{"paper":{"title":"NOVA: Fundamental Limits of Knowledge Discovery Through AI","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Under a Zipf-law assumption on discovery probabilities, the cost to gather D new AI discoveries grows as D to the power alpha.","cross_cats":["cs.IT","math.IT"],"primary_cat":"cs.AI","authors_text":"Ken Duffy, Muriel M\\'edard, Salman Avestimehr","submitted_at":"2026-05-12T21:37:09Z","abstract_excerpt":"Can AI systems discover genuinely new knowledge through iterative self improvement, and if so, at what cost? We introduce the NOVA framework, which models the common ``generate, verify, accumulate, retrain'' loop as an adaptive sampling process over a knowledge space. We identify sufficient conditions under which accumulated genuine knowledge eventually covers a finite domain, and show how their violations produce distinct failure modes: contamination, forgetting, exploration failure, and acceptance failure. We then analyze imperfect verification and identify a contamination trap: as easy-to-f"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Under a separate tail-equivalence assumption relating the model's effective discovery distribution to a Zipf law with exponent alpha greater than 1, we prove that the cumulative generation cost required to obtain D distinct genuine discoveries satisfies R_cum(D) = Theta(c_gen D^alpha).","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The tail-equivalence assumption relating the model's effective discovery distribution to a Zipf law with exponent alpha greater than 1, which is invoked to derive the asymptotic scaling of cumulative generation cost.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"NOVA models the generate-verify-accumulate-retrain loop as adaptive sampling and proves that cumulative generation cost to obtain D genuine discoveries scales as Theta(c_gen D^alpha) under a Zipf tail-equivalence assumption with alpha greater than 1.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Under a Zipf-law assumption on discovery probabilities, the cost to gather D new AI discoveries grows as D to the power alpha.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"dd8725f1e386d2e6938d20bfe51fe63ee845e958182a8c795d81ad45c23f040b"},"source":{"id":"2605.15219","kind":"arxiv","version":1},"verdict":{"id":"363da110-87dd-47b4-b006-be2c545d6fad","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-19T17:13:38.327509Z","strongest_claim":"Under a separate tail-equivalence assumption relating the model's effective discovery distribution to a Zipf law with exponent alpha greater than 1, we prove that the cumulative generation cost required to obtain D distinct genuine discoveries satisfies R_cum(D) = Theta(c_gen D^alpha).","one_line_summary":"NOVA models the generate-verify-accumulate-retrain loop as adaptive sampling and proves that cumulative generation cost to obtain D genuine discoveries scales as Theta(c_gen D^alpha) under a Zipf tail-equivalence assumption with alpha greater than 1.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The tail-equivalence assumption relating the model's effective discovery distribution to a Zipf law with exponent alpha greater than 1, which is invoked to derive the asymptotic scaling of cumulative generation cost.","pith_extraction_headline":"Under a Zipf-law assumption on discovery probabilities, the cost to gather D new AI discoveries grows as D to the power alpha."},"integrity":{"clean":false,"summary":{"advisory":2,"critical":0,"by_detector":{"doi_compliance":{"total":2,"advisory":2,"critical":0,"informational":0}},"informational":0},"endpoint":"/pith/2605.15219/integrity.json","findings":[{"note":"DOI in the printed bibliography is fragmented by whitespace or line breaks. A longer candidate (10.1109/TIT.2024.3440661.Prafulla) was visible in the surrounding text but could not be confirmed against doi.org as printed.","detector":"doi_compliance","severity":"advisory","ref_index":2,"audited_at":"2026-05-19T17:26:34.606049Z","detected_doi":"10.1109/TIT.2024.3440661.Prafulla","finding_type":"recoverable_identifier","verdict_class":"incontrovertible","detected_arxiv_id":null},{"note":"DOI in the printed bibliography is fragmented by whitespace or line breaks. A longer candidate (10.1109/ISIT45174.2021.9518169.11) was visible in the surrounding text but could not be confirmed against doi.org as printed.","detector":"doi_compliance","severity":"advisory","ref_index":10,"audited_at":"2026-05-19T17:26:34.606049Z","detected_doi":"10.1109/ISIT45174.2021.9518169.11","finding_type":"recoverable_identifier","verdict_class":"incontrovertible","detected_arxiv_id":null}],"available":true,"detectors_run":[{"name":"claim_evidence","ran_at":"2026-05-19T21:01:57.275998Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_title_agreement","ran_at":"2026-05-19T17:31:18.498417Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_compliance","ran_at":"2026-05-19T17:26:34.606049Z","status":"completed","version":"1.0.0","findings_count":2},{"name":"ai_meta_artifact","ran_at":"2026-05-19T13:33:22.837626Z","status":"skipped","version":"1.0.0","findings_count":0}],"snapshot_sha256":"3d275e4690e55ada9052eb1bd741698c0336e2d3de37d247e4e3c822f666d5d2"},"references":{"count":13,"sample":[{"doi":"10.1109/isit.2018.8437620","year":2018,"title":"Anna Ben-Hamou, Stéphane Boucheron, and Mesrob I","work_id":"ee8c9c83-3ac9-4d3c-9304-75320bf313fa","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"10.1109/tit","year":2024,"title":"2024.3440661","work_id":"4a747430-fdf9-4c15-9334-59f7e564e39c","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"10.1109/tsp.2022.3186176","year":2022,"title":"Bradley Efron and Ronald Thisted","work_id":"678dedb2-63aa-4614-be07-a673050ca7ac","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Reinforced Self-Training (ReST) for Language Modeling","work_id":"db054f5c-62a3-405b-aae8-43e02c39f672","ref_index":4,"cited_arxiv_id":"2308.08998","is_internal_anchor":true},{"doi":"","year":null,"title":"Besting Good–Turing: Optimality of non-parametric maximum likelihood for distribution estimation","work_id":"31e68224-2ac6-4f06-882b-a426258827cf","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":13,"snapshot_sha256":"8c75aee5565769e90388d16bed4dc74cd4b3ba376528f99fdf1f7d75a438454c","internal_anchors":2},"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"}