arxiv: 2604.05834 · v2 · submitted 2026-04-07 · 💻 cs.LG

Recognition: 2 theorem links

· Lean Theorem

Hidden in the Multiplicative Interaction: Uncovering Fragility in Multimodal Contrastive Learning

Tillmann Rheude , Stefan Hegselmann , Roland Eils , Benjamin Wild

Authors on Pith no claims yet

Pith reviewed 2026-05-10 19:06 UTC · model grok-4.3

classification 💻 cs.LG

keywords multimodal contrastive learningmultilinear inner productgating mechanismcross-modal retrievalSymilerobustness to noisetrimodal datamissing modalities

0 comments

The pith

A single unreliable modality distorts cross-modal retrieval scores through the multilinear inner product in multimodal contrastive learning.

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

The paper establishes that extending contrastive learning beyond two modalities by replacing the dot product with the multilinear inner product creates a fragility in which one weakly informative, misaligned, or missing modality propagates errors and warps similarity scores for all pairs. A reader would care because this limits reliable performance on real multimodal data that often contains noise or incomplete inputs. Gated Symile addresses the issue with an attention-based gating mechanism that, for each candidate, interpolates unreliable embeddings toward learnable neutral directions and can select an explicit NULL option when cross-modal alignment is unlikely. Across controlled synthetic tests designed to expose the fragility and three real trimodal datasets, this gated version improves top-1 retrieval accuracy over well-tuned baselines.

Core claim

Symile extends pairwise contrastive objectives such as CLIP by using the multilinear inner product over embeddings from three or more modalities to capture higher-order dependencies. Because the product is multiplicative, a single poor-quality modality embedding multiplies through the entire interaction and distorts the contrastive scores used for retrieval. Gated Symile adds a contrastive, attention-based gating module that adapts each modality's contribution on a per-candidate basis by shifting problematic embeddings toward neutral learnable vectors or invoking an explicit NULL direction when reliable alignment is improbable.

What carries the argument

The attention-based per-candidate gating mechanism that interpolates embeddings toward neutral directions or a NULL option to suppress unreliable modalities inside the multilinear inner product.

If this is right

Gated Symile reaches higher top-1 retrieval accuracy than well-tuned state-of-the-art baselines on synthetic and real trimodal datasets.
The same fragility appears whenever the multilinear inner product is used in the presence of noise, misalignment, or missing inputs.
Gating adapts modality contributions without requiring additional labels or global reweighting.
The approach provides a concrete route toward robust contrastive learning for more than two modalities.

Where Pith is reading between the lines

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

Similar per-candidate gating may improve robustness in other models that rely on multiplicative or higher-order feature interactions.
The method suggests that reliability should be assessed example-by-example rather than with fixed modality weights.
In sensor-fusion settings such as medical imaging or robotics, the gate could reduce errors caused by intermittently missing or degraded inputs.
Extending the gate to variable numbers of modalities would test whether the same mechanism scales beyond trimodal cases.

Load-bearing premise

The attention-based gate can reliably detect and suppress unreliable modalities on a per-candidate basis without extra supervision and without harming alignment when every modality is informative.

What would settle it

A controlled experiment on a dataset where all modalities are verifiably informative and aligned, yet Gated Symile produces lower retrieval accuracy than plain Symile, showing the gate introduces unnecessary distortion.

Figures

Figures reproduced from arXiv: 2604.05834 by Benjamin Wild, Roland Eils, Stefan Hegselmann, Tillmann Rheude.

**Figure 2.** Figure 2: Attention-based gate with sigmoid, NULL option, and neutral directions. We introduce a gate that modulates the contribution of each modality in Symile’s MIP. For a retrieval direction, the gate outputs gated embeddings e G 1 , . . . , eG M by using gate weights {wt→m}M m=1 that control how strongly each modality should influence the MIP score. Intuitively, the gate aims to suppress non-target modalities wh… view at source ↗

**Figure 3.** Figure 3: Analyses of well-tuned models on the Synthetic- [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

**Figure 4.** Figure 4: Scaling analyses of well-tuned models on the Synthetic- [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗

read the original abstract

Contrastive learning has become a standard approach for unsupervised learning from paired data, as demonstrated by CLIP for image-text matching. However, many domains involve more than two modalities and require objectives that capture higher-order dependencies beyond pairwise alignment. Symile extends CLIP to this setting by replacing the dot product with the multilinear inner product (MIP) over modality embeddings. In this work, we show that there is a fragility which ishidden in the multiplicative interaction: a single weakly informative, misaligned, or missing modality can propagate through the objective and distort cross-modal retrieval scores. We propose Gated Symile, a contrastive gating mechanism that adapts modality contributions on an attention-based, per-candidate basis. The gate suppresses unreliable inputs by interpolating embeddings toward learnable neutral directions with an explicit NULL option when reliable cross-modal alignment is unlikely. Across a controlled synthetic benchmark that uncovers this fragility and three real-world trimodal datasets, Gated Symile achieves higher top-1 retrieval accuracy than well-tuned state-of-the-art (sota) baselines. More broadly, our results highlight gating as a step toward robust multimodal contrastive learning beyond two modalities in the presence of noise, misalignment, or missing inputs.

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.

Desk Editor's Note private letter to a colleague

The paper identifies a genuine fragility in multilinear inner products for trimodal contrastive learning and shows that per-candidate gating can improve retrieval, but the gains may not be cleanly attributable to fragility mitigation.

read the letter

The core observation is that a single weak or misaligned modality can drag down the entire MIP-based score in Symile-style objectives, and their gated version counters this by learning to suppress or neutralize bad inputs on a per-example basis. That fragility point is real and worth noting for anyone scaling contrastive learning past two modalities. The synthetic benchmark they built to expose it is a useful addition, and the reported lifts on three real trimodal datasets are concrete enough to take seriously at first glance. The gating mechanism itself—attention-driven interpolation toward neutral directions plus an explicit NULL—is a direct, implementable response to the problem they describe. It does not require extra labels, which keeps the method practical. On the downside, the stress-test concern lands: there is no clear ablation confirming the gate stays near identity when all modalities are informative and aligned. Without that, the accuracy gains could come from the gating network simply adding capacity or smoothing rather than from targeted fragility repair. The abstract also stays light on experimental controls, baseline tuning details, and significance tests, so the headline comparisons feel a bit under-supported until the full tables are checked. The work is aimed at researchers extending contrastive methods to noisy or incomplete multimodal settings, such as medical imaging or sensor data. It is not a foundational theoretical advance, but the empirical angle and the synthetic diagnostic are solid enough that a serious referee should see it. I would send it out for review rather than desk-reject; the idea is testable and the limitation is fixable with one or two extra experiments.

Referee Report

2 major / 2 minor

Summary. The paper extends pairwise contrastive learning (e.g., CLIP) to trimodal settings by replacing the dot product with the multilinear inner product in Symile. It identifies a fragility in this multiplicative interaction whereby a single weakly informative, misaligned, or missing modality can distort cross-modal retrieval scores. To mitigate this, the authors propose Gated Symile, which introduces an attention-based gating network that, on a per-candidate basis, interpolates unreliable modality embeddings toward learnable neutral directions or an explicit NULL embedding. Experiments on a controlled synthetic benchmark and three real trimodal datasets report higher top-1 retrieval accuracy for Gated Symile compared with well-tuned baselines.

Significance. If the gating mechanism can be shown to selectively suppress unreliable modalities while remaining near-neutral on fully informative inputs, the work would provide a concrete step toward robust multimodal contrastive objectives beyond two modalities. The synthetic benchmark is a useful controlled testbed for isolating the fragility effect.

major comments (2)

[Experimental results (synthetic benchmark and real datasets)] The central attribution of performance gains to fragility mitigation rests on the assumption that the learned gate remains close to the identity (or neutral) when all modalities are aligned and informative. No ablation isolating this clean-input regime is reported on either the synthetic benchmark or the real datasets; without it, the improvements over untuned Symile and other baselines could arise from the gating network acting as additional capacity or smoothing rather than targeted suppression.
[Experiments] The experimental section provides insufficient detail on controls, baseline implementations, hyperparameter search procedures, number of random seeds, and statistical significance testing. This limits verification that the reported top-1 accuracy gains are robust and directly comparable.

minor comments (2)

[Abstract] Abstract contains a typographical error: 'which ishidden' should be 'which is hidden'.
[Method] The precise formulation of the attention-based gate (how queries/keys are formed from the three embeddings, how the NULL option is parameterized) should be stated explicitly with equations for reproducibility.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive comments, which help clarify the experimental validation needed to support our claims about fragility in multimodal contrastive learning. We address each major point below and outline revisions to strengthen the manuscript.

read point-by-point responses

Referee: The central attribution of performance gains to fragility mitigation rests on the assumption that the learned gate remains close to the identity (or neutral) when all modalities are aligned and informative. No ablation isolating this clean-input regime is reported on either the synthetic benchmark or the real datasets; without it, the improvements over untuned Symile and other baselines could arise from the gating network acting as additional capacity or smoothing rather than targeted suppression.

Authors: We agree that an explicit ablation on the clean-input regime is necessary to isolate the effect of targeted suppression from general capacity gains. In the revised manuscript, we will add this analysis on the synthetic benchmark by reporting gate interpolation weights and retrieval performance when all three modalities are fully informative and aligned, demonstrating that the gate remains near-neutral without degrading accuracy relative to ungated Symile. For the real datasets, we will include a similar per-modality reliability analysis on subsets where cross-modal alignment is strong, with quantitative metrics on gate behavior. revision: yes
Referee: The experimental section provides insufficient detail on controls, baseline implementations, hyperparameter search procedures, number of random seeds, and statistical significance testing. This limits verification that the reported top-1 accuracy gains are robust and directly comparable.

Authors: We acknowledge the need for greater experimental transparency. The revised version will expand the experimental section with: complete descriptions of baseline implementations (including any adaptations for trimodal settings), the full hyperparameter search procedure and ranges used for all methods, the number of random seeds (five seeds were used across all runs), standard deviations on reported metrics, and statistical significance tests (paired t-tests with p-values) comparing Gated Symile to baselines. These additions will ensure reproducibility and direct comparability. revision: yes

Circularity Check

0 steps flagged

No circularity; empirical proposal grounded in experiments

full rationale

The paper identifies fragility in the multilinear inner product of Symile via a controlled synthetic benchmark and proposes Gated Symile with attention-based per-candidate gating learned end-to-end from the contrastive loss. Improvements are shown through direct accuracy comparisons on synthetic and real trimodal datasets. No load-bearing step reduces by construction to a fitted parameter, self-definition, or self-citation chain; the claims rest on observable experimental outcomes rather than tautological equivalence to inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract provides insufficient technical detail to enumerate specific free parameters, axioms or invented entities; gating appears to introduce learnable neutral directions and NULL option but their exact parameterization is not described.

pith-pipeline@v0.9.0 · 5522 in / 960 out tokens · 49790 ms · 2026-05-10T19:06:28.827145+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

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

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

the gate interpolates embeddings toward learnable neutral directions nm ... ˜em = wt→m em + (1−wt→m)nm
IndisputableMonolith/Foundation/BranchSelection.lean branch_selection unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

Symile replaces CLIP’s dot product with the multilinear inner product (MIP) ... g(x(1),…,x(M))=⟨e1,…,eM⟩/τMIP

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

68 extracted references · 33 canonical work pages · 5 internal anchors

[1]

Acosta, G.J

Julián N. Acosta, Guido J. Falcone, Pranav Rajpurkar, and Eric J. Topol. Multimodal biomedical ai.Nature Medicine, 28(9):1773–1784, September 2022. ISSN 1078-8956, 1546-170X. doi: 10.1038/s41591-022-01981-2

work page doi:10.1038/s41591-022-01981-2 2022
[2]

Sanity checks for saliency maps

Julius Adebayo, Justin Gilmer, Michael Muelly, Ian Goodfellow, Moritz Hardt, and Been Kim. Sanity checks for saliency maps. In S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. Cesa- Bianchi, and R. Garnett, editors,Advances in Neural Information Processing Systems, volume 31. Curran Associates, Inc., 2018. URL https://proceedings.neurips.cc/paper_file...

2018
[3]

Abhijit Bendale and Terrance E. Boult. Towards open set deep networks. In2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV , USA, June 27-30, 2016, page 1563–1572. IEEE Computer Society, 2016. doi: 10.1109/CVPR.2016.173. URLhttps://doi.org/10.1109/CVPR.2016.173

work page doi:10.1109/cvpr.2016.173 2016
[4]

Braunger, Benjamin Wild, Scott T

Thore Buergel, Jakob Steinfeldt, Greg Ruyoga, Maik Pietzner, Daniele Bizzarri, Dina V ojinovic, Julius Upmeier Zu Belzen, Lukas Loock, Paul Kittner, Lara Christmann, Noah Hollmann, Henrik Strangalies, Jana M. Braunger, Benjamin Wild, Scott T. Chiesa, Joachim Spranger, Fabian Klostermann, Erik B. Van Den Akker, Stella Trompet, Simon P. Mooijaart, Naveed Sa...

2022
[5]

Why do we need large batchsizes in contrastive learning? a gradient- bias perspective

Changyou Chen, Jianyi Zhang, Yi Xu, Liqun Chen, Jiali Duan, Yiran Chen, Son Tran, Belinda Zeng, and Trishul Chilimbi. Why do we need large batchsizes in contrastive learning? a gradient- bias perspective. In S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho, and A. Oh, editors,Advances in Neural Information Processing Systems, volume 35, page 33860–3...

2022
[6]

Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey E. Hinton. A simple framework for contrastive learning of visual representations. InProceedings of the 37th International Conference on Machine Learning, ICML 2020, 13-18 July 2020, Virtual Event, volume 119 ofProceedings of Machine Learning Research, page 1597–1607. PMLR, 2020. URL http: //procee...

2020
[7]

Breaking the memory barrier: Near infinite batch size scaling for contrastive loss

Zesen Cheng, Hang Zhang, Kehan Li, Sicong Leng, Zhiqiang Hu, Fei Wu, Deli Zhao, Xin Li, and Lidong Bing. Breaking the memory barrier: Near infinite batch size scaling for contrastive loss. arXiv:2410.17243 [cs], October 2024. URLhttp://arxiv.org/abs/2410.17243

work page arXiv 2024
[8]

On the properties of neural machine translation: Encoder-decoder approaches

Kyunghyun Cho, Bart van Merrienboer, Dzmitry Bahdanau, and Yoshua Bengio. On the properties of neural machine translation: Encoder-decoder approaches. In Dekai Wu, Marine Carpuat, Xavier Carreras, and Eva Maria Vecchi, editors,Proceedings of SSST@EMNLP 10 2014, Eighth Workshop on Syntax, Semantics and Structure in Statistical Translation, Doha, Qatar, 25 ...

work page doi:10.3115/v1/w14-4012 2014
[9]

C. Chow. On optimum recognition error and reject tradeoff.IEEE Transactions on Information Theory, 16(1):41–46, 1970. doi: 10.1109/TIT.1970.1054406

work page doi:10.1109/tit.1970.1054406 1970
[10]

A triangle enables multi- modal alignment beyond cosine similarity

Giordano Cicchetti, Eleonora Grassucci, and Danilo Comminiello. A triangle enables multi- modal alignment beyond cosine similarity. InThe Thirty-ninth Annual Conference on Neu- ral Information Processing Systems, 2025. URL https://openreview.net/forum?id= 3Hjfzh5Eyk

2025
[11]

Gramian mul- timodal representation learning and alignment

Giordano Cicchetti, Eleonora Grassucci, Luigi Sigillo, and Danilo Comminiello. Gramian mul- timodal representation learning and alignment. InThe Thirteenth International Conference on Learning Representations, 2025. URL https://openreview.net/forum?id=ftGnpZrW7P

2025
[12]

Vision transformers need registers

Timothée Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. InThe Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net, 2024. URL https://openreview. net/forum?id=2dnO3LLiJ1

2024
[13]

The road less scheduled

Aaron Defazio, Xingyu Yang, Ahmed Khaled, Konstantin Mishchenko, Harsh Mehta, and Ashok Cutkosky. The road less scheduled. In Amir Globersons, Lester Mackey, Danielle Belgrave, Angela Fan, Ulrich Paquet, Jakub M. Tomczak, and Cheng Zhang, editors,Ad- vances in Neural Information Processing Systems 38: Annual Conference on Neural In- formation Processing S...

2024
[14]

BERT : Pre-training of Deep Bidirectional Transformers for Language Understanding

Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Jill Burstein, Christy Doran, and Thamar Solorio, editors,Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAAC...

work page doi:10.18653/v1/n19-1423 2019
[15]

Jesse Dodge, Suchin Gururangan, Dallas Card, Roy Schwartz, and Noah A. Smith. Show your work: Improved reporting of experimental results. In Kentaro Inui, Jing Jiang, Vincent Ng, and Xiaojun Wan, editors,Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Pro...

work page doi:10.18653/v1/d19-1224 2019
[16]

An image is worth 16x16 words: Transformers for image recognition at scale

Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In9th International Conference on Learning Representations, ICLR 2021, V...

2021
[17]

What to align in multimodal contrastive learning? InThe Thirteenth International Conference on Learning Representations, 2025

Benoit Dufumier, Javiera Castillo Navarro, Devis Tuia, and Jean-Philippe Thiran. What to align in multimodal contrastive learning? InThe Thirteenth International Conference on Learning Representations, 2025. URLhttps://openreview.net/forum?id=Pe3AxLq6Wf

2025
[18]

On the foundations of noise-free selective classification.Journal of Machine Learning Research, 11(53):1605–1641, 2010

Ran El-Yaniv and Yair Wiener. On the foundations of noise-free selective classification.Journal of Machine Learning Research, 11(53):1605–1641, 2010

2010
[19]

Selectivenet: A deep neural network with an integrated reject option

Yonatan Geifman and Ran El-Yaniv. Selectivenet: A deep neural network with an integrated reject option. In Kamalika Chaudhuri and Ruslan Salakhutdinov, editors,Proceedings of the 36th International Conference on Machine Learning, ICML 2019, 9-15 June 2019, Long Beach, California, USA, volume 97 ofProceedings of Machine Learning Research, page 2151–2159. P...

2019
[20]

Imagebind: One embedding space to bind them all

Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), page 15180–15190, June 2023

2023
[21]

Understanding the difficulty of training deep feedforward neural networks

Xavier Glorot and Yoshua Bengio. Understanding the difficulty of training deep feedforward neural networks. In Yee Whye Teh and Mike Titterington, editors,Proceedings of the Thirteenth International Conference on Artificial Intelligence and Statistics, volume 9 ofProceedings of Machine Learning Research, page 249–256, Chia Laguna Resort, Sardinia, Italy, ...

2010
[22]

Dahl, Justin Gilmer, Christopher J

Varun Godbole, George E. Dahl, Justin Gilmer, Christopher J. Shallue, and Zachary Nado. Deep learning tuning playbook, 2023. URL http://github.com/google-research/tuning_ playbook. Version 1.0

2023
[24]

Audioclip: Extending clip to image, text and audio

Andrey Guzhov, Federico Raue, Jörn Hees, and Andreas Dengel. Audioclip: Extending clip to image, text and audio. InICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), page 976–980, 2022. doi: 10.1109/ICASSP43922. 2022.9747631

work page doi:10.1109/icassp43922 2022
[25]

Delving deep into rectifiers: Surpassing human-level performance on imagenet classification

Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. InProceedings of the 2015 IEEE International Conference on Computer Vision (ICCV), ICCV ’15, page 1026–1034, USA,

2015
[27]

Deep Residual Learning for Image Recognition , isbn =

Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV , USA, Jun 27-30, 2016, page 770–778. IEEE Computer Society, 2016. doi: 10.1109/CVPR.2016.90. URLhttps://doi.org/10.1109/CVPR.2016.90

work page doi:10.1109/cvpr.2016.90 2016
[28]

Momentum contrast for unsupervised visual representation learning

Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020

2020
[29]

Large Language Models are Powerful Electronic Health Record Encoders

Stefan Hegselmann, Georg von Arnim, Tillmann Rheude, Noel Kronenberg, David Sontag, Gerhard Hindricks, Roland Eils, and Benjamin Wild. Large language models are powerful electronic health record encoders. arXiv:2502.17403 [cs], October 2025. URL http://arxiv. org/abs/2502.17403

work page internal anchor Pith review Pith/arXiv arXiv 2025
[30]

A baseline for detecting misclassified and out-of-distribution examples in neural networks

Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. In5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings. OpenReview.net,

2017
[31]

URLhttps://openreview.net/forum?id=Hkg4TI9xl
[32]

Geoffrey E. Hinton. Training products of experts by minimizing contrastive divergence.Neural Comput., 14(8):1771–1800, 2002. doi: 10.1162/089976602760128018

work page doi:10.1162/089976602760128018 2002
[33]

Long short-term memory.Neural Computation, 9 (8):1735–1780, 1997

Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory.Neural Computation, 9 (8):1735–1780, 1997

1997
[34]

Squeeze-and-excitation networks

Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2018

2018
[35]

Jacobs, Michael I

Robert A. Jacobs, Michael I. Jordan, Steven J. Nowlan, and Geoffrey E. Hinton. Adaptive mixtures of local experts.Neural Computation, 3(1):79–87, 1991. doi: 10.1162/neco.1991.3.1. 79. 12

work page doi:10.1162/neco.1991.3.1 1991
[36]

Sarthak Jain and Byron C. Wallace. Attention is not explanation. In Jill Burstein, Christy Doran, and Thamar Solorio, editors,Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), page 3543–3556, Minneapolis, Minnesota, June 2019. As...

work page doi:10.18653/v1/n19-1357 2019
[37]

Cr-moe: Consistent routed mixture-of-experts for scaling contrastive learning.Transactions on Machine Learning Research, 2024

Ziyu Jiang, Guoqing Zheng, Yu Cheng, Ahmed Hassan Awadallah, and Zhangyang Wang. Cr-moe: Consistent routed mixture-of-experts for scaling contrastive learning.Transactions on Machine Learning Research, 2024. ISSN 2835-8856. URL https://openreview.net/ forum?id=qKIvn9xL1R

2024
[38]

Jordan and Robert A

Michael I. Jordan and Robert A. Jacobs. Hierarchical mixtures of experts and the em algorithm. Neural Computation, 6(2):181–214, 1994. doi: 10.1162/neco.1994.6.2.181

work page doi:10.1162/neco.1994.6.2.181 1994
[39]

Reasoning models sometimes output illegible chains of thought

Arun Jose. Reasoning models sometimes output illegible chains of thought. InThe Thirty- ninth Annual Conference on Neural Information Processing Systems, 2025. URL https: //openreview.net/forum?id=w1TjXJk846

2025
[40]

u tt, Kristof T. and D \

Pieter-Jan Kindermans, Sara Hooker, Julius Adebayo, Maximilian Alber, Kristof T. Schütt, Sven Dähne, Dumitru Erhan, and Been Kim.The (Un)reliability of Saliency Methods, volume 11700 ofLecture Notes in Computer Science, page 267–280. Springer International Publishing, Cham, 2019. ISBN 978-3-030-28953-9. doi: 10.1007/978-3-030-28954-6_14. URL http: //link....

work page doi:10.1007/978-3-030-28954-6_14 2019
[41]

Shiyu Liang, Yixuan Li, and R. Srikant. Enhancing the reliability of out-of-distribution image detection in neural networks. InInternational Conference on Learning Representations, 2018. URLhttps://openreview.net/forum?id=H1VGkIxRZ

2018
[42]

Zachary C. Lipton. The mythos of model interpretability.Commun. ACM, 61(10):36–43, September 2018. ISSN 0001-0782. doi: 10.1145/3233231

work page doi:10.1145/3233231 2018
[43]

Beyond global similarity: Towards fine-grained, multi-condition multimodal retrieval, 2026 c

Xuan Lu, Kangle Li, Haohang Huang, Rui Meng, Wenjun Zeng, and Xiaoyu Shen. Beyond global similarity: Towards fine-grained, multi-condition multimodal retrieval. arXiv:2603.01082 [cs], March 2026. URLhttp://arxiv.org/abs/2603.01082

work page arXiv 2026
[44]

A unified approach to interpreting model predictions

Scott M Lundberg and Su-In Lee. A unified approach to interpreting model predictions. In I. Guyon, U. V on Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors,Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc., 2017. URL https://proceedings.neurips.cc/paper_files/paper/2017/file/ 8a20a8621...

2017
[45]

2024.00913

Jian Meng, Li Yang, Jinwoo Shin, Deliang Fan, and Jae-Sun Seo. Contrastive dual gating: Learning sparse features with contrastive learning. In2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), page 12247–12255, 2022. doi: 10.1109/CVPR52688. 2022.01194

work page doi:10.1109/cvpr52688 2022
[46]

Representation Learning with Contrastive Predictive Coding

Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv:1807.03748 [cs], January 2019. URL http://arxiv.org/abs/ 1807.03748

work page internal anchor Pith review Pith/arXiv arXiv 2019
[47]

In: McIlraith, S.A., Weinberger, K.Q

Ethan Perez, Florian Strub, Harm De Vries, Vincent Dumoulin, and Aaron Courville. Film: Visual reasoning with a general conditioning layer.Proceedings of the AAAI Conference on Artificial Intelligence, 32(1), April 2018. ISSN 2374-3468, 2159-5399. doi: 10.1609/aaai.v32i1. 11671. URLhttps://ojs.aaai.org/index.php/AAAI/article/view/11671

work page doi:10.1609/aaai.v32i1 2018
[48]

Learning transferable visual models from natural language supervi- sion

Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervi- sion. In Marina Meila and Tong Zhang, editors,Proceedings of the 38th International Conference on Machin...

2021
[49]

Leveraging cam algorithms for explaining medical semantic segmentation.Machine Learning for Biomedical Imaging, 2(iMIMIC 2023 special issue):2089–2102, 2024

Tillmann Rheude, Andreas Wirtz, Arjan Kuijper, and Stefan Wesarg. Leveraging cam algorithms for explaining medical semantic segmentation.Machine Learning for Biomedical Imaging, 2(iMIMIC 2023 special issue):2089–2102, 2024. ISSN 2766-905X. doi: https://doi.org/10. 59275/j.melba.2024-ebd3

2023
[50]

Fusion or Confusion? Multimodal Complexity Is Not All You Need

Tillmann Rheude, Roland Eils, and Benjamin Wild. Fusion or confusion? multimodal complexity is not all you need. arXiv:2512.22991 [cs], December 2025. URL http: //arxiv.org/abs/2512.22991

work page internal anchor Pith review Pith/arXiv arXiv 2025
[51]

Cohort-Based Active Modality Acquisition

Tillmann Rheude, Roland Eils, and Benjamin Wild. Cohort-based active modality acquisition. arXiv:2505.16791 [cs], December 2025. URLhttp://arxiv.org/abs/2505.16791

work page internal anchor Pith review Pith/arXiv arXiv 2025
[52]

Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead.Nat

Cynthia Rudin. Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead.Nat. Mach. Intell., 1(5):206–215, 2019. doi: 10.1038/ S42256-019-0048-X

2019
[53]

Contrasting with symile: Simple model-agnostic representation learning for unlimited modalities

Adriel Saporta, Aahlad Puli, Mark Goldstein, and Rajesh Ranganath. Contrasting with symile: Simple model-agnostic representation learning for unlimited modalities. InAdvances in Neural Information Processing Systems, 2024. URLhttps://arxiv.org/pdf/2411.01053

work page arXiv 2024
[54]

Scheirer, Anderson de Rezende Rocha, Archana Sapkota, and Terrance E

Walter J. Scheirer, Anderson de Rezende Rocha, Archana Sapkota, and Terrance E. Boult. Toward open set recognition.IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(7):1757–1772, 2013. doi: 10.1109/TPAMI.2012.256

work page doi:10.1109/tpami.2012.256 2013
[55]

When explanations lie: Why many modified bp attributions fail

Leon Sixt, Maximilian Granz, and Tim Landgraf. When explanations lie: Why many modified bp attributions fail. InProceedings of the 37th International Conference on Machine Learning, ICML 2020, 13-18 July 2020, Virtual Event, volume 119 ofProceedings of Machine Learning Research, page 9046–9057. PMLR, 2020. URL http://proceedings.mlr.press/v119/ sixt20a.html

2020
[56]

Prototypical networks for few-shot learning

Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. In I. Guyon, U. V on Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors,Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc., 2017. URL https://proceedings.neurips.cc/paper_files/paper/2017/file/ cb8d...

2017
[57]

Medical history predicts phenome-wide disease onset and enables the rapid response to emerging health threats.Nature Communications, 16(1):585, January 2025

Jakob Steinfeldt, Benjamin Wild, Thore Buergel, Maik Pietzner, Julius Upmeier Zu Belzen, Andre Vauvelle, Stefan Hegselmann, Spiros Denaxas, Harry Hemingway, Claudia Langenberg, Ulf Landmesser, John Deanfield, and Roland Eils. Medical history predicts phenome-wide disease onset and enables the rapid response to emerging health threats.Nature Communications...

work page doi:10.1038/s41467-025-55879-x 2025
[58]

Uk biobank: an open access resource for identifying the causes of a wide range of complex diseases of middle and old age.PLOS Medicine, 12(3):e1001779, 2015

Cathie Sudlow, John Gallacher, Naomi Allen, Valerie Beral, Paul Burton, John Danesh, Paul Downey, Paul Elliott, Jane Green, Martin Landray, Bette Liu, Paul Matthews, Giok Ong, Jill Pell, Alan Silman, Alan Young, Tim Sprosen, Tim Peakman, and Rory Collins. Uk biobank: An open access resource for identifying the causes of a wide range of complex diseases of...

work page doi:10.1371/journal.pmed.1001779 2015
[59]

Axiomatic attribution for deep networks

Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep networks. In Doina Precup and Yee Whye Teh, editors,Proceedings of the 34th International Conference on Machine Learning, volume 70 ofProceedings of Machine Learning Research, page 3319–3328. PMLR, August 2017. URL https://proceedings.mlr.press/v70/sundararajan17a. html

2017
[60]

Garomsa, Anna Zapaishchykova, Tafadzwa L

Divyanshu Tak, Biniam A. Garomsa, Anna Zapaishchykova, Tafadzwa L. Chaunzwa, Juan Car- los Climent Pardo, Zezhong Ye, John Zielke, Yashwanth Ravipati, Suraj Pai, Sri Vajapeyam, Maryam Mahootiha, Mitchell Parker, Luke R. G. Pike, Ceilidh Smith, Ariana M. Familiar, Kevin X. Liu, Sanjay Prabhu, Omar Arnaout, Pratiti Bandopadhayay, Ali Nabavizadeh, Sabine Mue...

work page doi:10.1038/s41593-026-02202-6 2026
[61]

Neural discrete representation learning

Aaron van den Oord, Oriol Vinyals, and koray kavukcuoglu. Neural discrete representation learning. In I. Guyon, U. V on Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors,Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc., 2017. URL https://proceedings.neurips.cc/paper_files/paper/ 2017/f...

2017
[62]

Gomez, Lukasz Kaiser, and Illia Polosukhin

Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Isabelle Guyon, Ulrike von Luxburg, Samy Bengio, Hanna M. Wallach, Rob Fergus, S. V . N. Vishwanathan, and Roman Garnett, editors,Advances in Neural Information Processing Systems 30: Annual Conference...

2017
[63]

Clamr: Contex- tualized late-interaction for multimodal content retrieval

David Wan, Han Wang, Elias Stengel-Eskin, Jaemin Cho, and Mohit Bansal. Clamr: Contex- tualized late-interaction for multimodal content retrieval. arXiv:2506.06144 [cs], June 2025. URLhttp://arxiv.org/abs/2506.06144

work page arXiv 2025
[64]

Selvaraju, Michael Cogswell, Ab- hishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra

Zihao Wang, Xihui Liu, Hongsheng Li, Lu Sheng, Junjie Yan, Xiaogang Wang, and Jing Shao. Camp: Cross-modal adaptive message passing for text-image retrieval. In2019 IEEE/CVF International Conference on Computer Vision, ICCV 2019, Seoul, Korea (South), October 27 - November 2, 2019, page 5763–5772. IEEE, 2019. doi: 10.1109/ICCV .2019.00586. URL https://doi...

work page doi:10.1109/iccv 2019
[65]

Information Theoretical Analysis of Multivariate Correlation

Satosi Watanabe. Information theoretical analysis of multivariate correlation.IBM Journal of Research and Development, 4(1):66–82, 1960. doi: 10.1147/rd.41.0066

work page doi:10.1147/rd.41.0066 1960
[66]

Attention is not not explanation

Sarah Wiegreffe and Yuval Pinter. Attention is not not explanation. In Kentaro Inui, Jing Jiang, Vincent Ng, and Xiaojun Wan, editors,Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), page 11–20, Hong Kong, China, November 2019. A...

work page doi:10.18653/v1/d19-1002 2019
[67]

Sigmoid loss for language image pre-training

Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), page 11975–11986, October 2023

2023
[68]

Qwen3 Embedding: Advancing Text Embedding and Reranking Through Foundation Models

Yanzhao Zhang, Mingxin Li, Dingkun Long, Xin Zhang, Huan Lin, Baosong Yang, Pengjun Xie, An Yang, Dayiheng Liu, Junyang Lin, Fei Huang, and Jingren Zhou. Qwen3 embedding: Advancing text embedding and reranking through foundation models. arXiv:2506.05176 [cs], June 2025. URLhttp://arxiv.org/abs/2506.05176

work page internal anchor Pith review Pith/arXiv arXiv 2025
[69]

Learning deep features for discriminative localization.2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), page 2921–2929, 2015

Bolei Zhou, Aditya Khosla, Àgata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization.2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), page 2921–2929, 2015

2016
[71]

uniform" 10optimizer.lr: 11min: 0.00001 12max: 0.01 13distribution:

URLhttps://arxiv.org/abs/2512.12678. 15 A Relation to the Cauchy-Schwarz Bound To quantify the sensitivity of the MIP critic to corruption in a single modality, we compare its score on a clean tuple and on a corrupted tuple and study the score deviation ∆g:=g corr −g clean. This difference isolates the effect of the corruption and admits a simple closed f...

work page arXiv 2048