{"paper":{"title":"CSLibPremiseBench: Structure-Guided Premise Retrieval and Label Robustness for Lean 4 Computer-Science Theorems","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"CSLibPremiseBench shows structure-guided reranking gives only modest early-rank gains over lexical baselines in Lean 4 computer-science premise retrieval.","cross_cats":[],"primary_cat":"cs.LO","authors_text":"Junye Ji","submitted_at":"2026-05-14T08:28:57Z","abstract_excerpt":"CSLib is an emerging Lean 4 library for computer-science formalization, but its premise-retrieval behavior is not well represented by broad mathematical theorem-proving benchmarks. We introduce CSLibPremiseBench, a reproducible CSLib-specific benchmark and empirical study for source-visible premise retrieval over Lean 4 theorem and lemma declarations. The benchmark pins CSLib v4.29.0 at commit 0d37cc7fcc985cfc53b155e7eef2453f846c6da2, builds with Lean 4.29.0, and evaluates a strict import/source-order task set with 801 proxy-labelable tasks and 1875 CSLib candidate declarations. The labels are"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"CSG-Rerank gives a modest early-rank MRR gain over lexical BM25 under the strict policy, but does not reliably outperform BM25+symbol and does not improve Recall@10.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The source-visible proxy labels from proof references accurately stand in for the premises that would actually be needed in a realistic proof-generation or proof-repair setting.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"CSLibPremiseBench shows that a structure-guided reranker provides only modest MRR gains over BM25 for premise retrieval in Lean 4 CS theorems and does not improve Recall@10.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"CSLibPremiseBench shows structure-guided reranking gives only modest early-rank gains over lexical baselines in Lean 4 computer-science premise retrieval.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"5af6aa15496abeae1536dd27d6e1b7a699fbbb412810a8fbd768c42e3c39994c"},"source":{"id":"2605.14549","kind":"arxiv","version":1},"verdict":{"id":"a064f180-b8bc-4063-b740-90574b937d56","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T01:27:08.827243Z","strongest_claim":"CSG-Rerank gives a modest early-rank MRR gain over lexical BM25 under the strict policy, but does not reliably outperform BM25+symbol and does not improve Recall@10.","one_line_summary":"CSLibPremiseBench shows that a structure-guided reranker provides only modest MRR gains over BM25 for premise retrieval in Lean 4 CS theorems and does not improve Recall@10.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The source-visible proxy labels from proof references accurately stand in for the premises that would actually be needed in a realistic proof-generation or proof-repair setting.","pith_extraction_headline":"CSLibPremiseBench shows structure-guided reranking gives only modest early-rank gains over lexical baselines in Lean 4 computer-science premise retrieval."},"references":{"count":16,"sample":[{"doi":"","year":2025,"title":"Leanexplore: A search engine for lean 4 declarations, 2025","work_id":"caf84e7b-650a-41e6-9de3-fe9a418e9ba8","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"Ayers, Dragomir Radev, and Jeremy Avigad","work_id":"fa9350a1-82e0-40b6-84c8-9b63af97ad6b","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2026,"title":"Cslib: The lean computer science library, 2026","work_id":"183e90e1-b20e-409a-9cab-8a2fe5aebed8","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2021,"title":"The lean 4 theorem prover and programming language, 2021","work_id":"44987b9c-b745-4852-aee6-d3d9e9f8ccf0","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2026,"title":"Computer science as infrastructure: the spine of the lean computer science library (cslib), 2026","work_id":"e7f1f2b8-0e8d-4475-8ada-1ab1b1d021a1","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":16,"snapshot_sha256":"cce310dbdecb00c476262cbad61cd6fd9f61c4398322b1bf2ed559efb5058bcd","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"}