This page organizes papers related to artificial intelligence published around the world. This page is summarized using Google Gemini and is operated on a non-profit basis. The copyright of the paper belongs to the author and the relevant institution. When sharing, simply cite the source.
Self-Evolving LLMs via Continual Instruction Tuning
Created by
Haebom
Author
Jiazheng Kang, Le Huang, Cheng Hou, Zhe Zhao, Zhenxiang Yan, Chuan Shi, Ting Bai
Outline
MoE-CL is a parameter-efficient adversarial mixed-expert framework for continuous training of large-scale language models (LLMs) that must self-evolve to adapt to dynamic data distributions in large-scale industrial environments. MoE-CL utilizes a dual-expert design, preserving task-specific knowledge through LoRA experts for each task to mitigate forgetting, while shared LoRA experts enable inter-task transfer. To prevent task-related noise transmission through shared paths, a task-aware discriminator is integrated within the GAN.
Takeaways, Limitations
•
MoE-CL utilizes dedicated LoRA experts with independent parameters for task-specific knowledge maintenance and shared LoRA experts for inter-task transfer.
•
Through adversarial learning, a shared expert learns generalized representations that mimic the discriminator, while a dedicated expert maintains task-specific details, balancing knowledge preservation and cross-task generalization.
•
The effectiveness of MoE-CL has been proven in the MTL5 public benchmark and the industry Tencent3 benchmark.
•
In an A/B test for content compliance review on the Tencent Video platform, MoE-CL reduced manual review costs by 15.3%.
•
This paper does not make any specific reference to MoE-CL's Limitations.
