No Gaussian Required: Contrastive Inverse Dynamics for JEPA World Models
本文提出了一种无需高斯分布的对比逆动力学方法AC-MTM，通过行动对比任务防止JEPA世界模型训练时的编码崩溃问题。
0 citationsRead paper
Selected work
Representative Papers
NegotiationGym: Self-Optimizing Agents in a Multi-Agent Social Simulation Environment
This study addresses the limited flexibility in modeling negotiation and cooperation, as well as weak strategy evolution mechanisms, in multi-agent social simulation. Methodologically, we propose a configuration-driven multi-agent social simulation framework: (1) agents are modeled via utility functions, enabling multi-round interaction and feedback-driven autonomous strategy optimization; (2) a modular API and interactive visualization interface support low-code scenario configuration and dynamic runtime execution. Our key contribution lies in the deep integration of self-optimizing agent architectures with a configurable simulation framework—achieving, for the first time within a unified platform, the organic unification of social behavior modeling, dynamic strategy evolution, and human-in-the-loop experimentation. Empirical evaluation demonstrates that agents consistently improve both collective cooperation efficiency and individual utility across repeated negotiation episodes.
0 citationsRead paper
GLiREL -- Generalist Model for Zero-Shot Relation Extraction
To address the challenges of identifying unseen relations and low inference efficiency in zero-shot relation extraction, this paper proposes GLiREL, a lightweight and general-purpose model. Methodologically, GLiREL introduces three key innovations: (1) the first zero-shot multi-entity relation classification architecture enabling single forward pass inference; (2) a scalable relation label synthesis protocol that generalizes across diverse relation definitions; and (3) a prompt-driven contrastive learning framework jointly optimizing relation label embedding alignment and a lightweight cross-modal encoder. Evaluated on FewRel and WikiZSL benchmarks, GLiREL achieves state-of-the-art performance in zero-shot relation classification, significantly outperforming existing methods. It delivers substantial improvements in both classification accuracy and inference speed, demonstrating strong generalization to unseen relations while maintaining computational efficiency.
0 citationsRead paper
Recent publications
Latest Papers
No Gaussian Required: Contrastive Inverse Dynamics for JEPA World Models
本文提出了一种无需高斯分布的对比逆动力学方法AC-MTM，通过行动对比任务防止JEPA世界模型训练时的编码崩溃问题。
0 citationsRead paper
NegotiationGym: Self-Optimizing Agents in a Multi-Agent Social Simulation Environment
This study addresses the limited flexibility in modeling negotiation and cooperation, as well as weak strategy evolution mechanisms, in multi-agent social simulation. Methodologically, we propose a configuration-driven multi-agent social simulation framework: (1) agents are modeled via utility functions, enabling multi-round interaction and feedback-driven autonomous strategy optimization; (2) a modular API and interactive visualization interface support low-code scenario configuration and dynamic runtime execution. Our key contribution lies in the deep integration of self-optimizing agent architectures with a configurable simulation framework—achieving, for the first time within a unified platform, the organic unification of social behavior modeling, dynamic strategy evolution, and human-in-the-loop experimentation. Empirical evaluation demonstrates that agents consistently improve both collective cooperation efficiency and individual utility across repeated negotiation episodes.
0 citationsRead paper
GLiREL -- Generalist Model for Zero-Shot Relation Extraction
To address the challenges of identifying unseen relations and low inference efficiency in zero-shot relation extraction, this paper proposes GLiREL, a lightweight and general-purpose model. Methodologically, GLiREL introduces three key innovations: (1) the first zero-shot multi-entity relation classification architecture enabling single forward pass inference; (2) a scalable relation label synthesis protocol that generalizes across diverse relation definitions; and (3) a prompt-driven contrastive learning framework jointly optimizing relation label embedding alignment and a lightweight cross-modal encoder. Evaluated on FewRel and WikiZSL benchmarks, GLiREL achieves state-of-the-art performance in zero-shot relation classification, significantly outperforming existing methods. It delivers substantial improvements in both classification accuracy and inference speed, demonstrating strong generalization to unseen relations while maintaining computational efficiency.
0 citationsRead paper