UIT-AMMC at SemEval-2026 Task 13: Exploiting Structural Formatting Signatures for Robust AI-Generated Code Detection

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

Summary

UIT-AMMC's Structure-Aware Contrastive Cascade system, presented at SemEval-2026 Task 13, addresses the challenge of distinguishing human-authored from AI-generated code. This multi-stage architecture integrates generative reasoning with explicit structural linguistic features, specifically exploiting formatting signatures that emerge in AI-generated code due to post-training alignment. The system's pipeline employs a Qwen-2.5-Coder 14B model, fine-tuned using QLoRA, and incorporates stochastic data augmentation to ensure robustness across diverse programming languages. Classification is finalized via a late-fusion mechanism that combines contrastive probability scores with code presentation density metrics. For cases of high epistemic uncertainty, a multi-agent adversarial debate step is used to refine the decision. This approach achieved a Macro F1 score of 0.802, securing 3rd place on the official leaderboard for Subtask A.

Key takeaway

For AI Security Engineers tasked with verifying code origin, UIT-AMMC's approach highlights the importance of structural formatting signatures. You should integrate analysis of these signatures, alongside generative reasoning, into your code detection pipelines. This method achieved a Macro F1 of 0.802. It suggests focusing on subtle AI-induced patterns significantly improves robustness in distinguishing human from machine-generated code across diverse languages. Consider implementing multi-agent debate for high-uncertainty cases.

Key insights

AI-generated code can be robustly detected by exploiting structural formatting signatures and combining generative reasoning with statistical code presentation metrics.

Principles

• AI-generated code leaves structural formatting signatures.
• Combine generative reasoning with explicit structural features.
• Stochastic data augmentation enhances cross-language robustness.

Method

Fine-tune Qwen-2.5-Coder 14B via QLoRA with data augmentation. Classify using late-fusion of contrastive probabilities and code presentation density. Refine uncertain verdicts with multi-agent adversarial debate.

In practice

• Identify AI-generated code for integrity checks.
• Verify code authorship in professional settings.
• Detect LLM-specific formatting patterns.

Topics

• AI-Generated Code Detection
• Structural Formatting Signatures
• SemEval-2026 Task 13
• Qwen-2.5-Coder
• QLoRA
• Multi-agent Systems

Best for: Research Scientist, CTO, VP of Engineering/Data, AI Scientist, Machine Learning Engineer, AI Security Engineer

Related on AIssential

• How Generation Architecture Shapes Code Complexity in Multi-Agent LLM Systems: A Paired Study on HumanEval — Multi-agent LLM code generation architectures significantly impact code complexity without necessarily improving functional correctness.
• Sun Finance automates ID extraction and fraud detection with generative AI on AWS — Combining specialized OCR with LLM structuring significantly enhances ID extraction accuracy and efficiency.
• Citation Discipline in Spec-Driven Development: A Cross-Model Empirical Study of Output Determinism and Automated Hallucination Detection in LLM-Generated Code — Citation annotations in LLM-generated code trade output determinism for critical verifiability and automated hallucination detection.
• On the troubling rise of generative AI suspicion in academic publishing — Overextended suspicion of GenAI in academic publishing leads to false positives and an "AI witch hunt."
• Contract-Coding: Towards Repo-Level Generation via Structured Symbolic Paradigm — Contract-Coding uses a formal "Language Contract" to enable parallel, intent-driven repository generation, mitigating context and hallucination issues.
• Code-Augur: Agentic Vulnerability Detection via Specification Inference — Agentic vulnerability detection improves by making LLM assumptions explicit and continuously validating them with fuzzing.

Open in AIssential →

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