CodeDet-NITS at SemEval-2026 Task 13: AI Code Authorship Detection Beyond Truncation

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

Summary

CodeDet-NITS presents a lightweight system for SemEval 2026 Task 13 Subtask B, focusing on AI code authorship detection. This system automatically determines if source code is human-written or produced by one of ten specific large language model (LLM) families. This is crucial for assessment, provenance tracking, and dataset curation. The method repurposes Qwen2.5 Coder series backbones as sequence classifiers, adapting them with QLoRA. This combines frozen low-precision weights with low-rank trainable adapters, reducing memory and compute overhead. A core design uses an overlapping sliding window strategy to process long code snippets without truncation. During training, long examples expand into multiple fixed-length windows sharing the same label. For validation and testing, windows are generated on the fly. Their evidence is aggregated by averaging logits to produce a single, stable prediction per snippet. The final submission achieved an 8th place ranking on the official Subtask B test set leaderboard.

Key takeaway

For Machine Learning Engineers developing code analysis tools, this research offers a robust approach to AI code authorship detection. If you are struggling with context loss in long code snippets, consider implementing an overlapping sliding window strategy. This method, combined with QLoRA for efficient model adaptation, allows you to process full inputs and aggregate evidence for stable, accurate predictions. It provides a competitive framework for attributing code to human or specific LLM families, enhancing provenance tracking and dataset curation.

Key insights

AI code authorship detection can be improved by processing full snippets via sliding windows and QLoRA-adapted models.

Principles

• Avoid code truncation to preserve full contextual evidence.
• QLoRA adapts large models efficiently for specific tasks.
• Aggregate window-level evidence for stable snippet predictions.

Method

Repurpose Qwen2.5 Coder backbones as sequence classifiers. Adapt with QLoRA for efficiency. Apply an overlapping sliding window strategy for long snippets, expanding them into multiple windows during training. Aggregate window logits by averaging for final prediction.

In practice

• Use QLoRA for memory-efficient fine-tuning.
• Implement sliding windows for long sequence processing.
• Average logits from windows for robust classification.

Topics

• AI Code Authorship Detection
• Large Language Models
• QLoRA
• Sliding Window Strategy
• Code Classification
• SemEval 2026

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

Related on AIssential

• HybridCodeAuthorship: A Benchmark Dataset for Line-Level Code Authorship Detection — Industry codebases require fine-grained, line-level detection of AI-generated code for risk and productivity analysis.
• Cross-lingual Relation Extraction with Large Language Models: Zero-Shot, Few-Shot, and Fine-Tuned Evaluation on Romanian — QLoRA fine-tuning significantly boosts cross-lingual relation extraction performance for low-resource languages, outperforming prompt-only LLMs.
• 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.
• Adapting TrOCR for Printed Tigrinya Text Recognition: Word-Aware Loss Weighting for Cross-Script Transfer Learning — Adapting TrOCR for Tigrinya text requires tokenizer extension and Word-Aware Loss Weighting to overcome script-specific challenges.
• Detecting Functional Memorization in Code Language Models — Code LLMs can functionally memorize logic even when textually dissimilar, requiring advanced auditing.
• Operation-Guided Progressive Human-to-AI Text Transformation Benchmark for Multi-Granularity AI-Text Detection — AI-text detectability is non-monotonic, influenced by edit operations and revision history, not just AI content proportion.

Open in AIssential →

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