Chinese idioms (成语, Chengyu) are concise four-character expressions steeped in history and culture, whose literal translations often fail to capture their full meaning.This complexity makes them challenging for language models to interpret and use correctly.Existing benchmarks focus on narrow tasks-multiplechoice cloze tests, isolated translation, or simple paraphrasing.We introduce CHENGYU-BENCH, a comprehensive benchmark featuring three tasks: (1) Evaluative Connotation, classifying idioms as positive or negative; (2) Appropriateness, detecting incorrect idiom usage in context; and (3) Open Cloze, filling blanks in longer passages without options.CHENGYU-BENCH comprises 2,937 human-verified examples covering 1,765 common idioms sourced from diverse corpora.We evaluate leading LLMs and find they achieve over 95% accuracy on Evaluative Connotation, but only ~85% on Appropriateness and ~40% top-1 accuracy on Open Cloze.Error analysis reveals that most mistakes arise from fundamental misunderstandings of idiom meanings.CHENGYU-BENCH demonstrates that while LLMs can reliably gauge idiom sentiment, they still struggle to grasp the cultural and contextual nuances essential for proper usage.The benchmark and source code are available at: https: //github.com/sofyc/ChengyuBench.

Vision-Language Models (VLMs) extend the capabilities of Large Language Models (LLMs) by incorporating visual information, yet they remain vulnerable to jailbreak attacks, especially when processing noisy or corrupted images. Although existing VLMs adopt security measures during training to mitigate such attacks, vulnerabilities associated with noise-augmented visual inputs are overlooked. In this work, we identify that missing noise-augmented training causes critical security gaps: many VLMs are susceptible to even simple perturbations such as Gaussian noise. To address this challenge, we propose Robust-VLGuard, a multimodal safety dataset with aligned / misaligned image-text pairs, combined with noise-augmented fine-tuning that reduces attack success rates while preserving functionality of VLM. For stronger optimization-based visual perturbation attacks, we propose DiffPure-VLM, leveraging diffusion models to convert adversarial perturbations into Gaussian-like noise, which can be defended by VLMs with noise-augmented safety fine-tuning. Experimental results demonstrate that the distribution-shifting property of diffusion model aligns well with our fine-tuned VLMs, significantly mitigating adversarial perturbations across varying intensities. The dataset and code are available at https://github.com/JarvisUSTC/DiffPure-RobustVLM.

Query-based pricing enables personalized data acquisition for data buyers, exhibiting potential in data markets. The state-of-the-art SQL query pricing strategy tackles the #P-hard arbitrage-free pricing task with the quadratic computational complexity, far from promptly fulfilling customer demands. In this paper, we propose a novel arbitrage-free and scalable pricing framework ARIA to calculate the prices for various query types in linear time, including select-project-join and simple aggregate (SPJA) queries. For the first time, we model what the query answer tells about the value of each tuple and formulate the tuple-level information of selection, projection, and simple aggregation queries. We develop several price functions based on the total information gain of all tuples. The containing relationship between the query information prevents possible arbitrage arising from query determinacy. We present efficient price computation algorithms to derive the prices of different types of queries with linear time complexity, which scan the common possible value set of tuples one time. In ARIA, the join query is decomposed as multiple single-relation queries for pricing in linear time. Extensive experiments on real and synthetic datasets demonstrate that, ARIA performs 3x faster than the state of the arts while enjoying desirable pricing characteristics.

Online data marketplaces emerge in diverse data-driven applications, where dynamically arriving consumers pur-chase the data at posted prices. The data value decays over time in many tasks, such as machine learning predictions and realtime systems. Existing query pricing methods do not consider the time-discounting data value. In this paper, we study the query feature-based data pricing problem with unknown time-discounting data valuation. We propose an effective online data pricing mechanism Pride to maximize the cumulative sales revenue. It leverages the powerful property of the ellipsoid method to efficiently solve online optimization via exploration and exploitation. Based on Thompson sampling, we present a novel non-stationary MAB algorithm Biased-TS to determine a suitable discount factor and attain the dynamic posted price. It is theoretically proved that, the regret upper bound order of Pride is dominated by the discretization error$O(\frac{T}{k})$, where$K$and$T$are the numbers of discount candidates and total trading rounds, respectively. Biased-TS gets a sub-linear regret upper bound$O(K^{3}\sqrt{T\ln T}+K\exp\{4\sqrt{\ln T}\})$. Extensive experiments using both synthetic and real datasets demonstrate that Pride yields around 90% of the optimal cumulative revenue, and it substantially outperforms the state-of-the-art methods.

We introduce SwiftSage, a novel agent framework inspired by the dual-process theory of human cognition, designed to excel in action planning for complex interactive reasoning tasks. SwiftSage integrates the strengths of behavior cloning and prompting large language models (LLMs) to enhance task completion performance. The framework comprises two primary modules: the Swift module, representing fast and intuitive thinking, and the Sage module, emulating deliberate thought processes. The Swift module is a small encoder-decoder LM fine-tuned on the oracle agent's action trajectories, while the Sage module employs LLMs such as GPT-4 for subgoal planning and grounding. We develop a heuristic method to harmoniously integrate the two modules, resulting in a more efficient and robust problem-solving process. In 30 tasks from the ScienceWorld benchmark, SwiftSage significantly outperforms other methods such as SayCan, ReAct, and Reflexion, demonstrating its effectiveness in solving complex interactive tasks.