Search arXiv⌕ Search
Home/Archive/arXiv

arXiv · 2609.29754

SWE-PolyVision: Benchmarking Cross-Image Abductive Reasoning for Repository-Level Software Engineering

Jiajun Wu·Leixin Sun·Zihan Tan·Yitao Liu·Shuo Li·Jiaru Qian·Shanghaoran Quan·Chuangxin Zhao·Yangxu Liao·Yang Liu·Bin Chong·Guancheng Wan
View original source ↗Read source document ↗Export metadata ↓

Abstract

Current multimodal software-engineering benchmarks expose images as additional context, but do not test whether an agent can integrate evidence distributed across images into a verified repository-level repair. We present SWE-PolyVision, an executable benchmark of 92 real tasks from 36 open-source organizations, with 48 public tasks and 44 private holdouts. The release contains 402 static images and 6 videos, with at least two visual inputs per task. Each task pairs a fixed pre-fix repository with an isolated verifier and is evaluated under the supported conditions among three access modes: Text-only, Native Vision, and Tool-mediated Vision. Across eleven coding models, visual access changes which tasks are solved, but effects depend on both model and task. Two trace-linked Native Vision cases illustrate how complementary visual and textual clues can lead to source-localized, verified repairs; controlled interventions show that this conversion is not yet stable across inputs. SWE-PolyVision thus separates the availability of multi-image evidence from its successful use in repository-level repair, without treating patch success alone as proof of explicit reasoning.

Explore related subjects

cs.SE

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Jiajun Wu, Leixin Sun, Zihan Tan, Yitao Liu, Shuo Li, Jiaru Qian, Shanghaoran Quan, Chuangxin Zhao, Yangxu Liao, Yang Liu, Bin Chong, Guancheng Wan. 2026-09-24. SWE-PolyVision: Benchmarking Cross-Image Abductive Reasoning for Repository-Level Software Engineering. https://arxiv.org/abs/2609.29754

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

Structural Enforcement of Statistical Rigor in AI-Driven Discovery: A Functional Architecture

AI-Scientist systems risk manufacturing spurious discoveries through uncontrolled multiple testing. We present a functional architecture that enforces statistical rigor at two levels: a Haskell embedded domain-specific language (the Research monad) that makes it impossible to test a hypothesis without updating the error budget, and a declarative scaffold that fixes the data flow and the statistical test, together with an OS-level sandbox that makes validation data physically absent from the environment in which LLM-generated code runs. We treat FDR control as a formal requirement and trace it to the implementation. We ground the design in a machine-checked Lean~4 formalization of LORD online false-discovery-rate (FDR) control: we derive its error budget and prove marginal FDR control, and full FDR control when thresholds do not adapt to earlier rejections. We then verify in SPARK/Ada that the LORD thresholds, computed in IEEE~754 arithmetic, never exceed the available wealth, given a margin condition that our configurations meet by a factor of at least eight; without a margin the property fails. To our knowledge this is the first machine-checked proof of an online FDR control theorem. In simulation, the architecture holds the false discovery rate near 1\% against a 5\% target, where a naive approach reaches 41\%. In end-to-end case studies, a valid test avoids the false discoveries a flawed one produces, yet still finds real effects when the data allow. An adversarial evaluation confirms that, inside the sandbox, generated code cannot read the held-out data even when given its exact path.

cs.SE↗
arXiv2026

LLM-based Vulnerability Detection at Project Scale: An Empirical Study

As software complexity grows, automated vulnerability detection becomes increasingly important. Recent LLM-based detectors combine semantic reasoning with static analysis for project-scale scanning, but their practical effectiveness and failure causes remain unclear. We present the first comprehensive empirical study of specialized LLM-based detectors at project scale, evaluating five specialized methods, two general-purpose agents, and four traditional static analyzers on 265 known C/C++ and Java vulnerabilities and 24 active open-source projects/modules. Using Codex-assisted labeling with stratified manual validation, we analyze 6,442 sampled warnings and classify 5,896 false positives using a taxonomy derived from 355 manually inspected reports. Our study yields three findings. First, general-purpose agents achieve the highest recall but struggle with complex code, while both LLM-based and traditional methods suffer from incomplete API modeling. Second, many tools exhibit high false discovery rates on real-world projects/modules. Truncated interprocedural context and incorrect program-point classification are the dominant false-positive causes, with LLM-based tools exhibiting additional reasoning and prompt-compliance failures. Third, LLM-based methods incur substantial costs: hundreds of thousands to hundreds of millions of tokens, API charges exceeding $1,000 per project/module, and runtimes ranging from hours to days. These findings reveal limitations in the robustness, reliability, and scalability of current detectors and inform directions for more effective and practical project-scale vulnerability detection.

cs.SE↗
arXiv2026

A Framework for Identifying, Categorizing, and Explaining Bias in AI-Generated Code

As Large Language Models (LLMs) become integrated into software development workflows, concerns regarding unintentional biases in AI-generated code. Although evidence suggests these biases exist, limited research has systematically identified, categorized, and explained them. This study investigates bias in AI-generated code and evaluates whether LLMs can reliably identify and explain it through a taxonomy-driven framework. We extended an existing dataset of biased AI-generated Python code and manually annotated snippets with bias categories and human-authored justifications to establish a ground-truth dataset. Using this dataset, we evaluated proprietary and open-source LLMs as automated bias detection and justification systems through ICL. Finally, we analyzed similarity between LLM-generated explanations and human-authored justifications using structured justification and code identification metrics. Our findings demonstrate that LLMs can effectively support code bias identification and explanation. Gemini achieved 80.14% classification accuracy, with 84.0% precision and 95.7% recall, while the best open-source alternative, Qwen3-coder, achieved 82.45% accuracy, 68.64% precision, and 80.22% recall. Additionally, the models achieved justification similarity scores of 80.4% and 80.14%, respectively, relative to human-authored reasoning, and code identification similarity scores of 86.0% and 87.82%. These results suggest that LLMs can detect biased logic in generated Python code and produce explanations that substantially align with expert interpretations.

cs.SE↗