{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:4EFBXV5QKAG6FJKN4MYSUZTYUI","short_pith_number":"pith:4EFBXV5Q","schema_version":"1.0","canonical_sha256":"e10a1bd7b0500de2a54de3312a6678a22032d124fd31dd4a1c8825d904fbab84","source":{"kind":"arxiv","id":"2509.21474","version":4},"attestation_state":"computed","paper":{"title":"d2: Improving Reasoning in Diffusion Language Models via Trajectory Likelihood Estimation","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Gilad Turok, Guanghan Wang, Marianne Arriola, Volodymyr Kuleshov, Yair Schiff","submitted_at":"2025-09-25T19:40:36Z","abstract_excerpt":"While diffusion language models (DLMs) have achieved competitive performance in text generation, improving their reasoning ability with reinforcement learning remains an active research area. Here, we introduce d2, a reasoning framework tailored for masked DLMs. Central to our framework is a new policy gradient algorithm that relies on accurate estimates of the sampling trajectory likelihoods. Because computing these likelihoods naively is computationally expensive for masked DLMs, we develop a family of estimators tailored to distinct model classes. For DLMs that support a sampling algorithm "},"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":false,"formal_links_present":false},"canonical_record":{"source":{"id":"2509.21474","kind":"arxiv","version":4},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2025-09-25T19:40:36Z","cross_cats_sorted":[],"title_canon_sha256":"e462750fe38134c39ae25fd21966e1e44481aca2a0389a24997f434b550e7dcb","abstract_canon_sha256":"11889dd113b5bf57af7664ad9732b6293265b27677d669440d78a09010c7f1ec"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-06-02T01:03:33.832496Z","signature_b64":"oCAui1DsSNPW2+6bF2CQro3VswAg5/PEzuoGB0HWkaW5YtqrDBx2+nuyk8EJUXHjV9U5cOP4ZbUD5dSlBRXJDg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"e10a1bd7b0500de2a54de3312a6678a22032d124fd31dd4a1c8825d904fbab84","last_reissued_at":"2026-06-02T01:03:33.831928Z","signature_status":"signed_v1","first_computed_at":"2026-06-02T01:03:33.831928Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"d2: Improving Reasoning in Diffusion Language Models via Trajectory Likelihood Estimation","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Gilad Turok, Guanghan Wang, Marianne Arriola, Volodymyr Kuleshov, Yair Schiff","submitted_at":"2025-09-25T19:40:36Z","abstract_excerpt":"While diffusion language models (DLMs) have achieved competitive performance in text generation, improving their reasoning ability with reinforcement learning remains an active research area. Here, we introduce d2, a reasoning framework tailored for masked DLMs. Central to our framework is a new policy gradient algorithm that relies on accurate estimates of the sampling trajectory likelihoods. Because computing these likelihoods naively is computationally expensive for masked DLMs, we develop a family of estimators tailored to distinct model classes. For DLMs that support a sampling algorithm "},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2509.21474","kind":"arxiv","version":4},"verdict":{"id":null,"model_set":{},"created_at":null,"strongest_claim":"","one_line_summary":"","pipeline_version":null,"weakest_assumption":"","pith_extraction_headline":""},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2509.21474/integrity.json","findings":[],"available":true,"detectors_run":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938"},"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"},"aliases":[{"alias_kind":"arxiv","alias_value":"2509.21474","created_at":"2026-06-02T01:03:33.832003+00:00"},{"alias_kind":"arxiv_version","alias_value":"2509.21474v4","created_at":"2026-06-02T01:03:33.832003+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2509.21474","created_at":"2026-06-02T01:03:33.832003+00:00"},{"alias_kind":"pith_short_12","alias_value":"4EFBXV5QKAG6","created_at":"2026-06-02T01:03:33.832003+00:00"},{"alias_kind":"pith_short_16","alias_value":"4EFBXV5QKAG6FJKN","created_at":"2026-06-02T01:03:33.832003+00:00"},{"alias_kind":"pith_short_8","alias_value":"4EFBXV5Q","created_at":"2026-06-02T01:03:33.832003+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":8,"internal_anchor_count":8,"sample":[{"citing_arxiv_id":"2605.22939","citing_title":"Learnability-Informed Fine-Tuning of Diffusion Language Models","ref_index":13,"is_internal_anchor":true},{"citing_arxiv_id":"2605.11726","citing_title":"Block-R1: Rethinking the Role of Block Size in Multi-domain Reinforcement Learning for Diffusion Large Language Models","ref_index":61,"is_internal_anchor":true},{"citing_arxiv_id":"2605.11726","citing_title":"Block-R1: Rethinking the Role of Block Size in Multi-domain Reinforcement Learning for Diffusion Large Language Models","ref_index":61,"is_internal_anchor":true},{"citing_arxiv_id":"2605.04647","citing_title":"ReflectDrive-2: Reinforcement-Learning-Aligned Self-Editing for Discrete Diffusion Driving","ref_index":123,"is_internal_anchor":true},{"citing_arxiv_id":"2605.09291","citing_title":"dFlowGRPO: Rate-Aware Policy Optimization for Discrete Flow Models","ref_index":131,"is_internal_anchor":true},{"citing_arxiv_id":"2605.04647","citing_title":"ReflectDrive-2: Reinforcement-Learning-Aligned Self-Editing for Discrete Diffusion Driving","ref_index":123,"is_internal_anchor":true},{"citing_arxiv_id":"2604.18739","citing_title":"Discrete Tilt Matching","ref_index":25,"is_internal_anchor":true},{"citing_arxiv_id":"2605.02263","citing_title":"Break the Block: Dynamic-size Reasoning Blocks for Diffusion Large Language Models via Monotonic Entropy Descent with Reinforcement Learning","ref_index":16,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/4EFBXV5QKAG6FJKN4MYSUZTYUI","json":"https://pith.science/pith/4EFBXV5QKAG6FJKN4MYSUZTYUI.json","graph_json":"https://pith.science/api/pith-number/4EFBXV5QKAG6FJKN4MYSUZTYUI/graph.json","events_json":"https://pith.science/api/pith-number/4EFBXV5QKAG6FJKN4MYSUZTYUI/events.json","paper":"https://pith.science/paper/4EFBXV5Q"},"agent_actions":{"view_html":"https://pith.science/pith/4EFBXV5QKAG6FJKN4MYSUZTYUI","download_json":"https://pith.science/pith/4EFBXV5QKAG6FJKN4MYSUZTYUI.json","view_paper":"https://pith.science/paper/4EFBXV5Q","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2509.21474&json=true","fetch_graph":"https://pith.science/api/pith-number/4EFBXV5QKAG6FJKN4MYSUZTYUI/graph.json","fetch_events":"https://pith.science/api/pith-number/4EFBXV5QKAG6FJKN4MYSUZTYUI/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/4EFBXV5QKAG6FJKN4MYSUZTYUI/action/timestamp_anchor","attest_storage":"https://pith.science/pith/4EFBXV5QKAG6FJKN4MYSUZTYUI/action/storage_attestation","attest_author":"https://pith.science/pith/4EFBXV5QKAG6FJKN4MYSUZTYUI/action/author_attestation","sign_citation":"https://pith.science/pith/4EFBXV5QKAG6FJKN4MYSUZTYUI/action/citation_signature","submit_replication":"https://pith.science/pith/4EFBXV5QKAG6FJKN4MYSUZTYUI/action/replication_record"}},"created_at":"2026-06-02T01:03:33.832003+00:00","updated_at":"2026-06-02T01:03:33.832003+00:00"}