Formally verifying software is a highly desirable but labor-intensive task. Recent work has developed methods to automate formal verification using proof assistants, such as Coq and Isabelle/HOL, e.g., by training a model to predict one proof step at a time and using that model to search through the space of possible proofs. This paper introduces a new method to automate formal verification: We use large language models, trained on natural language and code and fine-tuned on proofs, to generate whole proofs at once. We then demonstrate that a model fine-tuned to repair generated proofs further increasing proving power. This paper: (1) Demonstrates that whole-proof generation using transformers is possible and is as effective but more efficient than search-based techniques. (2) Demonstrates that giving the learned model additional context, such as a prior failed proof attempt and the ensuing error message, results in proof repair that further improves automated proof generation. (3) Establishes, together with prior work, a new state of the art for fully automated proof synthesis. We reify our method in a prototype, Baldur, and evaluate it on a benchmark of 6,336 Isabelle/HOL theorems and their proofs, empirically showing the effectiveness of whole-proof generation, repair, and added context. We also show that Baldur complements the state-of-the-art tool, Thor, by automatically generating proofs for an additional 8.7% of the theorems. Together, Baldur and Thor can prove 65.7% of the theorems fully automatically. This paper paves the way for new research into using large language models for automating formal verification.

Autoformalization is the process of automatically translating from natural language mathematics to formal specifications and proofs. A successful autoformalization system could advance the fields of formal verification, program synthesis, and artificial intelligence. While the long-term goal of autoformalization seemed elusive for a long time, we show large language models provide new prospects towards this goal. We make the surprising observation that LLMs can correctly translate a significant portion ($25.3\%$) of mathematical competition problems perfectly to formal specifications in Isabelle/HOL. We demonstrate the usefulness of this process by improving a previously introduced neural theorem prover via training on these autoformalized theorems. Our methodology results in a new state-of-the-art result on the MiniF2F theorem proving benchmark, improving the proof rate from $29.6\%$ to $35.2\%$.

In this work, we demonstrate the feasibility and usefulness of autoformalization in the context of the newly introduced MiniF2F [10] benchmark. We use large language models to translate several thousands of informal problems into Isabelle and use them to improve our neural theorem prover. We find that transformer-based [7] language models trained on a large amount of web data are capable of formalizing mathematical competition problem statements with a relatively high success rate and the resulting statements can be used for creating new correct proofs that can be used for fine-tuning a neural theorem prover for improved proof automation. Using this methodology, we achieve a new state of the art on the MiniF2F benchmark.

Language models typically need to be trained or finetuned in order to acquire new knowledge, which involves updating their weights. We instead envision language models that can simply read and memorize new data at inference time, thus acquiring new knowledge immediately. In this work, we extend language models with the ability to memorize the internal representations of past inputs. We demonstrate that an approximate kNN lookup into a non-differentiable memory of recent (key, value) pairs improves language modeling across various benchmarks and tasks, including generic webtext (C4), math papers (arXiv), books (PG-19), code (Github), as well as formal theorems (Isabelle). We show that the performance steadily improves when we increase the size of memory up to 262K tokens. On benchmarks including code and mathematics, we find that the model is capable of making use of newly defined functions and theorems during test time.

In this paper, we demonstrate how to do automated theorem proving in the presence of a large knowledge base of potential premises without learning from human proofs. We suggest an exploration mechanism that mixes in additional premises selected by a tf-idf (term frequency-inverse document frequency) based lookup in a deep reinforcement learning scenario. This helps with exploring and learning which premises are relevant for proving a new theorem. Our experiments show that the theorem prover trained with this exploration mechanism outperforms provers that are trained only on human proofs. It approaches the performance of a prover trained by a combination of imitation and reinforcement learning. We perform multiple experiments to understand the importance of the underlying assumptions that make our exploration approach work, thus explaining our design choices.

We present a very simple algorithm for attention that requires $O(1)$ memory with respect to sequence length and an extension to self-attention that requires $O(\log n)$ memory. This is in contrast with the frequently stated belief that self-attention requires $O(n^2)$ memory. While the time complexity is still $O(n^2)$, device memory rather than compute capability is often the limiting factor on modern accelerators. Thus, reducing the memory requirements of attention allows processing of longer sequences than might otherwise be feasible. We provide a practical implementation for accelerators that requires $O(\sqrt{n})$ memory, is numerically stable, and is within a few percent of the runtime of the standard implementation of attention. We also demonstrate how to differentiate the function while remaining memory-efficient. For sequence length 16384, the memory overhead of self-attention is reduced by 59X for inference and by 32X for differentiation.

Over the recent years deep learning has found successful applications in mathematical reasoning. Today, we can predict fine-grained proof steps, relevant premises, and even useful conjectures using neural networks. This extended abstract summarizes recent developments of machine learning in mathematical reasoning and the vision of the N2Formal group at Google Research to create an automatic mathematician. The second part discusses the key challenges on the road ahead.

While designing inductive bias in neural architectures has been widely studied, we hypothesize that transformer networks are flexible enough to learn inductive bias from suitable generic tasks. Here, we replace architecture engineering by encoding inductive bias in the form of datasets. Inspired by Peirce's view that deduction, induction, and abduction are the primitives of reasoning, we design three synthetic tasks that are intended to require the model to have these three abilities. We specifically design these tasks to be synthetic and devoid of mathematical knowledge to ensure that only the fundamental reasoning biases can be learned from these tasks. This defines a new pre-training methodology called "LIME" (Learning Inductive bias for Mathematical rEasoning). Models trained with LIME significantly outperform vanilla transformers on four very different large mathematical reasoning benchmarks. Unlike dominating the computation cost as traditional pre-training approaches, LIME requires only a small fraction of the computation cost of the typical downstream task. The code for generating LIME tasks is available at https://github.com/tonywu95/LIME.

We train hierarchical Transformers on the task of synthesizing hardware circuits directly out of high-level logical specifications in linear-time temporal logic (LTL). The LTL synthesis problem is a well-known algorithmic challenge with a long history and an annual competition is organized to track the improvement of algorithms and tooling over time. New approaches using machine learning might open a lot of possibilities in this area, but suffer from the lack of sufficient amounts of training data. In this paper, we consider a method to generate large amounts of additional training data, i.e., pairs of specifications and circuits implementing them. We ensure that this synthetic data is sufficiently close to human-written specifications by mining common patterns from the specifications used in the synthesis competitions. We show that hierarchical Transformers trained on this synthetic data solve a significant portion of problems from the synthesis competitions, and even out-of-distribution examples from a recent case study.

We demonstrate that self-supervised language modeling applied to mathematical formulas enables logical reasoning. To measure the logical reasoning abilities of language models, we formulate several evaluation (downstream) tasks, such as inferring types, suggesting missing assumptions and completing equalities. For training language models for formal mathematics, we propose a novel skip-tree task. We find that models trained on the skip-tree task show surprisingly strong mathematical reasoning abilities, and outperform models trained on standard skip-sequence tasks. We also analyze the models' ability to formulate new conjectures by measuring how often the predictions are provable and useful in other proofs.

We study two fundamental questions in neuro-symbolic computing: can deep learning tackle challenging problems in logics end-to-end, and can neural networks learn the semantics of logics. In this work we focus on linear-time temporal logic (LTL), as it is widely used in verification. We train a Transformer on the problem to directly predict a solution, i.e. a trace, to a given LTL formula. The training data is generated with classical solvers, which, however, only provide one of many possible solutions to each formula. We demonstrate that it is sufficient to train on those particular solutions to formulas, and that Transformers can predict solutions even to formulas from benchmarks from the literature on which the classical solver timed out. Transformers also generalize to the semantics of the logics: while they often deviate from the solutions found by the classical solvers, they still predict correct solutions to most formulas.

We demonstrate how to learn efficient heuristics for automated reasoning algorithms for quantified Boolean formulas through deep reinforcement learning. We focus on a backtracking search algorithm, which can already solve formulas of impressive size - up to hundreds of thousands of variables. The main challenge is to find a representation of these formulas that lends itself to making predictions in a scalable way. For a family of challenging problems, we learned a heuristic that solves significantly more formulas compared to the existing handwritten heuristics.

We design and conduct a simple experiment to study whether neural networks can perform several steps of approximate reasoning in a fixed dimensional latent space. The set of rewrites (i.e. transformations) that can be successfully performed on a statement represents essential semantic features of the statement. We can compress this information by embedding the formula in a vector space, such that the vector associated with a statement can be used to predict whether a statement can be rewritten by other theorems. Predicting the embedding of a formula generated by some rewrite rule is naturally viewed as approximate reasoning in the latent space. In order to measure the effectiveness of this reasoning, we perform approximate deduction sequences in the latent space and use the resulting embedding to inform the semantic features of the corresponding formal statement (which is obtained by performing the corresponding rewrite sequence using real formulas). Our experiments show that graph neural networks can make non-trivial predictions about the rewrite-success of statements, even when they propagate predicted latent representations for several steps. Since our corpus of mathematical formulas includes a wide variety of mathematical disciplines, this experiment is a strong indicator for the feasibility of deduction in latent space in general.

This paper presents the first use of graph neural networks (GNNs) for higher-order proof search and demonstrates that GNNs can improve upon state-of-the-art results in this domain. Interactive, higher-order theorem provers allow for the formalization of most mathematical theories and have been shown to pose a significant challenge for deep learning. Higher-order logic is highly expressive and, even though it is well-structured with a clearly defined grammar and semantics, there still remains no well-established method to convert formulas into graph-based representations. In this paper, we consider several graphical representations of higher-order logic and evaluate them against the HOList benchmark for higher-order theorem proving.

We examine whether self-supervised language modeling applied to mathematical formulas enables logical reasoning. We suggest several logical reasoning tasks that can be used to evaluate language models trained on formal mathematical statements, such as type inference, suggesting missing assumptions and completing equalities. To train language models for formal mathematics, we propose a novel skip-tree task. We find that models trained on the skip-tree task show surprisingly strong mathematical reasoning abilities, and outperform models trained on standard skip-sequence tasks. We also analyze the models' ability to formulate new conjectures by measuring how often the predictions are provable and useful in other proofs.

We show that neural networks can learn the semantics of propositional and linear-time temporal logic (LTL) from imperfect training data. Instead of only predicting the truth value of a formula, we use a Transformer architecture to predict the solution for a given formula, e.g., a variable assignment for a formula in propositional logic. Most formulas have many solutions and the training data thus depends on the particularities of the generator. We make the surprising observation that while the Transformer does not perfectly predict the generator's output, it still produces correct solutions to almost all formulas, even when its prediction deviates from the generator. It appears that it is easier to learn the semantics of the logics than the particularities of the generator. We observe that the Transformer preserves this semantic generalization even when challenged with formulas of a size it has never encountered before. Surprisingly, the Transformer solves almost all LTL formulas in our test set including those for which our generator timed out.

In this paper, we systematize the modeling of probabilistic systems for the purpose of analyzing them with model counting techniques. Starting from unbiased coin flips, we show how to model biased coins, correlated coins, and distributions over finite sets. From there, we continue with modeling sequential systems, such as Markov chains, and revisit the relationship between weighted and unweighted model counting. Thereby, this work provides a conceptual framework for deriving #SAT encodings for probabilistic inference.

Dependency quantified Boolean formulas (DQBF) is a logic admitting existential quantification over Boolean functions, which allows us to elegantly state synthesis problems in verification such as the search for invariants, programs, or winning regions of games. In this paper, we lift the clausal abstraction algorithm for quantified Boolean formulas (QBF) to DQBF. Clausal abstraction for QBF is an abstraction refinement algorithm that operates on a sequence of abstractions that represent the different quantifier levels. For DQBF we need to generalize this principle to partial orders of abstractions. The two challenges to overcome are: (1) Clauses may contain literals with incomparable dependencies, which we address by the recently proposed proof rule called Fork Extension, and (2) existential variables may have spurious dependencies, which we prevent by tracking consistency requirements during the execution. Our implementation \(\textsc {dCAQE}\) solves significantly more formulas than the existing DQBF algorithms.

Quantifier elimination and its cousin functional synthesis are fundamental problems in automated reasoning that could be used in many applications of formal methods. But, effective algorithms are still elusive. In this paper, we suggest a simple modification to a QBF algorithm to adapt it for quantifier elimination and functional synthesis. We demonstrate that the approach significantly outperforms previous algorithms for functional synthesis.