CLaS Shared Task at SemEval 2027

We are happy to announce the CLaS: Cross-Lingual Alignment and Steering of Large Language Models Shared Task at SemEval 2027.

The CLaS shared task focuses on controlling the output language of large language models (LLMs) through inference-time internal interventions. Given a question in a source language, participants steer a model to generate a relevant and coherent answer in a specified target language.

We encourage participants to develop new approaches to language steering, improve existing methods, and combine complementary techniques. We are particularly interested in methods that achieve reliable language control while preserving the relevance and quality of model responses.

The task also provides an opportunity to investigate how language information is represented within multilingual LLMs and whether steering methods generalize across languages, including low-resource languages.

Participate in CLaS

We welcome researchers and teams interested in participating in the CLaS Shared Task at SemEval 2027.

Register your team Submissions open January 10, 2027

Registration is now available through our website. System submissions will open on January 10, 2027, at the beginning of the evaluation period. Detailed submission instructions will be published closer to that date.

Resources

  • Task description: Tracks, models, permitted methods, and evaluation.
  • Important dates: Evaluation and publication timeline.
  • GitHub repository: Code for obtaining training and validation data, running baseline generations, and evaluating results.