{"paper":{"title":"Self-Distilled Trajectory-Aware Boltzmann Modeling: Bridging the Training-Inference Discrepancy in Diffusion Language Models","license":"http://creativecommons.org/licenses/by/4.0/","headline":"A new optimization framework for diffusion language models uses self-distilled inference trajectories and Boltzmann modeling of entropies to close the gap with standard supervised fine-tuning.","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Dandan Tu, Haoliang Li, Hui Liu, Kecheng Chen, Lingpeng Kong, Rui Liu, Shi Wu, Suiyun Zhang, Xijia Tao, Xinyu Fu, Yibing Liu, Ziru Liu","submitted_at":"2026-05-12T09:39:06Z","abstract_excerpt":"Diffusion Language Models (DLMs) have recently emerged as a promising alternative to autoregressive language models, offering stronger global awareness and highly parallel generation. However, post-training DLMs with standard Negative Evidence Lower Bound (NELBO)-based supervised fine-tuning remains inefficient: training reconstructs randomly masked tokens in a single step, whereas inference follows a confidence-guided, multi-step easy-to-hard denoising trajectory. Recent trajectory-based self-distillation methods exploit such inference trajectories mainly for sampling-step compression and acc"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"TABOM achieves substantial gains in new domains, expands the effective knowledge boundary of DLMs, and significantly mitigates catastrophic forgetting compared with standard SFT.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That modeling the inference unmasking preference as a Boltzmann distribution over predictive entropies and deriving a pairwise ranking objective from it will produce genuine knowledge acquisition rather than marginal or illusory gains.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"TABOM models inference unmasking preferences as a Boltzmann distribution over predictive entropies and derives a ranking loss to align DLM training with observed trajectories, yielding gains in new domains and reduced catastrophic forgetting versus standard SFT.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"A new optimization framework for diffusion language models uses self-distilled inference trajectories and Boltzmann modeling of entropies to close the gap with standard supervised fine-tuning.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"39504c315db187e1d938a3635f6991d019c81cf91b39a2672696461f8066250d"},"source":{"id":"2605.11854","kind":"arxiv","version":2},"verdict":{"id":"d6c29f0f-8efb-4c95-87c2-0163a8e682bc","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-13T05:48:46.216040Z","strongest_claim":"TABOM achieves substantial gains in new domains, expands the effective knowledge boundary of DLMs, and significantly mitigates catastrophic forgetting compared with standard SFT.","one_line_summary":"TABOM models inference unmasking preferences as a Boltzmann distribution over predictive entropies and derives a ranking loss to align DLM training with observed trajectories, yielding gains in new domains and reduced catastrophic forgetting versus standard SFT.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That modeling the inference unmasking preference as a Boltzmann distribution over predictive entropies and deriving a pairwise ranking objective from it will produce genuine knowledge acquisition rather than marginal or illusory gains.","pith_extraction_headline":"A new optimization framework for diffusion language models uses self-distilled inference trajectories and Boltzmann modeling of entropies to close the gap with standard supervised fine-tuning."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2605.11854/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"ai_meta_artifact","ran_at":"2026-05-19T11:36:39.823964Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_title_agreement","ran_at":"2026-05-19T09:01:18.338110Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_compliance","ran_at":"2026-05-19T08:08:14.829608Z","status":"completed","version":"1.0.0","findings_count":0}],"snapshot_sha256":"3e31ee0a46068f7e82e0ecf3caef4096db9c92f52d8a0fa93e0f2efbbc1ea3ad"},"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"}