Dream at SemEval-2026 Task 13: SALSA for Single-Pass Machine-Generated Code Detection

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

Summary

The paper "Dream at SemEval-2026 Task 13: SALSA for Single-Pass Machine-Generated Code Detection" introduces SALSA (Single-pass Autoregressive LLM Structured Classification), a novel method for detecting machine-generated code. This approach addresses concerns regarding authorship and software trust arising from large language models' code generation capabilities. SALSA formulates detection as a binary classification task, mapping each class to a dedicated output token for single-token label emission. The core innovation lies in delegating authorship decisions to the model rather than relying on hand-crafted features. To ensure robustness against out-of-distribution (OOD) data, particularly across unseen programming languages and domains, the system employs balanced sampling, parameter-efficient fine-tuning, and conservative training with a low learning rate over a single epoch. This strategy resulted in an OOD F1 score of 0.789, substantially outperforming the CodeBERT baseline's F1 of 0.305 on the SemEval-2026 Task 13 leaderboard.

Key takeaway

For Machine Learning Engineers developing code authorship detection systems, this SALSA approach offers a robust strategy for out-of-distribution generalization. You should consider implementing a structured classification model that delegates authorship decisions to the LLM, rather than relying on hand-crafted features. Employing balanced sampling, parameter-efficient fine-tuning, and conservative training with a single epoch can significantly improve your system's performance on unseen languages and domains, as demonstrated by its F1 score of 0.789.

Key insights

SALSA enables robust, single-pass detection of machine-generated code, achieving strong out-of-distribution generalization.

Principles

• Delegate authorship decisions to the model.
• Prioritize OOD robustness in training.
• Avoid overfitting with conservative fine-tuning.

Method

SALSA formulates code detection as structured classification, mapping each class to a dedicated output token. It uses balanced sampling, parameter-efficient fine-tuning, and conservative training (low learning rate, single epoch) for OOD robustness.

In practice

• Apply SALSA for binary code authorship classification.
• Use balanced sampling for language diversity.
• Employ PEFT with conservative training.

Topics

• Machine-Generated Code Detection
• Large Language Models
• Out-of-Distribution Generalization
• Parameter-Efficient Fine-Tuning
• SemEval-2026 Task 13
• Code Authorship

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

Related on AIssential

• HybridCodeAuthorship: A Benchmark Dataset for Line-Level Code Authorship Detection — The increasing hybrid nature of codebases necessitates fine-grained, line-level AI authorship detection benchmarks.
• Detecting Functional Memorization in Code Language Models — Code LLMs can functionally memorize logic even when textually dissimilar, requiring advanced auditing.
• 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.
• 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.
• Intro to NLP with spaCy (4): Detecting programming languages — Training a custom spaCy Named Entity Recognition (NER) model automates programming language detection by learning from labeled data.
• Beyond Surface Forms: A Comprehensive, Mechanism-Oriented Taxonomy of Indirect Linguistic Encoding for LLM-Based Coded Language Detection — A mechanism-oriented taxonomy of indirect linguistic encoding significantly enhances LLM-based coded language detection.

Open in AIssential →

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