mlsanityData to 7 Oct 2026

Relational Composition in Neural Networks: A Survey and Call to Action

arXiv 2407.14662 ↗First posted 19 Jul 2024cs.AI cs.LGcited by 23

Not on the map. Why?

Abstract (arXiv)

Many neural nets appear to represent data as linear combinations of "feature vectors." Algorithms for discovering these vectors have seen impressive recent success. However, we argue that this success is incomplete without an understanding of relational composition: how (or whether) neural nets combine feature vectors to represent more complicated relationships. To facilitate research in this area, this paper offers a guided tour of various relational mechanisms that have been proposed, along with preliminary analysis of how such mechanisms might affect the search for interpretable features. We end with a series of promising areas for empirical research, which may help determine how neural networks represent structured data.

Key facts. Cited by 23 papers in the map (data through 7 Oct 2026). Builds on 26 of its 56 references that are in the map. First posted on arXiv 19 Jul 2024.

Builds on

26 of 56 references.

References · 26

  1. 2013Distributed Representations of Words and Phrases and their Compositionalitycited by 3,462
  2. 2023Sparse Autoencoders Find Highly Interpretable Features in Language Modelscited by 1,359
  3. 2023Inference-Time Intervention: Eliciting Truthful Answers from a Language Modelcited by 1,259
  4. 2023Representation Engineering: A Top-Down Approach to AI Transparencycited by 1,239
  5. 2022Toy Models of Superpositioncited by 928
  6. 2023The Geometry of Truth: Emergent Linear Structure in Large Language Model Representations of True/False Datasetscited by 673
  7. 2019Visualizing and Measuring the Geometry of BERTcited by 215
  8. 2024Not All Language Model Features Are One-Dimensionally Linearcited by 201
  9. 2023Linearity of Relation Decoding in Transformer Language Modelscited by 190
  10. 2024Improving Dictionary Learning with Gated Sparse Autoencoderscited by 171
  11. 2024Fine-Tuning Enhances Existing Mechanisms: A Case Study on Entity Trackingcited by 140
  12. 2020Finding Universal Grammatical Relations in Multilingual BERTcited by 108
14 more
  1. 2023How do Language Models Bind Entities in Context?cited by 102
  2. 2022Does CLIP Bind Concepts? Probing Compositionality in Large Image Modelscited by 101
  3. 2019Do Attention Heads in BERT Track Syntactic Dependencies?cited by 87
  4. 2019Measuring Compositionality in Representation Learningcited by 84
  5. 2023Is This the Subspace You Are Looking for? An Interpretability Illusion for Subspace Activation Patchingcited by 75
  6. 2020A comparison of Vector Symbolic Architecturescited by 63
  7. 2024Designing a Dashboard for Transparency and Control of Conversational AIcited by 43
  8. 2024Talking Heads: Understanding Inter-layer Communication in Transformer Language Modelscited by 39
  9. 2024Monitoring Latent World States in Language Models with Propositional Probescited by 39
  10. 2018RNNs Implicitly Implement Tensor Product Representationscited by 25
  11. 2015Representing Objects, Relations, and Sequencescited by 24
  12. 2022Unit Testing for Concepts in Neural Networkscited by 23
  13. 2021Attention Can Reflect Syntactic Structure (If You Let It)cited by 22
  14. 2022Compositionality as Lexical Symmetrycited by 8

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2024: 2; 2025: 11; 2026: 10.

23 papers.

Sort:

By area:

  1. 2025Learning Multi-Level Features with Matryoshka Sparse Autoencoderscited by 136
  2. 2025Archetypal SAE: Adaptive and Stable Dictionary Learning for Concept Extraction in Large Vision Modelscited by 67
  3. 2025Universal Sparse Autoencoders: Interpretable Cross-Model Concept Alignmentcited by 62
  4. 2025From Flat to Hierarchical: Extracting Sparse Representations with Matching Pursuitcited by 45
  5. 2026When Models Manipulate Manifolds: The Geometry of a Counting Taskcited by 43
  6. 2024Towards Unifying Interpretability and Control: Evaluation via Interventioncited by 28
  7. 2025One-shot Optimized Steering Vectors Mediate Safety-relevant Behaviors in LLMscited by 27
  8. 2026Do Sparse Autoencoders Capture Concept Manifolds?cited by 25
  9. 2024Residual Stream Analysis with Multi-Layer SAEscited by 18
  10. 2025Transformer Dynamics: A neuroscientific approach to interpretability of large language modelscited by 11
  11. 2026Cross-Modal Redundancy and the Geometry of Vision-Language Embeddingscited by 9
  12. 2025How Do Language Models Compose Functions?cited by 7
11 more
  1. 2025Addressing divergent representations from causal interventions on neural networkscited by 6
  2. 2026From Atoms to Trees: Building a Structured Feature Forest with Hierarchical Sparse Autoencoderscited by 6
  3. 2026Position: Anthropomorphic Misalignment Research Needs Stronger Evidencecited by 6
  4. 2025Attribution-Guided Distillation of Matryoshka Sparse Autoencoderscited by 4
  5. 2025But what is your honest answer? Aiding LLM-judges with honest alternatives using steering vectorscited by 3
  6. 2025Insights into a radiology-specialised multimodal large language model with sparse autoencoderscited by 2
  7. 2026Evaluating the Interpretability of Sparse Autoencoders with Concept Annotationscited by 2
  8. 2026Tensor Product Representation Probes Reveal Shared Structure Across Linear Directionscited by 0
  9. 2026Inverted Detection and Control in Steering Vectorscited by 0
  10. 2026Activation Flow: Manufacturing Activations for Steeringcited by 0
  11. 2026Verifying the Linear Representation Hypothesis: How Interpretable Are Vision SAEs?cited by 0