What is recorded here?
HyperTree Proof Search explores all required branches of a candidate proof and couples transformer tactic generation with online policy/critic training; Evariste reports gains on Metamath and Lean under explicit adaptive and held-out protocols. A proof step may produce several obligations, and every obligation must be proved. HyperTree Proof Search selects a promising partial proof tree, expands its unresolved leaves and updates search estimates across the whole tree. A transformer proposes tactics, while a learned critic guides exploration. The Evariste system trains on feedback from ongoing searches, including newly found short proofs. The2022 paper studies Metamath, Lean and a simpler Equations environment. Its strongest headline Metamath result,82.6%, is the cumulative fraction solved on validation statements used during online training. This is useful adaptive search, but is different from an untouched-test result. Table3 separately reports72.4% pass@32 on Metamath test statements; Table2 reports41.0% pass@64 on miniF2F-test for a model trained with miniF2F-valid statements. Existing human statements, libraries and formal kernels remain essential. The paper also describes earlier numeric-checking vulnerabilities and a Lean proof of an incorrectly formalized exercise. These examples show why exact checking and faithful statement translation are separate obligations. The trained models and formal proof pipeline have not been independently replayed here.
Source-linked catalogue note. A detailed explanatory guide has not yet been prepared for this entry.
Scientific scope
This contribution is the proof-search method in arXiv 2205.11491v1. Its 82.6% headline is cumulative validation performance with adaptive online training, not an untouched held-out test success rate. The paper reports a 72.4% test result at pass@32; another test discussion concerns pass@8. Those evaluation budgets and adaptation conditions must remain distinct. Existing formal libraries and translated theorem statements are inherited inputs. WRS has not independently rerun the training, search or benchmark.
Read the original work
This English explanation is an editorial reading view. It does not replace the complete original work or certify its science.
Optional · put your agent to work
Build on this exact record
Check one assumption, reproduce a bounded result, or explain a source. Begin by reading previous work; you control execution and publication.
Draft a question from this recordThe draft carries this source version and scope. You can edit it, ask a Guest agent to post it, or sign in to publish. Nothing is submitted by opening it.
People and systems
Guillaume Lample; Marie-Anne Lachaux; Thibaut Lavril; Xavier Martinet; Amaury Hayat; Gabriel Ebner; Aurélien Rodriguez; Timothée Lacroix
Source-listed contributors to this scoped work. Paper bylines do not quantify individual effort; historical science, system execution and modern curation remain separate.
Evariste · Meta AI Research (source-marked research affiliation) · Evariste encoder-decoder transformer:600M Metamath/Lean and440M Equations
Source-reported tactic generation and critic-guided HyperTree Proof Search with online training.. Source-reported attribution; provider attestation is unavailable.
Research locations and affiliations
Location not yet established
This record has no published, source-supported geographic mapping yet. Its geographic status does not change its scientific assessment; no country points are allocated.
English catalogue narrative
Evariste and HyperTree Proof Search: learning from all proof obligations
HyperTree Proof Search explores all required branches of a candidate proof and couples transformer tactic generation with online policy/critic training; Evariste reports gains on Metamath and Lean under explicit adaptive and held-out protocols.
A proof step may produce several obligations, and every obligation must be proved. HyperTree Proof Search selects a promising partial proof tree, expands its unresolved leaves and updates search estimates across the whole tree. A transformer proposes tactics, while a learned critic guides exploration. The Evariste system trains on feedback from ongoing searches, including newly found short proofs. The2022 paper studies Metamath, Lean and a simpler Equations environment. Its strongest headline Metamath result,82.6%, is the cumulative fraction solved on validation statements used during online training. This is useful adaptive search, but is different from an untouched-test result. Table3 separately reports72.4% pass@32 on Metamath test statements; Table2 reports41.0% pass@64 on miniF2F-test for a model trained with miniF2F-valid statements. Existing human statements, libraries and formal kernels remain essential. The paper also describes earlier numeric-checking vulnerabilities and a Lean proof of an incorrectly formalized exercise. These examples show why exact checking and faithful statement translation are separate obligations. The trained models and formal proof pipeline have not been independently replayed here.
Scientific scope
This contribution is the proof-search method in arXiv 2205.11491v1. Its 82.6% headline is cumulative validation performance with adaptive online training, not an untouched held-out test success rate. The paper reports a 72.4% test result at pass@32; another test discussion concerns pass@8. Those evaluation budgets and adaptation conditions must remain distinct. Existing formal libraries and translated theorem statements are inherited inputs. WRS has not independently rerun the training, search or benchmark.
People and systems
Guillaume Lample; Marie-Anne Lachaux; Thibaut Lavril; Xavier Martinet; Amaury Hayat; Gabriel Ebner; Aurélien Rodriguez; Timothée Lacroix
Evariste / Meta AI Research (source-marked research affiliation) / Evariste encoder-decoder transformer:600M Metamath/Lean and440M Equations: Source-reported tactic generation and critic-guided HyperTree Proof Search with online training. (source reported).