Zhitian Xie

Yao Zhao, Zhitian Xie, Chen Liang et al.

As Large Language Models (LLMs) have made significant advancements across various tasks, such as question answering, translation, text summarization, and dialogue systems, the need for accuracy in information becomes crucial, especially for serious financial products serving billions of users like Alipay. However, for a real-world product serving millions of users, the inference speed of LLMs becomes a critical factor compared to a mere experimental model. Hence, this paper presents a generic framework for accelerating the inference process, resulting in a substantial increase in speed and cost reduction for our LLM-based scenarios, with lossless generation accuracy. In the traditional inference process, each token is generated sequentially by the LLM, leading to a time consumption proportional to the number of generated tokens. To enhance this process, our framework, named \textit{lookahead}, introduces a \textit{multi-branch} strategy. Instead of generating a single token at a time, we propose a Trie-based retrieval and verification mechanism to be able to accept several tokens at a forward step. Our strategy offers two distinct advantages: (1) it guarantees absolute correctness of the output, avoiding any approximation algorithms, and (2) the worst-case performance of our approach is equivalent to the conventional process. We conduct extensive experiments to demonstrate the significant improvements achieved by applying our inference acceleration framework. Our framework is widely deployed in Alipay since April 2023, and obtain remarkable 2.66x to 6.26x speedup. Our code is available at https://github.com/alipay/PainlessInferenceAcceleration.

Zhitian Xie, Qintong Wu, Chengyue Yu et al.

The rapid advancement of large language models (LLMs) has empowered intelligent agents to leverage diverse external tools for solving complex real-world problems. However, this reliance introduces new challenges, as extended contexts and noisy tool outputs can undermine system reliability. To address this, we propose a dynamic Multi-Agent System (MAS) in our AWorld framework, where an Execution Agent is supervised by a Guard Agent that provides on-demand dynamic maneuvering, verifying and correcting the reasoning process to improve robustness over single-agent systems. To move beyond this generic supervision, we enhance the architecture with a methodology inspired by System Identification from control theory. This method first profiles the Execution Agent offline on a benchmark dataset to create a "performance fingerprint" of its unique weaknesses. The Guard Agent then leverages this fingerprint online to deliver profile-aware supervision, making targeted interventions based on known failure patterns rather than merely reacting to immediate logical flaws. Extensive experiments on the GAIA dataset demonstrate that this profile-aware MAS significantly improves both effectiveness and stability, outperforming not only single-agent systems but also its naive counterpart. This superior performance led our system to achieve first place among open-source projects on the prestigious GAIA leaderboard. These findings highlight that building truly trustworthy intelligent systems requires not just collaboration, but a deep, empirically-grounded understanding of each agent's unique capabilities and limitations.

Zhitian Xie, Yinger Zhang, Chenyi Zhuang et al.

The application of mixture-of-experts (MoE) is gaining popularity due to its ability to improve model's performance. In an MoE structure, the gate layer plays a significant role in distinguishing and routing input features to different experts. This enables each expert to specialize in processing their corresponding sub-tasks. However, the gate's routing mechanism also gives rise to narrow vision: the individual MoE's expert fails to use more samples in learning the allocated sub-task, which in turn limits the MoE to further improve its generalization ability. To effectively address this, we propose a method called Mixture-of-Distilled-Expert (MoDE), which applies moderate mutual distillation among experts to enable each expert to pick up more features learned by other experts and gain more accurate perceptions on their original allocated sub-tasks. We conduct plenty experiments including tabular, NLP and CV datasets, which shows MoDE's effectiveness, universality and robustness. Furthermore, we develop a parallel study through innovatively constructing "expert probing", to experimentally prove why MoDE works: moderate distilling knowledge can improve each individual expert's test performances on their assigned tasks, leading to MoE's overall performance improvement.