  50. [50]

    GLM-5V-Turbo: Toward a Native Foundation Model for Multimodal Agents

    GLM-5V-Turbo: Toward a Native Foundation Model for Multimodal Agents , author=. arXiv preprint arXiv:2604.26752 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  51. [51]

    Thinking with Visual Primitives , author=

  52. [52]

    MiMo-V2-Flash Technical Report

    Mimo-v2-flash technical report , author=. arXiv preprint arXiv:2601.02780 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  53. [53]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    Videoespresso: A large-scale chain-of-thought dataset for fine-grained video reasoning via core frame selection , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  54. [54]

    arXiv preprint arXiv:2602.10675 , year=

    TwiFF (Think With Future Frames): A Large-Scale Dataset for Dynamic Visual Reasoning , author=. arXiv preprint arXiv:2602.10675 , year=

    work page arXiv

  55. [55]

    arXiv preprint arXiv:2507.16746 , year=

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

    work page arXiv

  56. [56]

    arXiv preprint arXiv:2510.27492 , year=

    Thinkmorph: Emergent properties in multimodal interleaved chain-of-thought reasoning , author=. arXiv preprint arXiv:2510.27492 , year=

    work page arXiv

  57. [57]

    DeepEyes: Incentivizing "Thinking with Images" via Reinforcement Learning

    Deepeyes: Incentivizing" thinking with images" via reinforcement learning , author=. arXiv preprint arXiv:2505.14362 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  58. [58]

    arXiv preprint arXiv:2601.05175 , year=

    VideoAuto-R1: Video Auto Reasoning via Thinking Once, Answering Twice , author=. arXiv preprint arXiv:2601.05175 , year=

    work page arXiv

  59. [59]

    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=

  60. [60]

    Training Large Language Models to Reason in a Continuous Latent Space

    Training large language models to reason in a continuous latent space , author=. arXiv preprint arXiv:2412.06769 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  61. [61]

    arXiv preprint arXiv:2509.20317 , year=

    SIM-CoT: Supervised Implicit Chain-of-Thought , author=. arXiv preprint arXiv:2509.20317 , year=

    work page arXiv

  62. [62]

    Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    Codi: Compressing chain-of-thought into continuous space via self-distillation , author=. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    2025

  63. [63]

    arXiv preprint arXiv:2601.10129 , year=

    LaViT: Aligning Latent Visual Thoughts for Multi-modal Reasoning , author=. arXiv preprint arXiv:2601.10129 , year=

    work page arXiv

  64. [64]

    Latent Visual Reasoning

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

    work page internal anchor Pith review Pith/arXiv arXiv

  65. [65]

    Machine Mental Imagery: Empower Multimodal Reasoning with Latent Visual Tokens

    Machine mental imagery: Empower multimodal reasoning with latent visual tokens , author=. arXiv preprint arXiv:2506.17218 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  66. [66]

    arXiv preprint arXiv:2511.21395 , year=

    Monet: Reasoning in latent visual space beyond images and language , author=. arXiv preprint arXiv:2511.21395 , year=

    work page arXiv

  67. [67]

    arXiv preprint arXiv:2512.16584 , year=

    Sketch-in-latents: Eliciting unified reasoning in mllms , author=. arXiv preprint arXiv:2512.16584 , year=

    work page arXiv

  68. [68]

    arXiv preprint arXiv:2602.06040 , year=

    SwimBird: Eliciting Switchable Reasoning Mode in Hybrid Autoregressive MLLMs , author=. arXiv preprint arXiv:2602.06040 , year=

    work page arXiv

  69. [69]

    Xiaomi OneVL: One-Step Latent Reasoning and Planning with Vision-Language Explanation

    OneVL: One-step latent reasoning and planning with vision-language explanation , author=. arXiv preprint arXiv:2604.18486 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  70. [70]

    European conference on computer vision , pages=

    A hierarchical representation for future action prediction , author=. European conference on computer vision , pages=. 2014 , organization=

    2014

  71. [71]

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

    Predicting the future: A jointly learnt model for action anticipation , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  72. [72]

    Video (language) modeling: a baseline for generative models of natural videos

    Video (language) modeling: a baseline for generative models of natural videos , author=. arXiv preprint arXiv:1412.6604 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  73. [73]

    Advances in neural information processing systems , volume=

    Generating videos with scene dynamics , author=. Advances in neural information processing systems , volume=

  74. [74]

    arXiv preprint arXiv:2603.14935 , year=

    Video-CoE: Reinforcing Video Event Prediction via Chain of Events , author=. arXiv preprint arXiv:2603.14935 , year=

    work page arXiv

  75. [75]

    arXiv preprint arXiv:2511.16669 , year=

    Video-as-Answer: Predict and Generate Next Video Event with Joint-GRPO , author=. arXiv preprint arXiv:2511.16669 , year=

    work page arXiv

  76. [76]

    Proceedings of the 33rd ACM International Conference on Multimedia , pages=

    EventFormer: A Node-graph Hierarchical Attention Transformer for Action-centric Video Event Prediction , author=. Proceedings of the 33rd ACM International Conference on Multimedia , pages=

  77. [77]

    arXiv preprint arXiv:2505.22457 , year=

    Fostering video reasoning via next-event prediction , author=. arXiv preprint arXiv:2505.22457 , year=

    work page arXiv

  78. [78]

    arXiv preprint arXiv:2505.01583 , year=

    Tempura: Temporal event masked prediction and understanding for reasoning in action , author=. arXiv preprint arXiv:2505.01583 , year=

    work page arXiv

  79. [79]

    Proceedings of the 2020 conference on empirical methods in natural language processing (EMNLP) , pages=

    What is more likely to happen next? video-and-language future event prediction , author=. Proceedings of the 2020 conference on empirical methods in natural language processing (EMNLP) , pages=

    2020

  80. [80]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    VidEvent: A Large Dataset for Understanding Dynamic Evolution of Events in Videos , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

Showing first 80 references.