A revised arXiv preprint argues that although large language models produce fluent-looking code, they offer no formal guarantee that their output is syntactically valid, and they usually ignore the hierarchical grammar of the target language. The work, titled "Weave of Formal Thought," proposes incorporating that structure directly rather than relying on surface-level fluency alone.
A new arXiv preprint reviews how diffusion-based large language models can be applied to code generation, an area currently dominated by left-to-right autoregressive decoding. The authors examine the limitations of standard autoregressive generation and assess whether diffusion approaches offer advantages for producing source code. The work is a replacement submission (v3) to the cs.AI category.
Researchers released cvxgenrust, an open-source tool that produces tailored Rust code for solving families of parameterized convex optimization problems written in CVXPY. The system canonicalizes a problem family and derives affine mappings onto Clarabel cone programs. The paper targets use cases where a solver must be recompiled for many related parameter values rather than solved from scratch each time.
A new arXiv paper introduces ModiGen, a large language model-driven workflow designed to handle several Modelica code generation tasks at once. Modelica is used to simulate complex physical systems, but building and tuning such models typically demands deep domain knowledge. The work targets lowering that barrier by having LLMs assist with code creation for these simulation models.
A new arXiv preprint introduces FVSpec, a benchmark that repurposes property-based tests drawn from real software projects as proof challenges in the Lean theorem prover. The work targets the growing need to verify machine-generated code, arguing that AI systems themselves could take on much of that verification work. It also notes that the field still lacks a clear picture of how well current tools and models handle such tasks.
A new arXiv paper introduces IWC-Bench, a benchmark that assesses LLM-generated web applications from a software testing angle. The authors argue that existing static benchmarks can reward functionality that appears in source code but does not actually work when the app runs. Their approach aims to make automated evaluation of generated web apps track human judgments more closely.
A new arXiv paper introduces SemVerBench, a benchmark that tests how well large language model coding agents interpret version constraints such as caret and tilde ranges. The authors report that confusing caret with tilde semantics (for example, treating ^1.2.3 as >=1.3.0) trips up nearly every model on Cargo, while standard PEP 440 prefix matching is handled more consistently. The work suggests dependency resolution remains a weak spot for agents that automate package management.
Researchers have introduced DiffuTester, a technique to accelerate unit test generation using diffusion large language models. The method extracts structural patterns from code to guide parallel generation, aiming to make automated testing faster and more scalable. The work addresses the need for efficient large-scale software testing.
A new arXiv paper introduces MTAC-IFBench, a benchmark aimed at measuring how well large language model agents follow instructions across multi-turn coding sessions. The work targets agentic software engineering, where models plan, run code, and call external tools over successive steps rather than producing a single answer. It addresses evaluation beyond functional correctness, focusing on whether agents keep to the constraints given to them.
A new arXiv paper addresses the gap between code that runs correctly and code that is secure when generated by large language models. The authors argue that progress has been limited by existing benchmarks that are small and not executable, making security flaws hard to measure reliably. Their approach combines task-adaptive modeling of vulnerabilities with an execution-based benchmark intended to evaluate both functional correctness and security.
A research paper posted to arXiv examines how large language models struggle when asked to write code against APIs from private, non-public libraries. The authors note that most existing methods focus on public code and propose teaching models to work with private library documentation. The work is a replacement submission on the cs.AI listing.
A new arXiv paper proposes SAILOR, an interactive approach in which a language model turns natural-language optimization problems into solver-ready code. Because such descriptions often omit or leave vague the numerical values a solver needs, the method draws on solver feedback to recover missing costs, capacities, demands, bounds and penalties. The work targets the gap between fluent problem descriptions and the complete, precise inputs that optimization software requires.
A new arXiv paper introduces CodeTS, a method that turns natural-language descriptions into time series by generating and running executable code rather than sampling outputs directly. This design makes the resulting synthetic data verifiable and suited to scenarios where real observations are scarce or expensive to collect. The work appears in the cs.LG and cs.AI listings.
A new arXiv paper introduces CIT-CAD, a method that converts natural-language design intent into executable, editable parametric CAD programs. The approach relies on a constraint intent tree to structure the design specification, and it includes a verification step intended to keep generated code faithful to the original request. The authors frame this as progress toward CAD systems built on large language models that produce reliable, reusable output.
A newly posted arXiv thesis investigates whether large language models can gauge the reliability of the code they produce, addressing the problem of fluent but functionally incorrect output. The work focuses on introspective uncertainty estimation as a way to flag low-confidence generations in software engineering workflows. The abstract is truncated, so the full methods and results are not yet detailed in the listing.
A Hacker News discussion raised the question of how much of the F-Droid Android app repository, including its code and app listings, was produced with large language models. Commenters debated methods for estimating AI-written content and what such provenance questions mean for open-source projects. No verified measurement was reached in the thread.
A German-language podcast episode features host Richard Seidl interviewing Dr. Benjamin Hummel about how programmers' daily work and accountability change as AI tools take over code generation. The discussion centers on what remains for developers to do and who is responsible for AI-written code. It offers a practical look at shifting skill demands in software testing and development.
A new arXiv paper investigates the natural-language comments that large language models produce while writing code, which end up as part of the context for generating subsequent code. It remains unclear which properties of these comments actually affect the quality of the generated code. The study sets out to identify what makes such self-generated comments helpful rather than distracting.
A newly posted arXiv paper outlines a software delivery model called A-JIT, in which AI agents assemble applications on the fly rather than shipping pre-built, fixed artifacts. The authors argue that construction should move from a static step completed before execution to a dynamic process performed at the moment code is actually needed.
A new arXiv paper investigates whether large language models can add exception-related code to existing software, covering throw statements, the conditions that guard them, and try/catch blocks. The goal is to automatically equip codebases with better handling of exceptional states, an area the authors identify as essential but often underdeveloped. The work is aimed at developers seeking AI-assisted ways to strengthen error handling in mature systems.
A new arXiv paper proposes SGA, a plug-and-play geometric verification method for educational videos generated by large language models. Existing frameworks that turn LLM output into executable animation code, such as Manim, often produce spatially incorrect or hard-to-read visuals. The approach aims to check and correct spatial accuracy and legibility without redesigning the underlying generation pipeline.
A new arXiv paper argues that existing benchmarks for LLM-generated scientific code rely too heavily on functional correctness or task completion, which is inadequate for software built on production HPC libraries. The authors introduce an agentic evaluation framework that assesses AI-generated code in the PETSc numerical library across a broader set of criteria. The approach aims to give researchers a more thorough way to judge whether generated scientific software is fit for real-world use.