PACIFIC: a framework for generating benchmarks to check Precise Automatically Checked Instruction Following In Code

PACIFIC: a framework for generating benchmarks to check Precise Automatically Checked Instruction Following In Code Itay Dreyfuss Antonio Abu Nassar Samuel Ackerman Axel Ben David Eitan Farchi Rami Katan Orna Raz Marcel Zalmanovici IBM Research Haifa, Israel {Itay.Dreyfuss,Antonio.Abu.Nassar,Samuel.Ackerman,Axel.Bendavid}@ibm.com {Rami.Katan,farchi,ornar,marcel}@il.ibm.com Abstract Large Language Model (LLM)-based code assistants have emerged as a powerful application of generative AI, demonstrating impres- sive capabilities in code generation and comprehension. A key requirement for these systems is their ability to accurately follow user instructions. We present Precise Automatically Checked In- struction Following In Code (PACIFIC), a novel framework designed to automatically generate benchmarks that rigorously assess se- quential instruction-following and code dry-running capabilities in LLMs, while allowing control over benchmark difficulty. PA- CIFIC produces benchmark variants with clearly defined expected outputs, enabling straightforward and reliable evaluation through simple output comparisons. In contrast to existing approaches that often rely on tool usage or agentic behavior, our work isolates and evaluates the LLM’s intrinsic ability to reason through code behavior step-by-step without execution (dry running) and to fol- low instructions. Furthermore, our framework mitigates training data contamination by facilitating effortless generation of novel benchmark variations. We validate our framework by generating a suite of benchmarks spanning a range of difficulty levels and evaluating multiple state-of-the-art LLMs. Our results demonstrate that PACIFIC can produce increasingly challenging benchmarks that effectively differentiate instruction-following and dry running capabilities, even among advanced models. Overall, our framework offers a scalable, contamination-resilient methodology for assessing core competencies of LLMs in code-related tasks. Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for components of this work owned by others than the author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Request permissions from permissions@acm.org. AI-SQE, Rio de Janeiro, Brazil © 2025 Copyright held by the owner/author(s). Publication rights licensed to ACM. ACM ISBN 978-x-xxxx-xxxx-x/YYYY/MM https://doi.org/10.1145/nnnnnnn.nnnnnnn CCS Concepts • Software and its engineering →Empirical software valida- tion. Keywords Code instruction following benchmark, automated expected results generation, controlled level of benchmark difficulty, automated benchmark generation ACM Reference Format: Itay Dreyfuss, Antonio Abu Nassar, Samuel Ackerman, Axel Ben David, Eitan Farchi, Rami Katan, Orna Raz, and Marcel Zalmanovici. 2026. PACIFIC: a framework for generating benchmarks to check Precise Automatically Checked Instruction Following In Code. In Proceedings of The 1st Interna- tional Workshop on AI for Software Quality Evaluation: Judgment, Metrics, Benchmarks, and Beyond (AI-SQE). ACM, New York, NY, USA, 14 pages. https://doi.org/10.1145/nnnnnnn.nnnnnnn Accepted at AI-SQE 2026. 1 Introduction Large Language Models (LLMs) have demonstrated remarkable ca- pabilities in natural language understanding and generation, lead- ing to the development of highly effective code assistants. These systems rely heavily on their ability to follow user instructions accurately, a trait that is central to their utility in real-world soft- ware engineering tasks. While high-quality benchmarks exist for evaluating instruction-following in natural language domains, such as IFEval[21], there remains a major gap in the availability of bench- marks tailored to code-related instruction following. Existing bench- marks in this space often depend on complex evaluation methods, such as using LLMs themselves as judges, and typically lack deter- ministic metrics and a focus on code dry-running. To address these limitations, we present PACIFIC, a framework designed to automatically generate benchmarks that evaluate both sequential instruction-following and code dry-running capabilities of LLMs. The framework emphasizes the ability of LLMs to simulate arXiv:2512.10713v2 [cs.SE] 22 Dec 2025 AI-SQE, April 14, 2026, Rio de Janeiro, Brazil Dreyfuss, et al. code execution (i.e. dry-running) without relying on external tools or agentic mechanisms. The framework operates on a set of instructions that are used as building blocks. Each instruction describes actions to perform on the output of a given algorithm when dry-running it with a given input. Below are a couple of

…(Full text truncated)…

This content is AI-processed based on ArXiv data.