{"paper":{"title":"Demystifying Mergeability: Interpretable Properties to Predict Model Merging Success","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Gradient alignment between fine-tuned models is the strongest predictor of successful merging across methods and tasks.","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Bo Zhao, Emanuele Rodol\\`a, Luca Zhou, Rose Yu","submitted_at":"2026-01-29T20:00:26Z","abstract_excerpt":"Model merging combines knowledge from separately fine-tuned models, yet the factors driving its success remain poorly understood. While recent work treats mergeability as an intrinsic property of the models, we show with an architecture-agnostic framework that it fundamentally depends on both the merging method and the partner tasks. Using L1-regularized linear optimization over a set of interpretable pairwise metrics (e.g., gradient L_2 distance), we uncover properties correlating with post-merge normalized accuracy across five merging methods. We find architecture- and method-specific variat"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Crucially, however, gradient alignment metrics consistently emerge as the most fundamental signals of compatibility.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the chosen set of interpretable pairwise metrics (such as gradient L2 distance) plus L1-regularized linear optimization is sufficient to identify the true drivers of merge success without missing key unmeasured factors.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Mergeability is not intrinsic to models but depends on merging method and tasks, with gradient alignment as the key predictive signal uncovered via interpretable metrics.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Gradient alignment between fine-tuned models is the strongest predictor of successful merging across methods and tasks.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"00cb05c533c0e179389859adb011782384da3ab79ca3fe96afc3fd518f657d7d"},"source":{"id":"2601.22285","kind":"arxiv","version":7},"verdict":{"id":"dbf9c059-828e-416f-a712-6c617498264b","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-16T09:47:05.919474Z","strongest_claim":"Crucially, however, gradient alignment metrics consistently emerge as the most fundamental signals of compatibility.","one_line_summary":"Mergeability is not intrinsic to models but depends on merging method and tasks, with gradient alignment as the key predictive signal uncovered via interpretable metrics.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the chosen set of interpretable pairwise metrics (such as gradient L2 distance) plus L1-regularized linear optimization is sufficient to identify the true drivers of merge success without missing key unmeasured factors.","pith_extraction_headline":"Gradient alignment between fine-tuned models is the strongest predictor of successful merging across methods and tasks."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2601.22285/integrity.json","findings":[],"available":true,"detectors_run":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}