Dream at SemEval-2026 Task 13: SALSA for Single-Pass Machine-Generated Code Detection
Quick Answer
This paper shows that The SemEval-2026 Task 13 introduces SALSA, a single-pass autoregressive LLM for detecting machine-generated code, achieving an OOD F1 score of 0.789, significantly surpassing CodeBERT's 0.305.
Quick Take
The SemEval-2026 Task 13 introduces SALSA, a single-pass autoregressive LLM for detecting machine-generated code, achieving an OOD F1 score of 0.789, significantly surpassing CodeBERT's 0.305. This method emphasizes OOD generalization and avoids overfitting through balanced sampling and conservative training techniques.
Key Points
- SALSA maps each class to a dedicated output token for structured classification.
- The model is designed to emit a single-token label without hand-crafted features.
- Balanced sampling and low learning rate training enhance OOD robustness.
- Best system outperformed CodeBERT by a significant margin on the leaderboard.
- Focus on unseen programming languages and application domains for broader applicability.
Paper Resources
Article Excerpt
From source RSS / original summaryarXiv:2606. 25102v1 Announce Type: new Abstract: Large language models have transformed code generation, raising concerns around authorship, assessment integrity, and software trust. SemEval-2026 Task 13 Subtask A operationalizes detection as binary classification over code snippets, with a particular emphasis on out-of-distribution (OOD) generalization across unseen programming languages and application domains.
We propose a SALSA-style formulation, Single-pass Autoregressive LLM Structured Classification, that maps each class to a dedicated output token and trains the model to emit a single-token label in a structured response. Rather than engineering hand-crafted features or decision rules, this formulation delegates the authorship decision to the model.
To improve OOD robustness, we combine balanced sampling across languages with parameter-efficient fine-tuning and conservative training (low learning rate, single epoch) to avoid overfitting to the training domain. Our best system achieves OOD $F_1 = 0. 789$ on the official leaderboard, substantially outperforming the CodeBERT baseline ($F_1 = 0. 305$).
Want this in your inbox every morning?
Daily brief at your local 8am — bilingual EN/中文, free.
More from arXiv cs.CL
See more →Quantifying Prior Dominance in Systems
The study introduces the Normalized Context Utilization (NCU) metric to evaluate Retrieval-Augmented Generation (RAG) systems, revealing that Small Language Models (SLMs) outperform larger models in factual extraction. The findings indicate that traditional scaling laws yield diminishing returns, with a commercial API frequently failing against adversarial evidence due to systemic confidence collapse.