LLM agents are increasingly expected to support enterprise workflows, where tasks often involve missing information, uncertainty, feedback, and long-term trade-offs. However, existing enterprise and financial benchmarks mainly test static capabilities such as information extraction, numerical calculation, domain knowledge, and financial QA, leaving interactive and long-horizon decision-making underexplored. To bridge this gap, we introduce EnterpriseBench, a benchmark that evaluates LLM agents across this spectrum, from static question answering to dynamic decision-making. Specifically, EnterpriseBench reorganizes existing enterprise and financial QA datasets into a unified foundational suite annotated by capability and difficulty, and introduces three professional interactive settings: Consulting, based on management-consulting-style business cases for client problem diagnosis through multi-turn information seeking; the Beer Game, adapted from a classic supply-chain management simulation for inventory control under delayed feedback; and Enterprise Digital Twin, a project-based business simulator for workforce, risk, and project planning. Experiments with nine agent methods under two backbone models show that current agents have not yet achieved stable, comprehensive, and cross-task reliability in enterprise scenarios. These results show that EnterpriseBench provides a practical benchmark for evaluating LLM agents in realistic enterprise strategic reasoning and decision-making. Our codes and benchmark implementation are publicly available at https://github.com/sduyangmin/FirmBench.

Low-Rank Adaptation (LoRA) is widely used for parameter-efficient fine-tuning, yet its robustness to textual adversarial perturbations remains understudied. Existing adversarial training methods often improve robustness by solving costly min-max objectives with inner perturbation optimization with suboptimal results. In this research, we propose Expectation-Regularized LoRA (ER-LoRA), a simple expectation-based regularization approach that improves the robustness of LoRA without requiring inner maximization or adversarial example generation. ER-LoRA smooths predictions over a LoRA-aware low-dimensional perturbation distribution and admits a closed-form curvature-aware objective via a local Gaussian approximation. It adds no trainable parameters and keeps the LoRA architecture unchanged. Experiments on adversarial NLP benchmarks and textual attacks show that ER-LoRA improves robustness while preserving clean accuracy and parameter efficiency.

Model editing aims to update large language models (LLMs) with new knowledge without full retraining, but existing methods mainly focus on fact-level knowledge. Rules that support mathematical and scientific reasoning must remain consistent across formulas, descriptions, and instances, making localized updates less effective for rule-level editing. In this paper, we construct RuleEdit-200, a manually verified benchmark of 200 mathematical and physical rules, each instantiated in formula, description, and instance forms. Using this benchmark, we conduct fine-grained causal tracing on four LLMs and identify a form-specific layer-wise separation of rule knowledge: formulas and descriptions are associated with earlier layers, whereas instances rely on a distinct set of middle-layer representations. This finding suggests that rule-level editing requires separate updates for different forms. Based on this insight, we propose Distributed Multi-Layer Editing (DMLE), which updates formulas, descriptions and instances at their corresponding layer ranges. Experiments on RuleEdit-200 show that DMLE remains competitive on standard editing metrics while substantially improving rule-level editing, with average gains of 18.22% and 61.00% in instance portability and rule understanding over the best baseline on each model.

Jailbreak attacks serve as a pivotal technique for evaluating the safety alignment of Large language models. Current token-level attacks have shown remarkable efficacy on open-source models by leveraging gradient-based optimization. However, these attacks suffer from poor cross-model transferability, severely limiting their utility on proprietary ones. To address this limitation, we propose Reparameterization Invariance Gradient-based Jailbreak (RIGJ), a natural gradient based framework designed to improve cross-model transferability. Unlike prior token-level methods whose optimization paths are constrained by model-specific Euclidean geometry, RIGJ defines update directions according to differences in output distributions rather than parameter-space distances. Since language models are trained to capture similar dependency structures of natural language, their output distributions share common geometry across architectures, yielding intrinsically model-agnostic optimization trajectories and substantially stronger jailbreak transferability. Extensive experiments demonstrate superior performance, increasing the cross-model Attack Success Rate and Average Harmfulness Score by 14.9 and 1.23, respectively. Our code is provided https://github.com/nohuma/AISafety_transfer_jailbreak_RIGJ_2026.

Large Language Models (LLMs) have advanced rapidly in recent years, with their applications in software engineering expanding to more complex repository-level tasks. GitHub issue resolving is a key challenge among these tasks. While recent approaches have made progress on this task, they focus on textual data within issues, neglecting visual data. However, this visual data is crucial for resolving issues as it conveys additional knowledge that text alone cannot. We propose CodeV, the first approach to leveraging visual data to enhance the issue-resolving capabilities of LLMs. CodeV resolves each issue by following a two-phase process: data processing and patch generation. To evaluate CodeV, we construct a benchmark for visual issue resolving, namely Visual SWE-bench. Through extensive experiments, we demonstrate the effectiveness of CodeV, as well as provide valuable insights into leveraging visual data to resolve GitHub issues.

In the era of code large language models (code LLMs), data engineering plays a pivotal role during the instruction fine-tuning phase. To train a versatile model, previous efforts devote tremendous efforts into crafting instruction data covering all the downstream scenarios. Nonetheless, this will incur significant expenses in constructing data and training model. Therefore, this paper introduces CodeM, a novel data construction strategy, which can efficiently train a versatile model using less data via our newly proposed ability matrix. CodeM uses ability matrix to decouple code LLMs’ abilities into two dimensions, constructing a lightweight training corpus that only covers a subset of target scenarios. Extensive experiments on HumanEvalPack and MultiPL-E imply that code LLMs can combine the single-dimensional abilities to master composed abilities, validating the effectiveness of CodeM.