{"paper":{"title":"Discrete Stochastic Localization for Non-autoregressive Generation","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Discrete Stochastic Localization makes one network handle any per-token noise path for sequence generation.","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Evangelos E. Papalexakis, Greg Ver Steeg, Jiayi Cheng, Longxuan Yu, Partha Thakuria, Rob Brekelmans, Yunshu Wu","submitted_at":"2026-05-13T00:12:24Z","abstract_excerpt":"Continuous diffusion is a natural framework for non-autoregressive generation but has generally lagged behind masked discrete diffusion models (MDMs) on discrete sequence generation. We argue that the bottleneck is not continuity itself, but a representation in which denoising depends on timestep-indexed noise regimes. We introduce \\emph{Discrete Stochastic Localization} (DSL), a continuous-state framework with unit-sphere token embeddings whose Bayes-optimal denoiser is invariant to the nominal signal-to-noise ratio (SNR) under the localization channel. One trained network then supports an en"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"One trained network then supports an entire family of per-token SNR paths, with endpoint masked-diffusion paths as a special case. Fine-tuning a pretrained MDLM checkpoint with DSL substantially improves distributional faithfulness (MAUVE) on OpenWebText across all step budgets from T=128 to T=1024.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The Bayes-optimal denoiser is invariant to the nominal signal-to-noise ratio under the localization channel when using unit-sphere token embeddings; this invariance is presented as enabling the single-network property but its validity depends on the specific channel definition.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Discrete Stochastic Localization provides a continuous-state framework with SNR-invariant denoisers on unit-sphere embeddings, enabling one network to support multiple per-token noise paths and improving MAUVE on OpenWebText.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Discrete Stochastic Localization makes one network handle any per-token noise path for sequence generation.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"10729db5cff6818924a6b4002a7a20884ebef865cc28f4581ea58ba497816d62"},"source":{"id":"2605.12836","kind":"arxiv","version":1},"verdict":{"id":"f955fe31-21a5-49a6-a0d6-115335a495a3","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-14T20:41:54.247134Z","strongest_claim":"One trained network then supports an entire family of per-token SNR paths, with endpoint masked-diffusion paths as a special case. Fine-tuning a pretrained MDLM checkpoint with DSL substantially improves distributional faithfulness (MAUVE) on OpenWebText across all step budgets from T=128 to T=1024.","one_line_summary":"Discrete Stochastic Localization provides a continuous-state framework with SNR-invariant denoisers on unit-sphere embeddings, enabling one network to support multiple per-token noise paths and improving MAUVE on OpenWebText.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The Bayes-optimal denoiser is invariant to the nominal signal-to-noise ratio under the localization channel when using unit-sphere token embeddings; this invariance is presented as enabling the single-network property but its validity depends on the specific channel definition.","pith_extraction_headline":"Discrete Stochastic Localization makes one network handle any per-token noise path for sequence generation."},"references":{"count":34,"sample":[{"doi":"","year":2021,"title":"Structured denoising diffusion models in discrete state-spaces.Advances in Neural Information Processing Systems, 34:17981–17993, 2021","work_id":"01cd60b5-f8aa-49d7-83d5-799c939c4e5b","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2022,"title":"Continuous diffusion for categor- ical data","work_id":"c0904b65-a618-46bd-85ef-53635f43ea5c","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"Likelihood-based diffusion language models","work_id":"7d177713-ac55-4588-a16c-319c11dc3492","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2005,"title":"Mutual information and minimum mean- square error in gaussian channels.IEEE transactions on information theory, 51(4):1261–1282, 2005","work_id":"d56fa707-beb6-4734-b8af-0d70488578e8","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2020,"title":"Denoising diffusion probabilistic models.Advances in Neural Information Processing Systems, 33:6840–6851","work_id":"b707e044-45b2-45ad-a843-aed07db67eff","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":34,"snapshot_sha256":"4f61f814b2974b5c57081b51ca471f3afbb7322fb4408cfde8f180a4ecf49625","internal_anchors":6},"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"}