On the Effect of Token Merging on Pre-trained Models for Code

• URL: http://arxiv.org/abs/2507.14423v1
• Date: Sat, 19 Jul 2025 00:48:20 GMT
• Title: On the Effect of Token Merging on Pre-trained Models for Code
• Authors: Mootez Saad, Hao Li, Tushar Sharma, Ahmed E. Hassan,
• Abstract summary: We investigate the effect of merging the hidden representations of subtokens that belong to the same semantic unit.<n>We propose two strategies: one based on averaging the representations and another that leverages a learning-based approach.<n>Results show that these strategies can reduce the number of floating-point operations by $1%$ to $19%$.
• Score: 11.029842116504726
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Tokenization is a fundamental component of language models for code. It involves breaking down the input into units that are later passed to the language model stack to learn high-dimensional representations used in various contexts, from classification to generation. However, the output of these tokenizers is often longer than that traditionally used in compilers and interpreters. This could result in undesirable effects, such as increased computational overhead. In this work, we investigate the effect of merging the hidden representations of subtokens that belong to the same semantic unit, such as subtokens that form a single identifier. We propose two strategies: one based on averaging the representations and another that leverages a learning-based approach. Both methods can be seamlessly integrated with existing language models for code. We conduct experiments using six language models for code: CodeBERT, GraphCodeBERT, UniXCoder, CdoeT5, CodeT5+ (220M), and CodeT5+ (770M), across three software engineering tasks: vulnerability detection, code classification, and code translation. Results show that these strategies can reduce the number of floating-point operations by $1\%$ to $19\%$. Regarding downstream performance, the most significant degradation was observed in the vulnerability detection task, where the F1 score decreased by $1.82$ points compared to the baseline. In contrast, for code translation, we observed an improvement of $2.47$ points in CodeBLEU. This work contributes to the broader effort of improving language models for code across multiple dimensions, including both computational efficiency and downstream performance.

Related papers

• Type-Constrained Code Generation with Language Models [51.03439021895432]
  We introduce a type-constrained decoding approach that leverages type systems to guide code generation.<n>For this purpose, we develop novel prefix automata and a search over inhabitable types, forming a sound approach to enforce well-typedness on LLM-generated code.<n>Our approach reduces compilation errors by more than half and significantly increases functional correctness in code synthesis, translation, and repair tasks.
  arXiv  Detail & Related papers  (2025-04-12T15:03:00Z)
• MoSE: Hierarchical Self-Distillation Enhances Early Layer Embeddings [2.1262605464247812]
  Self-Distillation is a principled approach to trading inference cost for accuracy across various code understanding tasks.<n>Our architecture improves text-to-code and code-to-code search by targeting specific encoder layers as exit heads.<n>We release a new dataset created through code translation that extends text-to-code benchmarks with cross-language code-to-code pairs.
  arXiv  Detail & Related papers  (2025-03-04T21:08:17Z)
• Code Representation Learning At Scale [75.04686476303436]
  We fuel code representation learning with a vast amount of code data via a two-stage pretraining scheme. We first train the encoders via a mix that leverages both randomness in masking language modeling and the structure aspect of programming language. We then enhance the representations via contrastive learning with hard negative and hard positive constructed in an unsupervised manner.
  arXiv  Detail & Related papers  (2024-02-02T22:19:15Z)
• CodeT5+: Open Code Large Language Models for Code Understanding and Generation [72.1638273937025]
  Large language models (LLMs) pretrained on vast source code have achieved prominent progress in code intelligence. CodeT5+ is a family of encoder-decoder LLMs for code in which component modules can be flexibly combined to suit a wide range of downstream code tasks. We extensively evaluate CodeT5+ on over 20 code-related benchmarks in different settings, including zero-shot, finetuning, and instruction-tuning.
  arXiv  Detail & Related papers  (2023-05-13T14:23:07Z)
• Enriching Source Code with Contextual Data for Code Completion Models: An Empirical Study [4.438873396405334]
  We aim to answer whether making code easier to understand through using contextual data improves the performance of pre-trained code language models for the task of code completion. For comments, we find that the models perform better in the presence of multi-line comments.
  arXiv  Detail & Related papers  (2023-04-24T17:09:14Z)
• Interactive Code Generation via Test-Driven User-Intent Formalization [60.90035204567797]
  Large language models (LLMs) produce code from informal natural language (NL) intent. It is hard to define a notion of correctness since natural language can be ambiguous and lacks a formal semantics. We describe a language-agnostic abstract algorithm and a concrete implementation TiCoder.
  arXiv  Detail & Related papers  (2022-08-11T17:41:08Z)
• Enhancing Semantic Code Search with Multimodal Contrastive Learning and Soft Data Augmentation [50.14232079160476]
  We propose a new approach with multimodal contrastive learning and soft data augmentation for code search. We conduct extensive experiments to evaluate the effectiveness of our approach on a large-scale dataset with six programming languages.
  arXiv  Detail & Related papers  (2022-04-07T08:49:27Z)
• An Exploratory Study on Code Attention in BERT [8.488193857572211]
  We investigate the attention behavior of PLM on code and compare it with natural language. We show that BERT pays more attention to syntactic entities, specifically identifiers and separators, in contrast to the most attended token in NLP. The findings can benefit the research community by using code-specific representations instead of applying the common embeddings used in NLP.
  arXiv  Detail & Related papers  (2022-04-05T21:23:10Z)
• CLSEBERT: Contrastive Learning for Syntax Enhanced Code Pre-Trained Model [23.947178895479464]
  We propose CLSEBERT, a Constrastive Learning Framework for Syntax Enhanced Code Pre-Trained Model. In the pre-training stage, we consider the code syntax and hierarchy contained in the Abstract Syntax Tree (AST) We also introduce two novel pre-training objectives. One is to predict the edges between nodes in the abstract syntax tree, and the other is to predict the types of code tokens.
  arXiv  Detail & Related papers  (2021-08-10T10:08:21Z)
• GraphCodeBERT: Pre-training Code Representations with Data Flow [97.00641522327699]
  We present GraphCodeBERT, a pre-trained model for programming language that considers the inherent structure of code. We use data flow in the pre-training stage, which is a semantic-level structure of code that encodes the relation of "where-the-value-comes-from" between variables. We evaluate our model on four tasks, including code search, clone detection, code translation, and code refinement.
  arXiv  Detail & Related papers  (2020-09-17T15:25:56Z)

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.