{"paper":{"title":"Darwin Family: MRI-Trust-Weighted Evolutionary Merging for Training-Free Scaling of Language-Model Reasoning","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Evolutionary merging of existing language model checkpoints produces superior reasoning performance without any training.","cross_cats":["cs.AI"],"primary_cat":"cs.NE","authors_text":"Jaewon Jang, Junghoon Shin, Minseo Kim, Minsik Kim, Sunyoung Choi, Taebong Kim, Youngsik Hong","submitted_at":"2026-05-14T05:09:12Z","abstract_excerpt":"We present Darwin Family, a framework for training-free evolutionary merging of large language models via gradient-free weight-space recombination. We ask whether frontier-level reasoning performance can be improved without additional training, by reorganizing latent capabilities already encoded in existing checkpoints. Darwin introduces three key ideas: (i) a 14-dimensional adaptive merge genome enabling fine-grained component- and block-level recombination; (ii) MRI-Trust Fusion, which adaptively balances diagnostic layer-importance signals with evolutionary search through a learnable trust "},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"The flagship Darwin-27B-Opus achieves 86.9% on GPQA Diamond, ranking #6 among 1,252 evaluated models, and outperforming its fully trained foundation model without any gradient-based training.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the evolutionary recombination guided by the 14-dimensional merge genome and MRI-Trust Fusion can reliably reorganize latent reasoning capabilities already present in existing checkpoints without loss of coherence or introduction of new failure modes.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Evolutionary merging with a 14-dimensional genome and MRI-Trust Fusion produces models that outperform their trained parents on reasoning benchmarks without any gradient updates.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Evolutionary merging of existing language model checkpoints produces superior reasoning performance without any training.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"1af84823694f7a281e48c186708ad9561a9f66045af0732bb7e0fd0dc86cd152"},"source":{"id":"2605.14386","kind":"arxiv","version":1},"verdict":{"id":"8c66b9a3-a779-4915-a734-ece30b7ba831","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T02:09:01.351243Z","strongest_claim":"The flagship Darwin-27B-Opus achieves 86.9% on GPQA Diamond, ranking #6 among 1,252 evaluated models, and outperforming its fully trained foundation model without any gradient-based training.","one_line_summary":"Evolutionary merging with a 14-dimensional genome and MRI-Trust Fusion produces models that outperform their trained parents on reasoning benchmarks without any gradient updates.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the evolutionary recombination guided by the 14-dimensional merge genome and MRI-Trust Fusion can reliably reorganize latent reasoning capabilities already present in existing checkpoints without loss of coherence or introduction of new failure modes.","pith_extraction_headline":"Evolutionary merging of existing language model checkpoints produces superior reasoning performance without any training."},"references":{"count":31,"sample":[{"doi":"","year":2022,"title":"Chain-of-thought prompting elicits reason- ing in large language models","work_id":"59e16d6c-e56a-4398-a4a3-f1d086439c8d","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2022,"title":"Takeshi Kojima, Shixiang Gu, M. Reid, et al. Large language models are zero-shot reasoners. InNeural Information Processing Systems, 2022","work_id":"5c4ef5d3-b983-49db-86d1-2ef46e73e9d8","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"Self-consistency improves chain-of-thought reasoning in language models","work_id":"f088816e-1c6c-4ac4-8a4e-0b209da8ab9a","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"Least-to-most prompting enables complex rea- soning in large language models","work_id":"87a0882f-6eb8-47cf-aa5e-41759cd0d615","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2019,"title":"Bert rediscovers the classical nlp pipeline","work_id":"c9de2a65-5b4c-48bc-98de-2b89d1846e02","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":31,"snapshot_sha256":"26e9066885c73f9ad8cefc80e611f0979829f320fdeeb33805dd4f4b2726617a","internal_anchors":3},"formal_canon":{"evidence_count":2,"snapshot_sha256":"d02bb810824e18a774590e3eb05bec50099351a9e375f14abcd36dd734e52090"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}