United States

This paper describes the system submitted by team Archaeology to SemEval-2026 Task~13 on AI-generated code detection. The shared task consists of three subtasks; we participate in Subtask-A (binary classification: human-written vs.\ AI-generated code) and Subtask-B (11-class attribution of the generating model).
Starting from a TF-IDF and Logistic Regression baseline, we fine-tune four pre-trained code models (CodeBERT, GraphCodeBERT, UniXcoder, and CodeT5+) with separate strategies for each subtask.
For Subtask-A, we use leave-one-language-out cross-validation, code augmentation, chunked inference with trimmed-mean aggregation, and threshold calibration on a difficult dataset.
For Subtask-B, we use sandwich token packing, class-balanced loss, and multi-seed ensembling with test-time augmentation. Our best submissions obtain macro-F1 scores of 0.737 on Subtask-A and 0.422 on Subtask-B.

ACL 2026

Archaeology at SemEval-2026 Task 13: Fine-Tuning Pre-Trained Code Models for AI-Generated Code Detection

code authorship attribution

ai-generated code detection

pre-trained code models

It is my great pleasure and honor to welcome you to the 64th Annual Meeting of the Association for Computational Linguistics (ACL 2026), held in San Diego, California, USA, July 2–7, 2026. ACL 2026 continues our fields tradition of excellence in scholarship, innovation, and inclusivity, and I am deeply grateful to the many volunteers who have worked tirelessly to bring this event to life.

It is exciting to see the many developments in the field reflected at our conference, such as the advancement of large language models (born out of our community) into many application directions. This is not without challenges, partly due to unsolved problems, and partly due to misuse of the technology. All this is reflected well at ACL 2026. The conference also reflects that the field is not only about advancing technology, but also the science of fundamentally understanding language and how it is use by many different speaker communities. <br>
[Full Message from General Chair Philipp Koehn](https://drive.google.com/file/d/1QiJFOmMjvDgsDo3i1XzlQnImxe3RU_Cy/view?usp=sharing) <br><br><br>
<html><button style="display: inline-flex; align-items: center; justify-content: center; white-space: nowrap; border-radius: 9999px; font-weight: bold; background: #7c3aed; color: white; font-family: 'Space Grotesk', sans-serif; height: 40px; font-size: 16px; padding: 0 20px; border: none; cursor: pointer" onclick="window.open('https://drive.google.com/file/d/1QiJFOmMjvDgsDo3i1XzlQnImxe3RU_Cy/view?usp=sharing')">Conference Handbook</button></html>
<br><br><br><br>

To access this event page, you need to log in with the **email address you registered with**. <br>Access credentials will be sent to your email from Underline -  subject line "Welcome to ACL 2026". Please be sure to check your spam email folder if you do not see an email confirmation right away.

To access this event page, you are required to register.
Please complete your registration to continue.

We recommend reading [**the registration information**](https://2026.aclweb.org/registration/) first.

**Online Registration Form**: https://acl.swoogo.com/acl2026

Registration Required

The 64th Annual Meeting of the Association for Computational Linguistics (ACL 2026) will take place in San Diego, California, from July 2 - 7, 2026.

workshop paper

We describe our system for SemEval-2026 Task 13, Subtask B, which focuses on multi-class authorship attribution for code: given a code snippet, the goal is to predict whether it is human-written or generated by one of ten LLM families. The task presents two central challenges: severe class imbalance and long input sequences that frequently exceed the context length of encoder-based Transformers. To address these issues, we adopt a window-based fine-tuning and inference framework. During training, we randomly sample 512-token windows from each snippet and optimize a class-weighted cross-entropy objective with label smoothing. At inference time, we apply a sliding-window strategy and aggregate window-level logits to obtain a snippet-level prediction. We fine-tune three pretrained code encoders (CodeBERT, UniXcoder, and StarEncoder) under this framework and combine their outputs via majority voting. On the official validation split, our best single model (StarEncoder) achieves 0.60 macro F1. On the final test set, the three-model ensemble reaches 0.41 macro F1, ranking 10th on the leaderboard. Our results demonstrate that window-based modeling combined with imbalance-aware optimization provides a robust and reproducible baseline for multi-class LLM attribution under distribution shift.

ICI Innolabs at SemEval-2026 Task 13: Sliding Windows Meet Code Transformers

In this paper, we present the solution submitted by TUCNLP at SemEval-2026 Task~11: Disentangling Content and Formal Reasoning in Large Language Models. The task requires predicting the formal validity of categorical syllogisms while minimizing susceptibility to content-driven biases in English and 11 additional languages. We show that a modestly-sized model (Qwen3-8B) can achieve near-perfect logical reasoning on the English validity-only subtask, and large reductions in content effect on multilingual and premise-retrieval variants, when augmented with a multi-stage neuro-symbolic pipeline: LLM-based content stripping with iterative error correction converts natural language to abstract categorical forms, a classical symbolic parser validates against the twenty-four Aristotelian syllogistic forms, and asymmetric confidence thresholds mediate between symbolic and neural decisions. Across the four subtasks (ST1 to ST4), our system achieves accuracy ranging from 91.1\% to 100\% and bias-penalized ranking scores ($\mathcal{M}$) from 31.8 to 100.0, with the main bottleneck being overconfident neural predictions that bypass symbolic verification.

TUCNLP at SemEval-2026 Task 11: Neuro-Symbolic Content Stripping for Debiased Syllogistic Reasoning

We describe our system for SemEval-2026 Task 9 (POLAR), a shared task on detecting multilingual, multicultural, and multievent online polarization. POLAR includes three subtasks: polarization detection (ST1), polarization type classification (ST2), and polarization manifestation identification (ST3). We participated in ST1 and ST2 for Arabic. We compare three method families, encoder fine-tuning, zero-shot prompting, and retrieval-augmented in-context learning (RAG-ICL), across six LLMs and six Arabic encoder models. For ST1, RAG-ICL with Gemma-3-27B reaches F1\textsubscript{macro}~=~0.83 on the test set, a $+0.33$ absolute gain over zero-shot prompting. For ST2, a pipeline that uses the best ST1 encoder as a hard filter before RAG-ICL classification reduces hallucination from 58\% to 19\% and yields F1\textsubscript{macro}~=~0.62. English prompts consistently outperform Arabic prompts, and advanced prompting strategies (chain-of-thought, self-refinement) do not improve over standard RAG-ICL.

NAMAA at SemEval-2026 Task 9: Comparing Generative, Retrieval-Augmented, and Discriminative Methods for Arabic Online Polarization Detection and Type Classification

We describe our system for SemEval-2026 Task 13, Subtask B: code classification into 11 categories (human-written or generated by one of 10 LLM families). The task presents extreme class imbalance and distribution shift across multiple generators provided in the dataset (31 in training, 59 in test, with 36 unseen). On that focus, we approached with two components: (1) UniXcoder as the encoder with Label-Distribution-Aware Margin (LDAM) loss for handling class imbalance, which provides a +7% absolute improvement over the cross-entropy baseline; and (2) a diverse ensemble of 12 models trained with different objectives and architectures which is detailed in the appendix, combined with hard voting. Our system achieves 41.28% Macro F1 on the official test set. We find that loss engineering and ensemble diversity matter more than domain adaptation techniques, which consistently degraded test performance.

Codexa at SemEval-2026 Task 13: Loss Engineering and Diverse Ensemble Strategies for Multi-Class Code Authorship Attribution

We present the results and the main findings of SemEval-2026 Task 13: Detecting Machine-Generated Code with Multiple Programming Languages, Generators, and Application Scenarios. Our task featured three subtasks. Subtask A is a binary classification task
that determines whether a given code snippet is written by a human or generated by a machine. This subtask focuses on the development of robust methods for AI-generated code identification, since the training and the test data splits have code in different languages and cover diverse usage domains. Subtask B focuses on defining synthetic code smells and requires participants to identify the provenance of the generator family of the model that generated the given code snippet. Subtask C aims at more fine-grained attribution of the written code: whether it was fully AI-generated, fully human-written, produced in human-AI collaboration (hybrid) or by a model tuned or prompted to give human-like code. The task attracted a large number of team members: subtask A (81), subtask B (34), and subtask C (32). In this study, we present the task, analyze the results and discuss the submissions of the system and the methods they used.

SemEval-2026 Task 13: Detecting Machine-Generated Code with Multiple Programming Languages, Generators, and Application Scenarios

We present SemEval-2026 Task 9, a shared task on online polarization detection, covering 22 languages and comprising over 110K annotated instances. Each data instance is multi-labeled with the presence of polarization, polarization type, and polarization manifestation. Participants were asked to predict labels in three subtasks: (1) detecting the presence of polarization, (2) identifying the type of polarization, and (3) recognizing the polarization manifestation. The three tasks attracted over 1,000 participants worldwide and more than 10k submissions on Codabench. We received final submissions from 67 teams and 69 system description papers. We report the baseline results and analyze the performance of the best-performing systems, highlighting the most common approaches and the most effective methods across different subtasks and languages. The dataset and other resources for this task are publicly available.

SemEval-2026 Task 9: Detecting Multilingual, Multicultural and Multievent Online Polarization

This paper presents our system developed for the SemEval-2026 Task 10: PsyCoMark Psycholinguistic Conspiracy Marker Extraction and Detection. on Subtask 1: Conspiracy Marker Extraction. on Subtask 2: Conspiracy Detection. To this end, we focus on English language use four different pre-trained languages models: models–distilbert–distilbert-base uncased, models–distilbert–distilbert-base-multilingual-cased, models–lxyuan–distilbert-base-multilingual-cased-sentiments-student, and models–microsoft–deberta-v3-base. We experiment with 1) the training set data is analyzed visually, 2) use the gemma-3-27b-it generative model to perform data augmentation on the training dataset through prompts for Subtask 2: Conspiracy Detection, and 3) multiple numbers of single models are trained on the training set data. We further study the influence of different hyperparameters on the single model and select the best single model for the prediction of the test set. Our submission achieved the good ranking place in the test set leaderboard. For Subtask 1, the evaluation criteria for this task mainly consist of the aggregate results of the four markers: Actor, Action, Effect, and Victim, and they are measured using the Macro F1 score. For Subtask 2, this task is essentially a binary classification task for text. Performance will be evaluated using macro-averaged F1 score. In other words, this subtask evaluated using Weighted F1 score across different sentences and cultural contexts. For Subtask 1 and Subtask 2, our best approach is to obtain the results are Macro F1 score 0.1587 and Weighted F1 score 0.7411 separately. For the final ranking, organizers will use the aggregate results of Macro F1 score and Weighted F1 score. Even so, our approach has yielded good results.

wangkongqiang at SemEval-2026 Task 10: PsyCoMark- Psycholinguistic Conspiracy Marker Extraction and Detection

This paper describes the system developed by team for SemEval-2026 Task 2, Subtask 1: Longitudinal Affect Assessment. Our goal is to predict Valence and Arousal from ecological essays and feeling words over time. We propose an efficient hybrid framework that uses quantized 7B-scale language models as deterministic meta-feature extractors and combines them with an ensemble of DeBERTa, RoBERTa, and DistilBERT encoders. The system is designed to run on a single consumer-grade RTX 5060 Ti (16GB) GPU while remaining competitive. To bridge discrete supervision and continuous evaluation, we train the model as an ordinal classification problem and decode class probabilities into continuous scores through expected-value decoding. Our best system achieved an overall V&A average of 0.587, with per-dimension composite correlations of 0.647 for Valence and 0.527 for Arousal, ranking 3rd out of 31 teams. The results show that lightweight SLM-derived priors and multi-encoder fusion provide a strong performance--efficiency trade-off, especially for Arousal, where contextual anchoring is crucial.

cclin at SemEval-2026 Task 2 : SLM-Enhanced Lightweight Multi-BERT Ensemble for Longitudinal Affect Assessment

Multi-turn retrieval-augmented generation (RAG) is challenging due to evolving user intent, conversational noise, and strict context limits. We propose a training-free hybrid retrieval pipeline for SemEval-2026 Task 8 that combines dense and sparse retrieval with controlled query rewriting and cross-encoder reranking. Our system achieves 0.5453 nDCG@5 on the official test set of Task A, ranking 3rd out of 38 teams and outperforming the strongest baseline (0.4795). For Task C, we reuse the Task A retrieved documents in a lightweight generation pipeline based on the official prompt, achieving 0.5312 (harmonic mean of quality and faithfulness) and ranking 15th out of 29 teams. All retrieval components are open-source, while rewriting and generation use LLM APIs. Code and scripts are available on GitHub (https://github.com/mengsifei/Multi_turn_RAG).

Sifei at SemEval-2026 Task 8: Hybrid Retrieval and Query Rewriting for Multi-Turn RAG

We describe our research to SemEval-2026 Task 4 on Narrative Story Similarity and Narrative Representation Learning (NSNRL). The shared task defines narrative similarity through comparative judgments over triples consisting of an anchor story and two candidates, where systems determine which candidate is narratively closer (Track A), and must output story embeddings whose cosine distances reproduce the same ordering under withheld evaluation triples (Track B). We implement a unified representation-learning approach based on a Sentence-BERT bi-encoder trained with triple-derived metric learning objectives, combining in-batch contrastive learning with explicit triplet and margin-ranking constraints. Track A is solved by direct cosine comparison between the anchor embedding and each candidate embedding, while Track B outputs normalized story vectors from the same encoder without any additional test-time modelling. During evaluation, we achieve 65.00% accuracy on Track A and 65.50% on Track B. These results suggest that a single, well-aligned bi-encoder can perform competitively across both tracks while remaining computationally efficient.

Downloads

Next from ACL 2026

ICI Innolabs at SemEval-2026 Task 13: Sliding Windows Meet Code Transformers

Similar lecture

SemEval-2026 Task 13: Detecting Machine-Generated Code with Multiple Programming Languages, Generators, and Application Scenarios

.css-70qvj9{display:-webkit-box;display:-webkit-flex;display:-ms-flexbox;display:flex;-webkit-align-items:center;-webkit-box-align:center;-ms-flex-align:center;align-items:center;}Downloads

Next from ACL 2026

ICI Innolabs at SemEval-2026 Task 13: Sliding Windows Meet Code Transformers

Similar lecture

SemEval-2026 Task 13: Detecting Machine-Generated Code with Multiple Programming Languages, Generators, and Application Scenarios

Downloads