We introduce FinanceReasoning, a novel benchmark designed to evaluate the reasoning capabilities of large reasoning models (LRMs) in financial numerical reasoning problems. Compared to existing benchmarks, our work provides three key advancements. (1) Credibility: We update 15.6% of the questions from four public datasets, annotating 908 new questions with detailed Python solutions and rigorously refining evaluation standards. This enables an accurate assessment of the reasoning improvements of LRMs. (2) Comprehensiveness: FinanceReasoning covers 67.8% of financial concepts and formulas, significantly surpassing existing datasets. Additionally, we construct 3,133 Python-formatted functions, which enhances LRMs' financial reasoning capabilities through refined knowledge (e.g., 83.2% $\rightarrow$ 91.6% for GPT-4o). (3) Challenge: Models are required to apply multiple financial formulas for precise numerical reasoning on 238 Hard problems. The best-performing model (i.e., OpenAI o1 with PoT) achieves 89.1% accuracy, yet LRMs still face challenges in numerical precision. We demonstrate that combining Reasoner and Programmer models can effectively enhance LRMs' performance (e.g., 83.2% $\rightarrow$ 87.8% for DeepSeek-R1). Our work paves the way for future research on evaluating and improving LRMs in domain-specific complex reasoning tasks.

Sarcasm is a complex linguistic phenomenon that involves a disparity between literal and intended meanings, making it challenging for sentiment analysis and other emotion-sensitive tasks. While traditional sarcasm detection methods primarily focus on text, recent approaches have incorporated multimodal information. However, the application of Large Visual Language Models (LVLMs) in Multimodal Sarcasm Analysis (MSA) remains underexplored. In this paper, we evaluate LVLMs in MSA tasks, specifically focusing on Multimodal Sarcasm Detection and Multimodal Sarcasm Explanation. Through comprehensive experiments, we identify key limitations, such as insufficient visual understanding and a lack of conceptual knowledge. To address these issues, we propose a training-free framework that integrates in-depth object extraction and external conceptual knowledge to improve the model's ability to interpret and explain sarcasm in multimodal contexts. The experimental results on multiple models show the effectiveness of our proposed framework. The code is available at https://github.com/cp-cp/LVLM-MSA.

Fine-tuning diffusion policies with reinforcement learning (RL) presents significant challenges. The long denoising sequence for each action prediction impedes effective reward propagation. Moreover, standard RL methods require millions of real-world interactions, posing a major bottleneck for practical fine-tuning. Although prior work frames the denoising process in diffusion policies as a Markov Decision Process to enable RL-based updates, its strong dependence on environment interaction remains highly inefficient. To bridge this gap, we introduce DiWA, a novel framework that leverages a world model for fine-tuning diffusion-based robotic skills entirely offline with reinforcement learning. Unlike model-free approaches that require millions of environment interactions to fine-tune a repertoire of robot skills, DiWA achieves effective adaptation using a world model trained once on a few hundred thousand offline play interactions. This results in dramatically improved sample efficiency, making the approach significantly more practical and safer for real-world robot learning. On the challenging CALVIN benchmark, DiWA improves performance across eight tasks using only offline adaptation, while requiring orders of magnitude fewer physical interactions than model-free baselines. To our knowledge, this is the first demonstration of fine-tuning diffusion policies for real-world robotic skills using an offline world model. We make the code publicly available at https://diwa.cs.uni-freiburg.de.

Belief systems are often treated as globally consistent sets of propositions or as scalar-valued probability distributions. Such representations tend to obscure the internal structure of belief, conflate external credibility with internal coherence, and preclude the modeling of fragmented or contradictory epistemic states. This paper introduces a minimal formalism for belief systems as directed, weighted graphs. In this framework, nodes represent individual beliefs, edges encode epistemic relationships (e.g., support or contradiction), and two distinct functions assign each belief a credibility (reflecting source trust) and a confidence (derived from internal structural support). Unlike classical probabilistic models, our approach does not assume prior coherence or require belief updating. Unlike logical and argumentation-based frameworks, it supports fine-grained structural representation without committing to binary justification status or deductive closure. The model is purely static and deliberately excludes inference or revision procedures. Its aim is to provide a foundational substrate for analyzing the internal organization of belief systems, including coherence conditions, epistemic tensions, and representational limits. By distinguishing belief structure from belief strength, this formalism enables a richer classification of epistemic states than existing probabilistic, logical, or argumentation-based approaches.

Defending large language models (LLMs) against jailbreak attacks is essential for their safe and reliable deployment. Existing defenses often rely on shallow pattern matching, which struggles to generalize to novel and unseen attack strategies. To address this challenge, we propose the Cognitive-Driven Defense (CDD) framework, which targets the underlying structure of jailbreak prompts by applying meta-operations, defined as basic manipulations that conceal harmful intent.CDD emulates human cognitive reasoning through a structured reasoning chain. It begins with a global perception of the prompt and follows with a localized analysis to uncover hidden manipulations. By applying supervised fine-tuning on this structured chain, the model learns to identify and reason about known manipulation patterns. To enhance generalization to unseen threats, an entropy-guided reinforcement learning algorithm (EG-GRPO) is introduced to encourage exploration of new types and variants of meta-operations. Experiments demonstrate that CDD can achieve state-of-the-art defense performance and exhibit strong generalization to unseen jailbreak attacks.

--Despite significant advancements, current large language models (LLMs) and vision-language models (L VLMs) continue to struggle with complex, multi-step, cross-modal common sense reasoning tasks, often exhibiting a lack of "deliberative thinking." They tend to rely on superficial associations rather than deep, chained inference, particularly when integrating visual information with abstract concepts. T o address this, we propose the Coherent Multimodal Reasoning Framework (CMRF), a novel approach that enhances L VLMs' common sense reasoning capabilities through an iterative, self-evaluating inference mechanism. CMRF mimics human problem-solving by decomposing complex queries, generating step-by-step inferences, and self-correcting errors. Coupled with an Adaptive Iterative Refinement strategy, CMRF systematically refines its reasoning paths. Built upon LLaV A-1.6-34B and trained on a novel Multimodal Daily Activity Reasoning (MDAR) dataset, CMRF achieves state-of-the-art performance among open-source L VLMs on challenging benchmarks like VCR, A-OKVQA, and DailyLife-MRC. Extensive ablation studies and human evaluations confirm the critical contributions of each module and the effectiveness of iterative refinement in fostering more coherent and accurate reasoning. The remarkable advancements in large language models (LLMs) [1], [2] and vision-language models (L VLMs) have revolutionized various aspects of artificial intelligence, demonstrating unprecedented capabilities in understanding, generating, and processing information across modalities [3]. These models excel in tasks ranging from complex question answering to creative content generation, largely due to their extensive pre-training on vast amounts of data.

The Uniform Appraisal Dataset (UAD) 3.6's mandatory 2026 implementation transforms residential property valuation from narrative reporting to structured, machine-readable formats. This paper provides the first comprehensive analysis of this regulatory shift alongside concurrent AI advances in computer vision, natural language processing, and autonomous systems. We develop a three-layer framework for AI-augmented valuation addressing technical implementation and institutional trust requirements. Our analysis reveals how regulatory standardization converging with AI capabilities enables fundamental market restructuring with profound implications for professional practice, efficiency, and systemic risk. We make four key contributions: (1) documenting institutional failures including inter-appraiser variability and systematic biases undermining valuation reliability; (2) developing an architectural framework spanning physical data acquisition, semantic understanding, and cognitive reasoning that integrates emerging technologies while maintaining professional oversight; (3) addressing trust requirements for high-stakes financial applications including regulatory compliance, algorithmic fairness, and uncertainty quantification; (4) proposing evaluation methodologies beyond generic AI benchmarks toward domain-specific protocols. Our findings indicate successful transformation requires not merely technological sophistication but careful human-AI collaboration, creating systems that augment rather than replace professional expertise while addressing historical biases and information asymmetries in real estate markets.

We investigate how feature correlations influence the capacity of Dense Associative Memory (DAM), a Transformer attention-like model. Practical machine learning scenarios involve feature-correlated data and learn representations in the input space, but current capacity analyses do not account for this. We develop an empirical framework to analyze the effects of data structure on capacity dynamics. Specifically, we systematically construct datasets that vary in feature correlation and pattern separation using Hamming distance from information theory, and compute the model's corresponding storage capacity using a simple binary search algorithm. Our experiments confirm that memory capacity scales exponentially with increasing separation in the input space. Feature correlations do not alter this relationship fundamentally, but reduce capacity slightly at constant separation. This effect is amplified at higher polynomial degrees in the energy function, suggesting that Associative Memory is more limited in depicting higher-order interactions between features than patterns. Our findings bridge theoretical work and practical settings for DAM, and might inspire more data-centric methods.

World models have become increasingly popular in acting as learned traffic simulators. Recent work has explored replacing traditional traffic simulators with world models for policy training. In this work, we explore the robustness of existing metrics to evaluate world models as traffic simulators to see if the same metrics are suitable for evaluating a world model as a pseudo-environment for policy training. Specifically, we analyze the metametric employed by the Waymo Open Sim-Agents Challenge (WOSAC) and compare world model predictions on standard scenarios where the agents are fully or partially controlled by the world model (partial replay). Furthermore, since we are interested in evaluating the ego action-conditioned world model, we extend the standard WOSAC evaluation domain to include agents that are causal to the ego vehicle. Our evaluations reveal a significant number of scenarios where top-ranking models perform well under no perturbation but fail when the ego agent is forced to replay the original trajectory. To address these cases, we propose new metrics to highlight the sensitivity of world models to uncontrollable objects and evaluate the performance of world models as pseudo-environments for policy training and analyze some state-of-the-art world models under these new metrics.

Transformers are the basis of modern large language models, but relatively little is known about their precise expressive power on graphs. We study the expressive power of graph transformers (GTs) by Dwivedi and Bresson (2020) and GPS-networks by Rampásek et al. (2022), both under soft-attention and average hard-attention. Our study covers two scenarios: the theoretical setting with real numbers and the more practical case with floats. With reals, we show that in restriction to vertex properties definable in first-order logic (FO), GPS-networks have the same expressive power as graded modal logic (GML) with the global modality. With floats, GPS-networks turn out to be equally expressive as GML with the counting global modality. The latter result is absolute, not restricting to properties definable in a background logic. We also obtain similar characterizations for GTs in terms of propositional logic with the global modality (for reals) and the counting global modality (for floats).