{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2023:BP76C335XI4PH3WPZIAIKHC6FR","short_pith_number":"pith:BP76C335","schema_version":"1.0","canonical_sha256":"0bffe16f7dba38f3eecfca00851c5e2c4e46745330313c365e79c3e0fe6ff20a","source":{"kind":"arxiv","id":"2312.09390","version":1},"attestation_state":"computed","paper":{"title":"Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervision","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Adrien Ecoffet, Bowen Baker, Collin Burns, Ilya Sutskever, Jan Hendrik Kirchner, Jan Leike, Jeff Wu, Leo Gao, Leopold Aschenbrenner, Manas Joglekar, Pavel Izmailov, Yining Chen","submitted_at":"2023-12-14T23:07:33Z","abstract_excerpt":"Widely used alignment techniques, such as reinforcement learning from human feedback (RLHF), rely on the ability of humans to supervise model behavior - for example, to evaluate whether a model faithfully followed instructions or generated safe outputs. However, future superhuman models will behave in complex ways too difficult for humans to reliably evaluate; humans will only be able to weakly supervise superhuman models. We study an analogy to this problem: can weak model supervision elicit the full capabilities of a much stronger model? We test this using a range of pretrained language mode"},"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":"2312.09390","kind":"arxiv","version":1},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CL","submitted_at":"2023-12-14T23:07:33Z","cross_cats_sorted":[],"title_canon_sha256":"4f29988dba1f8b2e21f7f1e637e3590a3b117e65c0b37159323c00aa7f28b086","abstract_canon_sha256":"a3d464ead577611f817b00164caee52dfb06214d3958bfa17d0baa31a6bbe016"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T07:24:24.257760Z","signature_b64":"XOTES0aOcoiMFCHDwSvAFAzKtupTQ1g1PJQZGf+zeFNpEVDBd5hiy/rA44TVqHb1yBsDDMfK7ctxuQmsSMiPDA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"0bffe16f7dba38f3eecfca00851c5e2c4e46745330313c365e79c3e0fe6ff20a","last_reissued_at":"2026-07-05T07:24:24.257258Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T07:24:24.257258Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervision","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Adrien Ecoffet, Bowen Baker, Collin Burns, Ilya Sutskever, Jan Hendrik Kirchner, Jan Leike, Jeff Wu, Leo Gao, Leopold Aschenbrenner, Manas Joglekar, Pavel Izmailov, Yining Chen","submitted_at":"2023-12-14T23:07:33Z","abstract_excerpt":"Widely used alignment techniques, such as reinforcement learning from human feedback (RLHF), rely on the ability of humans to supervise model behavior - for example, to evaluate whether a model faithfully followed instructions or generated safe outputs. However, future superhuman models will behave in complex ways too difficult for humans to reliably evaluate; humans will only be able to weakly supervise superhuman models. We study an analogy to this problem: can weak model supervision elicit the full capabilities of a much stronger model? We test this using a range of pretrained language mode"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2312.09390","kind":"arxiv","version":1},"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/2312.09390/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":"2312.09390","created_at":"2026-07-05T07:24:24.257322+00:00"},{"alias_kind":"arxiv_version","alias_value":"2312.09390v1","created_at":"2026-07-05T07:24:24.257322+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2312.09390","created_at":"2026-07-05T07:24:24.257322+00:00"},{"alias_kind":"pith_short_12","alias_value":"BP76C335XI4P","created_at":"2026-07-05T07:24:24.257322+00:00"},{"alias_kind":"pith_short_16","alias_value":"BP76C335XI4PH3WP","created_at":"2026-07-05T07:24:24.257322+00:00"},{"alias_kind":"pith_short_8","alias_value":"BP76C335","created_at":"2026-07-05T07:24:24.257322+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":39,"internal_anchor_count":1,"sample":[{"citing_arxiv_id":"2607.05394","citing_title":"Weak-to-Strong Generalization via Direct On-Policy Distillation","ref_index":70,"is_internal_anchor":true},{"citing_arxiv_id":"2606.22873","citing_title":"SingGuard: A Policy-Adaptive Multimodal LLM Guardrail with Dynamic Reasoning","ref_index":219,"is_internal_anchor":false},{"citing_arxiv_id":"2606.10456","citing_title":"The Distributed Detectability Band Against Marginal-Preserving Attacks","ref_index":6,"is_internal_anchor":false},{"citing_arxiv_id":"2606.09475","citing_title":"Emergent alignment and the projectability of ethical personas","ref_index":14,"is_internal_anchor":false},{"citing_arxiv_id":"2606.01249","citing_title":"Trust Region On-Policy Distillation","ref_index":42,"is_internal_anchor":false},{"citing_arxiv_id":"2606.00424","citing_title":"Weak Critics Make Strong Learners: On-Policy Critique Distillation for Scalable Oversight","ref_index":5,"is_internal_anchor":false},{"citing_arxiv_id":"2606.22873","citing_title":"SingGuard: A Policy-Adaptive Multimodal LLM Guardrail with Dynamic Reasoning","ref_index":218,"is_internal_anchor":false},{"citing_arxiv_id":"2606.28166","citing_title":"Tandem Reinforcement Learning with Verifiable Rewards","ref_index":2,"is_internal_anchor":false},{"citing_arxiv_id":"2606.31023","citing_title":"Certified Speculative Execution for Untrusted AI Agents","ref_index":43,"is_internal_anchor":false},{"citing_arxiv_id":"2605.17314","citing_title":"Weak-to-Strong Elicitation via Mismatched Wrong Drafts","ref_index":4,"is_internal_anchor":false},{"citing_arxiv_id":"2605.25183","citing_title":"Knowledge Graph-Driven Expert-Level Reasoning for Neuroscience","ref_index":9,"is_internal_anchor":false},{"citing_arxiv_id":"2605.28421","citing_title":"DenoiseRL: Bootstrapping Reasoning Models to Recover from Noisy Prefixes","ref_index":4,"is_internal_anchor":false},{"citing_arxiv_id":"2606.10747","citing_title":"The Arbiter Agent: Continually Monitoring Multi-Agent Conversations to Detect Emergent Misalignment","ref_index":7,"is_internal_anchor":false},{"citing_arxiv_id":"2402.03300","citing_title":"DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models","ref_index":5,"is_internal_anchor":false},{"citing_arxiv_id":"2408.15549","citing_title":"WildFeedback: Aligning LLMs With In-situ User Interactions And Feedback","ref_index":5,"is_internal_anchor":false},{"citing_arxiv_id":"2505.03631","citing_title":"Generalizable Video Quality Assessment via Weak-to-Strong Learning","ref_index":2,"is_internal_anchor":false},{"citing_arxiv_id":"2605.21984","citing_title":"Echo: Learning from Experience Data via User-Driven Refinement","ref_index":4,"is_internal_anchor":false},{"citing_arxiv_id":"2604.05002","citing_title":"Learning Stable Predictors from Weak Supervision under Distribution Shift","ref_index":3,"is_internal_anchor":false},{"citing_arxiv_id":"2605.20994","citing_title":"Towards Context-Invariant Safety Alignment for Large Language Models","ref_index":71,"is_internal_anchor":false},{"citing_arxiv_id":"2605.12413","citing_title":"Beyond Localization: A Comprehensive Diagnosis of Perspective-Conditioned Spatial Reasoning in MLLMs from Omnidirectional Images","ref_index":3,"is_internal_anchor":false},{"citing_arxiv_id":"2605.12991","citing_title":"Not Just RLHF: Why Alignment Alone Won't Fix Multi-Agent Sycophancy","ref_index":6,"is_internal_anchor":false},{"citing_arxiv_id":"2605.17314","citing_title":"Weak-to-Strong Elicitation via Mismatched Wrong Drafts","ref_index":4,"is_internal_anchor":false},{"citing_arxiv_id":"2509.23542","citing_title":"On the Shelf Life of Fine-Tuned LLM-Judges: Future-Proofing, Backward-Compatibility, and Question Generalization","ref_index":3,"is_internal_anchor":false},{"citing_arxiv_id":"2510.13830","citing_title":"Users as Annotators: LLM Preference Learning from Comparison Mode","ref_index":4,"is_internal_anchor":false},{"citing_arxiv_id":"2509.08827","citing_title":"A Survey of Reinforcement Learning for Large Reasoning Models","ref_index":47,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/BP76C335XI4PH3WPZIAIKHC6FR","json":"https://pith.science/pith/BP76C335XI4PH3WPZIAIKHC6FR.json","graph_json":"https://pith.science/api/pith-number/BP76C335XI4PH3WPZIAIKHC6FR/graph.json","events_json":"https://pith.science/api/pith-number/BP76C335XI4PH3WPZIAIKHC6FR/events.json","paper":"https://pith.science/paper/BP76C335"},"agent_actions":{"view_html":"https://pith.science/pith/BP76C335XI4PH3WPZIAIKHC6FR","download_json":"https://pith.science/pith/BP76C335XI4PH3WPZIAIKHC6FR.json","view_paper":"https://pith.science/paper/BP76C335","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2312.09390&json=true","fetch_graph":"https://pith.science/api/pith-number/BP76C335XI4PH3WPZIAIKHC6FR/graph.json","fetch_events":"https://pith.science/api/pith-number/BP76C335XI4PH3WPZIAIKHC6FR/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/BP76C335XI4PH3WPZIAIKHC6FR/action/timestamp_anchor","attest_storage":"https://pith.science/pith/BP76C335XI4PH3WPZIAIKHC6FR/action/storage_attestation","attest_author":"https://pith.science/pith/BP76C335XI4PH3WPZIAIKHC6FR/action/author_attestation","sign_citation":"https://pith.science/pith/BP76C335XI4PH3WPZIAIKHC6FR/action/citation_signature","submit_replication":"https://pith.science/pith/BP76C335XI4PH3WPZIAIKHC6FR/action/replication_record"}},"created_at":"2026-07-05T07:24:24.257322+00:00","updated_at":"2026-07-05T07:24:24.257322+00:00"}