Vision-Language Models (VLMs) have demonstrated impressive capabilities in code generation across various domains. However, their ability to replicate complex, multi-panel visualizations from real-world data remains largely unassessed. To address this gap, we introduce RealChart2Code, a new large-scale benchmark with over 2,800 instances grounded in authentic datasets and featuring tasks with clear analytical intent. Crucially, it is the first benchmark to systematically evaluate chart generation from large-scale raw data and assess iterative code refinement in a multi-turn conversational setting. Our comprehensive evaluation of 14 leading VLMs on RealChart2Code reveals significant performance degradation compared to simpler benchmarks, highlighting their struggles with complex plot structures and authentic data. Our analysis uncovers a substantial performance gap between proprietary and open-weight models and confirms that even state-of-the-art VLMs often fail to accurately replicate intricate, multi-panel charts. These findings provide valuable insights into the current limitations of VLMs and guide future research directions.

Existing research in multi-hop questions has identified two reasoning modes: latent reasoning and factual shortcuts, but has not deeply investigated how these modes differ during inference. This impacts both model generalization ability and downstream reasoning tasks. In this work, we systematically examine these distinctions and propose a simple and efficient classification metric, Attribute Rate Ratio (ARR). First, we construct specialized datasets corresponding to the two reasoning modes based on our proposed criteria. Then, using reverse engineering methods, including attention knockout and logit lens techniques, we reveal that subject representations differ significantly across modes: latent reasoning encodes bridge-related information for final answer extraction, while factual shortcuts bypass intermediate reasoning and resemble single-hop factual queries. Finally, our proposed ARR achieves around 90% accuracy on our datasets and demonstrates effectiveness in RAG conflict scenarios, showing that model behavior under conflicting prompts is closely tied to its underlying reasoning mode. Our findings and proposed metric have significant potential for advancing LLM development and applications.

Open Information Extraction (OIE) seeks to extract structured information from raw text without the limitations of close ontology. Recently, the detection-based OIE methods have received great attention from the community due to their parallelism. However, as the essential step of those models, how to assign ground truth labels to the parallelly generated tuple proposals remains under-exploited. The commonly utilized Hungarian algorithm for this procedure is restricted to handling one-to-one assignment among the desired tuples and tuple proposals, which ignores the correlation between proposals and affects the recall of the models. To solve this problem, we propose a dynamic many-to-one label assignment strategy named IOT. Concretely, the label assignment process in OIE is formulated as an Optimal Transport (OT) problem. We leverage the intersection-over-union (IoU) as the assignment quality measurement, and convert the problem of finding the best assignment solution to the one of solving the optimal transport plan by maximizing the IoU values. To further utilize the knowledge from the assignment, we design an Assignment-guided Multi-granularity loss (AM) by simultaneously considering word-level and tuple-level information. Experiment results show the proposed method outperforms the state-of-the-art models on three benchmarks.