Denis Efremov

Long Phan, Alice Gatti, Ziwen Han et al. · amazon-science, apple-ml

Benchmarks are important tools for tracking the rapid advancements in large language model (LLM) capabilities. However, benchmarks are not keeping pace in difficulty: LLMs now achieve over 90\% accuracy on popular benchmarks like MMLU, limiting informed measurement of state-of-the-art LLM capabilities. In response, we introduce Humanity's Last Exam (HLE), a multi-modal benchmark at the frontier of human knowledge, designed to be the final closed-ended academic benchmark of its kind with broad subject coverage. HLE consists of 2,500 questions across dozens of subjects, including mathematics, humanities, and the natural sciences. HLE is developed globally by subject-matter experts and consists of multiple-choice and short-answer questions suitable for automated grading. Each question has a known solution that is unambiguous and easily verifiable, but cannot be quickly answered via internet retrieval. State-of-the-art LLMs demonstrate low accuracy and calibration on HLE, highlighting a significant gap between current LLM capabilities and the expert human frontier on closed-ended academic questions. To inform research and policymaking upon a clear understanding of model capabilities, we publicly release HLE at https://lastexam.ai.

Denis Efremov, Ilya Shchepetkov

The Linux kernel is one of the most important Free/Libre Open Source Software (FLOSS) projects. It is installed on billions of devices all over the world, which process various sensitive, confidential or simply private data. It is crucial to establish and prove its security properties. This work-in-progress paper presents a method to verify the Linux kernel for conformance with an abstract security policy model written in the Event-B specification language. The method is based on system call tracing and aims at checking that the results of system call execution do not lead to accesses that violate security policy requirements. As a basis for it, we use an additional Event-B specification of the Linux system call interface that is formally proved to satisfy all the requirements of the security policy model. In order to perform the conformance checks we use it to reproduce intercepted system calls and verify accesses.

Grigoriy Volkov, Mikhail Mandrykin, Denis Efremov

This paper describes the development of an auto-active verification technique in the Frama-C framework. We outline the lemma functions method and present the corresponding ACSL extension, its implementation in Frama-C, and evaluation on a set of string-manipulating functions from the Linux kernel. We illustrate the benefits our approach can bring concerning the effort required to prove lemmas, compared to the approach based on interactive provers such as Coq. Current limitations of the method and its implementation are discussed.

Denis Efremov, Mikhail Mandrykin, Alexey Khoroshilov

This paper presents results from the development and evaluation of a deductive verification benchmark consisting of 26 unmodified Linux kernel library functions implementing conventional memory and string operations. The formal contract of the functions was extracted from their source code and was represented in the form of preconditions and postconditions. The correctness of 23 functions was completely proved using AstraVer toolset, although success for 11 functions was achieved using 2 new specification language constructs. Another 2 functions were proved after a minor modification of their source code, while the final one cannot be completely proved using the existing memory model. The benchmark can be used for the testing and evaluation of deductive verification tools and as a starting point for verifying other parts of the Linux kernel.