publications

2025

1. arXiv

   SAE-RNA: A Sparse Autoencoder Model for Interpreting RNA Language Model Representations

   Taehan Kim and Sangdae Nam

   2025

   Abs Bib HTML

   We introduce SAE-RNA, a tool for interpreting RiNALMo RNA language model representations. It maps internal representations to known biological features and frames RNA interpretability as concept discovery without requiring end-to-end retraining. This approach enables researchers to examine what RNA language models encode about ncRNA families and supports hypothesis generation regarding previously unrecognized relationships between RNA groups.

   @article{kim2025saerna,
     title = {{SAE-RNA}: A Sparse Autoencoder Model for Interpreting {RNA} Language Model Representations},
     author = {Kim, Taehan and Nam, Sangdae},
     year = {2025},
     archiveprefix = {arXiv},
     primaryclass = {q-bio.BM},
   }

2024

1. arXiv

   Multi-Task Inference: Can Large Language Models Follow Multiple Instructions at Once?

   Guijin Son, Sangwon Baek, Sangdae Nam, and 2 more authors

   2024

   Abs Bib HTML

   Large Language Models can handle multiple instructions simultaneously in a single inference call. We introduce MTI Bench, a benchmark with 5,000 instances across 25 tasks, each containing 2-3 sub-tasks. Multi-task inference reduces latency by approximately 1.46x, and advanced models like Llama-2-Chat-70B and GPT-4 show improvements of up to 7.3% and 12.4% respectively over single-task processing.

   @article{son2024multitask,
     title = {Multi-Task Inference: Can Large Language Models Follow Multiple Instructions at Once?},
     author = {Son, Guijin and Baek, Sangwon and Nam, Sangdae and Jeong, Ilgyun and Kim, Seungone},
     year = {2024},
     archiveprefix = {arXiv},
     primaryclass = {cs.CL},
   }