{"paper":{"title":"Efficient Adjoint Matching for Fine-tuning Diffusion Models","license":"http://creativecommons.org/licenses/by/4.0/","headline":"By reformulating the stochastic optimal control problem with linear base drift, Efficient Adjoint Matching speeds up reward fine-tuning of diffusion models up to 4x while matching performance.","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Dongsoo Shin, Jaemoo Choi, Jaewoong Choi, Jeongwoo Shin, Joonseok Lee, Wei Guo, Yongxin Chen, Yuchen Zhu","submitted_at":"2026-05-12T03:55:12Z","abstract_excerpt":"Reward fine-tuning has become a common approach for aligning pretrained diffusion and flow models with human preferences in text-to-image generation. Among reward-gradient-based methods, Adjoint Matching (AM) provides a principled formulation by casting reward fine-tuning as a stochastic optimal control (SOC) problem. However, AM inevitably requires a substantial computational cost: it requires (i) stochastic simulation of full generative trajectories under memoryless dynamics, resulting in a large number of function evaluations, and (ii) backward ODE simulation of the adjoint state along each"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"On standard text-to-image reward fine-tuning benchmarks, EAM converges up to 4x faster than AM and matches or surpasses it across various metrics including PickScore, ImageReward, HPSv2.1, CLIPScore and Aesthetics.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That reformulating the SOC problem with linear base drift and modified terminal cost preserves the original alignment objective and solution quality without introducing bias or loss of expressivity.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"EAM speeds up adjoint matching for diffusion model reward fine-tuning by switching to linear base drift, allowing deterministic few-step solvers and closed-form adjoints with up to 4x faster convergence on text-to-image benchmarks.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"By reformulating the stochastic optimal control problem with linear base drift, Efficient Adjoint Matching speeds up reward fine-tuning of diffusion models up to 4x while matching performance.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"961c7d085c483c52d90646cb223e319207fdb86ce2ae1cd1ed15150b6693adca"},"source":{"id":"2605.11480","kind":"arxiv","version":2},"verdict":{"id":"c0d06026-e430-4267-9e7d-ca052735b6cd","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-13T01:48:08.979614Z","strongest_claim":"On standard text-to-image reward fine-tuning benchmarks, EAM converges up to 4x faster than AM and matches or surpasses it across various metrics including PickScore, ImageReward, HPSv2.1, CLIPScore and Aesthetics.","one_line_summary":"EAM speeds up adjoint matching for diffusion model reward fine-tuning by switching to linear base drift, allowing deterministic few-step solvers and closed-form adjoints with up to 4x faster convergence on text-to-image benchmarks.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That reformulating the SOC problem with linear base drift and modified terminal cost preserves the original alignment objective and solution quality without introducing bias or loss of expressivity.","pith_extraction_headline":"By reformulating the stochastic optimal control problem with linear base drift, Efficient Adjoint Matching speeds up reward fine-tuning of diffusion models up to 4x while matching performance."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2605.11480/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"ai_meta_artifact","ran_at":"2026-05-19T12:35:13.612761Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_title_agreement","ran_at":"2026-05-19T09:31:19.317994Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_compliance","ran_at":"2026-05-19T08:21:51.152392Z","status":"completed","version":"1.0.0","findings_count":0}],"snapshot_sha256":"d69ab342a34e23f0fa3d59dbf3f9abaaeb4205df9b42b33269cb10e1ad50b6d3"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":2,"snapshot_sha256":"8405a874ae7dfca82d3f39c9630be4e76245a017cfbd2d5282edded404646bb0"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}