Results

First OAEI Bio-ML edition (superseded LargeBio). Unsupervised (90% test maps) and semi-supervised (70% test maps): equivalence by P/R/F + local ranking, subsumption by local ranking. Source: krr-oxford.github.io.

Published OAEI Bio-ML 2022 results, transcribed from the campaign result pages. Footnote marks (∗ † ‡) and the “Use Train(ing) Maps” flag are those of the original tables. Blank cells were not reported. Authoritative copy: krr-oxford.github.io.

Unsupervised (no train maps) and semi-supervised (optional 30% train maps). Subsumption is local-ranking; "Use Training Maps" (✔/✘) marks whether the 30% train maps were used. Source: krr-oxford.github.io.

Published OAEI Bio-ML 2023 results, transcribed from the campaign result pages. Footnote marks (∗ † ‡) and the “Use Train(ing) Maps” flag are those of the original tables. Blank cells were not reported. Authoritative copy: krr-oxford.github.io.

Unsupervised (no train maps) and semi-supervised (optional 30% train maps). Subsumption is local-ranking; "Use Training Maps" (✔/✘) marks whether the 30% train maps were used. Source: krr-oxford.github.io.

Published OAEI Bio-ML 2024 results, transcribed from the campaign result pages. Footnote marks (∗ † ‡) and the “Use Train(ing) Maps” flag are those of the original tables. Blank cells were not reported. Authoritative copy: krr-oxford.github.io.

Unsupervised (no train maps) and semi-supervised (optional 30% train maps). Equivalence matching only this edition. Source: liseda-lab.github.io.

Published OAEI Bio-ML 2025 results, transcribed from the campaign result pages. Footnote marks (∗ † ‡) and the “Use Train(ing) Maps” flag are those of the original tables. Blank cells were not reported. Authoritative copy: liseda-lab.github.io.

Track 1 · Subtracks 1 and 2

Submitted systems

MacroNCIT–DOIDSNOMED–FMASNOMED–NCIT

		Repaired			Ranking				Standard (unrepaired)
System	Use Train Maps	Precision	Recall	F1	MRR	H@1	H@5	H@10	Precision	Recall	F1
Matcha-ML	✔	0.699	0.632	0.636	0.930	0.906	0.958	0.970	0.722	0.529	0.570
MetaMatch	✔	0.607	0.667	0.634	0.327	0.324	0.330	0.332	0.661	0.580	0.615
LogMapLLM	✘	0.507	0.812	0.623					0.557	0.683	0.611
LogMapBio-LLM	✘	0.497	0.826	0.621					0.551	0.711	0.617
Matcha	✘	0.544	0.736	0.612					0.623	0.700	0.642
GenOM	✘	0.507	0.654	0.554					0.566	0.616	0.566
LogMap	✘	0.402	0.812	0.534					0.450	0.690	0.544
mosaic	✘	0.450	0.564	0.487					0.508	0.508	0.504
davidlopez	-	0.124	0.953	0.216					0.160	0.924	0.271
CEDAR	✔				0.962	0.949	0.978	0.983
NeSyMatch	✔				0.132	0.085	0.140	0.188

Repaired and Standard: global alignment, scored against the repaired (headline) and the standard reference. Ranking: local equivalence ranking. Macro: the unweighted mean over the three pairs. A blank cell means not submitted. See evaluation metrics.

Track 1 · Subtrack 1

Systems coherence

Each submitted alignment is also checked with a reasoner: does the system's own alignment leave the merged ontology satisfiable? Each cell gives the number of unsatisfiable named classes and, in parentheses, the incoherence degree (that count over the counted named classes). Lower is better; 0 is coherent. Reported per pair rather than macro-averaged, because 5 systems have a cell that could not be computed. Rows follow the table above.

System	NCIT–DOID		SNOMED–FMA	SNOMED–NCIT
	ELK	HermiT	ELK	ELK
Matcha-ML	0 (0%)	0 (0%)	90,104 (18.4%)	14,250 (2.41%)
MetaMatch	11,735 (5.33%)	11,825 (5.37%)	—	288,574 (48.7%)
LogMapLLM	0 (0%)	0 (0%)	11 (0.00224%)	426 (0.0719%)
LogMapBio-LLM	14 (0.00636%)	14 (0.00636%)	11 (0.00224%)	926 (0.156%)
Matcha	16,782 (7.62%)	16,874 (7.66%)	—	406,291 (68.6%)
GenOM	23,915 (10.9%)	23,933 (10.9%)	—	358,925 (60.6%)
LogMap	9 (0.00409%)	9 (0.00409%)	98,593 (20.1%)	173 (0.0292%)
mosaic	10,766 (4.89%)	10,857 (4.93%)	—	233,199 (39.4%)
davidlopez	83,915 (38.1%)	83,917 (38.1%)	—	—

Computed organiser-side on each system's whole submitted alignment M, not only the test slice, merged with both ontologies (O_src ∪ O_tgt ∪ M, each correspondence asserted as an equivalence). Every ELK figure is a sound lower bound; HermiT is exact but is run on NCIT–DOID only. — means ELK exhausted a 384 GB heap, so the value is unknown. The conventions are those of the baseline coherence table. Coherence is not a quality ranking and does not track F1.

Track 1 · Subtracks 1 and 2

Resource use and wall-clock time

ℹ We are waiting on further correspondence from participants before publishing this table.

Final results, frozen on 3 October 2026. All scores were computed by the organisers with the unmodified CodaBench scoring programs and reproduce the CodaBench leaderboards. Data: participant_results.json.