pith. machine review for the scientific record. sign in

arxiv: 2605.11695 · v1 · submitted 2026-05-12 · 💻 cs.CV · cs.AI

Recognition: 2 theorem links

· Lean Theorem

Emergent Communication between Heterogeneous Visual Agents through Decentralized Learning

Mikako Ochiai , Masatoshi Nagano , Tadahiro Taniguchi
Authors on Pith no claims yet

Pith reviewed 2026-05-13 06:31 UTC · model grok-4.3

classification 💻 cs.CV cs.AI
keywords emergent communicationheterogeneous agentsvisual agentsdecentralized learningMetropolis-Hastingscaptioning gameshared tokensprivate perception
0
0 comments X

The pith

Agents with mismatched visual encoders develop shared descriptive token sequences through local acceptance-rejection exchanges alone.

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

The paper investigates whether visual agents with private and differing perceptual representations can still produce a usable common set of discrete tokens. It sets up a decentralized game in which one agent proposes sequences and the other accepts or rejects them according to a Metropolis-Hastings rule applied only to its own visual features. Experiments on MS-COCO demonstrate that the resulting sequences support cross-agent alignment, feature prediction, and retrieval better than a no-communication baseline. Performance and the symmetry of the sequences both decrease as the visual encoders become more dissimilar. The work therefore isolates the effect of representational similarity on the content and structure of an emergent language that arises without any shared perceptual access or external communicative loss.

Core claim

In the Metropolis-Hastings Captioning Game, two agents with frozen but heterogeneous visual encoders exchange proposed token sequences; each listener accepts or rejects a proposal by evaluating it against its private visual features via an MH-style criterion. Starting from random text modules, the agents converge on shared sequences that carry visual information, as measured by improved cross-agent alignment, visual-feature prediction, and image-text retrieval over a no-communication baseline. Moderate encoder mismatch reduces the number of shared sequences while preserving their per-sequence visual specificity; stronger mismatch produces fewer, coarser, and more asymmetric sequences. Ablgtl

What carries the argument

The Metropolis-Hastings Captioning Game (MHCG), in which a speaker proposes discrete token sequences and a listener accepts or rejects each proposal using an MH criterion computed solely from its own visual features.

If this is right

  • Shared token sequences outperform a no-communication baseline on cross-agent alignment, visual-feature prediction, and image-text retrieval tasks.
  • All cross-agent performance metrics decline monotonically as visual-encoder mismatch increases.
  • Moderate encoder heterogeneity reduces the total number of shared sequences while preserving their visual specificity.
  • Stronger encoder heterogeneity yields fewer, coarser-grained, and more asymmetric sequences.
  • Listener-side MH acceptance is required to avoid degenerate token formation.

Where Pith is reading between the lines

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

  • The same local-acceptance mechanism could be tested in multi-agent groups where each pair has a different degree of visual mismatch.
  • If representational similarity controls language symmetry, then deliberately aligning low-level visual features across agents should increase the symmetry of the resulting token sequences.
  • Under high heterogeneity the emergent language may become limited to coarse scene-level descriptions, suggesting a natural trade-off between shared vocabulary size and descriptive precision.
  • The approach could be extended to sequential decision tasks where agents must coordinate actions using only the shared tokens they have already converged upon.

Load-bearing premise

Listener-side Metropolis-Hastings acceptance is sufficient to prevent degenerate token formation and to ground symbols in private visual features without any shared perceptual access.

What would settle it

If cross-agent alignment, feature-prediction, and retrieval metrics remain flat or improve as the visual encoders are made increasingly dissimilar, or if removing the listener-side MH acceptance step still yields non-degenerate shared sequences.

Figures

Figures reproduced from arXiv: 2605.11695 by Masatoshi Nagano, Mikako Ochiai, Tadahiro Taniguchi.

Figure 1
Figure 1. Figure 1: Metropolis–Hastings Captioning Game for heterogeneous visual agents. Two agents [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Matched image-set concepts across encoder conditions. C1–C3 are token-sequence triplets [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Self-consistency diagnostic for the main [PITH_FULL_IMAGE:figures/full_fig_p015_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Single-agent BLIP-style architecture. A frozen ViT image encoder produces visual features [PITH_FULL_IMAGE:figures/full_fig_p016_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Token-sequence diversity, training dynamics, and token-sequence alignment over training [PITH_FULL_IMAGE:figures/full_fig_p023_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: HETERO1 acceptance-rule ablation dynamics. Curves compare MHCG, NoCom, AllAccept, Random accept, and ITM-based accept over training. Shaded bands indicate mean ± SD across three seeds. (A) Cross-agent top-1 retrieval, averaged over A→B and B→A and over i2t/t2i retrieval modes; (B) inter-agent text–text RSA based on normalized Hamming RDMs; (C) unique token sequences, averaged over agents; (D) mean acceptan… view at source ↗
Figure 7
Figure 7. Figure 7: C1–C3 cross-assignment matrices for all condition pairs. Each cell shows representative [PITH_FULL_IMAGE:figures/full_fig_p026_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Additional nearest-neighbor examples for [PITH_FULL_IMAGE:figures/full_fig_p027_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Additional nearest-neighbor examples for [PITH_FULL_IMAGE:figures/full_fig_p028_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Additional nearest-neighbor examples for [PITH_FULL_IMAGE:figures/full_fig_p029_10.png] view at source ↗
read the original abstract

Symbols are shared, but perception is private. We study emergent communication between heterogeneous visual agents through decentralized learning, asking what visual information can become shareable when agents have different visual representations. Instead of optimizing messages through a shared external communicative objective, our agents exchange only discrete token sequences and update their own models using local perceptual evidence. This setting focuses on an underexplored aspect of emergent communication, examining whether common symbols can arise without shared perceptual access, and how the similarity between private visual spaces constrains the content and symmetry of the resulting language. We instantiate this setting in the Metropolis-Hastings Captioning Game (MHCG), where two agents collaboratively form shared captions by exchanging proposed token sequences that a listener accepts or rejects using an MH-style criterion evaluated against its own visual features. We compare three pairings of frozen visual encoders, with agents starting from randomly initialized text modules. Experiments on MS-COCO show that MHCG produces visually informative shared token sequences that outperform a no-communication baseline in cross-agent alignment, visual-feature prediction, and image-text retrieval; all cross-agent metrics decline as encoder mismatch increases. Moderate encoder heterogeneity reduces the number of shared sequences while preserving per-sequence visual specificity, whereas stronger encoder heterogeneity yields fewer, coarser, and more asymmetric sequences. Ablations show that listener-side MH acceptance is critical for avoiding degenerate token formation. These results suggest that shared symbols can arise from local perceptual evaluation alone, with visual representational similarity across encoders shaping both the content and symmetry of the resulting language.

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

2 major / 3 minor

Summary. The paper introduces the Metropolis-Hastings Captioning Game (MHCG), a decentralized protocol in which two agents with heterogeneous frozen visual encoders exchange discrete token sequences and accept or reject proposals according to a listener-side Metropolis-Hastings criterion evaluated solely on the listener’s private visual features. Starting from random text modules, the agents are shown to produce shared captions on MS-COCO that outperform a no-communication baseline on cross-agent alignment, visual-feature prediction, and image-text retrieval; all metrics decline monotonically with increasing encoder mismatch. Moderate heterogeneity reduces the number of shared sequences while preserving per-sequence specificity, whereas stronger mismatch yields fewer, coarser, and more asymmetric sequences. Ablations indicate that the MH acceptance rule is necessary to avoid degenerate token formation.

Significance. If the reported effects are reproducible, the work provides concrete evidence that grounded, non-degenerate symbols can emerge from purely local perceptual evaluation without any shared perceptual access or external communicative loss. The systematic dependence of language content and symmetry on encoder similarity offers a falsifiable prediction for heterogeneous multi-agent systems and strengthens the case that decentralized learning suffices for visual grounding.

major comments (2)
  1. [§4.2] §4.2 and Table 2: the claim that MH acceptance is critical for avoiding degeneracy rests on a single ablation; without quantitative results for the always-accept and random-accept controls (including token entropy, cross-agent alignment scores, and retrieval R@1), it is impossible to isolate the contribution of the MH criterion from the mere presence of any acceptance rule.
  2. [§3.3] §3.3: the acceptance probability is defined using the listener’s own visual features, yet no derivation or bound is given showing that this rule guarantees convergence to a non-degenerate equilibrium; the experimental results therefore remain the sole support for the grounding claim.
minor comments (3)
  1. [§4.1] The abstract and §4.1 omit vocabulary size, training schedule, and number of independent runs; these details are required to assess statistical significance of the reported metric declines.
  2. [Figure 3] Figure 3 caption should state the exact encoder pairs and the quantitative mismatch measure used on the x-axis.
  3. [§3.2] Notation for the proposal distribution q(·) and the acceptance ratio is introduced without an explicit equation number; adding Eq. (X) would improve readability.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the positive evaluation and constructive comments. We address each major point below and have revised the manuscript to strengthen the presentation of the ablation results and clarify the empirical scope of the work.

read point-by-point responses

  1. Referee: [§4.2] §4.2 and Table 2: the claim that MH acceptance is critical for avoiding degeneracy rests on a single ablation; without quantitative results for the always-accept and random-accept controls (including token entropy, cross-agent alignment scores, and retrieval R@1), it is impossible to isolate the contribution of the MH criterion from the mere presence of any acceptance rule.

    Authors: We agree that the original ablation was insufficiently detailed to fully isolate the MH criterion. In the revised manuscript we have expanded §4.2 and Table 2 to report full quantitative results for always-accept and random-accept controls, including token entropy, cross-agent alignment scores, and image-text retrieval R@1. These additional metrics show that both alternative rules produce high-entropy, low-alignment outputs, confirming that the Metropolis-Hastings acceptance step is necessary to avoid degeneracy. revision: yes

  2. Referee: [§3.3] §3.3: the acceptance probability is defined using the listener’s own visual features, yet no derivation or bound is given showing that this rule guarantees convergence to a non-degenerate equilibrium; the experimental results therefore remain the sole support for the grounding claim.

    Authors: The manuscript is explicitly empirical and does not claim a theoretical convergence guarantee. The listener-side MH rule is introduced as a local sampling mechanism that conditions proposals on the listener’s private visual features; no derivation of global convergence is provided because proving such a bound for heterogeneous, decentralized agents is beyond the scope of the present study. We have added a short paragraph in the discussion clarifying that the grounding evidence rests on the reported experiments and that formal analysis remains future work. revision: partial

Circularity Check

0 steps flagged

No significant circularity

full rationale

The paper's central results consist of empirical comparisons between MHCG agents and a no-communication baseline on MS-COCO, with metrics (cross-agent alignment, feature prediction, retrieval) measured directly from held-out data rather than defined in terms of the acceptance criterion or any fitted parameter. Ablations confirm the role of listener-side MH acceptance but do not reduce the reported performance deltas to the acceptance rule itself. No self-citation chain, ansatz smuggling, or uniqueness theorem is invoked to derive the observed degradation with encoder mismatch; the claims rest on observable experimental outcomes.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the domain assumption that local perceptual evaluation alone can ground shared symbols; no free parameters or invented entities are introduced in the abstract.

axioms (1)
  • domain assumption Agents update their own models using only local perceptual evidence without shared access to the other agent's visual features
    Stated as the core of the decentralized learning setting and required for the claim that symbols arise without shared perception.

pith-pipeline@v0.9.0 · 5570 in / 1174 out tokens · 60298 ms · 2026-05-13T06:31:45.628393+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

39 extracted references · 39 canonical work pages · 1 internal anchor

  1. [1]

    H. H. Clark.Using language. Cambridge university press, 1996

    work page 1996

  2. [2]

    H. H. Clark and S. E. Brennan. Grounding in communication. In L. B. Resnick, J. M. Levine, and S. D. Teasley, editors,Perspectives on Socially Shared Cognition, pages 127–149. American Psychological Association, 1991. 9

    work page 1991

  3. [3]

    Dosovitskiy, L

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby. An Image is Worth 16x16 Words: Transformers for image recognition at scale. InInternational Conference on Learning Representations, 2021. URLhttps://openreview.net/forum?id=YicbFdNTTy

    work page 2021

  4. [4]

    K. Friston. The free-energy principle: a unified brain theory?Nature Reviews Neuroscience, 11:127–138, 2010. doi: 10.1038/nrn2787

    work page doi:10.1038/nrn2787 2010

  5. [5]

    K. J. Friston and C. D. Frith. Active inference, communication and hermeneutics.Cortex, 68: 129–143, 2015. doi: 10.1016/j.cortex.2015.03.025

    work page doi:10.1016/j.cortex.2015.03.025 2015

  6. [6]

    K. J. Friston, T. Parr, C. Heins, A. Constant, D. Friedman, T. Isomura, C. Fields, T. Verbelen, M. Ramstead, J. Clippinger, and C. D. Frith. Federated inference and belief sharing.Neuro- science & Biobehavioral Reviews, 156:105500, 2024. doi: 10.1016/j.neubiorev.2023.105500

    work page doi:10.1016/j.neubiorev.2023.105500 2024

  7. [7]

    Havrylov and I

    S. Havrylov and I. Titov. Emergence of language with multi-agent games: Learning to com- municate with sequences of symbols. InAdvances in Neural Information Processing Systems,

    work page

  8. [8]

    Emergence of Language with Multi-agent Games: Learning to Communicate with Sequences of Symbols

    doi: 10.48550/arXiv.1705.11192. URL https://proceedings.neurips.cc/paper_ files/paper/2017/hash/70222949cc0db89ab32c9969754d4758-Abstract.html

    work page Pith review doi:10.48550/arxiv.1705.11192 2017

  9. [9]

    K. He, X. Chen, S. Xie, Y . Li, P. Dollár, and R. Girshick. Masked autoencoders are scal- able vision learners. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16000–16009, 2022. doi: 10.1109/CVPR52688.2022. 01553. URLhttps://openaccess.thecvf.com/content/CVPR2022/html/He_Masked_ Autoencoders_Are_Scalable_...

    work page doi:10.1109/cvpr52688.2022 2022

  10. [10]

    N. L. Hoang, Y . Matsui, Y . Hagiwara, A. Taniguchi, and T. Taniguchi. Compositionality and generalization in emergent communication using metropolis–hastings naming game. In IEEE International Conference on Development and Learning (ICDL), pages 1–7, 2024. doi: 10.1109/ICDL61372.2024.10644635. URL https://doi.org/10.1109/ICDL61372.2024. 10644635

    work page doi:10.1109/icdl61372.2024.10644635 2024

  11. [12]

    Inukai, T

    J. Inukai, T. Taniguchi, A. Taniguchi, and Y . Hagiwara. Recursive metropolis- hastings naming game: Symbol emergence in a multi-agent system based on prob- abilistic generative models.Frontiers in Artificial Intelligence, 6:1229127, 2023. doi: 10.3389/frai.2023.1229127. URL https://www.frontiersin.org/journals/ artificial-intelligence/articles/10.3389/fr...

    work page doi:10.3389/frai.2023.1229127 2023

  12. [13]

    Kriegeskorte, M

    N. Kriegeskorte, M. Mur, and P. A. Bandettini. Representational similarity analysis – con- necting the branches of systems neuroscience.Frontiers in Systems Neuroscience, 2:4,

    work page

  13. [14]

    URLhttps://doi.org/10.3389/neuro.06.004.2008

    doi: 10.3389/neuro.06.004.2008. URL https://www.frontiersin.org/journals/ systems-neuroscience/articles/10.3389/neuro.06.004.2008/full

    work page doi:10.3389/neuro.06.004.2008 2008

  14. [15]

    & Baroni, M

    A. Lazaridou and M. Baroni. Emergent multi-agent communication in the deep learning era.arXiv preprint arXiv:2006.02419, 2020. doi: 10.48550/arXiv.2006.02419. URL https: //arxiv.org/abs/2006.02419

    work page doi:10.48550/arxiv.2006.02419 2006

  15. [16]

    Lazaridou, K

    A. Lazaridou, K. M. Hermann, K. Tuyls, and S. Clark. Emergence of linguistic communication from referential games with symbolic and pixel input. InInternational Conference on Learning Representations, 2018. URLhttps://openreview.net/forum?id=HJGv1Z-AW

    work page 2018

  16. [17]

    J. Li, D. Li, C. Xiong, and S. Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. InProceedings of the 39th International Conference on Machine Learning, volume 162 ofProceedings of Machine Learning Research, pages 12888–12900, 2022. URLhttps://proceedings.mlr.press/v162/li22n.html. 10

    work page 2022

  17. [18]

    T.-Y . Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Dollár, and C. L. Zitnick. Microsoft coco: Common objects in context. InProceedings of ECCV, pages 740–755, 2014. doi: 10.1007/978-3-319-10602-1_48. URL https://www.ri.cmu.edu/publications/ microsoft-coco-common-objects-in-context/

    work page doi:10.1007/978-3-319-10602-1_48 2014

  18. [19]

    Loshchilov and F

    I. Loshchilov and F. Hutter. Decoupled weight decay regularization. InProceedings of ICLR,

    work page

  19. [20]

    URLhttps://openreview.net/forum?id=Bkg6RiCqY7

    work page

  20. [21]

    Mahaut, R

    M. Mahaut, R. Dessì, F. Franzon, and M. Baroni. Referential communication in heterogeneous communities of pre-trained visual deep networks.Transactions on Machine Learning Research,

    work page

  21. [22]

    URLhttps://openreview.net/forum?id=8L3khbpUJL

    work page

  22. [23]

    Matsui, R

    Y . Matsui, R. Yamaki, R. Ueda, S. Shinagawa, and T. Taniguchi. Metropolis-hastings captioning game: Knowledge fusion of vision language models via decentralized bayesian inference. arXiv preprint arXiv:2504.09620, 2025. doi: 10.48550/arXiv.2504.09620. URL https: //arxiv.org/abs/2504.09620

    work page doi:10.48550/arxiv.2504.09620 2025

  23. [24]

    Michel, M

    P. Michel, M. Rita, K. W. Mathewson, O. Tieleman, and A. Lazaridou. Revisiting populations in multi-agent communication. InProceedings of ICLR, 2023

    work page 2023

  24. [25]

    Mordatch and P

    I. Mordatch and P. Abbeel. Emergence of grounded compositional language in multi-agent populations. InProceedings of the AAAI conference on artificial intelligence, volume 32, pages 1495–1502, 2018. doi: 10.1609/aaai.v32i1.11492. URL https://ojs.aaai.org/index. php/AAAI/article/view/11492

    work page doi:10.1609/aaai.v32i1.11492 2018

  25. [26]

    Oquab, T

    M. Oquab, T. Darcet, T. Moutakanni, H. V o, M. Szafraniec, V . Khalidov, P. Fernandez, D. Haziza, F. Massa, A. El-Nouby, M. Assran, N. Ballas, W. Galuba, R. Howes, P.-Y . Huang, S.-W. Li, I. Misra, M. Rabbat, V . Sharma, G. Synnaeve, H. Xu, H. Jegou, J. Mairal, P. Labatut, A. Joulin, and P. Bojanowski. Dinov2: Learning robust visual features without super...

    work page 2024

  26. [27]

    Peters, C

    J. Peters, C. Waubert de Puiseau, H. Tercan, A. Gopikrishnan, G. A. L. De Carvalho, C. Bitter, and T. Meisen. Emergent language: a survey and taxonomy.Autonomous Agents and Multi- Agent Systems, 39(1):18, 2025. doi: 10.1007/s10458-025-09691-y. URL https://doi.org/ 10.1007/s10458-025-09691-y

    work page doi:10.1007/s10458-025-09691-y 2025

  27. [28]

    Pitzer and D

    N. Pitzer and D. Mihai. Learning to communicate across modalities: Perceptual heterogeneity in multi-agent systems.arXiv preprint arXiv:2601.22041, 2026. doi: 10.48550/arXiv.2601.22041. URLhttps://arxiv.org/abs/2601.22041

    work page doi:10.48550/arxiv.2601.22041 2026

  28. [29]

    M. Rita, F. Strub, J.-B. Grill, O. Pietquin, and E. Dupoux. On the role of population heterogeneity in emergent communication. InProceedings of ICLR, 2022

    work page 2022

  29. [30]

    Sucholutsky, L

    I. Sucholutsky, L. Muttenthaler, A. Weller, A. Peng, A. Bobu, P. de Haan, J. S. Bowers, and N. Kriegeskorte. Getting aligned on representational alignment.Transactions on Machine Learning Research, 2025. URLhttps://openreview.net/forum?id=Hiq7lUh4Yn

    work page 2025

  30. [31]

    Taniguchi

    T. Taniguchi. Collective predictive coding hypothesis: Symbol emergence as decentralized bayesian inference.Frontiers in Robotics and AI, 11:1353870, 2024. doi: 10.3389/frobt. 2024.1353870. URL https://www.frontiersin.org/journals/robotics-and-ai/ articles/10.3389/frobt.2024.1353870/full

    work page doi:10.3389/frobt 2024

  31. [32]

    Taniguchi, Y

    T. Taniguchi, Y . Yoshida, Y . Matsui, N. Le Hoang, A. Taniguchi, and Y . Hagiwara. Emer- gent communication through metropolis-hastings naming game with deep generative models. Advanced Robotics, 37(19):1266–1282, 2023. doi: 10.1080/01691864.2023.2260856. URL https://www.tandfonline.com/doi/full/10.1080/01691864.2023.2260856

    work page doi:10.1080/01691864.2023.2260856 2023

  32. [33]

    Taniguchi, M

    T. Taniguchi, M. Oizumi, N. Saji, T. Horii, and N. Tsuchiya. Constructive approach to bidirectional influence between qualia structure and language emergence.Philosophy and the Mind Sciences, 6, 2025. doi: 10.33735/phimisci.2025.11765. 11

    work page doi:10.33735/phimisci.2025.11765 2025

  33. [34]

    Taniguchi, S

    T. Taniguchi, S. Takagi, J. Otsuka, Y . Hayashi, and H. T. Hamada. Collective predictive coding as model of science: Formalizing scientific activities towards generative science.Royal Society Open Science, 12(6), 2025

    work page 2025

  34. [35]

    Taniguchi, R

    T. Taniguchi, R. Ueda, T. Nakamura, M. Suzuki, and A. Taniguchi. Generative emergent commu- nication: Large language model is a collective world model.arXiv preprint arXiv:2501.00226,

    work page arXiv

  35. [36]

    Taniguchi, R

    doi: 10.48550/arXiv.2501.00226. URLhttps://arxiv.org/abs/2501.00226

    work page doi:10.48550/arxiv.2501.00226

  36. [37]

    Well-read students learn better: The impact of student initialization on knowledge distillation

    I. Turc, M.-W. Chang, K. Lee, and K. Toutanova. Well-read students learn better: On the importance of pre-training compact models.arXiv preprint arXiv:1908.08962, 2019. doi: 10.48550/arXiv.1908.08962. URLhttps://github.com/google-research/bert

    work page doi:10.48550/arxiv.1908.08962 1908

  37. [38]

    J. v. Uexküll.A foray into the worlds of animals and humans: With a theory of meaning, volume 12. University of Minnesota Press, 2013. URL https://www.upress.umn.edu/ book-division/books/a-foray-into-the-worlds-of-animals-and-humans

    work page 2013

  38. [39]

    2023 , url =

    U. Upadhyay, S. Karthik, M. Mancini, and Z. Akata. Probvlm: Probabilistic adapter for frozen vision-language models. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1899–1910, 2023. doi: 10.1109/ICCV51070.2023.00182. URL https://openaccess.thecvf.com/content/ICCV2023/html/Upadhyay_ProbVLM_ Probabilistic_Adapter_for_...

    work page doi:10.1109/iccv51070.2023.00182 1910

  39. [40]

    Representation Learning with Contrastive Predictive Coding

    A. van den Oord, Y . Li, and O. Vinyals. Representation learning with contrastive predictive coding.arXiv preprint arXiv:1807.03748, 2018. doi: 10.48550/arXiv.1807.03748. URL https://arxiv.org/abs/1807.03748. 12 A Correspondence between BLIP Training Losses and the Approximate M-step We relate the MHCG training procedure to a Monte Carlo EM-like update fo...

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1807.03748 2018