{"paper":{"title":"Compile to Compress: Boosting Formal Theorem Provers by Compiler Outputs","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Compiler outputs compress diverse proof attempts into a compact set of failure modes that support local corrections and stronger theorem proving.","cross_cats":["cs.AI","cs.LO","cs.PL"],"primary_cat":"cs.LG","authors_text":"Guchan Li, Hongning Wang, Rui Tian","submitted_at":"2026-03-13T01:33:20Z","abstract_excerpt":"Large language models (LLMs) have demonstrated significant potential in formal theorem proving, yet state-of-the-art performance often necessitates prohibitive test-time compute via massive roll-outs or extended context windows. In this work, we address this scalability bottleneck by exploiting an informative structure in formal verification: the observation that compilers map a vast space of diverse proof attempts to a compact set of structured failure modes. We introduce a learning-to-refine framework that leverages this compression to perform efficient learning and proof exploration. We per"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"our approach achieves state-of-the-art performance on PutnamBench among publicly reported ∼8B and ∼32B parameter models under comparable test-time budgets","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"that compilers map a vast space of diverse proof attempts to a compact set of structured failure modes that can be leveraged for efficient learning and proof exploration without accumulating long histories of attempts","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Compiler feedback compresses diverse proof attempts into structured failure modes for efficient local refinement, boosting LLM-based theorem provers to SOTA on PutnamBench for 8B and 32B models under fixed test-time budgets.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Compiler outputs compress diverse proof attempts into a compact set of failure modes that support local corrections and stronger theorem proving.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"9d8a70447989471a8167eb3fd4526f32fea3ebd23c3453067172bfec1ca0fe9b"},"source":{"id":"2604.18587","kind":"arxiv","version":2},"verdict":{"id":"751764c0-9b85-496d-adc3-362d9d142c74","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T12:38:53.288785Z","strongest_claim":"our approach achieves state-of-the-art performance on PutnamBench among publicly reported ∼8B and ∼32B parameter models under comparable test-time budgets","one_line_summary":"Compiler feedback compresses diverse proof attempts into structured failure modes for efficient local refinement, boosting LLM-based theorem provers to SOTA on PutnamBench for 8B and 32B models under fixed test-time budgets.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"that compilers map a vast space of diverse proof attempts to a compact set of structured failure modes that can be leveraged for efficient learning and proof exploration without accumulating long histories of attempts","pith_extraction_headline":"Compiler outputs compress diverse proof attempts into a compact set of failure modes that support local corrections and stronger theorem proving."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2604.18587/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":2,"snapshot_sha256":"3d67ba3a59f67d66b04218410b8324c85731d1a39d5919baadd0d43e80e5c0c2"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}