{"paper":{"title":"ACES: Who Tests the Tests? Leave-One-Out AUC Consistency for Code Generation","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"You can weight LLM-generated tests by leave-one-out ranking agreement, recovering each test’s power to separate correct from incorrect code without knowing which codes are right.","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Hui Sun, Ming Li, Ren-Biao Liu, Xin-Ye Li, Yali Du, Yun-Ji Zhang, Zheng Xie","submitted_at":"2026-04-05T01:15:57Z","abstract_excerpt":"Selecting LLM-generated code candidates using LLM-generated tests is challenging because the tests themselves may be incorrect. Existing methods either treat all tests equally or rely on ad-hoc heuristics to filter unreliable tests. Yet determining test correctness requires knowing which codes are correct, creating a \\emph{circular dependency}. Our key insight is that we need not determine test correctness at all: \\emph{test votes should rank, not merely count}. What matters is not how many codes pass a test, but whether the test can \\emph{distinguish} correct from incorrect code. We break the"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Expected leave-one-out AUC of test j equals 1/2 plus a positive coefficient times the test's discriminative power δ_j = α_j − β_j (Theorem 3); consequently the closed-form ACES-C weights recover the sign of δ_j in expectation under Assumption 4 and approximate the oracle signal-to-noise ratio (Theorem 6, Corollary 8), while both ACES variants achieve state-of-the-art Pass@k among execution-only methods on HumanEval, HumanEval+ and MBPP.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"Assumption 4: the average discriminative power of the generated test pool is positive and large enough that leaving any single test out still leaves the remaining ranking better than random (¯δ > 2√(ln 2 / m)). The closed-form guarantees of ACES-C and the positivity of the LOO-AUC coefficients rest on this average-quality condition; when it fails, only the iterative ACES-O variant is claimed to remain effective.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Leave-one-out AUC of each test against the ranking induced by the remaining tests is proportional to that test's latent discriminative power, yielding closed-form and optimized weights that raise Pass@k.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"You can weight LLM-generated tests by leave-one-out ranking agreement, recovering each test’s power to separate correct from incorrect code without knowing which codes are right.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"431ab1cad14b0bca270cb31997212d25aceabef2b6527325a1036421c7a2ef4f"},"source":{"id":"2604.03922","kind":"arxiv","version":2},"verdict":{"id":"0ce40174-c50b-4edb-967f-bbf357a35dce","model_set":{"reader":"grok-4.5"},"created_at":"2026-07-13T11:58:43.055488Z","strongest_claim":"Expected leave-one-out AUC of test j equals 1/2 plus a positive coefficient times the test's discriminative power δ_j = α_j − β_j (Theorem 3); consequently the closed-form ACES-C weights recover the sign of δ_j in expectation under Assumption 4 and approximate the oracle signal-to-noise ratio (Theorem 6, Corollary 8), while both ACES variants achieve state-of-the-art Pass@k among execution-only methods on HumanEval, HumanEval+ and MBPP.","one_line_summary":"Leave-one-out AUC of each test against the ranking induced by the remaining tests is proportional to that test's latent discriminative power, yielding closed-form and optimized weights that raise Pass@k.","pipeline_version":"pith-pipeline@v1.1.0-grok45","weakest_assumption":"Assumption 4: the average discriminative power of the generated test pool is positive and large enough that leaving any single test out still leaves the remaining ranking better than random (¯δ > 2√(ln 2 / m)). The closed-form guarantees of ACES-C and the positivity of the LOO-AUC coefficients rest on this average-quality condition; when it fails, only the iterative ACES-O variant is claimed to remain effective.","pith_extraction_headline":"You can weight LLM-generated tests by leave-one-out ranking agreement, recovering each test’s power to separate correct from incorrect code without knowing which codes are right."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2604.03922/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"}