Baidu has introduced Ernie Bot 3.5, an updated version of its chatbot that the company says performs substantially better than its predecessor. The release also adds plugins, giving users ways to connect the chatbot to search and document tasks.
Baidu describes faster processing and stronger results
Compared with version 3.0, Baidu says Ernie Bot 3.5 more than doubles training data throughput and increases inference throughput by 17 times. The company also claims model performance has improved by more than 50 percent.
Baidu points to logical reasoning, mathematical computing, and coding as areas where the updated model is expected to do better. These are company-reported claims; the source article says Baidu cites a China Science Daily report as evidence and notes comparisons with ChatGPT 3.5 and GPT-4 in some Chinese-language application scenarios and benchmarks.
Throughput describes how much data a system processes over time. For users, faster inference may mean quicker responses, while better performance in reasoning, math, and coding could make the chatbot more useful for tasks that require working through instructions or producing structured outputs.
Training changes and knowledge retrieval
Baidu attributes the gains to several changes in how the model is trained and optimized. It lists multi-type, multi-stage supervised fine-tuning, a reward system applied across multiple stages, and a combined approach that uses multiple loss functions.
The company also describes a technique called double-flywheel model optimization, which it says is intended to support continued efficiency improvements. The source does not provide detailed measurements for each technique, so the reported overall performance increase remains Baidu's claim.
Another addition is “Knowledge Snippet Enhancement.” The process breaks a user's query into relevant pieces of knowledge, then uses Baidu's internal knowledge graph and external search engines to locate useful information and form further prompts. Baidu says this improves how the model understands and uses world knowledge.
In practical terms, the feature is designed to let Ernie draw on retrieved information when responding, rather than relying only on what is encoded in the model. The article does not describe how the system selects sources or how users can inspect the information it retrieves.
Plugins extend what Ernie can do
Ernie Bot 3.5 also introduces plugin support, a feature Baidu compares with offerings from Google Bard, Microsoft Bing Chat, and OpenAI's ChatGPT. The initial standard plugins include Baidu Search for internet search and ChatFile for working with files.
According to Baidu, ChatFile can summarize long texts and generate frequently asked questions. The company says it plans to release more of its own plugins and let third-party developers contribute plugins as well. That could broaden the range of tasks available through the chatbot, though the source does not give a timeline for those additions.
Public beta, with broader ambitions
At the time described in the source, Ernie Bot was in public beta testing. Baidu said the model could potentially support applications involving language, text, or code, positioning it as a platform for more than conversation alone.
The company had previously presented an industrial version of Ernie at a closed event in late March. Demonstrated uses included summarizing financial statements, creating PowerPoint presentations, building personalized travel plans, and making virtual avatars for product advertisements.
Those examples suggest Baidu's ambitions span both everyday chatbot features and business applications. Ernie Bot 3.5 brings together claimed performance gains, knowledge retrieval, and plugins, but the source frames the headline improvements as Baidu's own account while the product is still being tested publicly.