MindFlayer at SemEval-2026 Task 13:LACR-ENS: Calibration-Aware Ensemble Routing for Cross-Language AI-Generated Code Detection
Year: 2026
Type: conference-paper
Abstract: We present LACR-ENS, our submission to SemEval-2026 Task 13 (Orel et al., 2026b) on detecting AI-generated code across eight programming languages.We frame the task as a calibration study and report an asymmetric out-of-distribution (OOD) failure mode in fine-tuned code transformers: expected calibration error roughly doubles when the model is moved from seen to unseen languages, and the model is wrong on a non-trivial fraction of its highest-confidence predictions.Our main contribution, Language-Aware Confidence Routing (LACR), applies language-conditioned thresholds to the ensemble logits and is formally equivalent to implicit per-language temperature scaling; it reduces OOD calibration error substantially while leaving raw F1 essentially unchanged.We further show that syntactic proximity to the training languages strongly predicts OOD behaviour, motivating a continuous extension of LACR.The final ensemble of UniXCoder (Guo et al., 2022) and GraphCodeBERT (Guo et al., 2021) attains macro-F1 0.538 on the official test set.Code: https://github.com/Jerin-Romijah-Tuli/ Mind Flayer Semeval2026 task 13 A.
Language: English
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
Funder: Rajshahi University
OpenAlex ID: W7166845542