{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:2DDT7GQNFROWPUW73U6X7Y5INL","short_pith_number":"pith:2DDT7GQN","schema_version":"1.0","canonical_sha256":"d0c73f9a0d2c5d67d2dfdd3d7fe3a86af9d4b1b9f5f46ece921496f831f4370f","source":{"kind":"arxiv","id":"2605.13935","version":1},"attestation_state":"computed","paper":{"title":"Beyond Mode-Seeking RL: Trajectory-Balance Post-Training for Diffusion Language Models","license":"http://creativecommons.org/licenses/by/4.0/","headline":"A trajectory-balance objective stops diffusion language models from locking onto narrow denoising paths during post-training.","cross_cats":["cs.CL"],"primary_cat":"cs.LG","authors_text":"Prasanna Parthasarathi, Saba Ahmadi, Yufei Cui","submitted_at":"2026-05-13T16:14:46Z","abstract_excerpt":"Diffusion language models are a promising alternative to autoregressive models, yet post-training methods for them largely adapt reward-maximizing objectives. We identify a central failure mode in this setting we call trajectory locking: sampled reward-driven updates over-concentrate probability mass onto a narrow set of denoising paths, reducing coverage of alternative correct solutions under repeated sampling. To address this, we propose TraFL (Trajectory Flow baLancing), a trajectory-balance objective that trains the policy toward a reward-tilted target distribution anchored to a frozen ref"},"verification_status":{"content_addressed":true,"pith_receipt":true,"author_attested":false,"weak_author_claims":0,"strong_author_claims":0,"externally_anchored":false,"storage_verified":false,"citation_signatures":0,"replication_records":0,"graph_snapshot":true,"references_resolved":true,"formal_links_present":true},"canonical_record":{"source":{"id":"2605.13935","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2026-05-13T16:14:46Z","cross_cats_sorted":["cs.CL"],"title_canon_sha256":"0845ff7196936e93f23f3c82f8bbcc355bf9f08909a3858c46573b9b7ca8c249","abstract_canon_sha256":"62d9efd6af57211c0680900539c73ebdea02a5deb0ebf6f19c97d75a22120c32"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-17T23:39:13.931038Z","signature_b64":"H+N55xp2aOcaG27CNeuTRp0HH7u3MH61IYPTUYbAlZa2uaNLAdGWG2FXNKuMJHayx1o9eohWFsElOND6RYxhCA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"d0c73f9a0d2c5d67d2dfdd3d7fe3a86af9d4b1b9f5f46ece921496f831f4370f","last_reissued_at":"2026-05-17T23:39:13.930271Z","signature_status":"signed_v1","first_computed_at":"2026-05-17T23:39:13.930271Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Beyond Mode-Seeking RL: Trajectory-Balance Post-Training for Diffusion Language Models","license":"http://creativecommons.org/licenses/by/4.0/","headline":"A trajectory-balance objective stops diffusion language models from locking onto narrow denoising paths during post-training.","cross_cats":["cs.CL"],"primary_cat":"cs.LG","authors_text":"Prasanna Parthasarathi, Saba Ahmadi, Yufei Cui","submitted_at":"2026-05-13T16:14:46Z","abstract_excerpt":"Diffusion language models are a promising alternative to autoregressive models, yet post-training methods for them largely adapt reward-maximizing objectives. We identify a central failure mode in this setting we call trajectory locking: sampled reward-driven updates over-concentrate probability mass onto a narrow set of denoising paths, reducing coverage of alternative correct solutions under repeated sampling. To address this, we propose TraFL (Trajectory Flow baLancing), a trajectory-balance objective that trains the policy toward a reward-tilted target distribution anchored to a frozen ref"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"TraFL is the only evaluated post-training method that improves over the base model in every benchmark-length setting, with gains that persist as the sampling budget increases.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The diffusion-compatible sequence-level surrogate and learned prompt-dependent normalization faithfully approximate the trajectory-balance objective without introducing new collapse modes or requiring task-specific tuning.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"TraFL applies trajectory flow balancing to post-train diffusion language models, preventing mode collapse and delivering consistent gains on reasoning tasks that hold under increased sampling.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"A trajectory-balance objective stops diffusion language models from locking onto narrow denoising paths during post-training.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"8e66d9b26f83d2f5f8ad7422e648b699deaff32db4069de169cb99ac3e95abec"},"source":{"id":"2605.13935","kind":"arxiv","version":1},"verdict":{"id":"e6e37a21-437f-4a2f-83f4-c3f8b182175b","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T04:56:08.731453Z","strongest_claim":"TraFL is the only evaluated post-training method that improves over the base model in every benchmark-length setting, with gains that persist as the sampling budget increases.","one_line_summary":"TraFL applies trajectory flow balancing to post-train diffusion language models, preventing mode collapse and delivering consistent gains on reasoning tasks that hold under increased sampling.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The diffusion-compatible sequence-level surrogate and learned prompt-dependent normalization faithfully approximate the trajectory-balance objective without introducing new collapse modes or requiring task-specific tuning.","pith_extraction_headline":"A trajectory-balance objective stops diffusion language models from locking onto narrow denoising paths during post-training."},"references":{"count":32,"sample":[{"doi":"","year":1998,"title":"Natural gradient works efficiently in learning.Neural computation, 10(2):251– 276, 1998","work_id":"b9a8c272-145a-4fe5-930b-43b05b08fcd2","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2021,"title":"Program Synthesis with Large Language Models","work_id":"fd241a05-03b9-4de2-9588-9d77ce176125","ref_index":2,"cited_arxiv_id":"2108.07732","is_internal_anchor":true},{"doi":"","year":2003,"title":"Mirror descent and nonlinear projected subgradient methods for convex optimization.Operations Research Letters, 31(3):167–175","work_id":"d83325bc-fa70-4661-b21c-a8416937436b","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"Hu, Mo Tiwari, and Emmanuel Bengio","work_id":"185d0c88-9be3-4617-a532-40e368019cfc","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2021,"title":"Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, Alex Ray, Raul Puri, Gretchen Krueger, Mich","work_id":"fbf5883a-dc87-439f-a0ae-fcc99bbb1643","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":32,"snapshot_sha256":"8812a7d525a5777a87bb5b86aff9793e63b2b78593e41e614a6fdb8a2b9f3a7c","internal_anchors":7},"formal_canon":{"evidence_count":2,"snapshot_sha256":"75089fe02f257850b4c229472635656a8527d8e1c2889c4207a565211692b69d"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"},"aliases":[{"alias_kind":"arxiv","alias_value":"2605.13935","created_at":"2026-05-17T23:39:13.930413+00:00"},{"alias_kind":"arxiv_version","alias_value":"2605.13935v1","created_at":"2026-05-17T23:39:13.930413+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.13935","created_at":"2026-05-17T23:39:13.930413+00:00"},{"alias_kind":"pith_short_12","alias_value":"2DDT7GQNFROW","created_at":"2026-05-18T12:33:37.589309+00:00"},{"alias_kind":"pith_short_16","alias_value":"2DDT7GQNFROWPUW7","created_at":"2026-05-18T12:33:37.589309+00:00"},{"alias_kind":"pith_short_8","alias_value":"2DDT7GQN","created_at":"2026-05-18T12:33:37.589309+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":0,"internal_anchor_count":0,"sample":[]},"formal_canon":{"evidence_count":2,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/2DDT7GQNFROWPUW73U6X7Y5INL","json":"https://pith.science/pith/2DDT7GQNFROWPUW73U6X7Y5INL.json","graph_json":"https://pith.science/api/pith-number/2DDT7GQNFROWPUW73U6X7Y5INL/graph.json","events_json":"https://pith.science/api/pith-number/2DDT7GQNFROWPUW73U6X7Y5INL/events.json","paper":"https://pith.science/paper/2DDT7GQN"},"agent_actions":{"view_html":"https://pith.science/pith/2DDT7GQNFROWPUW73U6X7Y5INL","download_json":"https://pith.science/pith/2DDT7GQNFROWPUW73U6X7Y5INL.json","view_paper":"https://pith.science/paper/2DDT7GQN","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2605.13935&json=true","fetch_graph":"https://pith.science/api/pith-number/2DDT7GQNFROWPUW73U6X7Y5INL/graph.json","fetch_events":"https://pith.science/api/pith-number/2DDT7GQNFROWPUW73U6X7Y5INL/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/2DDT7GQNFROWPUW73U6X7Y5INL/action/timestamp_anchor","attest_storage":"https://pith.science/pith/2DDT7GQNFROWPUW73U6X7Y5INL/action/storage_attestation","attest_author":"https://pith.science/pith/2DDT7GQNFROWPUW73U6X7Y5INL/action/author_attestation","sign_citation":"https://pith.science/pith/2DDT7GQNFROWPUW73U6X7Y5INL/action/citation_signature","submit_replication":"https://pith.science/pith/2DDT7GQNFROWPUW73U6X7Y5INL/action/replication_record"}},"created_at":"2026-05-17T23:39:13.930413+00:00","updated_at":"2026-05-17T23:39:13.930413+00:00"}