{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:F6QCW7GLFJAMDP5ZIWUITM533H","short_pith_number":"pith:F6QCW7GL","schema_version":"1.0","canonical_sha256":"2fa02b7ccb2a40c1bfb945a889b3bbd9c0a801c5b32458052d00a67bb04eb446","source":{"kind":"arxiv","id":"2605.16727","version":1},"attestation_state":"computed","paper":{"title":"PopuLoRA: Co-Evolving LLM Populations for Reasoning Self-Play","license":"http://creativecommons.org/licenses/by/4.0/","headline":"A population of specialized LoRA adapters in asymmetric self-play creates a co-evolutionary arms race that improves LLM reasoning over single-agent baselines.","cross_cats":[],"primary_cat":"cs.AI","authors_text":"Augustine N. Mavor-Parker, Geoffrey Bradway, Lorenz Wolf, Matthew James Sargent, Maxwill Lin, Roger Creus Castanyer","submitted_at":"2026-05-16T00:29:35Z","abstract_excerpt":"We introduce PopuLoRA, a population-based asymmetric self-play framework for reinforcement learning with verifiable rewards (RLVR) post-training of LLMs. Teachers and students are specialised LoRA adapters on a shared frozen base: teachers propose problems, matched students solve them under a programmatic verifier, and cross-evaluation between sub-populations replaces the self-calibration that limits single-agent self-play. A family of LoRA weight-space evolution operators (mutations and crossovers that produce same-rank population members in seconds) serves as the replacement step of a popula"},"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":true,"formal_links_present":true},"canonical_record":{"source":{"id":"2605.16727","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.AI","submitted_at":"2026-05-16T00:29:35Z","cross_cats_sorted":[],"title_canon_sha256":"86dbf4174048556332b84ef277c240c97dde3a1c50fbb125074d6e5dc82c2512","abstract_canon_sha256":"7853f28c755e217f94a5d2a661ff5d73c45763cba9f4bf99300ba3cf39976d92"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-20T00:02:38.695932Z","signature_b64":"xPuYyBEKSX+lykzgGPdATptz2kVprwUcVBpdfT+Cjy6ZJBnzuKoJwQLa/HSHI/C9E5gq0DoQjULwdHNrudB9Dw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"2fa02b7ccb2a40c1bfb945a889b3bbd9c0a801c5b32458052d00a67bb04eb446","last_reissued_at":"2026-05-20T00:02:38.695064Z","signature_status":"signed_v1","first_computed_at":"2026-05-20T00:02:38.695064Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"PopuLoRA: Co-Evolving LLM Populations for Reasoning Self-Play","license":"http://creativecommons.org/licenses/by/4.0/","headline":"A population of specialized LoRA adapters in asymmetric self-play creates a co-evolutionary arms race that improves LLM reasoning over single-agent baselines.","cross_cats":[],"primary_cat":"cs.AI","authors_text":"Augustine N. Mavor-Parker, Geoffrey Bradway, Lorenz Wolf, Matthew James Sargent, Maxwill Lin, Roger Creus Castanyer","submitted_at":"2026-05-16T00:29:35Z","abstract_excerpt":"We introduce PopuLoRA, a population-based asymmetric self-play framework for reinforcement learning with verifiable rewards (RLVR) post-training of LLMs. Teachers and students are specialised LoRA adapters on a shared frozen base: teachers propose problems, matched students solve them under a programmatic verifier, and cross-evaluation between sub-populations replaces the self-calibration that limits single-agent self-play. A family of LoRA weight-space evolution operators (mutations and crossovers that produce same-rank population members in seconds) serves as the replacement step of a popula"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Despite lower training-time reward, the population mean outperforms the baseline on three code benchmarks (HumanEval+, MBPP+, LiveCodeBench) and seven math benchmarks (AIME 24/25, AMC 23, MATH-500, Minerva, GSM8K, OlympiadBench), and even the weakest member of the population beats the baseline on aggregate.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That cross-evaluation between sub-populations reliably prevents self-calibration and sustains an expanding problem-space arms race rather than allowing the population to converge on a narrow set of solvable problems.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"PopuLoRA shows that co-evolving populations of LoRA adapters through cross-evaluated self-play can outperform compute-matched single-agent baselines on multiple code and math reasoning benchmarks.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"A population of specialized LoRA adapters in asymmetric self-play creates a co-evolutionary arms race that improves LLM reasoning over single-agent baselines.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"68213903ef3093e6a4ac822c77750a3aeb6726f19f02e3c7c6752ad8021cf605"},"source":{"id":"2605.16727","kind":"arxiv","version":1},"verdict":{"id":"da00c2b2-9610-46cd-8978-5de6a7220c0e","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-19T21:33:52.891464Z","strongest_claim":"Despite lower training-time reward, the population mean outperforms the baseline on three code benchmarks (HumanEval+, MBPP+, LiveCodeBench) and seven math benchmarks (AIME 24/25, AMC 23, MATH-500, Minerva, GSM8K, OlympiadBench), and even the weakest member of the population beats the baseline on aggregate.","one_line_summary":"PopuLoRA shows that co-evolving populations of LoRA adapters through cross-evaluated self-play can outperform compute-matched single-agent baselines on multiple code and math reasoning benchmarks.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That cross-evaluation between sub-populations reliably prevents self-calibration and sustains an expanding problem-space arms race rather than allowing the population to converge on a narrow set of solvable problems.","pith_extraction_headline":"A population of specialized LoRA adapters in asymmetric self-play creates a co-evolutionary arms race that improves LLM reasoning over single-agent baselines."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2605.16727/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"doi_title_agreement","ran_at":"2026-05-19T22:01:23.082836Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_compliance","ran_at":"2026-05-19T21:41:13.976443Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"claim_evidence","ran_at":"2026-05-19T19:01:56.346329Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"ai_meta_artifact","ran_at":"2026-05-19T18:33:26.471815Z","status":"skipped","version":"1.0.0","findings_count":0}],"snapshot_sha256":"0d052aa4fd4e5a29937f12b459e123d0d37810d7f4d92c469629f916a12185cd"},"references":{"count":67,"sample":[{"doi":"","year":2019,"title":"Vinyals, Oriol and Babuschkin, Igor and Czarnecki, Wojciech M. and Mathieu, Micha. Grandmaster Level in. Nature , volume =. 2019 , doi =","work_id":"ee92cbd6-af5b-4ba7-a386-fa02f0285ffa","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"International Conference on Learning Representations , year =","work_id":"bedd69d1-1a3c-4eaf-905b-ceb510e077e2","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"10.48550/arxiv.2402.03300","year":null,"title":"DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models","work_id":"c5006563-f3ec-438a-9e35-b7b484f34828","ref_index":3,"cited_arxiv_id":"2402.03300","is_internal_anchor":true},{"doi":"10.48550/arxiv.1707.06347","year":2017,"title":"Proximal Policy Optimization Algorithms","work_id":"240c67fe-d14d-4520-91c1-38a4e272ca19","ref_index":4,"cited_arxiv_id":"1707.06347","is_internal_anchor":true},{"doi":"","year":1992,"title":"Machine Learning , volume =","work_id":"71efaacf-fbce-4277-80c9-97d1cec481a0","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":67,"snapshot_sha256":"b763dc216a8b4e8da68cb3ff19d43fbbf8d86d182a0c637715d86df12d565dba","internal_anchors":16},"formal_canon":{"evidence_count":1,"snapshot_sha256":"2facb1975cfb3e53f1f86762d6dc880fdb2895583df84123cd1954a66a89ed47"},"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":"2605.16727","created_at":"2026-05-20T00:02:38.695197+00:00"},{"alias_kind":"arxiv_version","alias_value":"2605.16727v1","created_at":"2026-05-20T00:02:38.695197+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.16727","created_at":"2026-05-20T00:02:38.695197+00:00"},{"alias_kind":"pith_short_12","alias_value":"F6QCW7GLFJAM","created_at":"2026-05-20T00:02:38.695197+00:00"},{"alias_kind":"pith_short_16","alias_value":"F6QCW7GLFJAMDP5Z","created_at":"2026-05-20T00:02:38.695197+00:00"},{"alias_kind":"pith_short_8","alias_value":"F6QCW7GL","created_at":"2026-05-20T00:02:38.695197+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":0,"internal_anchor_count":0,"sample":[]},"formal_canon":{"evidence_count":1,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/F6QCW7GLFJAMDP5ZIWUITM533H","json":"https://pith.science/pith/F6QCW7GLFJAMDP5ZIWUITM533H.json","graph_json":"https://pith.science/api/pith-number/F6QCW7GLFJAMDP5ZIWUITM533H/graph.json","events_json":"https://pith.science/api/pith-number/F6QCW7GLFJAMDP5ZIWUITM533H/events.json","paper":"https://pith.science/paper/F6QCW7GL"},"agent_actions":{"view_html":"https://pith.science/pith/F6QCW7GLFJAMDP5ZIWUITM533H","download_json":"https://pith.science/pith/F6QCW7GLFJAMDP5ZIWUITM533H.json","view_paper":"https://pith.science/paper/F6QCW7GL","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2605.16727&json=true","fetch_graph":"https://pith.science/api/pith-number/F6QCW7GLFJAMDP5ZIWUITM533H/graph.json","fetch_events":"https://pith.science/api/pith-number/F6QCW7GLFJAMDP5ZIWUITM533H/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/F6QCW7GLFJAMDP5ZIWUITM533H/action/timestamp_anchor","attest_storage":"https://pith.science/pith/F6QCW7GLFJAMDP5ZIWUITM533H/action/storage_attestation","attest_author":"https://pith.science/pith/F6QCW7GLFJAMDP5ZIWUITM533H/action/author_attestation","sign_citation":"https://pith.science/pith/F6QCW7GLFJAMDP5ZIWUITM533H/action/citation_signature","submit_replication":"https://pith.science/pith/F6QCW7GLFJAMDP5ZIWUITM533H/action/replication_record"}},"created_at":"2026-05-20T00:02:38.695197+00:00","updated_at":"2026-05-20T00:02:38.695197+00:00"}