MALTO at SemEval-2026 Task 13: Detecting Human, AI, and Hybrid Code via Hard Negative Mining and Curriculum-Driven Ensembles

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

Summary

The MALTO team's submission for SemEval-2026 Task 13 addresses the challenge of determining source code origin and authenticity, focusing on Subtask B (Authorship Attribution among 11 classes) and Subtask C (Hybrid Code Detection). Their unified framework employs a soft-voting ensemble of UniXcoder and CodeT5. This approach integrates a Tree-sitter-based Universal Canonicalization strategy, Data Augmentation, and a 3-Phase Curriculum Training schedule enhanced by Hard Negative Mining to counter severe class imbalance and complex mixed human-machine code boundaries. UniXcoder excels at distinguishing semantically overlapping LLM families for Subtask B, while CodeT5's identifier-aware architecture is superior for detecting structural anomalies in hybrid and adversarial snippets for Subtask C. The ensemble aggregates these complementary strengths, demonstrating robustness against imbalanced distributions and effectively discriminating between purely human, machine, hybrid, and adversarial code.

Key takeaway

For machine learning engineers developing code origin attribution systems, this work highlights the effectiveness of combining specialized models. Your approach should leverage an ensemble, such as UniXcoder for semantic distinctions and CodeT5 for structural anomalies, to handle diverse code types. Implementing techniques like hard negative mining and curriculum training will enhance robustness against class imbalance, enabling accurate classification of human, AI, and hybrid code, which is critical for maintaining software integrity and trust.

Key insights

An ensemble of specialized models with advanced training techniques robustly detects human, AI, and hybrid code origins.

Principles

• Ensembling models with complementary strengths improves overall robustness.
• Hard Negative Mining and curriculum training mitigate severe class imbalance.
• Canonicalization and data augmentation enhance model generalization.

Method

A unified framework ensembles UniXcoder and CodeT5, applying Tree-sitter-based Universal Canonicalization, Data Augmentation, and a 3-Phase Curriculum Training schedule with Hard Negative Mining.

In practice

• Distinguish among 11 semantically overlapping LLM families.
• Detect subtle structural anomalies in hybrid and adversarial code.
• Classify purely human, machine, hybrid, and adversarial code snippets.

Topics

• SemEval-2026
• Code Authorship Attribution
• Hybrid Code Detection
• LLM Code Generation
• Model Ensembling
• Hard Negative Mining

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

Related on AIssential

• HybridCodeAuthorship: A Benchmark Dataset for Line-Level Code Authorship Detection — The increasing hybrid nature of codebases necessitates fine-grained, line-level AI authorship detection benchmarks.
• Code Is More Than Text: Uncertainty Estimation for Code Generation — Code generation uncertainty estimation requires specialized methods accounting for code's unique properties, outperforming natural language-derived baselines.
• mdok-style at SemEval-2026 Task 10: Finetuning LLMs for Conspiracy Detection — Data augmentation and self-training can effectively finetune LLMs for specialized text classification tasks with limited data.
• A Prototypical Signature Approach for Writer-Independent Offline Signature Verification — Prototypical signatures enhance negative sample generation for improved writer-independent offline signature verification.
• AlignCoder: Aligning Retrieval with Target Intent for Repository-Level Code Completion — AlignCoder enhances RAG for repository-level code completion by using multiple candidate completions to refine queries and training a retriever with reinforcement learning.
• 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.

Open in AIssential →

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