Are Arithmetic Heuristic Neurons Form-Invariant? A Mechanistic Analysis of Symbols, Text, and Code in LLMs

• URL: http://arxiv.org/abs/2607.16693v1
• Date: Sat, 18 Jul 2026 08:02:50 GMT
• Title: Are Arithmetic Heuristic Neurons Form-Invariant? A Mechanistic Analysis of Symbols, Text, and Code in LLMs
• Abstract summary: We study whether arithmetic neurons are form-invariant across symbolic subtraction, natural language word problems, and Python code.<n>A compact set of neurons is shared across all three formats, and targeted interventions show this shared circuit is both necessary and sufficient for late-layer arithmetic.
• Score: 1.1087735229999816
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Large language models often succeed on one formulation of a problem while failing on an equivalent formulation. Whether these failures arise from distinct internal circuits or different activation states of a shared circuit remains unknown. Recent mechanistic interpretability studies suggest that arithmetic in LLMs emerges from a "bag of heuristics," encoded by a sparse set of MLP neurons that represent distinct arithmetic strategies. We investigate whether arithmetic heuristic neurons are form-invariant across symbolic arithmetic, natural language word problems, and Python code in three Llama-3 models. In each format, we identify arithmetic heuristic neurons using a two-stage pipeline combining attribution patching and activation patching. A compact set of neurons is shared across all three formats, and targeted interventions show this shared circuit is both necessary and sufficient for late-layer arithmetic computation. Transferring the shared neurons' activations from a successful execution in one format to a failed execution in another recovers most incorrect predictions, exceeding 97% for addition and subtraction, indicating that cross-format failures arise from activation states rather than distinct circuits. Moreover, shared neurons consistently belong to the same heuristic families across formats, demonstrating that arithmetic computation in LLMs is largely form-invariant at the neuron level.

Related papers

• Language Model Circuits Are Sparse in the Neuron Basis [50.460651620833055]
  We show that textbfMLP neurons are as sparse a feature basis as SAEs.<n>This work advances automated interpretability of language models without additional training costs.
  arXiv  Detail & Related papers  (2026-01-30T05:41:19Z)
• Diagonal Symmetrization of Neural Network Solvers for the Many-Electron Schrödinger Equation [10.862915635270346]
  We study different ways of incorporating diagonal invariance in neural network ans"atze trained via variational Monte Carlo methods.<n>We show that, contrary to standard ML setups, in-training symmetrization destabilizes training and can lead to worse performance.<n>Our theoretical and numerical results indicate that this unexpected behavior may arise from a unique computational-statistical tradeoff not found in standard ML analyses of symmetrization.
  arXiv  Detail & Related papers  (2025-02-07T20:37:25Z)
• An analysis of optimization problems involving ReLU neural networks [38.258426534664046]
  We study approaches to analyze and improve the run time behavior of mixed-integer programming solvers.<n>We numerically compare these approaches for three benchmark problems from the literature.<n>As a major takeaway we observe and quantify a trade-off between the often desired redundancy of neural network models versus the computational costs for solving related optimization problems.
  arXiv  Detail & Related papers  (2025-02-05T09:18:07Z)
• Arithmetic Without Algorithms: Language Models Solve Math With a Bag of Heuristics [43.86518549293703]
  We show that large language models (LLMs) perform arithmetic using neither robust algorithms nor memorization.<n>Overall, our experimental results show that LLMs perform arithmetic using neither robust algorithms nor memorization.
  arXiv  Detail & Related papers  (2024-10-28T17:59:06Z)
• In-Context Language Learning: Architectures and Algorithms [73.93205821154605]
  We study ICL through the lens of a new family of model problems we term in context language learning (ICLL) We evaluate a diverse set of neural sequence models on regular ICLL tasks.
  arXiv  Detail & Related papers  (2024-01-23T18:59:21Z)
• LogicMP: A Neuro-symbolic Approach for Encoding First-order Logic Constraints [46.60806942245395]
  This paper proposes a novel neural layer, LogicMP, whose layers perform mean-field variational inference over an MLN.<n>It can be plugged into any off-the-shelf neural network to encode FOLCs while retaining modularity and efficiency.<n> Empirical results in three kinds of tasks over graphs, images, and text show that LogicMP outperforms advanced competitors in both performance and efficiency.
  arXiv  Detail & Related papers  (2023-09-27T07:52:30Z)
• Bayesian Neural Network Language Modeling for Speech Recognition [59.681758762712754]
  State-of-the-art neural network language models (NNLMs) represented by long short term memory recurrent neural networks (LSTM-RNNs) and Transformers are becoming highly complex. In this paper, an overarching full Bayesian learning framework is proposed to account for the underlying uncertainty in LSTM-RNN and Transformer LMs.
  arXiv  Detail & Related papers  (2022-08-28T17:50:19Z)
• Mixed Precision Low-bit Quantization of Neural Network Language Models for Speech Recognition [67.95996816744251]
  State-of-the-art language models (LMs) represented by long-short term memory recurrent neural networks (LSTM-RNNs) and Transformers are becoming increasingly complex and expensive for practical applications. Current quantization methods are based on uniform precision and fail to account for the varying performance sensitivity at different parts of LMs to quantization errors. Novel mixed precision neural network LM quantization methods are proposed in this paper.
  arXiv  Detail & Related papers  (2021-11-29T12:24:02Z)
• Provably Efficient Neural Estimation of Structural Equation Model: An Adversarial Approach [144.21892195917758]
  We study estimation in a class of generalized Structural equation models (SEMs) We formulate the linear operator equation as a min-max game, where both players are parameterized by neural networks (NNs), and learn the parameters of these neural networks using a gradient descent. For the first time we provide a tractable estimation procedure for SEMs based on NNs with provable convergence and without the need for sample splitting.
  arXiv  Detail & Related papers  (2020-07-02T17:55:47Z)
• Separation of Memory and Processing in Dual Recurrent Neural Networks [0.0]
  We explore a neural network architecture that stacks a recurrent layer and a feedforward layer that is also connected to the input. When noise is introduced into the activation function of the recurrent units, these neurons are forced into a binary activation regime that makes the networks behave much as finite automata.
  arXiv  Detail & Related papers  (2020-05-17T11:38:42Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.