{"paper":{"title":"MLPs are Efficient Distilled Generative Recommenders","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Distilling generative recommenders into MLPs preserves accuracy while speeding up inference by 8.74x","cross_cats":[],"primary_cat":"cs.IR","authors_text":"Clark Mingxuan Ju, Julian McAuley, Neil Shah, Yupeng Hou, Zitian Guo","submitted_at":"2026-05-12T18:05:55Z","abstract_excerpt":"Generative recommendation models employing Semantic IDs (SIDs) exhibit strong potential, yet their practical deployment is bottlenecked by the high inference latency of beam-expanded autoregressive decoding. In this work, we identify that standard attention-heavy Transformer decoders represent a structural overkill for this task: the hierarchical nature of SIDs makes prediction difficulty drops sharply after the first token, rendering repeated attention computations highly redundant. Driven by this insight, we propose SID-MLP, a lightweight MLP-centric distillation framework that fundamentally"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Extensive experiments demonstrate that SID-MLP matches the accuracy of teacher models while accelerating inference by 8.74x. This distillation strategy can serve as a plug-and-play accelerator for different backbones and tokenizer settings.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The hierarchical nature of SIDs makes prediction difficulty drop sharply after the first token, rendering repeated attention computations highly redundant.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"SID-MLP distills autoregressive generative recommenders into efficient position-specific MLP heads for Semantic ID tasks, achieving 8.74x faster inference with matching accuracy.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Distilling generative recommenders into MLPs preserves accuracy while speeding up inference by 8.74x","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"07bdc74cc6460efc6a67511efaa3b45150cdb3a996393ffa80da030f0551d571"},"source":{"id":"2605.12617","kind":"arxiv","version":1},"verdict":{"id":"c375d6d3-cc80-41db-bcf6-3aa50bdcf746","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-14T20:10:41.001869Z","strongest_claim":"Extensive experiments demonstrate that SID-MLP matches the accuracy of teacher models while accelerating inference by 8.74x. This distillation strategy can serve as a plug-and-play accelerator for different backbones and tokenizer settings.","one_line_summary":"SID-MLP distills autoregressive generative recommenders into efficient position-specific MLP heads for Semantic ID tasks, achieving 8.74x faster inference with matching accuracy.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The hierarchical nature of SIDs makes prediction difficulty drop sharply after the first token, rendering repeated attention computations highly redundant.","pith_extraction_headline":"Distilling generative recommenders into MLPs preserves accuracy while speeding up inference by 8.74x"},"references":{"count":77,"sample":[{"doi":"","year":2022,"title":"Transformer memory as a differentiable search index","work_id":"86233c16-b959-4114-9574-a43c62cdee35","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"Tran, Jonah Samost, Maciej Kula, Ed H","work_id":"ba5c0f0c-d332-440a-93c5-8606c39de2fc","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"Adapting large language models by integrating collaborative semantics for recommenda- tion","work_id":"f51fe37c-a7aa-419f-a196-bf60fb106255","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2025,"title":"OneRec: Unifying Retrieve and Rank with Generative Recommender and Iterative Preference Alignment","work_id":"d1a07d92-e045-4af2-a79f-c7b0112cf824","ref_index":4,"cited_arxiv_id":"2502.18965","is_internal_anchor":true},{"doi":"","year":2016,"title":"Session-based recommendations with recurrent neural networks","work_id":"52fdec4e-20e0-4895-a662-0d3deb4fbb2c","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":77,"snapshot_sha256":"95ab4ab1a560bd6719e2e877be70c7bae451ba16f2f3c5a52fad1cff53927393","internal_anchors":5},"formal_canon":{"evidence_count":2,"snapshot_sha256":"45ceee976e77c80dc467ccff50a920c0c08e893e6405a463602a79e5c62d41ea"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}