{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:VWH64TAKBKYBFU55DWD6YRNLQQ","short_pith_number":"pith:VWH64TAK","schema_version":"1.0","canonical_sha256":"ad8fee4c0a0ab012d3bd1d87ec45ab84379df1cc40a699d2cab11b32c2efc424","source":{"kind":"arxiv","id":"2407.13692","version":2},"attestation_state":"computed","paper":{"title":"Prover-Verifier Games improve legibility of LLM outputs","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Harri Edwards, Jan Hendrik Kirchner, Jan Leike, Nat McAleese, Yining Chen, Yuri Burda","submitted_at":"2024-07-18T16:58:18Z","abstract_excerpt":"One way to increase confidence in the outputs of Large Language Models (LLMs) is to support them with reasoning that is clear and easy to check -- a property we call legibility. We study legibility in the context of solving grade-school math problems and show that optimizing chain-of-thought solutions only for answer correctness can make them less legible. To mitigate the loss in legibility, we propose a training algorithm inspired by Prover-Verifier Game from Anil et al. (2021). Our algorithm iteratively trains small verifiers to predict solution correctness, \"helpful\" provers to produce corr"},"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":"2407.13692","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CL","submitted_at":"2024-07-18T16:58:18Z","cross_cats_sorted":[],"title_canon_sha256":"0e621d2ade5b1a3e6cc1cdb381ecf1f0139de3ff1297e6f4af0f0eafacd66e8d","abstract_canon_sha256":"178df820896ce72111bc10ff854a69d760dec85459c610e9cf8af7e8dc4cdd2c"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T08:50:57.742954Z","signature_b64":"YEgC1UEy0kjuzo6DvYROjPEpICFKxvLjNCDJlue/RphByWBbGgDs/SI5r1VWMVDL2Xt0CoJL4zXR+EIO7p+3Dg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"ad8fee4c0a0ab012d3bd1d87ec45ab84379df1cc40a699d2cab11b32c2efc424","last_reissued_at":"2026-07-05T08:50:57.742519Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T08:50:57.742519Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Prover-Verifier Games improve legibility of LLM outputs","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Harri Edwards, Jan Hendrik Kirchner, Jan Leike, Nat McAleese, Yining Chen, Yuri Burda","submitted_at":"2024-07-18T16:58:18Z","abstract_excerpt":"One way to increase confidence in the outputs of Large Language Models (LLMs) is to support them with reasoning that is clear and easy to check -- a property we call legibility. We study legibility in the context of solving grade-school math problems and show that optimizing chain-of-thought solutions only for answer correctness can make them less legible. To mitigate the loss in legibility, we propose a training algorithm inspired by Prover-Verifier Game from Anil et al. (2021). Our algorithm iteratively trains small verifiers to predict solution correctness, \"helpful\" provers to produce corr"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2407.13692","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/2407.13692/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":"2407.13692","created_at":"2026-07-05T08:50:57.742571+00:00"},{"alias_kind":"arxiv_version","alias_value":"2407.13692v2","created_at":"2026-07-05T08:50:57.742571+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2407.13692","created_at":"2026-07-05T08:50:57.742571+00:00"},{"alias_kind":"pith_short_12","alias_value":"VWH64TAKBKYB","created_at":"2026-07-05T08:50:57.742571+00:00"},{"alias_kind":"pith_short_16","alias_value":"VWH64TAKBKYBFU55","created_at":"2026-07-05T08:50:57.742571+00:00"},{"alias_kind":"pith_short_8","alias_value":"VWH64TAK","created_at":"2026-07-05T08:50:57.742571+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":17,"internal_anchor_count":1,"sample":[{"citing_arxiv_id":"2607.07690","citing_title":"Agon: Competitive Cross-Model RL with Implicit Rival Grading of Reasoning","ref_index":12,"is_internal_anchor":true},{"citing_arxiv_id":"2606.10684","citing_title":"Divide and Cooperate: Role-Decomposed Multi-Agent LLM Training with Cross-Agent Learning Signals","ref_index":6,"is_internal_anchor":false},{"citing_arxiv_id":"2606.09711","citing_title":"Proxy Reward Internalization and Mechanistic Exploitation: A Learned Precursor to Reward Hacking and Its Generalization","ref_index":92,"is_internal_anchor":false},{"citing_arxiv_id":"2607.01223","citing_title":"Theoria: Rewrite-Acceptability Verification over Informal Reasoning States","ref_index":12,"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":13,"is_internal_anchor":false},{"citing_arxiv_id":"2606.28166","citing_title":"Tandem Reinforcement Learning with Verifiable Rewards","ref_index":9,"is_internal_anchor":false},{"citing_arxiv_id":"2606.31748","citing_title":"Addressing Over-Refusal in LLMs with Competing Rewards","ref_index":104,"is_internal_anchor":false},{"citing_arxiv_id":"2605.20531","citing_title":"Pseudo-Formalization for Automatic Proof Verification","ref_index":14,"is_internal_anchor":false},{"citing_arxiv_id":"2605.27914","citing_title":"Does Capability Transfer to Subjective Behavior -- and Would Our Instruments Tell Us? A Self-Evolving, Trust-by-Construction Evaluation Paradigm","ref_index":62,"is_internal_anchor":false},{"citing_arxiv_id":"2605.28807","citing_title":"Calibrating Conservatism for Scalable Oversight","ref_index":6,"is_internal_anchor":false},{"citing_arxiv_id":"2605.30290","citing_title":"Self-Trained Verification for Training- and Test-Time Self-Improvement","ref_index":3,"is_internal_anchor":false},{"citing_arxiv_id":"2605.22211","citing_title":"CLORE: Content-Level Optimization for Reasoning Efficiency","ref_index":24,"is_internal_anchor":false},{"citing_arxiv_id":"2503.11926","citing_title":"Monitoring Reasoning Models for Misbehavior and the Risks of Promoting Obfuscation","ref_index":76,"is_internal_anchor":false},{"citing_arxiv_id":"2605.20531","citing_title":"Pseudo-Formalization for Automatic Proof Verification","ref_index":14,"is_internal_anchor":false},{"citing_arxiv_id":"2605.10810","citing_title":"Likelihood scoring for continuations of mathematical text: a self-supervised benchmark with tests for shortcut vulnerabilities","ref_index":46,"is_internal_anchor":false},{"citing_arxiv_id":"2605.13875","citing_title":"Common-agency Games for Multi-Objective Test-Time Alignment","ref_index":207,"is_internal_anchor":false},{"citing_arxiv_id":"2412.21187","citing_title":"Do NOT Think That Much for 2+3=? On the Overthinking of o1-Like LLMs","ref_index":17,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/VWH64TAKBKYBFU55DWD6YRNLQQ","json":"https://pith.science/pith/VWH64TAKBKYBFU55DWD6YRNLQQ.json","graph_json":"https://pith.science/api/pith-number/VWH64TAKBKYBFU55DWD6YRNLQQ/graph.json","events_json":"https://pith.science/api/pith-number/VWH64TAKBKYBFU55DWD6YRNLQQ/events.json","paper":"https://pith.science/paper/VWH64TAK"},"agent_actions":{"view_html":"https://pith.science/pith/VWH64TAKBKYBFU55DWD6YRNLQQ","download_json":"https://pith.science/pith/VWH64TAKBKYBFU55DWD6YRNLQQ.json","view_paper":"https://pith.science/paper/VWH64TAK","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2407.13692&json=true","fetch_graph":"https://pith.science/api/pith-number/VWH64TAKBKYBFU55DWD6YRNLQQ/graph.json","fetch_events":"https://pith.science/api/pith-number/VWH64TAKBKYBFU55DWD6YRNLQQ/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/VWH64TAKBKYBFU55DWD6YRNLQQ/action/timestamp_anchor","attest_storage":"https://pith.science/pith/VWH64TAKBKYBFU55DWD6YRNLQQ/action/storage_attestation","attest_author":"https://pith.science/pith/VWH64TAKBKYBFU55DWD6YRNLQQ/action/author_attestation","sign_citation":"https://pith.science/pith/VWH64TAKBKYBFU55DWD6YRNLQQ/action/citation_signature","submit_replication":"https://pith.science/pith/VWH64TAKBKYBFU55DWD6YRNLQQ/action/replication_record"}},"created_at":"2026-07-05T08:50:57.742571+00:00","updated_at":"2026-07-05T08:50:57.742571+00:00"}