Introducing ERNIE 3.5: Baidu’s Knowledge-Enhanced Foundation Model Takes a Giant Leap Forward


Back to list


-       ERNIE Bot v2.1.0, released on June 21, is powered by ERNIE 3.5.

-       One defining feature of ERNIE 3.5 is plugins, including “Baidu Search” and “ChatFile”.

-       ERNIE 3.5 dramatically boosts inference throughput by an astonishing 17-fold compared to ERNIE 3.0.


We are thrilled to announce the iteration of our foundation model, ERNIE, to version 3.5. ERNIE 3.5 has made significant strides in beta testing, surpassing ChatGPT (3.5) in comprehensive ability scores and outperforming GPT-4 in several Chinese language capabilities, as reported by China Science Daily.


Just three months after the beta release of ERNIE Bot, Baidu’s large language model (LLM) built on ERNIE 3.0, ERNIE 3.5 has achieved broad enhancements in efficacy, functionality, and performance. These improvements are evident in creative writing, Q&A, reasoning, and code generation, as well as in training performance and inference performance, said Dr. Haifeng Wang, CTO of Baidu.


ERNIE 3.5 dramatically boosts training throughput by two-fold and inference throughput by an astonishing 17-fold compared to ERNIE 3.0. These upgrades will substantially accelerate our model iteration upgrades, reduce training and usage costs, and enhance user experience.


New Plugins Expand the Capabilities of ERNIE


One defining feature of ERNIE 3.5 is plugins. For example, the default built-in plugin “Baidu Search” equips ERNIE Bot with the ability to generate real-time and precise information. Another ChatFile plugin enables long text summary and Q&A. “ERNIE 3.5 expands the model’s capabilities through plugins,” Dr. Wang explained.


In the future, ERNIE Bot will add more high-quality plugins from Baidu and third parties. We are also committed to opening the plugin ecosystem to third-party developers, empowering them to build unique applications based on ERNIE.


Constant Innovation: Expanding ERNIE's Knowledge Base


In ERNIE 3.5, we’ve implemented cutting-edge strategies from PaddlePaddle, including adaptive hybrid parallel training technology and mixed-precision computing, Dr. Wang explained. These enhancements, combined with optimized data sources and data distribution, have accelerated our model’s iteration speed, bolstered its efficacy, and ensured its safety.


We’ve further improved the model performance through the multi-type multi-stage supervised fine-tuning, the multi-level multi-grain reward model, the mixed optimization of multiple loss functions, and the double-flywheel model optimization.


On top of the previous Knowledge Enhancement and Retrieval Enhancement, ERNIE 3.5 has further implemented a technique called “Knowledge Snippet Enhancement.” Specifically, the model analyzes user queries and identifies relevant knowledge snippets. It then uses knowledge graph and search engine to find corresponding answers, subsequently using these snippets to write prompts. This technique significantly enhances the model’s understanding and utilizing of world knowledge, leading to remarkable task improvements.


Additionally, we have improved the reasoning of ERNIE 3.5 in logical reasoning, mathematical computation, and code generation, through large-scale logical data construction, logical knowledge modeling, the combination of coarse-grained and fine-grained semantic knowledge, and symbolic neural networks.


ERNIE Ready for Applications


ERNIE Bot, currently in public beta testing, has been upgraded to version 2.0 since May 23. The most recent update, ERNIE Bot v2.1.0, released on June 21, added the new ChatFile plugin and improved capabilities in mathematical computations and creative writing. These enhancements are all powered by ERNIE 3.5.


Dr. Wang said that test users can access the service at any time to experience the performance of ERNIE 3.5.


“Any applications involving language, text, or code can potentially utilize ERNIE Bot,” said Dr. Wang. He elaborated that many applications are already using ERNIE Bot, spanning fields such as smart offices, coding, marketing, media, education, and finance. For example, Baidu's smart work platform, Infoflow, has unveiled several new features such as “smart summary”, “smart insights”, and “super assistant”, all derived from ERNIE Bot.


In coding, Baidu’s smart coding assistant, Comate, leverages ERNIE Bot’s capabilities to generate corresponding code snippets based on natural language prompts. Additionally, it can auto-generate code within the code editor based on comments, thereby improving development efficiency.