arXiv subjects

Jacques Klein

Publications and source records attributed to Jacques Klein.

At least 19 recordsLinked to original sources

arXiv2026

Program Bindings Inside Code Models

Code language models complete, review, and repair programs without exposing the program structure encoded in their hidden states. In this paper, we test whether the hidden states of a code model can be described and edited using the information a compiler front end computes. For each token of a Python program, standard analyzers provide its position in the abstract syntax tree, its scope, its definition-use links, and its syntactic category. We fit a decomposition that expresses every hidden state of a frozen model as a sum of one term per token, each combining the token with this structural label, and we replace the model's own states with the decomposition at every layer while it predicts the output of a small program. If the model still answers correctly, the decomposition preserves the information the model uses. We apply this test to 14 open models, seven of them trained for code, and compare against a baseline decomposition of the same size built from token positions. The static-analysis decomposition preserves the model's answers better than the positional one in 10 of the 14 models, including six of the seven code models, and deleting the definition-use links from it removes most of that advantage on the two models tested. Editing the decomposition, for instance rebinding the argument of a call, makes StarCoder2-15B produce the answer of the edited program in 90 of 100 cases, compared with 19 for the positional baseline. The Qwen2.5 family is the exception on several measures. For most of the models tested, standard static analysis therefore provides both a description of what the hidden states encode about a program and a way to change it without retraining.

cs.SE↗

arXiv2026

Securing Agent-Generated Code with Its Own Development History

Pull requests written by coding agents can pass a project's test suite while omitting a guard or a validation step. To find the missing condition, a reviewer currently relies on static analysis of the final code. An agent-generated change, however, is accompanied by a development history that a human commit does not provide. The agent's trajectory records each edit together with the agent's stated assumptions, and structural differencing shows how each changed function evolved. We present TRACER, a workflow that uses this development history as evidence for security review. TRACER tracks changes at the symbol level, treats the agent's recorded rationale as claims to be verified against the current code, accepts a repair only after assessment and functional validation, and attaches to each accepted repair an executable check that later changes must satisfy. On the SusVibes benchmark, symbol-level tracking places the vulnerable symbol in the review scope for 94.6% of agent patches, compared with 13.4% for the primary scanner warning. With two investigator models, the configuration that includes the agent's rationale yields the largest number of verified security repairs while preserving functionality, and adding AST history to it does not increase that number. A funnel analysis identifies where the remaining candidates are lost. In a study of structural-biology software, the combination of repair memory and a boundary check allows a subsequent feature request to succeed in nine of ten tasks without reintroducing the repaired vulnerability, whereas functional feedback alone succeeds in four and reintroduces the vulnerability in three.

cs.SE↗

arXiv2026

Not All Similarity Is Leakage: Distribution-Aware Identity Leakage Assessment in Synthetic Face Datasets

Synthetic data provides the means for more ethical and less biased facial recognition. Nonetheless, the limitations of generative models and the conditioning process they rely on pose a risk of identity leakage, resulting in synthetic samples that are overly similar to those seen in a generator's training set, mitigating privacy benefits. In this work, we show that previously proposed approaches are affected by two key limitations: high classifier dependence and a tendency to identify leakage with respect to real datasets that were not used in training. To address these shortcomings, we propose a novel metric called Scaled Overlap (SO), which adapts a non-parametric data-copying test for facial recognition and extends it to the identity level. By leveraging a multi-model setup that compensates for errors in individual classifiers, it provides insights into the true strength of identity leakage. Our analysis identifies leaking datasets (CemiFace, SFace2, IDNet) with the most acute leakage introduced by generative model conditioning in CemiFace, demonstrates that real datasets do not contribute equally to generator training, and shows that the majority of synthetic images are drawn from multiple identities rather than relying on a single identity. Ultimately, our study provides new insights into the quality of synthetic datasets and gives more reliable techniques to help the community ensure the trustworthiness of synthetic data.

cs.CV↗

arXiv2026

Multi-Agent Strategies for Bridging Programming Language Gaps in Code Generation

Real-world software systems are inherently multilingual, but current large language models are not equally consistent across programming languages. This mismatch limits code generation usefulness, especially for underrepresented languages. Existing approaches improve code generation through fine-tuning, multi-agent reasoning, or translation, but remain language-specific, assume existing source code, or rely on per-language test suites. We introduce XL-CoGen, a three-stage multilingual code-generation pipeline that starts from a natural-language specification and a shared test list to generate correct implementations across multiple target languages. XL-CoGen first validates direct generation by constructing and correcting the test harness; when it fails, it transfers through empirically selected intermediate languages and translates validated solutions; it then repairs the best candidate through diagnosis and minimal patching. This design requires neither target-language training nor language-specific test suites. Across two benchmarks and multiple LLMs, XL-CoGen consistently improves over direct generation, with the largest gains on low-performing languages. In our Rust fine-tuning case study, XL-CoGen outperforms the best fine-tuned baseline by 22 percentage points and improves challenging languages by up to 33 points on multilingual benchmarks. Ablation results show that transfer and repair are complementary: repair suffices on easier tasks, whereas transfer becomes more important as difficulty increases, especially for weak target languages.

cs.SE↗

arXiv2026

RAGit: Turning Document Preprocessing into an Experimental Variable for RAG

Document preprocessing choices can substantially affect retrieval in Retrieval-Augmented Generation (RAG) pipelines, yet they are often selected heuristically and are difficult to compare under consistent conditions. We present RAGit, an open-source experimentation workbench for the systematic and reproducible evaluation of document preprocessing choices through downstream retrieval. RAGit represents parsing, chunking, indexing, retrieval, and evaluation as configurable experimental stages, enabling controlled comparisons while the remaining downstream settings are kept fixed. It preserves provenance from retrieved chunks to source pages and maps heterogeneous retrieval units to a common page-level ranking for evaluation. The current implementation supports multiple parsers and chunking strategies, dense and late-interaction retrieval, reusable intermediate artifacts, and multi-configuration benchmarking. We demonstrate RAGit through controlled experiments comparing alternative preprocessing configurations under a fixed retrieval setup, illustrating how preprocessing can be treated as an explicit experimental variable. RAGit and its documentation are publicly available on GitHub (https://github.com/OmarElbachyr/ragit). A demonstration video is available at https://youtu.be/tsst-Bhfz4I.

cs.IR↗

arXiv2026

VisR-RAG-Fin: Benchmarking Multimodal RAG over Visually Rich Financial Documents

Enterprise documents such as financial reports are increasingly processed by automated systems, yet they are primarily designed for human reading and exhibit complex visual structure, combining dense text with tables, charts, diagrams, and rich layouts. These characteristics pose significant challenges for Retrieval-Augmented Generation (RAG) systems, which must retrieve relevant pages and leverage visual and structural information during answer generation. Despite recent advances in multimodal retrieval and vision-language models, there remains a limited understanding of how design choices across retrieval and generation components affect end-to-end RAG performance on visually rich documents. In this paper, we present a comprehensive empirical study of multimodal RAG for page-level Question Answering (QA) over financial documents. We introduce VisR-RAG-Fin (Visually Rich RAG for Finance), an open-source benchmark constructed through a general pipeline for visually rich documents that preserves original layout and visual content. Using this benchmark, we systematically analyze retrieval and generation behavior under text-only, multimodal, and hybrid configurations, providing a foundation for rigorous evaluation of RAG systems on complex enterprise document collections. Overall, this work provides actionable insights for practitioners designing and evaluating multimodal RAG pipelines over complex, visually rich enterprise documents.

cs.IR↗

arXiv2026

Trajectory-Level Security Debt in LLM Coding Agents

LLM coding agents can traverse hundreds of intermediate code states before submitting a solution. Evaluating only the final artifact leaves the evolution of security findings unmeasured. We introduce the Security Debt Line Integral (SDLI), which accumulates static-analysis risk when an agent reaches a new best test pass ratio. We instantiate it with four static application security testing (SAST) tools and study artifacts from 830 passing SWE-bench runs, 712 ProgramBench final workspaces, and 13 public MirrorCode trajectories. The two large populations use the final-state special case of SDLI. Two-tool Common Weakness Enumeration (CWE) class agreement occurs in 3.9% of SWE-bench runs and 26.2% of the 80 ProgramBench runs passing at least 90% of official tests. These are scanner findings, not validated vulnerability rates. Excluding three advisory-heavy classes reduces the latter rate to 6.2%. Same-task runs differ in their measured scores, while one reconstructed ProgramBench run exposes persistent findings from its first implementation write. A repair case study reduces the scanner signal while preserving tested behavior, but also reveals sensitivity to equivalent API rewrites. SDLI offers a way to study progress and security findings together. Its value for steering agents and confirming exploitable vulnerabilities remains to be established.

cs.CR↗

arXiv2026

From Rookie to Pro: Social Engineering LLMs for Automated Vulnerability Exploitation in Enterprise Software

LLMs democratize software engineering by enabling non-programmers to create applications, but this same accessibility fundamentally undermines security assumptions that have guided software engineering for decades. We show in this work how publicly available LLMs can be socially engineered to transform novices into capable attackers, challenging the foundational principle that exploitation requires technical expertise. To that end, we propose RSA (Role-assignment, Scenario-pretexting, and Action-solicitation), a pretexting strategy that manipulates LLMs into generating functional exploits despite their safety mechanisms. Testing against Odoo---a widely used ERP platform, we evaluated five mainstream LLMs (GPT-4o, Gemini, Claude, Microsoft Copilot, and DeepSeek) and successfully exploited every tested CVE: at least one LLM produced a functional exploit for each within 3-5 prompting rounds. While prior work [13] found LLM-assisted attacks difficult and requiring manual effort, we demonstrate that this overhead can be eliminated entirely. Our findings invalidate core software engineering security principles: the distinction between technical and non-technical actors no longer provides valid threat models; technical complexity of vulnerability descriptions offers no protection when LLMs can abstract it away; and traditional security boundaries dissolve when the same tools that build software can be manipulated to break it. This represents a paradigm shift in software engineering---we must redesign security practices for an era where exploitation requires only the ability to craft prompts, not understand code. Artifacts available at: https://anonymous.4open.science/r/From-Rookie-to-Attacker-D8B3.

cs.SE↗

arXiv2026

Towards LLM-Enhanced Android Taint Analysis

Taint analysis is a fundamental technique for detecting sensitive data leaks in Android apps. However, traditional static tools, such as FlowDroid, still face well-known challenges due to the complexity of accurately modeling the Android framework. In this paper, we investigate whether off-the-shelf Large Language Models (LLMs) can effectively reason about taint flows in Android apps. Our preliminary approach relies on an agentic interaction strategy, enabling the LLM to iteratively explore code and reason about data flows. We conduct an initial evaluation on the DroidBench benchmark against FlowDroid, where our approach outperforms the baseline: Gemini-3 Flash achieves an F1-score of 0.96, compared to 0.55 for FlowDroid. In particular, we observe improvements in challenging categories such as inter-component communication (0.95 vs. 0.17), implicit flows (0.94 vs. 0.00), and reflection (1.00 vs. 0.50), where FlowDroid typically struggles. On a small set of real-world apps, the LLM-based approach also identifies additional potential data leaks not reported by FlowDroid. These preliminary findings suggest that LLM reasoning may effectively complement traditional static taint analysis, motivating future research on hybrid LLM-enhanced taint analysis pipelines.

cs.SE↗

arXiv2026

Do Not Copy/Paste: Soft Barriers for Copying in AI-Assisted Programming

Copying a function from a chat window into an editor takes less than a second. For many uses of AI coding tools, that speed is the point; in settings such as programming education, code review, and security-sensitive development, it can also be the problem. This paper frames copy-paste as an \emph{AI code handoff problem}: the moment model-generated text crosses from a conversational context into executable or committed software is a design boundary that current tools leave largely unmanaged. We argue that AI coding assistants should not only be evaluated by the code they generate, but also by how they mediate the transfer of that code into software artifacts. We propose \emph{soft barriers} as one class of handoff-aware mechanisms. Soft barriers preserve access to AI assistance while making unexamined transfer less frictionless. As an initial technical probe, we instantiate this idea using Unicode output perturbations that preserve visual readability but disrupt naive copy-paste execution. We introduce Copy-Paste Resistance (CPR), the fraction of functionally correct clean solutions that become syntactically invalid after perturbation. Across HumanEval and MBPP with four LLMs and four perturbation families, we find that output-level barriers can achieve high copy-paste resistance, but their effectiveness is highly model- and task-dependent. An exploratory pilot with 18 participants provides early evidence that soft barriers can shift users from direct transfer toward editing and reconstruction. We do not present Unicode perturbations as a deployment-ready solution; rather, we use them as a minimal probe for a broader research agenda on practical, transparent, and policy-aware AI code handoff.

cs.SE↗

arXiv2026

Evaluating Inference-Time Defenses Against Package Hallucination in LLM-Generated Code

LLMs are increasingly used for code generation, yet they frequently hallucinate non-existent software packages, creating exploitable entry points into the software supply chain. We make four contributions to this problem. First, we show that prior evaluation methodologies systematically inflate hallucination rates by misclassifying standard-library modules as hallucinations in some languages. For Python, the overestimation reaches 9.4 percentage points. Second, we evaluate seven inference-time defenses for mitigating package hallucinations, including five guided decoding strategies (Greedy, Contrastive, DoLa, Nudging, and Active Layer-Contrastive Decoding), an iterative self-refinement approach (Self-Refine), and a Retrieval-Augmented Generation (RAG)-based defense.. Across eight models spanning five families and four programming languages (Python, JavaScript, Ruby, Rust), RAG reduces the package hallucination rate (PHR) in 18 of 32 model--language configurations. Third, we introduce Package Utility (PU) to assess whether defenses preserve valid and task-relevant recommendations. Among strategies evaluated, Greedy decoding provides the strongest average mitigation--utility trade-off. Fourth, we stress-test all strategies under adversarial prompts seeded with fabricated package names and find that PHR surges by up to 45 percentage points relative to standard prompts, with Ruby consistently the most vulnerable language (80.9--95.2\%). Under adversarial conditions, RAG and Self-Refine outperform all decoding-only strategies, indicating that robust defense requires either external grounding or iterative self-verification when prompts are actively hostile. Our results recast package hallucination as both a measurement problem and a decoding-time control problem, and they demonstrate that the choice of defense must be matched to the threat model and recommendation utility.

cs.SE↗

arXiv2026

LëtzCross: A Cross-Lingual Page-Level Benchmark for Multimodal Retrieval over Luxembourgish Documents

Recent page-image retrievers such as ColPali have improved retrieval over visually rich documents, yet little is known about how they behave in cross-lingual, low-resource settings. We introduce LëtzCross, a benchmark for cross-lingual page-level retrieval over Luxembourgish PDF documents, with document pages indexed as images and queries provided in English, French, German, and Luxembourgish. The benchmark combines text-focused QA pairs with visually grounded QA pairs, covering both textual and visual retrieval needs in PDF-based RAG. We use LëtzCross to compare OCR-based text-only retrievers with ColPali-style page-image retrievers and find that the latter perform better across query languages in this system-level comparison. We also examine single-language and multilingual fine-tuning. Fine-tuning transfers across query languages, with French yielding the highest mean performance on Luxembourgish queries among the single-language settings. In the multilingual setting, including Luxembourgish gives the strongest results and substantially improves retrieval for Luxembourgish queries.

cs.CL↗

arXiv2026

Semantic Drift in Bug Resolution: How Behavioral Signals Propagate from Reports to Tests and Patches

Desc2Fix is a framework for measuring semantic alignment between bug reports, triggering tests, and developer-written fixes. Alignment is operationalized through structured behavioral anchors (e.g., reproduction steps, API/exception cues, expected vs. actual behavior), deterministic similarity metrics (ROUGE, SBERT, CodeBERT, OpenAI embeddings), and LLM-based judgments grounded in coverage, correctness, and specificity. Our analysis covers 2,857 report-test-patch triplets from Defects4J and SWT-Bench using two widely adopted instruction-tuned LLMs from distinct model families. LLMs reliably extract structured signals (up to 90% completeness) and exhibit strong cross-model consistency, yielding a stable semantic input contract for downstream reasoning. However, alignment is highly representation-sensitive: lexical similarity alone is insufficient, full diffs provide the most stable basis for judging report-patch correspondence, and structured summaries trade surface overlap for stronger correspondence at the level of individual actions and entities. Across more than 182,000 LLM-based alignment judgments, both models exhibit systematic optimism relative to humans (1-2 points on 5-point scales) and only modest rank agreement, motivating bias-aware evaluation. Behavioral alignment is measurable but not reducible to similarity, and structured anchors combined with embedding-based proxies provide reproducible signals for ranking and filtering tests and patches. Desc2Fix could support more reliable test generation, fault localization, patch ranking, and bug report authoring.

cs.SE↗

arXiv2026

Memorization Diagnostics for Code LLMs Should be Scale-Aware

The extent to which large language models for code rely on memorization over genuine understanding remains highly debated. While current literature frequently reports widespread memorization, evaluating the underlying probing techniques across dense architectures reveals a severe breakdown in their utility at scale. Traditional encoder-style probes using perturbations such as synonym fuzzing or dead-code insertion struggle to expose memorization in scaled models, even on known-contaminated benchmarks, and decoder-style probes that rely on log probabilities show similar performance degradation. The specific mode of failure for these probes, particularly why such techniques disrupt smaller models but fail to impact larger ones, motivates us to untangle representation load from memorization rather than treating them as a single phenomenon. By applying invertible mathematical transforms to numeric problems, we isolate these two factors and reveal that scaled encoders successfully absorb substantial representation load while still converging on the correct family of solutions. In practical software engineering, this ability to adapt to varying surface forms is what truly matters for usability and generalizability in LLM and agentic applications. Whether a specific solution was seen during training becomes a much less pressing question because although memorization inflates scores on contaminated benchmarks, factoring out representation load makes it debatable how much we should truly care if a functional answer was originally memorized. Future evaluations must therefore be built around separating these phenomena rather than relying on methodologies that quietly entangle them.

cs.SE↗

arXiv2026

GapFuzz: Cross-Plane Divergence Fuzzing for Distributed SDN Controllers

Distributed Software-Defined Networking (SDN) clusters replicate flow state asynchronously between a master node and its backups, leaving a window during which two backup nodes can each commit a contradictory rule, the master can serialize both into the data plane, and the kernel datapath can latch onto an action that no node believes authoritative. Existing SDN fuzzers miss this fault: they confine their oracle to the control plane, target a single controller, or do not steer concurrency to provoke replication races. We present GapFuzz, a stateful concurrency fuzzer for distributed SDN clusters. GapFuzz injects pairs of contradictory Northbound requests on two non-master nodes with controlled inter-injection delay $Δt$, and reconstructs the global cross-plane state by querying every replica and the kernel-datapath action through ovs-appctl ofproto/trace. A two-phase timing search detects whether a divergence exists, then doubles and bisects on $Δt$ to bound the injection-time window; a lifetime probe labels each verdict transient or persistent and assigns it to one of four cross-plane state classes derived from the ONOS 2.7 source. On a three-node ONOS 2.7 cluster, GapFuzz produces a divergent verdict in 81.7% of attempts ($N=50$, Wilson 95% CI $[77.3, 85.4]$%); every divergence sits between the cluster's authoritative state and the kernel datapath. Phase 2 separates a 5 ms race window for one template from a doubling-cap regime ($Δt_{\max}=10.24$ s) for six others, and 99.4% of divergences persist past 30 s. Replacing the kernel-datapath probe with the OpenFlow user-space probe used by prior fuzzers drops detection by 26.6 percentage points overall and by 46.5 points after excluding canonicalization-forced verdicts.

cs.SE↗

arXiv2026

Can LLMs Test Terminal User Interfaces?

Terminal User Interfaces (TUIs) combine the stateful, screen-oriented behaviour of GUIs with terminal deployment and are now common in developer tools. Yet they lack a dedicated testing methodology. We survey 197 real-world TUI applications: only 12% of test code exercises the interface, and 45% of those tests never send input, checking a static frame instead. We turn these applications into a headless benchmark spanning ratatui/Rust, bubbletea/Go, textual/Python, and ink/TypeScript, packaging each as an instrumented Docker image. We record line and widget coverage where reliable, rendered terminal states, and crashes. Under equal wall-clock budgets, we compare four frontier LLMs with random exploration. No model dominates. Random is a strong time-budgeted baseline, but its crash advantage comes from higher throughput: per interaction, LLM guidance is more efficient and uniquely reaches input-gated faults. Automatically deriving launch inputs yields the largest practical gain, enabling applications that otherwise never start. Line coverage poorly predicts crash discovery, weakening it as a proxy for test effectiveness. Automated TUI testing is feasible but far from solved, and honest baselines matter more than model choice. We release the coverage tool tuicov at https://github.com/tui-testing/tuicov and the testing framework tuibot at https://github.com/tui-testing/tuibot.

cs.SE↗

arXiv2026

On the Diffusion of Test Smells in LLM-Generated Unit Tests

LLMs promise to transform unit test generation from a manual burden into an automated solution. Yet, beyond metrics such as compilability or coverage, little is known about the quality of LLM-generated tests, particularly their susceptibility to test smells, design flaws that undermine readability and maintainability. This paper presents the first multi-benchmark, large-scale analysis of test smell diffusion in LLM-generated unit tests. We contrast LLM outputs with human-written suites (as the reference for real-world practices) and SBST-generated tests from EvoSuite (as the automated baseline), disentangling whether LLMs reproduce human-like flaws or artifacts of synthetic generation. Our study draws on 20,505 class-level suites from four LLMs (GPT-3.5, GPT-4, Mistral 7B, Mixtral 8x7B), 972 method-level cases from TestBench, 14,469 EvoSuite tests, and 779,585 human-written tests from 34,635 open-source Java projects. Using two complementary detection tools (TsDetect and JNose), we analyze prevalence, co-occurrence, and correlations with software attributes and generation parameters. Results show that LLM-generated tests consistently manifest smells such as Assertion Roulette and Magic Number Test, with patterns strongly influenced by prompting strategy, context length, and model scale. Comparisons reveal overlaps with human-written tests, raising concerns of potential data leakage from training corpora while EvoSuite exhibits distinct, generator-specific flaws. These findings highlight both the promise and the risks of LLM-based test generation, and call for the design of smell-aware generation frameworks, prompt engineering strategies, and enhanced detection tools to ensure maintainable, high-quality test code.

cs.SE↗

arXiv2026

Have I Seen You? Embedding Behavior Signals Synthetic Face Dataset Membership

Synthetic face datasets are increasingly used to reduce privacy exposure and data access constraints in biometric recognition. Yet the generators that produce these datasets are trained on real faces, so synthetic data may still reveal their real source data. We study this risk through a dataset-level membership inference attack that first identifies the synthetic dataset used to train a face recognizer and then infers the real dataset used to train the generator. Across 11 face recognition models, 11 synthetic datasets, and 7 real datasets, the attack recovers the synthetic training dataset in 100% of cases and identifies the generator's source dataset in 54.5% of cases. These results show that synthetic data can retain dataset-level traces of real training data and that privacy-preserving deployment requires stronger leakage mitigation.

cs.CV↗