Seven papers accepted at EMNLP 2026!

Our team will present several papers at EMNLP 2026:

• Energy-Aware Benchmarking of Generative Information Extraction
  We measure the energy needed to extract structured information from text and how system settings affect accuracy and energy use.

• SciEmbed
  We train a model to represent scientific papers using the sentences researchers write when citing them.

• Quantifying the Impact of Translation Errors on Multilingual LLM Evaluation
  We examine how well translated benchmarks capture language skills, knowledge, and local context.

• Vital Trace
  We combine structured patient memory with clinical rules to track changing conditions and predict risks over time, evaluated on intensive-care records.

• Does Modern Standard Arabic (MSA) Dominate Arabic Dialects in LLMs?
  We investigate whether models’ preference for standard Arabic is also reflected in their internal representations.

• Semantic Calibration Prevails Where Token Confidence Fails
  Across 685,000 responses, we find that consistency in meaning is more informative about the reliability of scientific answers than confidence in individual words.

• Analyzing Bias in False Refusal Behavior of Large Language Models for Hate Speech Detoxification
  We study how false refusals depend on language and the groups mentioned when models are asked to detoxify hateful text.

We look forward to sharing our work and connecting with the research community at EMNLP 2026!