{"paper":{"title":"LoMETab: Beyond Rank-1 Ensembles for Tabular Deep Learning","license":"http://creativecommons.org/licenses/by/4.0/","headline":"LoMETab generalizes rank-1 multiplicative ensembles to rank-r adapters for tabular models, strictly enlarging the hypothesis class and exposing tunable diversity controls.","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Changryeol Choi, Gowun Jeong, Hyewon Park, Yujin Kwon","submitted_at":"2026-05-14T04:47:16Z","abstract_excerpt":"Recent tabular learning benchmarks increasingly show a tight performance cluster rather than a clear hierarchy among leading methods, spanning gradient boosted decision trees, attention-based architectures, and implicit ensembles such as TabM. As benchmark gains plateau, a complementary goal is to understand and control the mechanisms that make simple neural tabular models competitive. We propose LoMETab, a rank-$r$ generalization of multiplicative implicit ensembles. LoMETab lifts the rank-1 BatchEnsemble/TabM modulation to a rank-$r$ identity-residual Hadamard family by parameterizing each m"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"We prove that for r >= 2 this generalization strictly enlarges BatchEnsemble's hypothesis class. Empirically, LoMETab sustains higher pairwise KL than an additive low-rank ablation, and (r, sigma_init) provides broad control over pairwise KL, varying by up to several orders of magnitude.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the added representational capacity and induced diversity will translate into practically useful predictive behavior rather than merely dataset-dependent variation; the abstract reports that performance over the (r, sigma_init) grid is dataset-dependent.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"LoMETab is a rank-r generalization of multiplicative implicit ensembles that strictly enlarges the hypothesis class for r >= 2 and supplies tunable control over predictive diversity via adapter rank and initialization scale.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"LoMETab generalizes rank-1 multiplicative ensembles to rank-r adapters for tabular models, strictly enlarging the hypothesis class and exposing tunable diversity controls.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"5d5426ace85027eeaa4289c21f1c5e1837e669502f77a42df925d3fe7a222bb0"},"source":{"id":"2605.14365","kind":"arxiv","version":1},"verdict":{"id":"fb9c7d3c-d33a-4766-8274-0424ae153871","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T02:49:04.351626Z","strongest_claim":"We prove that for r >= 2 this generalization strictly enlarges BatchEnsemble's hypothesis class. Empirically, LoMETab sustains higher pairwise KL than an additive low-rank ablation, and (r, sigma_init) provides broad control over pairwise KL, varying by up to several orders of magnitude.","one_line_summary":"LoMETab is a rank-r generalization of multiplicative implicit ensembles that strictly enlarges the hypothesis class for r >= 2 and supplies tunable control over predictive diversity via adapter rank and initialization scale.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the added representational capacity and induced diversity will translate into practically useful predictive behavior rather than merely dataset-dependent variation; the abstract reports that performance over the (r, sigma_init) grid is dataset-dependent.","pith_extraction_headline":"LoMETab generalizes rank-1 multiplicative ensembles to rank-r adapters for tabular models, strictly enlarging the hypothesis class and exposing tunable diversity controls."},"references":{"count":23,"sample":[{"doi":"","year":2019,"title":"Optuna: A next-generation hyperparameter optimization framework","work_id":"f11f4cd0-9ce5-4174-a976-f401f5e9ce28","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2016,"title":"XGBoost: A scalable tree boosting system","work_id":"bea565cd-2b99-41c7-8085-02eba78c6f8f","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2021,"title":"Masksembles for uncertainty estimation","work_id":"c8d2a3d6-20d5-47ec-955a-24064d06b6d3","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2022,"title":"On embeddings for numerical features in tabular deep learning","work_id":"aaa16ab9-7e1e-4cca-903a-566c4b099f74","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"TabR: Tabular deep learning meets nearest neighbors","work_id":"063c1227-1e8a-4626-9d08-87eb984116d0","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":23,"snapshot_sha256":"1cb42c6ab57edb06565febc01d19bb8e642782881483b3ce025f6df37199a497","internal_anchors":0},"formal_canon":{"evidence_count":2,"snapshot_sha256":"ebde908e23afe8710f643e2ac4f28152414efdc0a3796f700efb496812ff706b"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}