Expense reimbursement internal control is a high-stakes corporate finance task that combines multimodal, visually rich evidence (e.g., receipts and approval forms), firm-specific policy text, and multi-step compliance reasoning. No existing benchmark evaluates large multimodal models (LMMs) in this domain. We introduce ReimburseBench, the first benchmark targeting full-workflow multimodal audit reasoning in corporate internal control. It comprises 151 expert-constructed reimbursement workflows, each packaging an application, receipts and supporting documents, and a firm-specific policy, evaluated by a process-level rubric grounded in twelve auditing logic points. Benchmarking nine frontier and locally deployable systems, we find the strongest attains only 72.5%, and no model exceeds 47.7% on hard workflows. Beyond this aggregate score, risk-aware, process-level evaluation surfaces three structural findings: (i) a persistent gap between binary admissibility judgment and precise quantitative computation; (ii) a systematic lenient bias in monetary predictions that poses deployment risk; and (iii) a performance pattern consistent with compositional-generalization difficulties under realistic workflow load. We provide ReimburseBench as a testbed for advancing LMMs toward trustworthy financial systems. Data and code: https://github.com/ikerzhan/ReimburseBench.

Real-world data analysis often requires joint reasoning over heterogeneous artifacts such as tables, charts, and textual reports. Training reliable multimodal large language model (MLLM) agents for this setting remains challenging, as large-scale aligned training data are scarce, and outcome-only rewards are too coarse to distinguish grounded cross-source analysis from spurious success. We introduce OmniDA, a training recipe that addresses these two challenges with a scalable multimodal data-synthesis pipeline and an Evidence-Aware Process Reward (EAPR) that supervises both evidence recovery and procedural validity. Through this pipeline, we construct OmniDA-11K, a high-quality training corpus for multimodal data analysis. When instantiated with Qwen3-VL-8B, OmniDA improves over the corresponding agentic baseline by 19.2 points on average across five challenging benchmarks and achieves the strongest overall performance among existing open-source data-analysis agents. These results highlight the value of supervising evidence recovery and analytical procedure beyond final-answer correctness.

Existing tabular reasoning benchmarks typically assume that input tables are clean and error-free, which contrasts sharply with real-world data pipelines where noise and anomalies are pervasive. While humans routinely validate tables and reject conclusions when anomalies render a question unanswerable, it remains unclear whether large language models (LLMs) exhibit similar anomaly awareness. In this work, we introduce AnTBench, a benchmark designed to systematically evaluate LLM behavior under anomalous tabular data. AnTBench consists of 1,571 curated instances covering five common and verifiable anomaly types. Experiments on 14 representative LLMs reveal a pronounced capability–behavior gap: even the strongest model achieves 76.07% on explicit anomaly detection, yet only 19.03% on spontaneous anomaly perception during standard table QA. As an inference-time mitigation, we introduce the Multi-Path Thinker, a multi-path reasoning framework. While this approach improves explicit anomaly detection, enforcing proactive verification for every query is practically prohibitive. We urge the community to move beyond inference-time patches and instill spontaneous anomaly awareness directly into the training and alignment phases. Our code and data are available at https://github.com/Wei12097/AnTBench.

Parameter-efficient fine-tuning is usually framed as a question of how many parameters to update. Under a severe trainable-state budget, however, where those coefficients act is equally consequential. We study this choice through frozen-core adaptation: a calibration pass fixes left and right bases for each weight matrix, and fine-tuning optimizes only an r× r core. This removes the ability of trainable factors to repair a poor initial span and makes subspace quality directly observable. We introduce FCCA, which estimates the signed input–error cross-covariance, whitens it with diagonal Fisher moments, truncates it in the resulting local metric, maps the selected directions back, and applies thin QR to obtain stable core coordinates. Under a matched r² budget, we compare eight basis constructors on 11 tasks, four model settings, and three seeds. On Qwen2.5-3B, FCCA reaches an 83.0 macro-average, 2.3 points above the next-best matched-budget constructor, and exceeds its unwhitened RawGrad control on all 11 tasks. It ranks first at all three Qwen scales and finishes within 0.13 points of the best method on Llama-3.2-1B. Controlled ablations show gains of 2.7–17.2 points from whitening and identify QR as necessary for stable core optimization in the tested regime. Finally, FCCA comes within 0.32 and 0.23 average points of LoRA and DoRA while optimizing 36.9K rather than roughly 7.4M parameters. These results show that a carefully selected fixed span can recover most of the benefit of movable low-rank factors at a much smaller trainable and optimizer-state cost.

We introduce LongTableBench, a benchmark for evaluating long-context reasoning over semi-structured tables across diverse formats, tasks, and domains. It comprises 5,950 QA instances spanning 7 table formats (e.g., Markdown, HTML, SQL), 18 domains, and input lengths up to 128K tokens, including multi-turn and multi-table settings. To ensure data quality, we combine symbolic supervision, cross-model validation, and human review. Evaluating 52 LLMs—including general-purpose, table-specific, and reasoning-enhanced models—reveals that only the strongest models maintain robust performance under increasing context lengths and format diversity. We further show that end-to-end models outperform compression-based approaches, especially on tasks requiring semantic integration. LongTableBench provides a rigorous, scalable testbed for advancing long-context tabular understanding and highlights key limitations in current LLMs’ structural and reasoning capabilities. The code and data are available at https://github.com/liyaooi/LongTableBench.

With the continuous development of language models and the widespread availability of various types of accessible interfaces, large language models (LLMs) have been applied to an increasing number of fields. However, due to the vast amounts of data and computational resources required for model development, protecting the model’s parameters and training data has become an urgent and crucial concern. Due to the revolutionary training and application paradigms of LLMs, many new attacks on language models have emerged in recent years. In this paper, we define these attacks as “reverse engineering” (RE) techniques on LMs and aim to provide an in-depth analysis of reverse engineering of language models. We illustrate various methods of reverse engineering applied to different aspects of a model, while also providing an introduction to existing protective strategies. On the one hand, it demonstrates the vulnerabilities of even black box models to different types of attacks; on the other hand, it offers a more holistic perspective for the development of new protective strategies for models.

With the rapid advancement of Large Language Models (LLMs), there is an increasing need for challenging benchmarks to evaluate their capabilities in handling complex tabular data. However, existing benchmarks are either based on outdated data setups or focus solely on simple, flat table structures. In this paper, we introduce RealHiTBench, a comprehensive benchmark designed to evaluate the performance of both LLMs and Multimodal LLMs (MLLMs) across a variety of input formats for complex tabular data, including LaTeX, HTML, and PNG. RealHiTBench also includes a diverse collection of tables with intricate structures, spanning a wide range of task types. Our experimental results, using 25 state-of-the-art LLMs, demonstrate that RealHiTBench is indeed a challenging benchmark. Moreover, we also develop TreeThinker, a tree-based agent that organizes hierarchical headers into a tree structure for enhanced tabular reasoning, validating the importance of improving LLMs’ perception of table hierarchies. We hope that our work will inspire further research on tabular data reasoning and the development of more robust models. The code and data are available at https://github.com/cspzyy/RealHiTBench.

Instruction tuning is vital for aligning large language models (LLMs) with human intent, but current methods typically rely on costly human-annotated seed data or powerful external teacher models. While instruction back-translation techniques reduce this dependency, they remain fundamentally tethered to an initial seed set, which limits full automation, introduces biases, and can lead to inefficient use of unlabeled corpora. In this paper, we propose Cycle-Instruct, a novel framework that achieves fully seed-free instruction tuning. Inspired by cycle consistency, Cycle-Instruct employs a dual self-training loop where two models—an answer generator and a question generator—are bootstrapped solely from raw, unlabeled text. These models mutually supervise each other by reconstructing original text segments from their counterpart’s generated pseudo-labels, effectively learning from the intrinsic structure of the data without any human-provided seeds. We demonstrate Cycle-Instruct’s efficacy across four diverse data tracks, including general instruction-following, domain-specific tasks, dialogue logs, and plain text. Our extensive experiments show that Cycle-Instruct not only outperforms seed-driven back-translation baselines but also achieves performance comparable to strongly supervised methods.

The rapid advancements of Large Language Models (LLMs) tightly associate with the expansion of the training data size. However, the unchecked ultra-large-scale training sets introduce a series of potential risks like data contamination, i.e. the benchmark data is used for training. In this work, we propose a holistic method named Polarized Augment Calibration (PAC) along with a new to-be-released dataset to detect the contaminated data and diminish the contamination effect. PAC extends the popular MIA (Membership Inference Attack) — from machine learning community — by forming a more global target at detecting training data to Clarify invisible training data. As a pioneering work, PAC is very much plug-and-play that can be integrated with most (if not all) current white- and black-box LLMs. By extensive experiments, PAC outperforms existing methods by at least 4.5%, towards data contamination detection on more 4 dataset formats, with more than 10 base LLMs. Besides, our application in real-world scenarios highlights the prominent presence of contamination and related issues.

This study investigates the potential of Large Language Models (LLMs) for reconstructing and understanding the physical world based solely on textual knowledge. It explores the impact of model performance on spatial understanding abilities by introducing a set of geometric conventions and developing a workflow based on multi-layer graphs and multi-agent systems. The study examines how LLMs achieve multi-step and multi-objective geometric inference in a spatial environment, using unified geometric conventions and a graph-driven framework. A genetic algorithm, inspired by large-scale model knowledge, is employed to solve geometric constraint problems, enhancing the spatial reasoning capabilities of LLMs. This work innovatively explores the feasibility of using text-based LLMs as builders of the physical world and designs a workflow to enhance their spatial comprehension and construction capabilities.