Tanel Alumäe

Reinforcement learning against a concept-matching metric can enhance factual reliability in summarizing complex medical conversations without sacrificing performance.

Grounding LALM reasoning in diverse, reliability-weighted acoustic evidence blows away the competition in Audio Question Answering, proving that verifiable chains beat black boxes.

Continued pretraining with a carefully balanced data mixture can substantially improve a multilingual LLM's performance on a low-resource language like Estonian without compromising its English capabilities.