Results

Leaderboard

KROMA's F1 on six tracks, ranked against the best published baseline, its own ablations, and smaller or cheaper models.

KROMA has the highest F1 on all six tracks in the paper. The board also shows what each part of it contributes: removing knowledge retrieval or ontology refinement lowers F1 on every track reported, and with a small model it still beats the best baseline on some tracks.

How to read it

• F1 (%) is the alignment's F1 against the reference matches, as published (Table 3 of the paper).
• KROMA rows use Llama-3.3-70B unless a tag says otherwise. Ablation rows remove retrieval or refinement. Baseline is the best of MILA, OLaLa, and LLM4OM on that track, with their reported models.
• These are published numbers, not a run of this site. The paper has the protocol.