{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:66MAX7UCOADDI55SGR34MQRY4Y","short_pith_number":"pith:66MAX7UC","schema_version":"1.0","canonical_sha256":"f7980bfe8270063477b23477c64238e6313e3fd975fafb9a7ea2091792815ac3","source":{"kind":"arxiv","id":"2603.07079","version":2},"attestation_state":"computed","paper":{"title":"Entropy-Aware On-Policy Distillation of Language Models","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.CL"],"primary_cat":"cs.LG","authors_text":"Dennis Wei, Kimin Lee, Nathalie Baracaldo, Swanand Ravindra Kadhe, Taywon Min, Woogyeol Jin, Yi Zhou, Yongjin Yang","submitted_at":"2026-03-07T07:26:18Z","abstract_excerpt":"On-policy distillation is a promising approach for transferring knowledge between language models, where a student learns from dense token-level signals along its own trajectories. This framework typically uses reverse KL divergence, encouraging the student to match the teacher's high-confidence predictions. However, we show that the mode-seeking property of reverse KL reduces generation diversity and yields unstable learning signals when the teacher distribution has high entropy. To address this, we introduce Entropy-Aware On-Policy Distillation. Our key idea is augmenting the standard revers"},"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":"2603.07079","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2026-03-07T07:26:18Z","cross_cats_sorted":["cs.CL"],"title_canon_sha256":"bea500c72439d5997bfeefa2e7fd622a291f63b95b313bd066f2cb989c47dca3","abstract_canon_sha256":"e5d920a4e7581dc8284f5fee8fc6b9ed7635296d959262634cd1250d0df490b7"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-25T02:02:14.034893Z","signature_b64":"UUfL25837omQVc8w/43ykNHzQKRJGYQ+A0uZC8NQqs1P1/KsdHkCIYjfFOxJHPgZEk3jTUjGLyp9wNNnmwhIBQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"f7980bfe8270063477b23477c64238e6313e3fd975fafb9a7ea2091792815ac3","last_reissued_at":"2026-05-25T02:02:14.033926Z","signature_status":"signed_v1","first_computed_at":"2026-05-25T02:02:14.033926Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Entropy-Aware On-Policy Distillation of Language Models","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.CL"],"primary_cat":"cs.LG","authors_text":"Dennis Wei, Kimin Lee, Nathalie Baracaldo, Swanand Ravindra Kadhe, Taywon Min, Woogyeol Jin, Yi Zhou, Yongjin Yang","submitted_at":"2026-03-07T07:26:18Z","abstract_excerpt":"On-policy distillation is a promising approach for transferring knowledge between language models, where a student learns from dense token-level signals along its own trajectories. This framework typically uses reverse KL divergence, encouraging the student to match the teacher's high-confidence predictions. However, we show that the mode-seeking property of reverse KL reduces generation diversity and yields unstable learning signals when the teacher distribution has high entropy. To address this, we introduce Entropy-Aware On-Policy Distillation. Our key idea is augmenting the standard revers"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2603.07079","kind":"arxiv","version":2},"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/2603.07079/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":"2603.07079","created_at":"2026-05-25T02:02:14.034066+00:00"},{"alias_kind":"arxiv_version","alias_value":"2603.07079v2","created_at":"2026-05-25T02:02:14.034066+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2603.07079","created_at":"2026-05-25T02:02:14.034066+00:00"},{"alias_kind":"pith_short_12","alias_value":"66MAX7UCOADD","created_at":"2026-05-25T02:02:14.034066+00:00"},{"alias_kind":"pith_short_16","alias_value":"66MAX7UCOADDI55S","created_at":"2026-05-25T02:02:14.034066+00:00"},{"alias_kind":"pith_short_8","alias_value":"66MAX7UC","created_at":"2026-05-25T02:02:14.034066+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":27,"internal_anchor_count":27,"sample":[{"citing_arxiv_id":"2605.07711","citing_title":"SimCT: Recovering Lost Supervision for Cross-Tokenizer On-Policy Distillation","ref_index":33,"is_internal_anchor":true},{"citing_arxiv_id":"2605.06597","citing_title":"UniSD: Towards a Unified Self-Distillation Framework for Large Language Models","ref_index":43,"is_internal_anchor":true},{"citing_arxiv_id":"2605.21606","citing_title":"When Are Teacher Tokens Reliable? Position-Weighted On-Policy Self-Distillation for Reasoning","ref_index":17,"is_internal_anchor":true},{"citing_arxiv_id":"2605.21834","citing_title":"On-Policy Consistency Training Improves LLM Safety with Minimal Capability Degradation","ref_index":15,"is_internal_anchor":true},{"citing_arxiv_id":"2605.22263","citing_title":"Tailoring Teaching to Aptitude: Direction-Adaptive Self-Distillation for LLM Reasoning","ref_index":25,"is_internal_anchor":true},{"citing_arxiv_id":"2605.21924","citing_title":"Visual-Advantage On-Policy Distillation for Vision-Language Models","ref_index":30,"is_internal_anchor":true},{"citing_arxiv_id":"2605.08063","citing_title":"Flow-OPD: On-Policy Distillation for Flow Matching Models","ref_index":38,"is_internal_anchor":true},{"citing_arxiv_id":"2605.17862","citing_title":"$\\boldsymbol{f}$-OPD: Stabilizing Long-Horizon On-Policy Distillation with Freshness-Aware Control","ref_index":28,"is_internal_anchor":true},{"citing_arxiv_id":"2605.16826","citing_title":"Decoupling KL and Trajectories: A Unified Perspective for SFT, DAgger, Offline RL, and OPD in LLM Distillation","ref_index":16,"is_internal_anchor":true},{"citing_arxiv_id":"2605.08063","citing_title":"Flow-OPD: On-Policy Distillation for Flow Matching Models","ref_index":38,"is_internal_anchor":true},{"citing_arxiv_id":"2605.12652","citing_title":"Multi-Rollout On-Policy Distillation via Peer Successes and Failures","ref_index":23,"is_internal_anchor":true},{"citing_arxiv_id":"2605.09725","citing_title":"On-Policy Distillation with Best-of-N Teacher Rollout Selection","ref_index":21,"is_internal_anchor":true},{"citing_arxiv_id":"2605.13643","citing_title":"Prefix Teach, Suffix Fade: Local Teachability Collapse in Strong-to-Weak On-Policy Distillation","ref_index":16,"is_internal_anchor":true},{"citing_arxiv_id":"2605.13255","citing_title":"Respecting Self-Uncertainty in On-Policy Self-Distillation for Efficient LLM Reasoning","ref_index":4,"is_internal_anchor":true},{"citing_arxiv_id":"2605.08063","citing_title":"Flow-OPD: On-Policy Distillation for Flow Matching Models","ref_index":38,"is_internal_anchor":true},{"citing_arxiv_id":"2605.09725","citing_title":"On-Policy Distillation with Best-of-N Teacher Rollout Selection","ref_index":21,"is_internal_anchor":true},{"citing_arxiv_id":"2605.08737","citing_title":"The Extrapolation Cliff in On-Policy Distillation of Near-Deterministic Structured Outputs","ref_index":19,"is_internal_anchor":true},{"citing_arxiv_id":"2605.03677","citing_title":"Uni-OPD: Unifying On-Policy Distillation with a Dual-Perspective Recipe","ref_index":18,"is_internal_anchor":true},{"citing_arxiv_id":"2604.24005","citing_title":"TCOD: Exploring Temporal Curriculum in On-Policy Distillation for Multi-turn Autonomous Agents","ref_index":4,"is_internal_anchor":true},{"citing_arxiv_id":"2605.06597","citing_title":"UniSD: Towards a Unified Self-Distillation Framework for Large Language Models","ref_index":42,"is_internal_anchor":true},{"citing_arxiv_id":"2605.01347","citing_title":"MAD-OPD: Breaking the Ceiling in On-Policy Distillation via Multi-Agent Debate","ref_index":20,"is_internal_anchor":true},{"citing_arxiv_id":"2604.13016","citing_title":"Rethinking On-Policy Distillation of Large Language Models: Phenomenology, Mechanism, and Recipe","ref_index":12,"is_internal_anchor":true},{"citing_arxiv_id":"2605.08063","citing_title":"Flow-OPD: On-Policy Distillation for Flow Matching Models","ref_index":38,"is_internal_anchor":true},{"citing_arxiv_id":"2605.07396","citing_title":"Rubric-based On-policy Distillation","ref_index":25,"is_internal_anchor":true},{"citing_arxiv_id":"2605.07865","citing_title":"KL for a KL: On-Policy Distillation with Control Variate Baseline","ref_index":14,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/66MAX7UCOADDI55SGR34MQRY4Y","json":"https://pith.science/pith/66MAX7UCOADDI55SGR34MQRY4Y.json","graph_json":"https://pith.science/api/pith-number/66MAX7UCOADDI55SGR34MQRY4Y/graph.json","events_json":"https://pith.science/api/pith-number/66MAX7UCOADDI55SGR34MQRY4Y/events.json","paper":"https://pith.science/paper/66MAX7UC"},"agent_actions":{"view_html":"https://pith.science/pith/66MAX7UCOADDI55SGR34MQRY4Y","download_json":"https://pith.science/pith/66MAX7UCOADDI55SGR34MQRY4Y.json","view_paper":"https://pith.science/paper/66MAX7UC","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2603.07079&json=true","fetch_graph":"https://pith.science/api/pith-number/66MAX7UCOADDI55SGR34MQRY4Y/graph.json","fetch_events":"https://pith.science/api/pith-number/66MAX7UCOADDI55SGR34MQRY4Y/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/66MAX7UCOADDI55SGR34MQRY4Y/action/timestamp_anchor","attest_storage":"https://pith.science/pith/66MAX7UCOADDI55SGR34MQRY4Y/action/storage_attestation","attest_author":"https://pith.science/pith/66MAX7UCOADDI55SGR34MQRY4Y/action/author_attestation","sign_citation":"https://pith.science/pith/66MAX7UCOADDI55SGR34MQRY4Y/action/citation_signature","submit_replication":"https://pith.science/pith/66MAX7UCOADDI55SGR34MQRY4Y/action/replication_record"}},"created_at":"2026-05-25T02:02:14.034066+00:00","updated_at":"2026-05-25T02:02:14.034066+00:00"}