Pith. sign in

REVIEW 3 major objections 31 references

A 7B tactile-language model with dynamic contact encoding and chain-of-thought data beats a larger prior model on most tactile reasoning tasks.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-11 08:20 UTC pith:IKGJCUEZ

load-bearing objection Solid tactile-LLM systems paper: first CoT resource plus a simple dynamic encoder, with a real 7B-over-14B result that is useful even if the causal-reasoning story is oversold. the 3 major comments →

arxiv 2607.05131 v1 pith:IKGJCUEZ submitted 2026-07-06 cs.AI

TacReasoner: A Dynamic Tactile-Language Framework for Interactive Reasoning in Real-World Scenarios

Kailin Lyu , Di Wu , Long Xiao , Jianning Zeng , Jianwei He , Chang Lin , Lianyu Hu , Lin Shu
show 2 more authors
Jie Hao Ce Hao
This is my paper
classification cs.AI
keywords tactile-language modelsdynamic tactile encodingchain-of-thoughtTouchCoT-10kDynTac-Benchvisuo-tactile sensingembodied reasoningmultimodal LLMs
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.

Tactile sensing gives machines information about contact, material, and force that vision cannot, yet current tactile-language models still fail in real settings for two reasons: they treat touch as static frames or pixel reconstruction rather than continuous contact dynamics, and they answer from attribute templates without intermediate reasoning, so they hallucinate on new objects and sensors. TacReasoner answers both problems at once. It adds a Dynamic-aware Tactile Encoder that separately models appearance and inter-frame deformation (press, slide, shear) and fuses them under question guidance, then trains on TouchCoT-10k, a new 10k-sample chain-of-thought dataset whose structured think/answer traces supervise causal steps from contact evolution to physical conclusions. The resulting 7B model outperforms the 14B VTV-LLM baseline on most subtasks of physical-property prediction and commonsense reasoning, and it improves markedly on new dynamic discrimination tasks such as real-versus-fake fruit and contact-state estimation. A sympathetic reader cares because reliable touch-language reasoning is a missing piece for robots that must grasp, wipe, or distinguish objects when vision is occluded or ambiguous.

Core claim

Jointly modeling contact temporal dynamics with a Dynamic-aware Tactile Encoder and supervising intermediate reasoning steps with the TouchCoT-10k chain-of-thought dataset produces a tactile-language model that, at only 7B parameters, surpasses the 14B VTV-LLM on most physical-property and commonsense-reasoning subtasks and substantially improves dynamic discrimination tasks that prior attribute-template methods cannot handle.

What carries the argument

Dynamic-aware Tactile Encoder: a dual-branch module that freezes a pretrained appearance encoder, extracts inter-frame differences through a lightweight temporal branch conditioned on the question, and fuses them by cross-attention so the language model receives contact-evolution features rather than static patches; this representation is then aligned and fine-tuned on structured <think>…</think><answer>…</answer> traces from TouchCoT-10k.

Load-bearing premise

The LLM-generated then manually filtered reasoning chains in TouchCoT-10k supply genuine causal supervision of contact dynamics rather than fluent surface descriptions that the model can simply pattern-match.

What would settle it

On held-out objects and sensors, measure whether removing the intermediate <think> tokens (or replacing them with scrambled but fluent text of equal length) collapses accuracy on elasticity, friction, real-versus-fake, and contact-state tasks back to the levels of the non-CoT baseline; if performance stays high, the claimed benefit of structured causal supervision is not supported.

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

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 / 0 minor

Summary. TacReasoner addresses two claimed limitations of prior tactile-language models: weak modeling of temporal contact dynamics and hallucination from attribute/template supervision. It introduces a Dynamic-aware Tactile Encoder that freezes a VTV appearance encoder, adds a trainable temporal branch on frame differences with question-conditioned aggregation and cross-attention fusion (Eqs. 1–6), and a two-stage training pipeline (Stage I tactile–text alignment on VTV-150K; Stage II LoRA SFT on the new TouchCoT-10k CoT dataset). TouchCoT-10k is built by template-prompting an LLM (DeepSeek) on deformation/contact/sliding cues, followed by manual filtering into <think>/<answer> format. DynTac-Bench adds real-vs-fake fruit discrimination and contact-state estimation under a standardized robotic interaction protocol. On VTV-150K (Table I), TacReasoner-7B reaches 66.7 average accuracy and outperforms VTV-LLM-14B on most subtasks, with largest gains on elasticity, SOI, and OSC; ablations (Tables III–IV) and small DynTac results (Table II) attribute gains to DynEncoder and TouchCoT SFT.

Significance. If the gains are robust, the work is a useful systems contribution to embodied multimodal reasoning: it supplies the first tactile CoT resource, a practical dynamic encoder that improves elasticity/friction and higher-level reasoning tasks, and evidence that a 7B model can beat a 14B prior tactile-language baseline on the same evaluation suite. The two-stage paradigm, explicit CoT format, and DynTac real/fake protocol are concrete, reusable artifacts for the community. The paper does not claim theoretical novelty beyond the encoder design and data construction; its value is empirical and resource-oriented for real-world tactile interaction.

major comments (3)
  1. Section III-A Steps 2–3 and the interpretation of Stage-II gains: TouchCoT-10k chains are generated by template-prompting DeepSeek on the same video cues later used for training, then manually filtered for stage coverage and answer consistency. Nothing in the pipeline independently verifies that <think> tokens encode causal contact physics rather than fluent restatements of patterns already available to the frozen VTV encoder. Tables III–IV therefore show that longer, structured targets help, but do not yet establish that 'explicit reasoning mechanisms' reduce hallucination or that gains come from structured causal supervision. A load-bearing check is needed: e.g., human ratings of causal fidelity, comparison against non-CoT long-form targets of matched length, or evaluation of intermediate-step correctness on held-out contact events.
  2. Table II and DynTac-Bench construction (Section IV): RFOR and OCSE use only ~45 samples drawn from the same standardized press–rotate–slide protocol used for data collection. With n this small and no reported variance or significance tests, the 25% and 7% absolute lifts over VTV-LLM-7B cannot support the claim of reliable dynamic-aware reasoning in real-world scenarios. Expand the test set (more objects, sensors, and interaction styles), report confidence intervals, and include a clearer out-of-distribution split before treating DynTac as decisive evidence.
  3. Table I evaluation protocol: averages over three seeds are mentioned but standard deviations are not reported; several subtasks (Combined, OSC, TSA) remain low in absolute terms even for TacReasoner. Without variance or a statistical test against VTV-LLM-14B, the headline '7B outperforms 14B on most subtasks' is only partially supported. Please add per-seed or bootstrap intervals and clarify whether the test split is fully disjoint from any material used in CoT generation.

Circularity Check

0 steps flagged

No load-bearing circularity; empirical ML systems paper whose performance claims rest on held-out QA accuracy and a separately collected DynTac set rather than any result forced by construction.

full rationale

TacReasoner is an engineering contribution (Dynamic-aware Tactile Encoder + TouchCoT-10k CoT SFT + DynTac-Bench) evaluated by standard supervised metrics against external baselines (VTV-LLM, GPT-4o, Gemini, open VLMs) on VTV-150K held-out pairs and a newly robot-collected DynTac set. Equations (1)–(6) are ordinary frozen ViT appearance features plus a trainable temporal difference branch fused by cross-attention; the two-stage losses (7)–(8) are ordinary cross-entropy. Nothing equates a reported accuracy number to a fitted parameter or to a definition. TouchCoT-10k is LLM-generated then manually filtered, which introduces a mild style-matching risk, but final-answer accuracy (not CoT token match) is the reported metric, ablations isolate DynEncoder and CoT data, and DynTac uses a distinct real/fake protocol. Self-citations (Touchformer, TouchThinker) are peripheral and not used to import uniqueness theorems or force the architecture. Per the analyzer rules this is the expected non-finding for a self-contained empirical paper; score 1 only for the non-load-bearing CoT-generation detail.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 3 invented entities

Load-bearing content is architectural and data-construction choices rather than free physical constants. The claim rests on (1) the assumption that frame-difference temporal features plus question-guided attention capture contact dynamics better than pixel reconstruction, (2) that LLM-generated then filtered CoT traces are valid supervision for physical reasoning, and (3) that VTV-150K-derived and small DynTac test sets are representative enough to support 'real-world' claims. Free parameters are training and LoRA hyperparameters; invented entities are the named modules and datasets.

free parameters (3)
  • LoRA rank and scaling factor = rank=128, scale=256
    Rank 128 and scale 256 are chosen by hand for Stage-II SFT; performance depends on this capacity choice.
  • Stage I/II learning rate and step budget = 2e-4, 10000 steps
    AdamW 2e-4 and max 10k steps are fixed hyperparameters that control how much the adapter and LLM adapt to TouchCoT.
  • Temporal encoder and aggregator architecture
    Lightweight temporal branch on frame differences, aggregator, and cross-attention design are design choices not derived from first principles.
axioms (4)
  • domain assumption Inter-frame differences of tactile images plus question-conditioned attention are sufficient to encode contact dynamics (deformation, shear, slip) relevant to physical attributes.
    Stated as motivation from neurobiology and implemented in Eqs. (2)–(5); not independently validated against ground-truth force/shear sensors.
  • ad hoc to paper LLM-generated CoT traces (DeepSeek, template prompts) after manual filtering constitute reliable intermediate supervision for tactile causal reasoning.
    Section III-A Steps 2–3; quality of TouchCoT-10k and the anti-hallucination claim rest on this.
  • domain assumption Freezing the pretrained VTV appearance encoder preserves geometric semantics while the temporal branch adds dynamics without harming alignment.
    Stage I/II training design; ablation supports it but does not prove preservation of all relevant appearance features.
  • domain assumption Standard cross-entropy next-token loss on CoT-formatted outputs induces structured physical reasoning rather than only fluent imitation.
    Eqs. (7)–(8) and Stage-II description; common LLM assumption, unproven for tactile physics.
invented entities (3)
  • Dynamic-aware Tactile Encoder (DynEncoder) no independent evidence
    purpose: Decouple appearance (frozen VTV) from temporal dynamics (frame diffs + question-guided attention) to produce enhanced tactile tokens for the LLM.
    Core architectural contribution; independent evidence is only the paper's own ablations, not external measurements of contact physics.
  • TouchCoT-10k no independent evidence
    purpose: Provide the first tactile chain-of-thought supervision so the model learns intermediate reasoning steps before attribute answers.
    Dataset is constructed for this paper; no external validation that the chains match human tactile reasoning protocols.
  • DynTac-Bench no independent evidence
    purpose: Evaluate dynamic perception (real vs fake fruits, contact state estimation) beyond static attribute QA.
    New benchmark with limited object set; usefulness outside this paper depends on future public release and expansion.

pith-pipeline@v1.1.0-grok45 · 17170 in / 3744 out tokens · 38364 ms · 2026-07-11T08:20:10.054238+00:00 · methodology

0 comments
read the original abstract

Among the five primary human senses, tactile is arguably the most fundamental to survival, as it enables the perception of physical contact and interaction in real-world environments. In this paper, we explore two key challenges of integrating tactile sensing into intelligent systems for multimodal reasoning: (i) insufficient modeling of dynamic tactile signals, which restricts reasoning over temporally evolving properties, and (ii) hallucination in tactile foundation models caused by the absence of explicit reasoning mechanisms, leading to unstable real-world inference. To address these challenges, we propose TacReasoner, a dynamic tactile-language framework for interactive reasoning in real-world scenarios. First, TacReasoner incorporates a Dynamic-aware Tactile Encoder to enhance the perception and representation of dynamic tactile signals. More importantly, we introduce TouchCoT-10k, the first tactile chain-of-thought dataset for structured reasoning over tactile inputs. Upon it, we establish DynTac-Bench to systematically evaluate dynamic tactile perception and real-world commonsense reasoning. Experimental results demonstrate that TacReasoner achieves competitive performance against state-of-the-art models across multiple datasets. Notably, despite using only 7B parameters, TacReasoner outperforms the 14B VTV-LLM model on most subtasks, highlighting its effectiveness and efficiency in tactile commonsense reasoning.

Figures

Figures reproduced from arXiv: 2607.05131 by Ce Hao, Chang Lin, Di Wu, Jianning Zeng, Jianwei He, Jie Hao, Kailin Lyu, Lianyu Hu, Lin Shu, Long Xiao.

Figure 1
Figure 1. Figure 1: The core contributions of TacReasoner. (i) We propose a Dynamic-aware Tactile Encoder to strengthen dynamic tactile representation learning. Upon it, we build TouchCoT-10K, a tactile CoT dataset that provides explicit supervision for tactile reasoning. (ii) We further introduce DynTac-Bench to support dynamic perception and tactile interaction tasks. (iii) Across multiple tactile reasoning benchmarks, TacR… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the TouchCoT-10K construction pipeline and task taxonomy. Step 1: Data Preparation. We collect and organize tactile video samples acquired through a standardized exploration procedure, ensuring that each sequence includes the indenta￾tion to maximum contact and sliding interaction phases. Task metadata are systematically structured, including question type labels and step indicators, to disting… view at source ↗
Figure 3
Figure 3. Figure 3: The framework of TacReasoner. TacReasoner consists of a tactile encoder, a tactile language adapter, and a large language model. The training procedure follows a two stage strategy. In Stage I, we train the tactile encoder and the tactile￾language adapter to adapt the large language model to tactile inputs. In Stage II, we conduct end to end fine tuning on TouchCoT 10k to activate the reasoning capability … view at source ↗
Figure 4
Figure 4. Figure 4: DynTAC data acquisition process. The robotic arm initializes away from the object surface, performs pressing, rotation, and sliding upon contact, and retracts, generating a tactile video of approximately five seconds. It then repositions and repeats the procedure to obtain five videos in total. All remaining data are collected under the same protocol. The model is optimized by minimizing the cross￾entropy … view at source ↗
Figure 5
Figure 5. Figure 5: Overview of the DynTAC Dataset. (1) The top left panel shows five 1:1 real and fake fruit pairs. The bottom row presents detailed views, with red dashed circles marking fake objects that are indistinguishable from real ones under visual inspection or static touch. (2) The top right panel displays five categories of daily objects. fundamental tactile perception to complex reasoning. We introduce its task ta… view at source ↗
Figure 6
Figure 6. Figure 6: The task examples and the corresponding pre￾dictions from TacReasoner. Video1 shows a wrench, and Video2 shows an avocado. TacReasoner accurately answers questions across diverse sensor inputs by reasoning over intrinsic tactile physical states. paradigm. In Stage I, the Tactile-Language Adapter are trained using the raw instruction data from VTV-150K [7] with the AdamW [31] optimizer and a learning rate o… view at source ↗
Figure 7
Figure 7. Figure 7: The reasoning examples of TacReasoner on tactile dynamic-aware reasoning tasks. TacReasoner achieves ac￾curate reasoning across all tasks. [7]. Since most of the VLM models have parameters around 7B, we adopt TacReasoner-7B to ensure a fair compari￾son. To guarantee result robustness, we report the average performance over three runs with different random seeds. The experimental results are summarized in T… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

31 extracted references · 15 linked inside Pith

  1. [1]

    Why is there so much more research on vision than on any other sensory modality?

    F. Hutmacher, “Why is there so much more research on vision than on any other sensory modality?”Frontiers in psychology, vol. 10, p. 481030, 2019

    2019

  2. [2]

    The sense of touch,

    B. O’Shaughnessy, “The sense of touch,”Australasian journal of philosophy, vol. 67, no. 1, pp. 37–58, 1989

    1989

  3. [3]

    Human tactile perception as a standard for artificial tactile sensing—a review,

    J. Dargahi and S. Najarian, “Human tactile perception as a standard for artificial tactile sensing—a review,”The international journal of medical robotics and computer assisted surgery, vol. 1, no. 1, pp. 23–35, 2004

    2004

  4. [4]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clarket al., “Learning transferable visual models from natural language supervision,” inInternational conference on machine learning. PmLR, 2021, pp. 8748–8763

    2021

  5. [5]

    Octopi: Object property reasoning with large tactile-language models,

    S. Yu, K. Lin, A. Xiao, J. Duan, and H. Soh, “Octopi: Object property reasoning with large tactile-language models,”arXiv preprint arXiv:2405.02794, 2024

    Pith/arXiv arXiv 2024

  6. [6]

    Demonstrating the octopi-1.5 visual- tactile-language model,

    S. Yu, K. Lin, and H. Soh, “Demonstrating the octopi-1.5 visual- tactile-language model,”arXiv preprint arXiv:2507.09985, 2025

    Pith/arXiv arXiv 2025

  7. [7]

    Universal visuo-tactile video understanding for embodied interaction,

    Y . Xie, M. Li, S. Li, X. Li, G. Chen, F. Ma, F. R. Yu, and W. Ding, “Universal visuo-tactile video understanding for embodied interaction,”arXiv preprint arXiv:2505.22566, 2025

    Pith/arXiv arXiv 2025

  8. [8]

    Restoring tactile and proprioceptive sensation through a brain interface,

    G. A. Tabot, S. S. Kim, J. E. Winberry, and S. J. Bensmaia, “Restoring tactile and proprioceptive sensation through a brain interface,”Neuro- biology of disease, vol. 83, pp. 191–198, 2015

    2015

  9. [9]

    Binding touch to everything: Learning unified multimodal tactile representations,

    F. Yang, C. Feng, Z. Chen, H. Park, D. Wang, Y . Dou, Z. Zeng, X. Chen, R. Gangopadhyay, A. Owenset al., “Binding touch to everything: Learning unified multimodal tactile representations,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 26 340–26 353

    2024

  10. [10]

    Tactile sensors: A review,

    M. Meribout, N. A. Takele, O. Derege, N. Rifiki, M. El Khalil, V . Tiwari, and J. Zhong, “Tactile sensors: A review,”Measurement, vol. 238, p. 115332, 2024

    2024

  11. [11]

    Recent progress in tactile sensors and their applications in intelligent systems,

    Y . Liu, R. Bao, J. Tao, J. Li, M. Dong, and C. Pan, “Recent progress in tactile sensors and their applications in intelligent systems,”Science Bulletin, vol. 65, no. 1, pp. 70–88, 2020

    2020

  12. [12]

    Touch and go: Learning from human-collected vision and touch,

    F. Yang, C. Ma, J. Zhang, J. Zhu, W. Yuan, and A. Owens, “Touch and go: Learning from human-collected vision and touch,”arXiv preprint arXiv:2211.12498, 2022

    Pith/arXiv arXiv 2022

  13. [13]

    Touchformer: A robust transformer-based framework for multimodal material perception,

    K. Lyu, L. Xiao, J. Zeng, J. Dong, X. Liu, Z. Zou, H. Yang, L. Shu, and J. Hao, “Touchformer: A robust transformer-based framework for multimodal material perception,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 40, no. 22, 2026, pp. 18 496– 18 504

    2026

  14. [14]

    Learning of grasp adaptation through experience and tactile sensing,

    M. Li, Y . Bekiroglu, D. Kragic, and A. Billard, “Learning of grasp adaptation through experience and tactile sensing,” in2014 IEEE/RSJ International Conference on Intelligent Robots and Systems. Ieee, 2014, pp. 3339–3346

    2014

  15. [15]

    Robotic grasping and contact: A review,

    A. Bicchi and V . Kumar, “Robotic grasping and contact: A review,” in Proceedings 2000 ICRA. Millennium conference. IEEE international conference on robotics and automation. Symposia proceedings (Cat. No. 00CH37065), vol. 1. IEEE, 2000, pp. 348–353

    2000

  16. [16]

    Tactile-based insertion for dense box- packing,

    S. Dong and A. Rodriguez, “Tactile-based insertion for dense box- packing,” in2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2019, pp. 7953–7960

    2019

  17. [17]

    A survey of robot manipulation in contact,

    M. Suomalainen, Y . Karayiannidis, and V . Kyrki, “A survey of robot manipulation in contact,”Robotics and Autonomous Systems, vol. 156, p. 104224, 2022

    2022

  18. [18]

    Transferable tactile transformers for representation learning across diverse sensors and tasks,

    J. Zhao, Y . Ma, L. Wang, and E. H. Adelson, “Transferable tactile transformers for representation learning across diverse sensors and tasks,”arXiv preprint arXiv:2406.13640, 2024

    Pith/arXiv arXiv 2024

  19. [19]

    Anytouch: Learning unified static-dynamic representation across mul- tiple visuo-tactile sensors,

    R. Feng, J. Hu, W. Xia, T. Gao, A. Shen, Y . Sun, B. Fang, and D. Hu, “Anytouch: Learning unified static-dynamic representation across mul- tiple visuo-tactile sensors,”arXiv preprint arXiv:2502.12191, 2025

    Pith/arXiv arXiv 2025

  20. [20]

    Surveying the mllm landscape: A meta-review of current surveys,

    M. Li, K. Chen, Z. Bi, M. Liu, X. Song, Z. Jiang, T. Wang, B. Peng, Q. Niu, J. Liuet al., “Surveying the mllm landscape: A meta-review of current surveys,”arXiv preprint arXiv:2409.18991, 2024

    arXiv 2024

  21. [21]

    Qwen-image technical report,

    C. Wu, J. Li, J. Zhou, J. Lin, K. Gao, K. Yan, S.-m. Yin, S. Bai, X. Xu, Y . Chenet al., “Qwen-image technical report,”arXiv preprint arXiv:2508.02324, 2025

    Pith/arXiv arXiv 2025

  22. [22]

    Llava-onevision: Easy visual task transfer,

    B. Li, Y . Zhang, D. Guo, R. Zhang, F. Li, H. Zhang, K. Zhang, P. Zhang, Y . Li, Z. Liuet al., “Llava-onevision: Easy visual task transfer,”arXiv preprint arXiv:2408.03326, 2024

    Pith/arXiv arXiv 2024

  23. [23]

    Touchthinker: Scaling tactile commonsense reasoning to the open world with large-scale data and action-aware representation,

    K. Lyu, D. Wu, P. Zhang, Y . Zheng, Y . Lai, L. Xiao, K. Wu, P. Li, C. Gao, L. Huet al., “Touchthinker: Scaling tactile commonsense reasoning to the open world with large-scale data and action-aware representation,”arXiv preprint arXiv:2606.11637, 2026

    Pith/arXiv arXiv 2026

  24. [24]

    Exploring deepseek: A survey on advances, applications, challenges and future directions,

    Z. Deng, W. Ma, Q.-L. Han, W. Zhou, X. Zhu, S. Wen, and Y . Xiang, “Exploring deepseek: A survey on advances, applications, challenges and future directions,”IEEE/CAA Journal of Automatica Sinica, vol. 12, no. 5, pp. 872–893, 2025

    2025

  25. [25]

    When scaling meets llm finetuning: The effect of data, model and finetuning method,

    B. Zhang, Z. Liu, C. Cherry, and O. Firat, “When scaling meets llm finetuning: The effect of data, model and finetuning method,”arXiv preprint arXiv:2402.17193, 2024

    Pith/arXiv arXiv 2024

  26. [26]

    Gpt-4o system card,

    A. Hurst, A. Lerer, A. P. Goucher, A. Perelman, A. Ramesh, A. Clark, A. Ostrow, A. Welihinda, A. Hayes, A. Radfordet al., “Gpt-4o system card,”arXiv preprint arXiv:2410.21276, 2024

    Pith/arXiv arXiv 2024

  27. [27]

    Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities,

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

    Pith/arXiv arXiv 2025

  28. [28]

    Video instruction tuning with synthetic data, 2024,

    Y . Zhang, J. Wu, W. Li, B. Li, Z. Ma, Z. Liu, and C. Li, “Video instruction tuning with synthetic data, 2024,”URL https://arxiv. org/abs/2410.02713, vol. 17

    Pith/arXiv arXiv 2024

  29. [29]

    Expanding performance boundaries of open- source multimodal models with model, data, and test-time scaling,

    Z. Chen, W. Wang, Y . Cao, Y . Liu, Z. Gao, E. Cui, J. Zhu, S. Ye, H. Tian, Z. Liuet al., “Expanding performance boundaries of open- source multimodal models with model, data, and test-time scaling,” arXiv preprint arXiv:2412.05271, 2024

    Pith/arXiv arXiv 2024

  30. [30]

    Videollama 3: Frontier multimodal foundation models for image and video understanding,

    B. Zhang, K. Li, Z. Cheng, Z. Hu, Y . Yuan, G. Chen, S. Leng, Y . Jiang, H. Zhang, X. Liet al., “Videollama 3: Frontier multimodal foundation models for image and video understanding,”arXiv preprint arXiv:2501.13106, 2025

    Pith/arXiv arXiv 2025

  31. [31]

    Towards understanding conver- gence and generalization of adamw,

    P. Zhou, X. Xie, Z. Lin, and S. Yan, “Towards understanding conver- gence and generalization of adamw,”IEEE transactions on pattern analysis and machine intelligence, vol. 46, no. 9, pp. 6486–6493, 2024

    2024