At the 2025 World Artificial Intelligence Conference and Tencent Forum, Tencent Research Institute, together with Tencent Youtu Laboratory, Tencent Cloud Intelligence, and Tencent Technology jointly released the report "Symbiosis Partner: Top Ten Trends of Artificial Intelligence in 2025". Based on the long-term observations of many researchers from Tencent Research Institute on global technology and industry trends, the leap of the three major themes of the development of artificial intelligence in 2025, the rise of intelligent actors, and the move towards the physical world by 10 key trends, and the key leap of AI from "smart tools" to "symbiotic partners". The following is a detailed summary of these ten trends.
1. The transition of basic models
01 Reinforcement learning: leading new breakthroughs in big model reasoning and action capabilities
Reinforcement learning is becoming a key force in big model training. From the initial human feedback reinforcement learning (RLHF) to today’s reinforcement learning based on verifiable rewards (RLVR), AI’s inference ability has been significantly improved. This shift not only allows models to solve complex scientific and engineering problems, but also drives the transition from a "language generator" to a "task executor." For example, DeepSeek-R1-Zero demonstrates strong inference capabilities through pure reinforcement learning, providing new ideas for AI in the fields of industrial robot path optimization and complex logistics network scheduling.
02 Native multimodal generation: a new era of unified perception and generation
Native multimodal technology achieves the deep fusion of multimodal data such as images, speech and text by building a unified cross-modal representation space. This technological breakthrough allows AI to complete multimodal joint perception and generation within a unified framework. For example, OpenAI's GPT-4o and Sora models have been able to achieve seamless interaction between text, images and audio, greatly expanding the application boundaries of AI in the fields of creative industries, education and entertainment.
03 The broad evolution of sound model: moving towards emotional intelligence toward sentimentality
Phonetic synthesis technology is shifting from mechanized text reading to emotional expression based on contextual understanding. Models such as ElevenLabs V3 and Hume Octave not only support multilingual and dialects, but also automatically adjust the tone and emotional colors according to the context. In addition, music generation technology is also constantly maturing, moving from clip creation to the generation of complete works. These technological advances have enabled AI to have stronger capabilities in areas such as voice interaction, content creation and emotional companionship.
2. The rise of intelligent actors
04 Dual-track evolution of agents: orchestration and end-to-end progress together
The development of AI Agent is showing two major technical routes: orchestration and end-to-end. Orchestration Agent orchestrates the interaction between LLM and external tools through predefined code paths, which is suitable for complex tasks of enterprise-level workflow automation and multi-tool integration; while the end-to-end Agent model internalizes the inference, planning and tool usage capabilities into the model through reinforcement learning, which is suitable for professional tasks that require deep inference. The two routes have their own advantages and will develop in parallel in the long term in the future.
05LifeOS: AI has become an operating system for personalized life
<br//
The concept of LifeOS is gradually becoming a reality. AI will integrate users' multimodal data and become the "digital self" of users' lives through long-term memory and personalized reasoning capabilities. For example, ChatGPT's Memory feature has been able to preserve the user's writing style and long-term goals across sessions and automatically inject background knowledge based on user's instructions. This "lifelong memory" capability will enable AI to provide more personalized and proactive services and become a "life partner" for users' lives and work.
06 Intelligence as a Service: Intelligent workflow empowers industrial upgrading
AI is moving from "computer power-driven" to "intellectual power-driven" and becoming a native component of enterprise knowledge systems and process structure. Through the RAG (retrieval enhancement generation) architecture and data flywheel mechanism, enterprises can transform internal knowledge and data into real-time callable cognitive systems. For example, Microsoft 365 Copilot and FAW Toyota's customer service systems have significantly improved work efficiency and customer satisfaction by integrating large language models and domain knowledge. In the future, Agent will become a "digital employee" in the enterprise process, taking on the role of closed-loop process and task agent, and promoting the intelligent transformation of the enterprise.
07 Game Agent: The Immersive Evolution of AI in the Virtual World
Game Agent is transforming from simple tools to players’ smart partners. Through deep reinforcement learning and multimodal perception technology, AI teammates can understand player intentions, predict the direction of the battle, and make the best decisions at critical moments. For example, Tencent’s Honor of Kings AI Coaching system and the “Dark Zone Breakthrough” AI teammate system have been able to provide personalized training plans and tactical suggestions, greatly improving the gaming experience. In addition, game agents have made breakthroughs in emotional resonance and group intelligence, laying the foundation for the arrival of the meta-universe era.
3. AI moves to the physical world
08 The "GPT-2 Moment" of Embodied Intelligence: the coordinated evolution of basic models, data engineering and software platforms
2025 is expected to become the "GPT-2 Moment" of Embodied Intelligence, marking a major leap from virtual computing to physical execution. The basic model of embodied intelligence is evolving towards multimodality, and the vision-language-action (VLA) model has become the core of robots' advanced cognitive and execution capabilities. At the same time, large-scale real and synthetic data provide sufficient "fuel" for model training, while a unified software platform provides a solid foundation for development, deployment and ecological construction. For example, companies such as Tencent and Nvidia are building robot simulation and training platforms to accelerate the implementation of embodied intelligent technologies.
09 Space Intelligence: From being able to chat to truly understand the world
The rise of spatial intelligence means that AI moves from processing two-dimensional information to a new stage in processing three-dimensional space. Through technologies such as three-dimensional perception, spatial representation and scene generation, AI can predict three-dimensional space just like predicting the next sentence text. For example, World Labs has demonstrated a prototype model of "one image generates a 3D world", enabling the generation of fully simulated and interactive three-dimensional scenes from a single image or a sentence. Space intelligence not only brings new development opportunities to fields such as autonomous driving, robot manufacturing and XR mixed reality, but also provides AI with key physical common sense and causal reasoning capabilities to move towards general artificial intelligence (AGI).
10 Testing is transformed into mass production, and applications promote the accelerated maturity of embodied intelligent bodies
Embed intelligent robots are moving from laboratories to industrialization, ushering in a critical turning point from testing to mass production. In 2025, major embodied intelligent robot manufacturers began to increase their pilot efforts in the fields of industry, logistics, warehousing and retail, and continued to substitute and optimize hardware configuration based on pilot feedback. For example, manufacturers such as Tesla Optimus, 1X Neo and Agility Digit all have about 1,000 units in mass production plans. The movement systems, perception systems and infrastructure systems embodied in intelligent ontology are gradually being established, and their coordination and collaboration capabilities are also constantly improving, providing a strong supplement to the future labor market.
IV. Summary and Outlook
The top ten trends in artificial intelligence in 2025 show the transition of AI from the basic model to the rise of intelligent actors, and then to the deep integration of AI into the physical world. Reinforcement learning and multimodal fusion give AI stronger perception, understanding and generation capabilities, transforming it from a "tool" to a "symbiotic partner" of human beings. The concepts of LifeOS and intelligence as a service are reshaping the human-computer interaction model and enterprise operation processes, while the development of embodied intelligence and spatial intelligence has opened up new paths for AI to enter the physical world. With the continuous maturity and application of these technologies, AI will play an important role in more fields and promote the intelligent development of society. In the future, we should not only pay attention to the progress of AI technology, but also think about how to establish deeper trust and cooperation with the next generation of AI, and jointly write a future chapter of harmonious coexistence and sustainable development between human beings and AI.
