The recent DeepSeek-R1 release has demonstrated the immense potential of reinforcement learning (RL) in enhancing the general reasoning capabilities of large language models (LLMs). While DeepSeek-R1 and other follow-up work primarily focus on applying RL to competitive coding and math problems, this paper introduces SWE-RL, the first approach to scale RL-based LLM reasoning for real-world software engineering. Leveraging a lightweight rule-based reward (e.g., the similarity score between ground-truth and LLM-generated solutions), SWE-RL enables LLMs to autonomously recover a developer's reasoning processes and solutions by learning from extensive open-source software evolution data -- the record of a software's entire lifecycle, including its code snapshots, code changes, and events such as issues and pull requests. Trained on top of Llama 3, our resulting reasoning model, Llama3-SWE-RL-70B, achieves a 41.0% solve rate on SWE-bench Verified -- a human-verified collection of real-world GitHub issues. To our knowledge, this is the best performance reported for medium-sized (<100B) LLMs to date, even comparable to leading proprietary LLMs like GPT-4o. Surprisingly, despite performing RL solely on software evolution data, Llama3-SWE-RL has even emerged with generalized reasoning skills. For example, it shows improved results on five out-of-domain tasks, namely, function coding, library use, code reasoning, mathematics, and general language understanding, whereas a supervised-finetuning baseline even leads to performance degradation on average. Overall, SWE-RL opens up a new direction to improve the reasoning capabilities of LLMs through reinforcement learning on massive software engineering data.

We use Iris, an implementation of concurrent separation logic in the Coq proof assistant, to verify two queue data structures used for inter-process communication in an operating system under development. Our motivations are twofold. First, we wish to leverage formal verification to boost confidence in a delicate piece of industrial code that was subject to numerous revisions. Second, we aim to gain information on the cost-benefit tradeoff of applying a state-of-the-art formal verification tool in our industrial setting. On both fronts, our endeavor has been a success. The verification effort proved that the queue algorithms are correct and uncovered four algorithmic simplifications as well as bugs in client code. The simplifications involve the removal of two memory barriers, one atomic load, and one boolean check, all in a performance-sensitive part of the OS. Removing the redundant boolean check revealed unintended uses of uninitialized memory in multiple device drivers, which were fixed. The proof work was completed in person months, not years, by engineers with no prior familiarity with Iris. These findings are spurring further use of verification at Meta.

This paper presents a new static analysis for deriving upper bounds on the expected resource consumption of probabilistic programs. The analysis is fully automatic and derives symbolic bounds that are multivariate polynomials in the inputs. The new technique combines manual state-of-the-art reasoning techniques for probabilistic programs with an effective method for automatic resource-bound analysis of deterministic programs. It can be seen as both, an extension of automatic amortized resource analysis (AARA) to probabilistic programs and an automation of manual reasoning for probabilistic programs that is based on weakest preconditions. An advantage of the technique is that it combines the clarity and compositionality of a weakest-precondition calculus with the efficient automation of AARA. As a result, bound inference can be reduced to off-the-shelf LP solving in many cases and automatically-derived bounds can be interactively extended with standard program logics if the automation fails. Building on existing work, the soundness of the analysis is proved with respect to an operational semantics that is based on Markov decision processes. The effectiveness of the technique is demonstrated with a prototype implementation that is used to automatically analyze 39 challenging probabilistic programs and randomized algorithms. Experiments indicate that the derived constant factors in the bounds are very precise and even optimal for some programs.

This paper presents a new approach for automatically deriving worst-case resource bounds for C programs. The described technique combines ideas from amortized analysis and abstract interpretation in a unified framework to address four challenges for state-of-the-art techniques: compositionality, user interaction, generation of proof certificates, and scalability. Compositionality is achieved by incorporating the potential method of amortized analysis. It enables the derivation of global whole-program bounds with local derivation rules by naturally tracking size changes of variables in sequenced loops and function calls. The resource consumption of functions is described abstractly and a function call can be analyzed without access to the function body. User interaction is supported with a new mechanism that clearly separates qualitative and quantitative verification. A user can guide the analysis to derive complex non-linear bounds by using auxiliary variables and assertions. The assertions are separately proved using established qualitative techniques such as abstract interpretation or Hoare logic. Proof certificates are automatically generated from the local derivation rules. A soundness proof of the derivation system with respect to a formal cost semantics guarantees the validity of the certificates. Scalability is attained by an efficient reduction of bound inference to a linear optimization problem that can be solved by off-the-shelf LP solvers. The analysis framework is implemented in the publicly-available tool C4B. An experimental evaluation demonstrates the advantages of the new technique with a comparison of C4B with existing tools on challenging micro benchmarks and the analysis of more than 2900 lines of C code from the cBench benchmark suite.

Verified compilers guarantee the preservation of semantic properties and thus enable formal verification of programs at the source level. However, important quantitative properties such as memory and time usage still have to be verified at the machine level where interactive proofs tend to be more tedious and automation is more challenging.