Team Yuvan at SemEval-2026 Task 13: Task-Adaptive Ensemble Strategies for AI-Generated Code Detection

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

Team Yuvan's system for SemEval-2026 Task 13 addresses the detection of AI-generated code across eight programming languages and three distinct subtasks: binary human-vs-AI detection (Task A), 11-way source identification (Task B), and 4-way generator classification (Task C). Their approach integrates Qwen2.5-Coder-1.5B with LoRA fine-tuning, abstract syntax tree (AST) features, CodeBERT with head-tail chunking, and TF-IDF features. Experiments revealed that for Task A, neural detectors significantly degrade on the official test split, while AST-based structural features maintain stability, suggesting a substantial distribution shift. For Task B, inverse-frequency class weighting proved essential for extreme label imbalance, substantially improving macro-F1 scores. Task C demonstrated that combining neural and statistical models yields superior performance compared to single models, highlighting their complementary strengths. The final system achieved macro-F1 scores of 0.638 on Task A, 0.449 on Task B, and 0.714 on Task C.

Key takeaway

For Machine Learning Engineers developing AI-generated code detection systems, you should prioritize ensemble approaches that combine diverse model types. If facing distribution shifts, integrate abstract syntax tree (AST) features for greater stability, as neural models alone may degrade. When dealing with imbalanced datasets, implement inverse-frequency class weighting to significantly improve macro-F1 scores. Your strategy should leverage the complementary strengths of both neural and statistical models to achieve robust and accurate detection across various code generation tasks.

Key insights

Ensemble strategies combining neural and structural features enhance AI-generated code detection robustness and performance across diverse tasks.

Principles

• Distribution shifts impact neural detectors more than structural features.
• Inverse-frequency weighting is vital for imbalanced classification.
• Hybrid models leverage complementary strengths for better results.

Method

The system combines Qwen2.5-Coder-1.5B (LoRA fine-tuned), AST features, CodeBERT (head-tail chunking), and TF-IDF features in a task-specific ensemble.

In practice

• Prioritize AST features for distribution shift robustness.
• Apply inverse-frequency weighting for imbalanced datasets.
• Integrate neural and statistical models for improved detection.

Topics

• AI-Generated Code Detection
• Ensemble Strategies
• Abstract Syntax Trees
• Qwen2.5-Coder-1.5B
• CodeBERT
• Distribution Shift
• Class Imbalance Handling

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

Related on AIssential

• Model-Agnostic Meta Learning for Class Imbalance Adaptation — HAMR dynamically addresses class imbalance and data difficulty in NLP via adaptive weighting and neighborhood-aware resampling.
• Quantifying Sensitivity for Tree Ensembles: A symbolic and compositional approach — A new algorithm efficiently quantifies DTE sensitivity by discretizing input space and using Algebraic Decision Diagrams.
• Promoting Simple Agents: Ensemble Methods for Event-Log Prediction — N-gram models offer competitive accuracy for event-log prediction with significantly fewer resources than neural networks.
• Shapley Value-Guided Adaptive Ensemble Learning for Explainable Financial Fraud Detection with U.S. Regulatory Compliance Validation — Explainable AI models can achieve high fraud detection accuracy while meeting stringent U.S. financial regulations.
• AdaBoost - Explained — AdaBoost combines weak classifiers by adaptively weighting data points and their votes to form a strong ensemble.
• Enhancing Visual Feature Attribution via Weighted Integrated Gradients — Weighted Average Gradients (WG) improves feature attribution by fitness-weighting baselines, outperforming uniform averaging and reducing variance.

Open in AIssential →

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