TeamOmega at SemEval-2026 Task 13: Frozen vs. Trainable Representations for Out-of-Distribution AI-Generated Code Detection: A CodeBERT Fine-Tuning Study

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

Summary

TeamOmega developed a CodeBERT-based system for SemEval-2026 Task 13, focusing on detecting AI-generated code, particularly under significant cross-language and cross-domain distribution shifts. Their research compared two configurations: a fully frozen CodeBERT backbone and a partially fine-tuned setup where only the final transformer layer was unfrozen, utilizing discriminative learning rates. The study found that while partial fine-tuning significantly enhanced in-domain performance, the fully frozen backbone exhibited superior robustness when evaluated on out-of-distribution data. These results underscore a critical trade-off between achieving high task adaptation and maintaining strong cross-language generalization capabilities in the context of machine-generated code detection.

Key takeaway

For Machine Learning Engineers developing AI-generated code detection systems, if your application requires strong robustness against cross-language and cross-domain shifts, prioritize using a fully frozen CodeBERT backbone. While partial fine-tuning improves in-domain accuracy, your system's ability to generalize to out-of-distribution code will be stronger with a frozen model, directly impacting reliability in diverse real-world scenarios. Evaluate this trade-off carefully based on your specific deployment environment.

Key insights

Frozen CodeBERT backbones offer better out-of-distribution robustness for AI-generated code detection than partially fine-tuned models, despite lower in-domain performance.

Principles

• Fine-tuning improves in-domain task adaptation.
• Frozen backbones enhance OOD robustness.
• A trade-off exists between adaptation and generalization.

Method

The method involves comparing a fully frozen CodeBERT backbone against a partially fine-tuned configuration, unfreezing only the final transformer layer with discriminative learning rates for AI-generated code detection.

In practice

• Consider frozen backbones for OOD code detection.
• Use partial fine-tuning for in-domain tasks.
• Evaluate trade-offs for cross-language generalization.

Topics

• AI-Generated Code Detection
• CodeBERT Fine-Tuning
• Out-of-Distribution Robustness
• Cross-Language Generalization
• Transformer Models
• SemEval-2026

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

Related on AIssential

• Selection Without Signal, Recovery Through Expression: A Measurement Study of Post-Hoc Falsification Operators for Frozen Small Code Models — Semantic post-hoc operators fail to improve small code model accuracy; focus on extraction and harness issues instead.
• Is Fine-Tuning Still Valuable? — Effective LLM fine-tuning requires robust evaluation systems and iterative data curation, especially for domain-specific tasks.
• Modular Monolingual Adaptation using Pretrained Language Models — Modular PLM adaptation, freezing embeddings while tuning the rest, improves low-resource language performance.
• [AINews] GLM-5.2: the top Frontend Coding model in the world, IndexShare for Speculative Decoding — Open-weight models can achieve frontier-level coding and agentic performance with efficient long-context handling.
• Amplifying, Not Learning: Fine-Tuned AI Text Detectors Amplify a Pretrained Direction — Fine-tuned AI text detectors amplify a pretrained typicality axis, not construct a new AI-vs-human boundary.
• Dead Weights, Live Signals: Feedforward Graphs of Frozen Language Models — Frozen LLMs can form a trainable feedforward graph via linear projections in a shared latent space.

Open in AIssential →

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