Codexa at SemEval-2026 Task 13: Loss Engineering and Diverse Ensemble Strategies for Multi-Class Code Authorship Attribution

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

Summary

Codexa's system for SemEval-2026 Task 13, Subtask B, addresses the challenge of multi-class code authorship attribution, classifying code into 11 categories (human-written or from one of 10 LLM families). The task is complicated by extreme class imbalance and a significant distribution shift, with 31 code generators in training and 59 in testing, including 36 unseen ones. Their approach combines two main components: a UniXcoder encoder utilizing Label-Distribution-Aware Margin (LDAM) loss, which improved performance by +7% over a cross-entropy baseline by effectively handling class imbalance; and a diverse ensemble of 12 models trained with varied objectives and architectures, integrated via hard voting. The system achieved a 41.28% Macro F1 score on the official test set. A key finding was that loss engineering and ensemble diversity were more impactful than domain adaptation techniques, which consistently degraded test performance.

Key takeaway

For Machine Learning Engineers developing code authorship attribution systems, prioritize loss engineering and ensemble diversity over complex domain adaptation. If you face extreme class imbalance, implement Label-Distribution-Aware Margin (LDAM) loss, which showed a +7% improvement. Additionally, build diverse model ensembles with varied architectures and objectives, combining them with hard voting to enhance robustness. Be cautious with domain adaptation techniques, as they can degrade performance in scenarios with significant distribution shifts.

Key insights

Loss engineering and diverse model ensembles significantly improve multi-class code authorship attribution, outperforming domain adaptation.

Principles

• Class imbalance demands specialized loss functions.
• Diverse model ensembles boost classification robustness.
• Domain adaptation may hinder performance in distribution shifts.

Method

The system uses UniXcoder with LDAM loss for class imbalance, then combines 12 diverse models via hard voting for final classification.

In practice

• Apply LDAM loss for imbalanced code classification.
• Construct diverse ensembles using hard voting.
• Prioritize loss engineering over domain adaptation.

Topics

• Code Authorship Attribution
• SemEval-2026 Task 13
• Class Imbalance
• LDAM Loss
• Ensemble Learning
• UniXcoder
• LLM Classification

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

Related on AIssential

• Authorship Attribution in Multilingual Machine-Generated Texts — Multilingual authorship attribution for LLM-generated text is complex, with current monolingual methods showing limited cross-lingual transferability across diverse language families.
• CAMO: A Class-Aware Minority-Optimized Ensemble for Robust Language Model Evaluation on Imbalanced Data — CAMO is a hierarchical ensemble method that dynamically boosts minority class predictions using confidence and uncertainty.
• Domain-Specific Data Synthesis for LLMs via Minimal Sufficient Representation Learning — DOMINO inductively synthesizes domain-specific data for LLMs by learning minimal representations from reference examples, bypassing explicit domain descriptions.
• Distributional Loss for Robust Classification — Distributional Loss uses a bimodal Gaussian distribution as a softer target for classifier outputs, enhancing robustness and mitigating overfitting without extra labels.
• DIVERSED: Relaxed Speculative Decoding via Dynamic Ensemble Verification — Dynamic ensemble verification in speculative decoding significantly boosts token acceptance and inference speed without sacrificing quality.
• Aggregate Models, Not Explanations: Improving Feature Importance Estimation — Ensembling models directly improves feature importance estimation by reducing excess risk, especially for complex ML models.

Open in AIssential →

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