{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:T64LZBYPDQ4SE7RCWRJP7J3Q6Q","short_pith_number":"pith:T64LZBYP","schema_version":"1.0","canonical_sha256":"9fb8bc870f1c39227e22b452ffa770f4223fc82e447352543cb9502e25024a83","source":{"kind":"arxiv","id":"2512.19673","version":3},"attestation_state":"computed","paper":{"title":"Bottom-up Policy Optimization: Your Language Model Policy Secretly Contains Internal Policies","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI","cs.CL"],"primary_cat":"cs.LG","authors_text":"Chengfeng Zhao, Huanxuan Liao, Jun Zhao, Kang Liu, Minzheng Wang, Qiunan Lu, Shizhu He, Tian Liang, Yuqiao Tan","submitted_at":"2025-12-22T18:51:48Z","abstract_excerpt":"Existing reinforcement learning (RL) approaches treat large language models (LLMs) as a unified policy, overlooking their internal mechanisms. In this paper, we decompose the LLM-based policy into Internal Layer Policies and Internal Modular Policies via the Transformer's residual stream. Our entropy analysis of internal policy reveals distinct patterns: (1) universally, internal policies evolve from high-entropy exploration in early layers to deterministic refinement in the top layers; and (2) Qwen exhibits an explicit progressive reasoning structure, contrasting with the abrupt convergence i"},"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":"2512.19673","kind":"arxiv","version":3},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2025-12-22T18:51:48Z","cross_cats_sorted":["cs.AI","cs.CL"],"title_canon_sha256":"91e1005430805add82b33ced7dda6029f73e7743a960eb617cd373ecc8742a08","abstract_canon_sha256":"59022c9732b39dae29c8404680c1952b14c84bab2bed30330c9e03dc7546f5df"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-06-01T01:02:27.842207Z","signature_b64":"YqxAMQGYDygbEIot3NHrbwcwkJHsrttvJ2E3FOPZB7wTI/Otc8bT8exW9xs1LpCg0Td/UqUIXKFCNMYxzciIDA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"9fb8bc870f1c39227e22b452ffa770f4223fc82e447352543cb9502e25024a83","last_reissued_at":"2026-06-01T01:02:27.841128Z","signature_status":"signed_v1","first_computed_at":"2026-06-01T01:02:27.841128Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Bottom-up Policy Optimization: Your Language Model Policy Secretly Contains Internal Policies","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI","cs.CL"],"primary_cat":"cs.LG","authors_text":"Chengfeng Zhao, Huanxuan Liao, Jun Zhao, Kang Liu, Minzheng Wang, Qiunan Lu, Shizhu He, Tian Liang, Yuqiao Tan","submitted_at":"2025-12-22T18:51:48Z","abstract_excerpt":"Existing reinforcement learning (RL) approaches treat large language models (LLMs) as a unified policy, overlooking their internal mechanisms. In this paper, we decompose the LLM-based policy into Internal Layer Policies and Internal Modular Policies via the Transformer's residual stream. Our entropy analysis of internal policy reveals distinct patterns: (1) universally, internal policies evolve from high-entropy exploration in early layers to deterministic refinement in the top layers; and (2) Qwen exhibits an explicit progressive reasoning structure, contrasting with the abrupt convergence i"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2512.19673","kind":"arxiv","version":3},"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/2512.19673/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":"2512.19673","created_at":"2026-06-01T01:02:27.841282+00:00"},{"alias_kind":"arxiv_version","alias_value":"2512.19673v3","created_at":"2026-06-01T01:02:27.841282+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2512.19673","created_at":"2026-06-01T01:02:27.841282+00:00"},{"alias_kind":"pith_short_12","alias_value":"T64LZBYPDQ4S","created_at":"2026-06-01T01:02:27.841282+00:00"},{"alias_kind":"pith_short_16","alias_value":"T64LZBYPDQ4SE7RC","created_at":"2026-06-01T01:02:27.841282+00:00"},{"alias_kind":"pith_short_8","alias_value":"T64LZBYP","created_at":"2026-06-01T01:02:27.841282+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":5,"internal_anchor_count":5,"sample":[{"citing_arxiv_id":"2601.14004","citing_title":"Locate, Steer, and Improve: A Practical Survey of Actionable Mechanistic Interpretability in Large Language Models","ref_index":298,"is_internal_anchor":true},{"citing_arxiv_id":"2604.23505","citing_title":"Uncertainty Propagation in LLM-Based Systems","ref_index":43,"is_internal_anchor":true},{"citing_arxiv_id":"2605.05741","citing_title":"HyperLens: Quantifying Cognitive Effort in LLMs with Fine-grained Confidence Trajectory","ref_index":19,"is_internal_anchor":true},{"citing_arxiv_id":"2604.14142","citing_title":"From $P(y|x)$ to $P(y)$: Investigating Reinforcement Learning in Pre-train Space","ref_index":55,"is_internal_anchor":true},{"citing_arxiv_id":"2604.11297","citing_title":"The Past Is Not Past: Memory-Enhanced Dynamic Reward Shaping","ref_index":39,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/T64LZBYPDQ4SE7RCWRJP7J3Q6Q","json":"https://pith.science/pith/T64LZBYPDQ4SE7RCWRJP7J3Q6Q.json","graph_json":"https://pith.science/api/pith-number/T64LZBYPDQ4SE7RCWRJP7J3Q6Q/graph.json","events_json":"https://pith.science/api/pith-number/T64LZBYPDQ4SE7RCWRJP7J3Q6Q/events.json","paper":"https://pith.science/paper/T64LZBYP"},"agent_actions":{"view_html":"https://pith.science/pith/T64LZBYPDQ4SE7RCWRJP7J3Q6Q","download_json":"https://pith.science/pith/T64LZBYPDQ4SE7RCWRJP7J3Q6Q.json","view_paper":"https://pith.science/paper/T64LZBYP","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2512.19673&json=true","fetch_graph":"https://pith.science/api/pith-number/T64LZBYPDQ4SE7RCWRJP7J3Q6Q/graph.json","fetch_events":"https://pith.science/api/pith-number/T64LZBYPDQ4SE7RCWRJP7J3Q6Q/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/T64LZBYPDQ4SE7RCWRJP7J3Q6Q/action/timestamp_anchor","attest_storage":"https://pith.science/pith/T64LZBYPDQ4SE7RCWRJP7J3Q6Q/action/storage_attestation","attest_author":"https://pith.science/pith/T64LZBYPDQ4SE7RCWRJP7J3Q6Q/action/author_attestation","sign_citation":"https://pith.science/pith/T64LZBYPDQ4SE7RCWRJP7J3Q6Q/action/citation_signature","submit_replication":"https://pith.science/pith/T64LZBYPDQ4SE7RCWRJP7J3Q6Q/action/replication_record"}},"created_at":"2026-06-01T01:02:27.841282+00:00","updated_at":"2026-06-01T01:02:27.841282+00:00"}