Tencent Seeks AI Collaboration to Enhance Support for Vulnerable Users
- tech360.tv
- 4 hours ago
- 3 min read
Tencent Research Institute (TRI) is seeking collaboration with major artificial intelligence (AI) developers to improve how generative AI services, including chatbots, interact with vulnerable users. This includes the elderly and left-behind children, who increasingly rely on these services for emotional support and health assistance.
Senior researcher Lu Shiyu at TRI stated that specialised data sets can make AI services more beneficial for these groups. Lu’s team has been developing these data sets with vulnerable users in mind since 2024.
Data sets equip large language models (LLMs) with general knowledge before they are fine-tuned and deployed. TRI, the public strategy research unit of Shenzhen-based Tencent, seeks to build expert data sets for AI models that could broadly benefit the world’s second most populous nation.
Last year, Lu’s team and researchers at the University of Science and Technology Beijing tested leading US and Chinese-developed LLMs, including Tencent’s own Hunyuan model. They discovered that all models needed improvements in handling topics such as sex education and other subjects relevant to China’s 69 million left-behind children.
The team also collaborated with Chinese non-profit organisations that assist vulnerable groups to develop an “elderly data set.” This set was derived from thousands of example question-and-answer forms contributed by elderly respondents.
Lu emphasised the next step is to work with major AI model developers that have many users to make an impact. However, training is challenging because every new piece of data must be carefully judged as it affects a model’s overall performance.
The TRI team is now exploring the development of expert data sets, which consist of high-quality examples contributed by specialists in specific fields rather than end-users. These can adapt AI chatbots for specific use cases during deployment by instructing the LLM to retrieve and integrate external information, a technique known as retrieval-augmented generation.
Lu noted that expert data sets can be used with a technique known as retrieval-augmented generation. This shift towards more specialised data sets could fulfil a niche demand in the industry, even with tech companies having their own proprietary data.
According to Lu, companies require highly specific, narrow expert data sets to continue improving their models. He suggested that a highly authoritative expert data set, for example about elderly issues, might attract their interest.
A report by AliResearch, an institute under Alibaba Group Holding, found that more than 50% of Chinese aged over 50 were active AI service users. This report also showed 45% of respondents aged 76 and above were “using AI frequently every day and can’t live without it”.
This makes improving AI outputs related to prominent concerns of the elderly, such as dealing with loneliness and health problems, a natural priority in China, Lu said. Currently, the take-up of TRI’s elderly data set is low, as domestic non-profit organisations lack the technical know-how and resources to use the data sets on their own AI services.
Domestic non-profit organisations lack the technical know-how and resources to use the data sets on their own AI services. He Jingwen, an assistant researcher at TRI, added that proving the commercial value of these specialised data sets presents an additional challenge.
A study by researchers at Stanford University and Carnegie Mellon University noted some AI models in China were too flattering to users. This suggests users were not necessarily suffering from a lack of emotional support from their AI companions.
To drive adoption, TRI is considering open-sourcing its data sets, aligning with the country’s tech giants supporting the domestic open-source AI ecosystem. Even if these specialised data sets end up not being used by big AI companies, this open-source approach would enable individual developers and entrepreneurs to adopt them for their own products and applications.
Lu stated that the transparency and collaborative nature of the open-source ecosystem can increase awareness of the data sets. Ultimately, this could drive their adoption in practical applications.
Tencent Research Institute seeks collaboration with major AI developers to improve generative AI services for vulnerable users.
Specialised data sets are being developed to enhance AI interaction with groups such as the elderly and left-behind children.
Researchers found existing large language models needed improvement on specific topics relevant to vulnerable populations.
Source: SCMP
