Zihe Wang

Zheda Mai, Arpita Chowdhury, Zihe Wang et al.

The rise of vision foundation models (VFMs) calls for systematic evaluation. A common approach pairs VFMs with large language models (LLMs) as general-purpose heads, followed by evaluation on broad Visual Question Answering (VQA) benchmarks. However, this protocol has two key blind spots: (i) the instruction tuning data may not align with VQA test distributions, meaning a wrong prediction can stem from such data mismatch rather than a VFM' visual shortcomings; (ii) VQA benchmarks often require multiple visual abilities, making it hard to tell whether errors stem from lacking all required abilities or just a single critical one. To address these gaps, we introduce AVA-Bench, the first benchmark that explicitly disentangles 14 Atomic Visual Abilities (AVAs) -- foundational skills like localization, depth estimation, and spatial understanding that collectively support complex visual reasoning tasks. By decoupling AVAs and matching training and test distributions within each, AVA-Bench pinpoints exactly where a VFM excels or falters. Applying AVA-Bench to leading VFMs thus reveals distinctive "ability fingerprints," turning VFM selection from educated guesswork into principled engineering. Notably, we find that a 0.5B LLM yields similar VFM rankings as a 7B LLM while cutting GPU hours by 8x, enabling more efficient evaluation. By offering a comprehensive and transparent benchmark, we hope AVA-Bench lays the foundation for the next generation of VFMs.

Hongtao Liu, Zhicheng Du, Zihe Wang et al.

Game-playing ability serves as an indicator for evaluating the strategic reasoning capability of large language models (LLMs). While most existing studies rely on utility performance metrics, which are not robust enough due to variations in opponent behavior and game structure. To address this limitation, we propose \textbf{Cognitive Hierarchy Benchmark (CHBench)}, a novel evaluation framework inspired by the cognitive hierarchy models from behavioral economics. We hypothesize that agents have bounded rationality -- different agents behave at varying reasoning depths/levels. We evaluate LLMs' strategic reasoning through a three-phase systematic framework, utilizing behavioral data from six state-of-the-art LLMs across fifteen carefully selected normal-form games. Experiments show that LLMs exhibit consistent strategic reasoning levels across diverse opponents, confirming the framework's robustness and generalization capability. We also analyze the effects of two key mechanisms (Chat Mechanism and Memory Mechanism) on strategic reasoning performance. Results indicate that the Chat Mechanism significantly degrades strategic reasoning, whereas the Memory Mechanism enhances it. These insights position CHBench as a promising tool for evaluating LLM capabilities, with significant potential for future research and practical applications.

Zhenyang Feng, Zihe Wang, Jianyang Gu et al.

We study image segmentation in the biological domain, particularly trait segmentation from specimen images (e.g., butterfly wing stripes, beetle elytra). This fine-grained task is crucial for understanding the biology of organisms, but it traditionally requires manually annotating segmentation masks for hundreds of images per species, making it highly labor-intensive. To address this challenge, we propose a label-efficient approach, Static Segmentation by Tracking (SST), based on a key insight: while specimens of the same species exhibit natural variation, the traits of interest show up consistently. This motivates us to concatenate specimen images into a ``pseudo-video'' and reframe trait segmentation as a tracking problem. Specifically, SST generates masks for unlabeled images by propagating annotated or predicted masks from the ``pseudo-preceding'' images. Built upon recent video segmentation models, such as Segment Anything Model 2, SST achieves high-quality trait segmentation with only one labeled image per species, marking a breakthrough in specimen image analysis. To further enhance segmentation quality, we introduce a cycle-consistent loss for fine-tuning, again requiring only one labeled image. Additionally, we demonstrate the broader potential of SST, including one-shot instance segmentation in natural images and trait-based image retrieval.

Zihe Wang, Zhide Wei, Jie Zhang

The Probabilistic Serial mechanism is well-known for its desirable fairness and efficiency properties. It is one of the most prominent protocols for the random assignment problem. However, Probabilistic Serial is not incentive-compatible, thereby these desirable properties only hold for the agents' declared preferences, rather than their genuine preferences. A substantial utility gain through strategic behaviors would trigger self-interested agents to manipulate the mechanism and would subvert the very foundation of adopting the mechanism in practice. In this paper, we characterize the extent to which an individual agent can increase its utility by strategic manipulation. We show that the incentive ratio of the mechanism is $\frac{3}{2}$. That is, no agent can misreport its preferences such that its utility becomes more than 1.5 times of what it is when reports truthfully. This ratio is a worst-case guarantee by allowing an agent to have complete information about other agents' reports and to figure out the best response strategy even if it is computationally intractable in general. To complement this worst-case study, we further evaluate an agent's utility gain on average by experiments. The experiments show that an agent' incentive in manipulating the rule is very limited. These results shed some light on the robustness of Probabilistic Serial against strategic manipulation, which is one step further than knowing that it is not incentive-compatible.