YNU-HPCC at SemEval-2026 Task 13: Robust Machine-Generated Code Detection under Distribution Shifts

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

Summary

The YNU-HPCC team developed a system for SemEval-2026 Task 13, focusing on robust machine-generated code detection under distribution shifts, including cross-language, multi-generator, and hybrid scenarios. The system systematically examined three modeling paradigms: encoder-based fine-tuning, feature-based machine learning, and task-specific robustness strategies. For Subtask A (Binary Detection), frozen pre-trained encoders and shallow stylometric features demonstrated superior cross-domain robustness compared to full fine-tuning, with indentation entropy identified as a crucial discriminative signal. Subtask B (Multi-Class Attribution) utilized a hierarchical two-stage framework to separate human–machine discrimination from fine-grained generator attribution, effectively addressing class imbalance. For Subtask C (Hybrid Detection), a token-level splicing augmentation strategy, combined with Supervised Contrastive Learning and Focal Loss, was employed to model intra-sample stylistic variation. The system achieved notable rankings: 12th out of 81 teams in Subtask A, 14th out of 34 in Subtask B, and 8th out of 32 in Subtask C.

Key takeaway

For AI Engineers developing code origin detection systems, prioritize robust strategies against distribution shifts. You should consider using frozen pre-trained encoders and shallow stylometric features, like indentation entropy, for better cross-domain performance. When tackling multi-generator attribution, implement a hierarchical two-stage framework to manage class imbalance effectively. For hybrid code detection, explore token-level splicing augmentation with Supervised Contrastive Learning and Focal Loss to capture stylistic variations.

Key insights

Robust machine-generated code detection benefits from specialized strategies for distribution shifts and class imbalance.

Principles

• Frozen encoders and stylometric features enhance cross-domain robustness.
• Indentation entropy is a key discriminative signal for code origin.
• Decoupling attribution helps manage multi-class imbalance.

Method

A hierarchical two-stage framework for multi-class attribution, and token-level splicing augmentation with Supervised Contrastive Learning and Focal Loss for hybrid detection.

In practice

• Prioritize frozen encoders and stylometric features for cross-domain code detection.
• Implement indentation entropy as a core feature in code analysis.
• Use two-stage attribution for multi-generator code classification.

Topics

• Machine-Generated Code Detection
• Large Language Models
• Code Stylometry
• Distribution Shift Robustness
• SemEval-2026 Task 13
• Supervised Contrastive Learning

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

Related on AIssential

• SV-Detect: AI-generated Text Detection with Steering Vectors — SV-Detect uses steering vectors from frozen LMs to robustly detect AI-generated text, even under distribution shifts.
• AI Code Generation Ambiguity — AI code generation ambiguity stems from natural language complexity and non-determinism, impacting correctness and reliability.
• A Zipf-preserving, long-range correlated surrogate for written language and other symbolic sequences — A new model generates symbolic sequence surrogates preserving both symbol frequencies and long-range correlations.
• 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.
• Detecting AI Coding Agents in Open Source: A Validated Multi-Method Census of 180 Million Repositories — Accurate AI coding agent detection requires multi-method approaches, as single signals severely underestimate their widespread, diverse contributions.
• 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.

Open in AIssential →

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