Neuron-Guided Interpretation of Code LLMs: Where, Why, and How?

• URL: http://arxiv.org/abs/2512.19980v1
• Date: Tue, 23 Dec 2025 02:04:13 GMT
• Title: Neuron-Guided Interpretation of Code LLMs: Where, Why, and How?
• Abstract summary: We investigate code LLMs at the neuron level, localizing language-specific neurons (selectively responsive to one language) and concept layers.<n>We analyze Llama-3.1-8B and Qwen2.5-Coder-32B on multilingual inputs in C++, Java, Python, Go, and JavaScript.<n>We find (1) neurons specialized for individual languages alongside a universal subset supporting general-purpose generation; and (2) lower layers mainly encode language-specific syntax, while middle layers capture semantic abstractions shared across languages.
• Score: 9.839916751467134
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Code language models excel on code intelligence tasks, yet their internal interpretability is underexplored. Existing neuron interpretability techniques from NLP are suboptimal for source code due to programming languages formal, hierarchical, and executable nature. We empirically investigate code LLMs at the neuron level, localizing language-specific neurons (selectively responsive to one language) and concept layers (feed-forward layers encoding language-agnostic code representations). We analyze Llama-3.1-8B and Qwen2.5-Coder-32B on multilingual inputs in C++, Java, Python, Go, and JavaScript, measuring neuron selectivity and layerwise contributions during generation. We find (1) neurons specialized for individual languages alongside a universal subset supporting general-purpose generation; and (2) lower layers mainly encode language-specific syntax, while middle layers capture semantic abstractions shared across languages, emerging as concept layers. We demonstrate utility on three tasks: neuron-guided fine-tuning for code generation, clone detection via concept-layer embeddings, and concept-layer-guided transfer for code summarization, each yielding consistent gains in multilingual settings.

Related papers

• Neural FOXP2 -- Language Specific Neuron Steering for Targeted Language Improvement in LLMs [18.653643003491695]
  We argue that language defaultness is governed by a sparse, low-rank control circuit, language neurons.<n>We introduce Neural FOXP2, that makes a chosen language (Hindi or Spanish) primary in a model by steering language-specific neurons.
  arXiv  Detail & Related papers  (2026-02-01T00:23:25Z)
• Language Arithmetics: Towards Systematic Language Neuron Identification and Manipulation [9.518772041855923]
  We analyze language-specific neurons in Llama-3.1-8B, Mistral-Nemo-12B, and Aya-Expanse-8B & 32B across 21 typologically diverse languages.<n>We show that these neurons cluster in deeper layers, with non-Latin scripts showing greater specialization.<n>We steer models to deactivate unwanted languages and activate desired ones, outperforming simpler replacement approaches.
  arXiv  Detail & Related papers  (2025-07-30T12:23:39Z)
• How Programming Concepts and Neurons Are Shared in Code Language Models [55.22005737371843]
  We perform a few-shot translation task on 21 PL pairs using two Llama-based models.<n>We observe that the concept space is closer to English (including PL keywords) and assigns high probabilities to English tokens in the second half of the intermediate layers.<n>We analyze neuron activations for 11 PLs and English, finding that while language-specific neurons are primarily concentrated in the bottom layers, those exclusive to each PL tend to appear in the top layers.
  arXiv  Detail & Related papers  (2025-06-01T16:24:13Z)
• How does Alignment Enhance LLMs' Multilingual Capabilities? A Language Neurons Perspective [64.79894853375478]
  We propose a new finer-grained neuron identification algorithm, which detects language neurons(including language-specific neurons and language-related neurons) and language-agnostic neurons.<n>Based on the distributional characteristics of different types of neurons, we divide the LLMs' internal process for multilingual inference into four parts.<n>We systematically analyze the models before and after alignment with a focus on different types of neurons.
  arXiv  Detail & Related papers  (2025-05-27T17:59:52Z)
• Training Neural Networks as Recognizers of Formal Languages [87.06906286950438]
  We train and evaluate neural networks directly as binary classifiers of strings.<n>We provide results on a variety of languages across the Chomsky hierarchy for three neural architectures.<n>Our contributions will facilitate theoretically sound empirical testing of language recognition claims in future work.
  arXiv  Detail & Related papers  (2024-11-11T16:33:25Z)
• Sharing Matters: Analysing Neurons Across Languages and Tasks in LLMs [85.0284555835015]
  Large language models (LLMs) have revolutionized the field of natural language processing (NLP)<n>Few studies have attempted to explore the internal workings of LLMs in multilingual settings.<n>We classify neurons into four distinct categories based on their responses to a specific input across different languages.
  arXiv  Detail & Related papers  (2024-06-13T16:04:11Z)
• On the Multilingual Ability of Decoder-based Pre-trained Language Models: Finding and Controlling Language-Specific Neurons [37.32174349956148]
  We analyze the neuron-level internal behavior of multilingual decoder-based language models (PLMs) We show that language-specific neurons are unique, with a slight overlap ( 5%) between languages. We tamper with less than 1% of the total neurons in each model during inference and demonstrate that tampering with a few language-specific neurons drastically changes the probability of target language occurrence in text generation.
  arXiv  Detail & Related papers  (2024-04-03T03:37:22Z)
• IRCoder: Intermediate Representations Make Language Models Robust Multilingual Code Generators [49.903001442804594]
  This work investigates the prospect of leveraging compiler intermediate representations (IR) to improve the multilingual capabilities of Code-LMs. We first compile SLTrans, a parallel dataset consisting of nearly 4M self-contained source code files. Next, we carry out continued causal language modelling training on SLTrans, forcing the Code-LMs to learn the IR language. Our resulting models, dubbed IRCoder, display sizeable and consistent gains across a wide variety of code generation tasks and metrics.
  arXiv  Detail & Related papers  (2024-03-06T17:52:08Z)
• How do Large Language Models Handle Multilingualism? [81.15060972112563]
  This study explores how large language models (LLMs) handle multilingualism. LLMs initially understand the query, converting multilingual inputs into English for task-solving. In the intermediate layers, they employ English for thinking and incorporate multilingual knowledge with self-attention and feed-forward structures.
  arXiv  Detail & Related papers  (2024-02-29T02:55:26Z)
• Discovering Salient Neurons in Deep NLP Models [31.18937787704794]
  We present a technique called as Linguistic Correlation Analysis to extract salient neurons in the model. Our data-driven, quantitative analysis illuminates interesting findings. Our code is publicly available as part of the NeuroX toolkit.
  arXiv  Detail & Related papers  (2022-06-27T13:31:49Z)
• Multi-level Contrastive Learning for Cross-lingual Spoken Language Understanding [90.87454350016121]
  We develop novel code-switching schemes to generate hard negative examples for contrastive learning at all levels. We develop a label-aware joint model to leverage label semantics for cross-lingual knowledge transfer.
  arXiv  Detail & Related papers  (2022-05-07T13:44:28Z)
• Low-Dimensional Structure in the Space of Language Representations is Reflected in Brain Responses [62.197912623223964]
  We show a low-dimensional structure where language models and translation models smoothly interpolate between word embeddings, syntactic and semantic tasks, and future word embeddings. We find that this representation embedding can predict how well each individual feature space maps to human brain responses to natural language stimuli recorded using fMRI. This suggests that the embedding captures some part of the brain's natural language representation structure.
  arXiv  Detail & Related papers  (2021-06-09T22:59:12Z)

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.