mlsanityData to 13 Oct 2026

Revisiting Disentanglement in Downstream Tasks: A Study on Its Necessity for Abstract Visual Reasoning

arXiv 2403.00352 ↗First posted 1 Mar 2024cs.CV cs.LGcited by 1

On the map: LLM reasoning → Multimodal reasoning

Abstract (arXiv)

In representation learning, a disentangled representation is highly desirable as it encodes generative factors of data in a separable and compact pattern. Researchers have advocated leveraging disentangled representations to complete downstream tasks with encouraging empirical evidence. This paper further investigates the necessity of disentangled representation in downstream applications. Specifically, we show that dimension-wise disentangled representations are unnecessary on a fundamental downstream task, abstract visual reasoning. We provide extensive empirical evidence against the necessity of disentanglement, covering multiple datasets, representation learning methods, and downstream network architectures. Furthermore, our findings suggest that the informativeness of representations is a better indicator of downstream performance than disentanglement. Finally, the positive correlation between informativeness and disentanglement explains the claimed usefulness of disentangled representations in previous works. The source code is available at https://github.com/Richard-coder-Nai/disentanglement-lib-necessity.git.

Key facts. Cited by 1 paper in the map (data through 13 Oct 2026). Placed in Multimodal reasoning, one of 17 directions in LLM reasoning. Builds on 26 of its 45 references that are in the map. First posted on arXiv 1 Mar 2024.

Builds on

26 of 45 references.

Closest ancestors · 1

Same direction, newest first

  1. 2022Deep Learning Methods for Abstract Visual Reasoning: A Survey on Raven's Progressive Matrices28 Jan

Other references · 25

  1. 2017Attention Is All You Needcited by 60,106
  2. 2023DINOv2: Learning Robust Visual Features without Supervisioncited by 7,857
  3. 2019Momentum Contrast for Unsupervised Visual Representation Learningcited by 7,492
  4. 2021Masked Autoencoders Are Scalable Vision Learnerscited by 7,364
  5. 2021Emerging Properties in Self-Supervised Vision Transformerscited by 6,915
  6. 2020Bootstrap your own latent: A new approach to self-supervised Learningcited by 4,730
  7. 2020Exploring Simple Siamese Representation Learningcited by 2,860
  8. 2018Challenging Common Assumptions in the Unsupervised Learning of Disentangled Representationscited by 934
  9. 2016InfoGAN: Interpretable Representation Learning by Information Maximizing Generative Adversarial Netscited by 860
  10. 2018Disentangling by Factorisingcited by 614
  11. 2018Towards a Definition of Disentangled Representationscited by 295
  12. 2017Variational Inference of Disentangled Latent Concepts from Unlabeled Observationscited by 241
13 more
  1. 2019On the Fairness of Disentangled Representationscited by 144
  2. 2018Recent Advances in Autoencoder-Based Representation Learningcited by 131
  3. 2018Learning Deep Disentangled Embeddings with the F-Statistic Losscited by 101
  4. 2019Are Disentangled Representations Helpful for Abstract Visual Reasoning?cited by 100
  5. 2021Desiderata for Representation Learning: A Causal Perspectivecited by 76
  6. 2019Theory and Evaluation Metrics for Learning Disentangled Representationscited by 57
  7. 2022Disentanglement of Correlated Factors via Hausdorff Factorized Supportcited by 37
  8. 2020Unsupervised deep learning identifies semantic disentanglement in single inferotemporal neuronscited by 33
  9. 2022Disentanglement with Biological Constraints: A Theory of Functional Cell Typescited by 30
  10. 2018Improving Generalization for Abstract Reasoning Tasks Using Disentangled Feature Representationscited by 28
  11. 2019Attention on Abstract Visual Reasoningcited by 15
  12. 2021The Role of Pretrained Representations for the OOD Generalization of Reinforcement Learning Agentscited by 12
  13. 2022An Empirical Study on Disentanglement of Negative-free Contrastive Learningcited by 12

Built on by

Cited by year

Citing papers per year: 2024: 1.

1 paper.

  1. 2024Enhancing Feature Diversity Boosts Channel-Adaptive Vision Transformerscited by 10

Similar, not cited

Close in wording, no citation either way

  1. 2024Disentangling Representations through Multi-task Learning
  2. 2026Disentangled Representation Learning via Flow Matching
  3. 2026Intuition vectors
  4. 2024Beyond the Doors of Perception: Vision Transformers Represent Relations Between Objects
  5. 2025Disentanglement Beyond Static vs. Dynamic: A Benchmark and Evaluation Framework for Multi-Factor Sequential Representations
  6. 2025Unraveling the geometry of visual relational reasoning
  7. 2024Take A Step Back: Rethinking the Two Stages in Visual Reasoning
  8. 2024Graph-based Unsupervised Disentangled Representation Learning via Multimodal Large Language Models