Search for: All records

Formal verification using proof assistants, such as Coq, enables the creation of high-quality software. However, the verification process requires significant expertise and manual effort to write proofs. Recent work has explored automating proof synthesis using machine learning and large language models (LLMs). This work has shown that identifying relevant premises, such as lemmas and definitions, can aid synthesis. We present Rango, a fully automated proof synthesis tool for Coq that automatically identifies relevant premises and also similar proofs from the current project and uses them during synthesis. Rango uses retrieval augmentation at every step of the proof to automatically determine which proofs and premises to include in the context of its fine-tuned LLM. In this way, Rango adapts to the project and to the evolving state of the proof. We create a new dataset, CoqStoq, of 2,226 open-source Coq projects and 196,929 theorems from GitHub, which includes both training data and a curated evaluation benchmark of well-maintained projects. On this benchmark, Rango synthesizes proofs for 32.0% of the theorems, which is 29% more theorems than the prior state-of-the- art tool Tactician. Our evaluation also shows that Rango adding relevant proofs to its context leads to a 47% increase in the number of theorems proven. 

Recent advances in Automated Theorem Proving have shown the effectiveness of leveraging a (large) language model that generates tactics (i.e. proof steps) to search through proof states. The current model, while trained solely on successful proof paths, faces a discrepancy at the inference stage, as it must sample and try various tactics at each proof state until finding success, unlike its training which does not incorporate learning from failed attempts. Intuitively, a tactic that leads to a failed search path would indicate that similar tactics should receive less attention during the following trials. In this paper, we demonstrate the benefit of training models that additionally learn from failed search paths. Facing the lack of such trial-and-error data in existing open-source theorem-proving datasets, we curate a dataset on intuitionistic propositional logic theorems and formalize it in Lean, such that we can reliably check the correctness of proofs. We compare our model trained on relatively short trial-and-error information (TRIALMASTER) with models trained only on the correct paths and discover that the former solves more unseen theorems with lower trial searches. 

Interactive proofs of theorems often require auxiliary helper lemmas to prove the desired theorem. Existing approaches for automatically synthesizing helper lemmas fall into two broad categories. Some approaches are goal-directed, producing lemmas specifically to help a user make progress from a given proof state, but they have limited expressiveness in terms of the lemmas that can be produced. Other approaches are highly expressive, able to generate arbitrary lemmas from a given grammar, but they are completely undirected and hence not amenable to interactive usage. In this paper, we develop an approach to lemma synthesis that is both goal-directed and expressive. The key novelty is a technique for reducing lemma synthesis to a data-driven program synthesis problem, whereby examples for synthesis are generated from the current proof state. We also describe a technique to systematically introduce new variables for lemma synthesis, as well as techniques for filtering and ranking candidate lemmas for presentation to the user. We implement these ideas in a tool called lfind, which can be run as a Coq tactic. In an evaluation on four benchmark suites, lfind produces useful lemmas in 68% of the cases where a human prover used a lemma to make progress. In these cases lfind synthesizes a lemma that either enables a fully automated proof of the original goal or that matches the human-provided lemma. 

Regular expressions are a popular target for programming by example (PBE) systems, which seek to learn regexes from user-provided examples. Synthesizing from only positive examples remains an unsolved challenge, as the unrestricted search space makes it difficult to avoid over- and under- generalizing. Prior work has approached this in two ways: search-based techniques which require extra input, such as user feedback and/or a natural language description, and neural techniques. The former puts an extra burden on the user, while the latter requires large representative training data sets which are almost nonexistent for this domain. To tackle this challenge we present Regex+, a search-based syn- thesizer that infers regexes from just a few positive examples. Regex+ avoids over/under-generalization by using minimum description length (MDL) learning, adapted to version space algebras in order to efficiently search for an optimal regex according to a compositional MDL ranking function. Our evaluation shows that Regex+ more than triples the accu- racy of existing neural and search-based regex synthesizers on benchmarks with only positive examples