LLMs as a synthesis between symbolic and distributed approaches to language

Gemma BoledaEMNLP 2025

TL;DR

This paper argues that Large Language Models (LLMs) represent a synthesis of two long-standing, opposing approaches to language: symbolic (rule-based, discrete) and distributed (connectionist, continuous). By analyzing recent interpretability research, the authors show that LLMs can develop and utilize both types of representations, explaining their remarkable success.

What Problem Does This Paper Address?

For decades, the fields of linguistics and cognitive science have been divided by a debate between "symbolic" and "distributed" approaches to language and cognition. Symbolic approaches emphasize discrete, interpretable rules and categories (like grammar rules or parts of speech), while distributed approaches focus on uninterpretable, high-dimensional continuous representations (like neural network embeddings). This paper tackles the question of whether the recent success of LLMs, which are based on distributed approaches, signifies a victory for one camp or if they offer a more integrated perspective. The authors aim to bridge this divide by proposing that LLMs are not purely distributed but rather a synthesis that leverages the strengths of both traditions.

Methods

The paper is primarily a position paper that synthesizes existing research, particularly from the field of LLM interpretability. The authors do not introduce new algorithms or datasets but rather analyze how current LLMs, through their architecture and training, exhibit characteristics of both symbolic and distributed processing. They highlight:

  • Architectural Flexibility: Transformer architectures, the backbone of most LLMs, inherently support both continuous representations (vectors) and operations that can lead to near-discrete, interpretable processing.
  • Interpretability Studies: The core of the argument relies on findings from recent interpretability research. These studies use techniques to probe LLMs and understand how they represent and process linguistic information. This includes identifying specific neurons or "attention heads" that selectively respond to linguistic features (like parts of speech, tense, or syntactic relations) or even entire "circuits" that perform specific linguistic computations.
  • Focus on Emergent Behavior: The paper emphasizes that the near-symbolic behavior observed in LLMs is not explicitly programmed but emerges from training on vast amounts of language data.

Datasets

Not explicitly discussed as training or evaluation datasets for novel experiments. The paper refers to existing benchmarks and corpora that have been used in the interpretability studies it surveys, such as:

  • GLUE benchmark: Specifically mentioned for its Semantic Textual Similarity (STS) task.
  • Penn Treebank: Referenced in the context of challenges in manual POS tagging.

Key Insights and Ideas

  • LLMs as a Synthesis: The central thesis is that LLMs are not purely distributed but a powerful synthesis of symbolic and distributed approaches. This integration allows them to handle both the regularities and the "messiness" of language.
  • Emergent Near-Symbolic Representations: Interpretability research reveals that LLMs develop representations and processing mechanisms that are "near-symbolic" – meaning they exhibit discrete-like behavior for specific linguistic phenomena, even though the underlying representations are continuous.
  • Flexibility is Key to Success: The ability of LLMs to flexibly switch between distributed and near-symbolic modes, and everything in between, is proposed as a primary reason for their remarkable success in language tasks.
  • Language is Inherently Mixed: The paper argues that language itself is not purely symbolic or purely distributed but exists on a spectrum. LLMs, by mirroring this, are better equipped to model it.
  • Beyond Discrete/Fuzzy Dichotomy: The debate between symbolic and distributed approaches is framed as potentially sterile. The paper advocates for models that natively encompass both properties, moving beyond a strict dichotomy.

Key Results

The paper presents qualitative findings from interpretability studies, rather than quantitative performance metrics on specific tasks. Key findings include:

  • Neurons with Specific Linguistic Functions: Studies have identified individual neurons in LLMs that selectively respond to specific linguistic features like part of speech, tense, or gender.
  • Attention Heads for Syntactic Relations: Attention heads have been shown to specialize in tracking syntactic dependencies, such as possessive constructions or co-reference.
  • Interpretable Circuits: Researchers have identified specific "circuits" (subgraphs of the network) that perform interpretable linguistic computations, such as identifying indirect objects or handling subject-verb agreement.
  • Robustness and Generalization: Some of these identified circuits show robustness and generalization across different models and even languages, suggesting a high level of abstraction.
  • Polysemanticity and Redundancy: While near-symbolic units are found, the paper also notes that LLMs exhibit polysemanticity (units responding to multiple unrelated properties) and redundancy, which are characteristic of distributed systems and contribute to their robustness.

Limitations and Future Work

The authors acknowledge several limitations and point towards future research directions:

  • Fuzzy Definition of "Near-Symbolic": The paper admits that its definition of "near-symbolic" is itself somewhat fuzzy, reflecting the current infancy of mechanistic interpretability research. A more precise definition is needed.
  • Limited Scope of Current Interpretability: Most interpretability studies focus on morphosyntax and syntax. More research is needed on compositional semantics and the syntax-semantics interface.
  • Scaling Symbolic Behavior: While evidence for localized symbolic-like components exists, it's unclear how these mechanisms scale to capture generalizable rules or logical inferences in a way that fully matches traditional symbolic systems. More research is needed on the "addition circuitry" and its generalization capabilities.
  • Interplay with Reasoning: The interaction between linguistic processing and reasoning abilities in LLMs is an exciting area for further study.
  • Brain Analogies: Recent work suggests analogies between LLM and brain encoding of language, but more research is needed to understand the properties of linguistic representations and computations in both.
  • Developing New Methods: The paper calls for the development of new methods that go beyond analyzing specific phenomena and allow for systematic exploration of LLMs across the symbolic-to-distributed spectrum.

Summary updated

BibTeX
@inproceedings{boleda2025llms,
  author = {Gemma Boleda},
  title = {{LLMs as a synthesis between symbolic and distributed approaches to language}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2025},
}
Figure