{"paper":{"title":"ArcVQ-VAE: A Spherical Vector Quantization Framework with ArcCosine Additive Margin","license":"http://creativecommons.org/licenses/by/4.0/","headline":"ArcVQ-VAE adds a spherical angular-margin prior to VQ-VAE codebooks to increase utilization and dispersion.","cross_cats":["cs.AI","cs.LG"],"primary_cat":"cs.CV","authors_text":"Jaeyung Kim, Youngjoon Yoo","submitted_at":"2026-05-13T13:35:16Z","abstract_excerpt":"Vector Quantized Variational Autoencoder (VQ-VAE) has become a fundamental framework for learning discrete representations in image modeling. However, VQ-VAE models must tokenize entire images using a finite set of codebook vectors, and this capacity limitation restricts their ability to capture rich and diverse representations. In this paper, we propose ArcCosine Additive Margin VQ-VAE (ArcVQ-VAE), a novel vector quantization framework that introduces a spherical angular-margin prior (SAMP) for the codebook of a conventional VQ-VAE. The proposed SAMP consists of Ball-Bounded Norm Regularizati"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"The proposed SAMP consists of Ball-Bounded Norm Regularization... and ArcCosine Additive Margin Loss... This formulation promotes more discriminative and uniformly dispersed latent representations within the constrained space, thereby improving effective latent-space coverage and leading to improved codebook utilization.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the combination of time-dependent ball constraint and arc-cosine margin will reliably increase angular separability and utilization without harming training stability or reconstruction quality on the target datasets.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"ArcVQ-VAE constrains VQ-VAE codebook vectors inside a time-dependent ball and adds angular margin loss to increase separability and codebook utilization.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"ArcVQ-VAE adds a spherical angular-margin prior to VQ-VAE codebooks to increase utilization and dispersion.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"ca5047cb16a09534a4b8dec77d0a8b5f1beb6c5ce2965d20dd2a54b391bb53f8"},"source":{"id":"2605.13517","kind":"arxiv","version":1},"verdict":{"id":"156b36ee-cfdc-46d9-967c-285d79d6a76b","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-14T20:05:48.915274Z","strongest_claim":"The proposed SAMP consists of Ball-Bounded Norm Regularization... and ArcCosine Additive Margin Loss... This formulation promotes more discriminative and uniformly dispersed latent representations within the constrained space, thereby improving effective latent-space coverage and leading to improved codebook utilization.","one_line_summary":"ArcVQ-VAE constrains VQ-VAE codebook vectors inside a time-dependent ball and adds angular margin loss to increase separability and codebook utilization.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the combination of time-dependent ball constraint and arc-cosine margin will reliably increase angular separability and utilization without harming training stability or reconstruction quality on the target datasets.","pith_extraction_headline":"ArcVQ-VAE adds a spherical angular-margin prior to VQ-VAE codebooks to increase utilization and dispersion."},"references":{"count":15,"sample":[{"doi":"","year":null,"title":"Estimating or Propagating Gradients Through Stochastic Neurons for Conditional Computation","work_id":"1fe8c7c8-aff7-4b94-9096-e549d7e60789","ref_index":1,"cited_arxiv_id":"1308.3432","is_internal_anchor":true},{"doi":"","year":null,"title":"Hyper- spherical variational auto-encoders","work_id":"7164f439-c1a9-40af-8fd3-1df0d25bd03f","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2026,"title":"Fast decoding in se- quence models using discrete latent variables","work_id":"02e45462-cbbb-4c93-b2f8-83b5a42e882c","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2000,"title":"Crafting papers on machine learning","work_id":"d344ba9d-7725-491b-9cdd-eba5d0253623","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Unitok: A unified tokenizer for visual generation and understanding","work_id":"31d5c278-398d-4fee-8130-a864fb6b3717","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":15,"snapshot_sha256":"0e0eb4a237620c1ee7d3686bdf37a2c29fc4c0b4faf5f421aab50591aebe7482","internal_anchors":2},"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"}