Good news before the new semester: one paper has been accepted to the EMNLP 2026 Main Conference, first-authored by our Ph.D. student Hao An.
Teaching LLMs to Abstain via Fine-Grained Semantic Confidence Reward (FiSCoRe)
Standing out from last year’s fierce competition in RL fine-tuning for mitigating LLM hallucinations was not easy. Our core contribution — jointly fine-tuning the model with correctness and semantic-entropy uncertainty — remains of long-term value.
Arxiv: https://arxiv.org/abs/2510.24020
Wish everyone a pleasant trip to Budapest!