High-stakes domains such as finance, law, and biomedicine demand both accurate results and rigorous reasoning.Current reinforcement learning paradigms primarily rely on outcomebased rewards, often overlooking latent logical fallacies in intermediate steps.Leveraging the cognitive asymmetry where falsifying local errors is more efficient than generating global correctness, we propose RADO (Reasoning Audit-Driven Optimization).RADO introduces a specialized audit model augmented with external tools to identify local logical ruptures and calibrate reward signals.By integrating Direct Preference Optimization (DPO) with Group Relative Policy Optimization (GRPO), our framework enables explicit supervision over reasoning paths.Experimental results demonstrate that RADO consistently improves final accuracy while significantly enhancing logical rigor in high-stakes domains.

Graph-related tasks are traditionally addressed with Graph Neural Networks (GNNs) or graph transformers, but their task-specific training limits generalization.Large Language Models (LLMs) offer stronger generalization, yet encoding graphs as one-dimensional text struggles to capture multi-hop dependencies and two-dimensional topology.Vision-Language Models (VLMs) provide an alternative by visualizing graphs, but rendering large graphs in a single image causes clutter, occlusion, and distraction, hindering reasoning.We propose MuSe, a novel multi-stage graph reasoning framework based on VLMs.Instead of processing entire graphs at once, MuSe incrementally samples and visualizes task-relevant subgraphs, enabling progressive reasoning.The framework employs a two-stage training paradigm: supervised fine-tuning to acquire local sampling and reasoning skills, followed by reinforcement learning with GRPO to refine the sampling strategy and control dialog length.To support evaluation, we introduce LGVLQA, a new multimodal dataset with larger and more complex graph structures, addressing the scalability limitations of existing benchmarks.Experiments show that MuSe consistently outperforms leading LLM and VLM baselines, demonstrating improved structural understanding and reasoning ability.Our code and data are available at this url.

Multimedia Event Extraction (MEE) has become an important task in information extraction research as news today increasingly prefers to contain multimedia content. Current MEE works mainly face two challenges: (1) Inadequate extraction framework modeling for handling complex and flexible multimedia event structure; (2) The absence of multimodal-aligned training data for effective knowledge transfer to MEE task. In this work, we propose a Stepwise Schema-Guided Prompting Framework (SSGPF) using Multimodal Large Language Model (MLLM) as backbone for adaptive structure capturing to solve MEE task. At the initial step of SSGPF, we design Event Type Schema Guided Prompting (ETSGP) for event detection, then we devise Argument Role Schema Guided Prompting (ARSGP) that contains multi-step prompts with text-bridged grounding technique for argument extraction. We construct a weakly-aligned multimodal event labeled dataset based on existing unimodal event annotations, then conduct parameter efficient instruction tuning with LoRA on LLaVA-v1.5-7B under SSGPF. Experiments on the M2E2 benchmark demonstrate that SSGPF significantly outperforms current SOTA baselines by 5.8 percent F1 on event detection and 8.4 percent F1 on argument extraction.

Compositional Zero-Shot Learning (CZSL) aims at recognizing novel compositions by combining concepts learned from seen compositions. The key to tackle CZSL is disentangling highly coupled attribute-object compositions and learning exclusive concepts. Previous works mainly design networks to learn visual concepts from top-layer representations provided by visual backbones. As visual backbones progressively integrate information layer by layer, some low-level but critical information for concept learning may be lost, and the coupling between attribute and object features deepens. To address these issues, we propose to extract multi-scale visual features and fuse them in an adaptive way by Mixture of Experts (MoE) networks. We also employ feature-level similarity and a maximum entropy regularization term to constrain the model to effectively disentangle and learn concepts from multi-scale visual information. Comprehensive experiments on three CZSL benchmark datasets demonstrate that our method significantly outperforms previous SOTA methods in both closed-world and open-world settings.