Representational Atoms in Neural Networks

Chenhui Hu, Pengfei Cao, Kang Liu, Jun ZhaoarXiv 2026 · arXiv

This paper introduces a rigorous theoretical and practical framework for understanding the "periodic table" of neural network internal states. By defining Representational Atoms, the authors provide a principled way to identify the fundamental building blocks of AI models and show how these blocks compose to form human-like concepts.

TL;DR

The paper proposes "Representational Atoms" as the fundamental, distinguishable units of neural network representations, characterized by high fidelity and structural stability. Using a novel class of threshold-activated sparse autoencoders (TSAEs), the authors demonstrate that these atoms are more monosemantic than individual neurons and function as fine-grained components that compose across layers to form complex human concepts.

What Problem Does This Paper Address?

Despite the success of deep learning, the internal representations of neural networks remain "black boxes." We know they encode information, but we don't fully understand the basic units of that encoding.

Individual neurons are often polysemantic, meaning a single neuron might fire for completely unrelated concepts (e.g., "the color red" and "legal documents"). While Sparse Autoencoders (SAEs) have been used to find more interpretable "features," the field has lacked a formal definition of what makes a feature a "fundamental unit." Furthermore, existing units often face a trade-off: neurons have high fidelity (they represent the data well) but low stability (they are hard to distinguish), while many SAE features improve stability at the cost of fidelity.

Existing representational units like neurons and standard SAE features show a systematic gap in fidelity and stability compared to the ideal "atom" reference.

Methods

The authors develop a framework to define, evaluate, and identify Fundamental Representational Units (FRUs), which they term atoms.

  • Formal Definition: Atoms are defined by three properties: Representability (faithfully reconstructing original states), Sparsity (only a few atoms are active at once), and Separability (atoms are approximately orthogonal).
  • Atomic Inner Product (AIP): The authors introduce a new metric to characterize the relative geometry of atoms. Unlike standard Euclidean distance, AIP accounts for the specific geometry induced by the network's internal representations.
  • Evaluation Metrics: They propose two quantitative criteria: Fidelity ($R^2$), measuring reconstruction quality, and Stability ($q^*$), measuring the fraction of atoms that satisfy a theoretical distinguishability condition derived from compressed sensing.
  • BiJumpReLU SAE: To identify these atoms in practice, they introduce a novel TSAE architecture. It uses learnable bidirectional thresholds to detect atoms without being biased by the sign of the activation, aligning the architecture with their theoretical framework.
Overview of the framework for defining (a), evaluating (b), and identifying (c) representational atoms using threshold-activated SAEs.

Datasets

The researchers validated their framework across diverse models and modalities:

  • Language Models: Gemma-2 (2B/9B), Llama-3.1-8B, Qwen-3.5-9B, and Olmo-Hybrid-7B.
  • Diffusion Models: Stable Diffusion v1.5 and FLUX.2-klein-9B.
  • Data Sources: Wikipedia (text), Wikidata (entities), and LAION-COCO (images).
  • Knowledge Base: HowNet, an expert-annotated database of "sememes" (semantic primitives), used to evaluate the relationship between neural atoms and human conceptual structures.

Key Insights and Ideas

  • Stability is the Key to Meaning: The authors prove that structural distinguishability (stability) is a necessary condition for monosemanticity. A unit can only support a unique semantic interpretation if it remains distinguishable from other units or their combinations.
  • Atoms are "Sub-Conceptual": Interestingly, human concepts (sememes) are rarely represented by a single atom. Instead, they are typically represented in a distributed manner by multiple atoms, often spread across different layers of the network.
  • Compositional Organization: While single atoms capture fine-grained semantic components, their compositions recover higher-order human semantic structures with much higher alignment than any single unit. This suggests that neural networks organize knowledge as a hierarchy of atomic building blocks.

Key Results

  • State-of-the-Art Identification: The BiJumpReLU SAE consistently outperformed established architectures (like ReLU, TopK, and JumpReLU) in both fidelity and stability across all tested models. For instance, it maintained stability scores above 0.98 across all settings while improving reconstruction fidelity.
  • Empirical Validation: There is a strong positive correlation (Spearman $\rho = 0.71$) between the theoretical stability metric and empirical monosemanticity. More stable atoms exhibit clearer and more consistent semantics.
  • Recovering Human Concepts: By composing atoms across layers, the researchers achieved significantly higher semantic alignment with HowNet sememes (F1 scores > 0.80) compared to the best single-atom matches (~0.30). Most sememes required more than six atoms for accurate recovery.

Limitations and Future Work

  • Linearity: The framework currently models representations as sparse linear combinations. While effective, it may not capture potential non-linear interactions between atoms.
  • Computational Overhead: Estimating the AIP requires second-order statistics of representations, which may be computationally demanding for extremely large-scale models.
  • Intervention Research: The authors suggest that future work should explore how this compositional organization can be used for more precise model interventions, such as knowledge editing or representation steering, which currently often suffer from "collateral effects" because they target single directions rather than atomic compositions.

Summary updated

Abstract

Understanding the internal representations of neural networks and characterizing their relationship with human conceptual structure are two critical questions for elucidating model mechanisms. The former reveals how learned knowledge is organized in internal representations, whereas the latter provides a basis for interpreting these representations through human conceptual structure. Here, we formally define fundamental representational units (FRUs) and establish two evaluation criteria: fidelity and stability. We further prove that threshold-activated sparse autoencoders (TSAEs) can identify FRUs with high fidelity and stability under appropriate conditions, and validate this result through large-scale experiments. Further analyses reveal that FRUs with higher stability exhibit greater monosemanticity and can be composed across layers to recover semantic organization corresponding to human conceptual structure. These results provide a unified framework for defining, evaluating and identifying FRUs in neural networks, and establish a foundation for characterizing the relationship between neural network representations and human conceptual structure.
BibTeX
@misc{hu2026representational,
  author = {Chenhui Hu and Pengfei Cao and Kang Liu and Jun Zhao},
  title = {{Representational Atoms in Neural Networks}},
  year = {2026},
  eprint = {2611.14168},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.14168},
}
Figure