Farhan Nafis Rayhan at SemEval-2026 Task 13: Supervised Contrastive Learning Approach with Gated Multiclass Decomposition Ensemble Architecture for Code Authorship Identification

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Data Science & Analytics · Depth: Expert, medium

Summary

Farhan Nafis Rayhan and Fariska Ruskanda submitted a system for SemEval-2026 Task 13 Subtask B, focusing on multi-class attribution of code snippets to 10 distinct AI generator families and a human baseline. Their approach employs a three-stage ensemble architecture designed to handle extreme class imbalance and capture subtle stylometric fingerprints. Initially, Supervised Contrastive Learning fine-tunes UniXcoder and ModernBERT backbones to produce embeddings. These embeddings are then processed by five heterogeneous shallow experts, each using multiclass decomposition for specific generator lineages, alongside a "Human Shield" for human vs. machine classification. Finally, a Context-Aware Gated Meta-Learner aggregates these expert opinions. Experiments revealed that a streamlined system using only a UniXcoder backbone with supervised contrastive learning achieved a Macro-F1 score of 0.31389, surpassing the official CodeBERT baseline and securing the 26th rank overall.

Key takeaway

For Machine Learning Engineers developing code authorship identification systems, consider integrating supervised contrastive learning with models like UniXcoder. This approach demonstrated superior performance over CodeBERT, achieving a Macro-F1 of 0.31389 in distinguishing between 10 AI generator families and human code. You should explore ensemble architectures, including binary human-vs-machine layers, to enhance robustness against class imbalance and capture subtle stylometric differences in code.

Key insights

Supervised contrastive learning with UniXcoder effectively identifies code authorship from AI generators and humans.

Principles

• Ensemble architectures can manage class imbalance.
• Stylometric fingerprints aid code attribution.
• Streamlining models can improve performance.

Method

A three-stage ensemble: fine-tune UniXcoder/ModernBERT with Supervised Contrastive Learning, process embeddings with shallow experts and a Human Shield, then aggregate via a Gated Meta-Learner.

In practice

• Use UniXcoder for code authorship tasks.
• Apply contrastive learning for fine-tuning.
• Implement ensemble for imbalanced data.

Topics

• Code Authorship Identification
• Supervised Contrastive Learning
• Ensemble Architectures
• UniXcoder
• AI Code Generation
• SemEval-2026

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• HybridCodeAuthorship: A Benchmark Dataset for Line-Level Code Authorship Detection — Industry codebases require fine-grained, line-level detection of AI-generated code for risk and productivity analysis.
• How Far Can Classical NLP Go? From Bag-of-Words to Stacking on Spooky Author Identification — Classical NLP, with careful feature engineering and ensemble stacking, excels at stylistic text classification.
• Towards Knowledge Alignment in Code LLMs: Contrastive Unlearning for Evolving APIs — CURE uses contrastive unlearning to both suppress deprecated APIs and promote valid replacements in Code LLMs.
• Temporal Contrastive Transformer for Financial Crime Detection: Self-Supervised Sequence Embeddings via Predictive Contrastive Coding — TCT uses self-supervised contrastive learning to generate financial transaction embeddings, achieving strong predictive performance comparable to engineered features.
• Evaluation of AutoML Frameworks for IDS under Imbalanced Data Conditions of the NSL-KDD Dataset — AutoML frameworks require native imbalance-aware optimization and stratified evaluation for reliable intrusion detection on severely imbalanced datasets.
• Can AI tools assess coding assignments? — Generative AI can assist in coding assignment assessment when integrated thoughtfully with human expertise and context.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.