Zhaotian Weng

Chengzhi Liu, Yuzhe Yang, Sophia Xiao Pu et al.

Multimodal large language models are increasingly deployed as long-horizon agents, where memory must do more than recall: it must track an evolving world, revise what has gone stale, and surface the right evidence at decision time. Existing benchmarks measure recall over static dialogue, collapse memory into a single end-of-task accuracy, and reduce visual observations to captions, leaving us unable to localize failures to writing, maintenance, retrieval, or use. The rise of agent harnesses that author their own memory sharpens this gap, since we have no principled way to compare hand-designed pipelines with self-managing alternatives. To close these gaps, we formulate multimodal agent memory as an Action-World Interaction Loop with an observable four-stage lifecycle, and instantiate it in WorldMemArena: 400 multi-session multimodal tasks spanning Lifelong Evolution (evolving personal and task states) and Agentic Execution (memory from real observations, actions, and feedback), annotated with gold memory points, updates, distractors, and evidence chains for stage-level diagnosis. This enables the first head-to-head comparison of long-context, manually designed (RAG and external memory systems), and harness-based memory agents. Results show that: (1) better memory writing and storage do not guarantee better performance; (2) multimodal memory still struggles to fully use visual evidence; (3) systems are unstable across domains and degrade on realistic agentic trajectories; and (4) harness memory is more flexible but remains costly and less reliable.

Sophia Xiao Pu, Zhaotian Weng, Chengzhi Liu et al.

Self-play reinforcement learning trains language models on their own generated tasks, co-evolving a proposer and solver without human labels. Recent systems report strong reasoning gains, but collapse and instability are widely observed and poorly understood. The dominant response treats this as a reward-design problem. We argue instead that self-play stability is governed by two distinct levers: a data-level gate that decides which proposer-generated tasks enter the training pool, and the reward signal that updates the policy on tasks already admitted. Through controlled experiments on a Python output-prediction task and a deterministic-DSL twin task that strips pretraining priors, output ambiguity, and executor noise, we find the two levers are asymmetric. A strict gate is sufficient for stability under every reward variant we test, including a self-consistency reward with no access to ground truth; while no reward variant is sufficient once the gate is removed. This asymmetry exposes a counter-intuitive coupling we call the Grounded Proposer Paradox: a proposer with ground-truth access accelerates collapse faster than an ungrounded one when paired with a self-consistency solver, by concentrating training on clean tasks that form the fastest path to a spurious self-consistent attractor. Replacing the binary gate with a continuous strictness parameter $\varepsilon$ further reveals a two-stage phase transition: training-side metrics decouple at low $\varepsilon$, while validation accuracy holds until $\varepsilon$ is much higher. Data-level gating, not reward calibration, is the binding constraint on self-play stability.

Zhaotian Weng, Zijun Gao, Jerone Andrews et al.

Vision-language models (VLMs) pre-trained on extensive datasets can inadvertently learn biases by correlating gender information with specific objects or scenarios. Current methods, which focus on modifying inputs and monitoring changes in the model's output probability scores, often struggle to comprehensively understand bias from the perspective of model components. We propose a framework that incorporates causal mediation analysis to measure and map the pathways of bias generation and propagation within VLMs. This approach allows us to identify the direct effects of interventions on model bias and the indirect effects of interventions on bias mediated through different model components. Our results show that image features are the primary contributors to bias, with significantly higher impacts than text features, specifically accounting for 32.57% and 12.63% of the bias in the MSCOCO and PASCAL-SENTENCE datasets, respectively. Notably, the image encoder's contribution surpasses that of the text encoder and the deep fusion encoder. Further experimentation confirms that contributions from both language and vision modalities are aligned and non-conflicting. Consequently, focusing on blurring gender representations within the image encoder, which contributes most to the model bias, reduces bias efficiently by 22.03% and 9.04% in the MSCOCO and PASCAL-SENTENCE datasets, respectively, with minimal performance loss or increased computational demands.

Zhaotian Weng, Antonis Antoniades, Deepak Nathani et al.

Open-ended self-improving agents can autonomously modify their own structural designs to advance their capabilities and overcome the limits of pre-defined architectures, thus reducing reliance on human intervention. We introduce Group-Evolving Agents (GEA), a new paradigm for open-ended self-improvements, which treats a group of agents as the fundamental evolutionary unit, enabling explicit experience sharing and reuse within the group throughout evolution. Unlike existing open-ended self-evolving paradigms that adopt tree-structured evolution, GEA overcomes the limitation of inefficient utilization of exploratory diversity caused by isolated evolutionary branches. We evaluate GEA on challenging coding benchmarks, where it significantly outperforms state-of-the-art self-evolving methods (71.0% vs. 56.7% on SWE-bench Verified, 88.3% vs. 68.3% on Polyglot) and matches or exceeds top human-designed agent frameworks (71.8% and 52.0% on two benchmarks, respectively). Analysis reveals that GEA more effectively converts early-stage exploratory diversity into sustained, long-term progress, achieving stronger performance under the same number of evolved agents. Furthermore, GEA exhibits consistent transferability across different coding models and greater robustness, fixing framework-level bugs in 1.4 iterations on average, versus 5 for self-evolving methods.

Zhaotian Weng, Jianbo Hong, Lan Wang

Conditional Generative Adversarial Nets (CGAN) is often used to improve conditional image generation performance. However, there is little research on Representation learning with CGAN for causal inference. This paper proposes a new method for finding representation learning functions by adopting the adversarial idea. We apply the pattern of CGAN and theoretically emonstrate the feasibility of finding a suitable representation function in the context of two distributions being balanced. The theoretical result shows that when two distributions are balanced, the ideal representation function can be found and thus can be used to further research.

Haoxuan Li, Mingyu Derek Ma, Jen-tse Huang et al.

Detecting biases in structured data is a complex and time-consuming task. Existing automated techniques are limited in diversity of data types and heavily reliant on human case-by-case handling, resulting in a lack of generalizability. Currently, large language model (LLM)-based agents have made significant progress in data science, but their ability to detect data biases is still insufficiently explored. To address this gap, we introduce the first end-to-end, multi-agent synergy framework, BIASINSPECTOR, designed for automatic bias detection in structured data based on specific user requirements. It first develops a multi-stage plan to analyze user-specified bias detection tasks and then implements it with a diverse and well-suited set of tools. It delivers detailed results that include explanations and visualizations. To address the lack of a standardized framework for evaluating the capability of LLM agents to detect biases in data, we further propose a comprehensive benchmark that includes multiple evaluation metrics and a large set of test cases. Extensive experiments demonstrate that our framework achieves exceptional overall performance in structured data bias detection, setting a new milestone for fairer data applications.

Zhaotian Weng, Haoxuan Li, Kuan-Hao Huang et al.

Despite the impressive performance of vision-language models (VLMs) on downstream tasks, their ability to understand and reason about causal relationships in visual inputs remains unclear. Robust causal reasoning is fundamental to solving complex high-level reasoning tasks, yet existing benchmarks often include a mixture of reasoning questions, and VLMs can frequently exploit object recognition and activity identification as shortcuts to arrive at the correct answers, making it challenging to truly assess their causal reasoning abilities. To bridge this gap, we introduce VQA-Causal and VCR-Causal, two new benchmarks specifically designed to isolate and rigorously evaluate VLMs' causal reasoning abilities. Our findings reveal that while VLMs excel in object and activity recognition, they perform poorly on causal reasoning tasks, often only marginally surpassing random guessing. Further analysis suggests that this limitation stems from a severe lack of causal expressions in widely used training datasets, where causal relationships are rarely explicitly conveyed. We additionally explore fine-tuning strategies with hard negative cases, showing that targeted fine-tuning can improve model's causal reasoning while maintaining generalization and downstream performance. Our study highlights a key gap in current VLMs and lays the groundwork for future work on causal understanding.

Chenxiao Yu, Zhaotian Weng, Yuangang Li et al.

Can Large Language Models (LLMs) accurately predict election outcomes? While LLMs have demonstrated impressive performance in various domains, including healthcare, legal analysis, and creative tasks, their ability to forecast elections remains unknown. Election prediction poses unique challenges, such as limited voter-level data, rapidly changing political landscapes, and the need to model complex human behavior. To address these challenges, we introduce a multi-step reasoning framework designed for political analysis. Our approach is validated on real-world data from the American National Election Studies (ANES) 2016 and 2020, as well as synthetic personas generated by the leading machine learning framework, offering scalable datasets for voter behavior modeling. To capture temporal dynamics, we incorporate candidates' policy positions and biographical details, ensuring that the model adapts to evolving political contexts. Drawing on Chain of Thought prompting, our multi-step reasoning pipeline systematically integrates demographic, ideological, and time-dependent factors, enhancing the model's predictive power.

Zerun Wang, Ling Xiao, Liuyu Xiang et al.

Open-set semi-supervised object detection (OSSOD) task leverages practical open-set unlabeled datasets that comprise both in-distribution (ID) and out-of-distribution (OOD) instances for conducting semi-supervised object detection (SSOD). The main challenge in OSSOD is distinguishing and filtering the OOD instances (i.e., outliers) during pseudo-labeling since OODs will affect the performance. The only OSSOD work employs an additional offline OOD detection network trained solely with labeled data to solve this problem. However, the limited labeled data restricts the potential for improvement. Meanwhile, the offline strategy results in low efficiency. To alleviate these issues, this paper proposes an end-to-end online OSSOD framework that improves performance and efficiency: 1) We propose a semi-supervised outlier filtering method that more effectively filters the OOD instances using both labeled and unlabeled data. 2) We propose a threshold-free Dual Competing OOD head that further improves the performance by suppressing the error accumulation during semi-supervised outlier filtering. 3) Our proposed method is an online end-to-end trainable OSSOD framework. Experimental results show that our method achieves state-of-the-art performance on several OSSOD benchmarks compared to existing methods. Moreover, additional experiments show that our method is more efficient and can be easily applied to different SSOD frameworks to boost their performance.