Trend 01
Reinforcement Learning: Leading New Breakthroughs in Large Model Inference and Action Capabilities
Reinforcement Learning (RL) is triggering a profound paradigm shift in the field of large language models.Currently, the application of reinforcement learning in large language models is evolving from the initial human feedback reinforcement learning (RLHF)⸺primarily aimed at aligning model outputs with human preferences⸺to a large-scale evolution of reinforcement learning based on verifiable rewards (RLVR).RLVR directly binds reward signals to objective, verifiable outcomes (such as correct answers to programming or mathematical problems), shifting the optimization goal from “sounding correct” to “being correct,” significantly enhancing the core reasoning capabilities of large models. This shift is pushing large models beyond simple content generation towards advanced intelligence that solves real-world problems and achieves complex goals.
Behind this progress lies a profound insight:the “Asymmetry of Verification” in many complex tasks⸺that is, verifying solutions is far easier than finding the solutions themselves.As revealed by the “Verifier’s Law,” the efficiency of AI training is proportional to the verifiability of tasks, with verifiability providing efficient and high-quality feedback signals for reinforcement learning.Based on this, this report proposes four key trends driving the evolution of large models through reinforcement learning:
-
From Alignment to Creation: Reinforcement learning reshapes the reasoning capabilities of large models;
-
Agent Leap: Reinforcement learning builds bridges connecting models with the physical and digital worlds;
-
Deepening Vertical Domains: Reinforcement learning helps overcome data and security bottlenecks, empowering specialized models;
-
Emerging Collective Intelligence: Reinforcement learning drives multi-agent collaboration to solve complex systemic problems
Overall, reinforcement learning is comprehensively driving large models from linguistic intelligence towards action intelligence, embodied intelligence, and collective intelligence by leveraging task verifiability, heralding a significant leap for AI in solving real-world problems.
Trend 02
Natively Multimodal Generation: A New Era of Unified Perception and Generation
The early development of artificial intelligence primarily focused on single modalities, such as computer vision concentrating on image understanding and natural language processing focusing on text analysis.In the face of multimodal tasks, traditional methods often adopt “post-fusion” or “concatenation” strategies, training modality-independent models separately and then simply combining their outputs through additional modules. While this approach has achieved some degree of multimodal information utilization, the interaction between modalities is shallow, leading to information loss during transmission, making it difficult to capture deep semantic associations across modalities and achieve true “joint generation.”
Entering the era of deep learning, especially with the breakthrough success of the Transformer architecture, has paved the way for thorough multimodal integration.This has given rise to “Natively Multimodal Models,” whose core idea is to treat multiple modalities (such as text, images, audio, and video) as a unified input space from the very beginning of architectural design.By sharing or tightly coupling representation layers, models can achieve deep interaction, alignment, and fusion of cross-modal information.
This “native” design allows models to simultaneously perform joint perception of multiple modalities within a single framework and conduct multimodal generation based on a deep understanding of inter-modal associations.From OpenAI’s GPT-4o achieving seamless interaction among text, images, and audio, to breakthroughs in video generation by models like Sora and Veo3, these landmark achievements herald the comprehensive arrival of a new era of unified perception and generation, set to change paradigms across multiple industries.
Trend 03
Voice Models Advancing: Moving Towards Emotionally Intelligent AI
Voice models are rapidly advancing towards a new stage of emotional intelligence, becoming a core technological force driving the naturalization and personalization of human-computer interaction.From early mechanical reading to speech synthesis systems with contextual understanding and emotional expression capabilities, and now to multimodal voice intelligence capable of creating complete musical works and driving visual content generation, AI voice technology is making a leap from “tool” to “partner.” The real-time and emotional expression advantages of voice make it highly applicable in future Voice Agents, immersive content creation, education, and medical assistance.
With advancements in model personalization, low latency, and edge deployment, voice intelligence will evolve towards more user-friendly and inclusive interaction forms, ushering in an intelligent new era where “everyone can create, and everywhere can interact.”
The Rise of Intelligent Agents
Trend 04
Dual-Track Evolution of Agents: Orchestration-Based and End-to-End Approaches Progressing in Parallel
AI Agents, as a significant development direction in the field of artificial intelligence, are undergoing a critical transition from proof of concept to production application.Since exploratory development in 2023, AI Agents have gradually differentiated into two main technical routes:Orchestration-based Agents and End-to-End Agent Models.
-
Orchestration-based Agents adopt an “external” architecture, using large language models as central decision-makers, orchestrating interactions between LLMs and external tools or APIs through predefined code paths to achieve complex task decomposition and execution.
-
End-to-End Agent Models, on the other hand, adopt an “internal” architecture, directly training reasoning, planning, and tool usage capabilities into the model through techniques like reinforcement learning, allowing the model to dynamically guide its processes and tool usage. Represented by OpenAI’s o3 and DeepResearch, this route is still in its early stages but has shown breakthrough effects in specific professional fields.
Both routes have their advantages and applicable scenarios and will develop in parallel in the long term, jointly promoting the evolution of AI Agent technology towards more practical and powerful directions.
Trend 05
LifeOS: AI as the Operating System of Personalized Living
With the rapid maturation of generative AI technology, artificial intelligence is gradually evolving from an auxiliary tool to a “symbiotic partner” deeply embedded in human life.
OpenAI founder Sam Altman recently proposed a forward-looking vision: “LifeOS” envisions an AI future that transcends traditional tools. He pointed out that the way people use AI is shifting from sporadic single tasks to continuous intelligent interactions, where AI is no longer just a tool for answering questions but an intelligent companion that actively provides assistance throughout a user’s life.This vision suggests that AI will be more deeply integrated into our daily lives, becoming a “life operating system” with lifelong memory, personalized reasoning, and proactive action capabilities.The technological foundations behind this trend include long-sequence memory models, contextual understanding engines, and proactive decision-making engines, which are continuously breaking through and improving.
Understanding the development direction of LifeOS not only helps us gain insights into the evolutionary trajectory of next-generation AI applications but will also redefine the relationship between humans and machines, profoundly impacting future human life and social operation models.
Trend 06
Intelligence as a Service: Empowering Industrial Upgrades through Intelligent Workflows
As AI capabilities transition from “intelligence-driven” to “intelligence-as-a-service,” enterprises are entering the “intelligence as a service” phase, where Agents gradually become an integral part of enterprise knowledge systems, process structures, and organizational roles.
In terms of knowledge relationships, enterprises are moving from “having knowledge” to “being able to invoke knowledge.” The evolution of RAG, data flywheels, and knowledge structuring mechanisms is transforming enterprise knowledge from “silent assets” into cognitive systems that are dynamically scheduled by intelligent agents.
In terms of human relationships, Agents are evolving from passive tools to digital employees.They are beginning to undertake closed-loop processes, possess authority boundaries, and become native nodes within process systems. Enterprises are no longer deploying AI functions but are deploying “digital positions” with behavioral responsibilities.
In terms of process relationships, intelligence is no longer reliant on a single Agent but on a systemic network of Agents and scheduling platforms.Organizations are beginning to reconfigure process boundaries around task flows, perception flows, and control flows, moving towards a system intelligence era driven by Agent collaboration.
Trend 07
Game Agents: The Immersive Evolution of AI in Virtual Worlds
Game agents are redefining the boundaries of AI and human interaction in virtual worlds.From early simple scripted NPCs to today’s autonomous agents with deep learning capabilities, this evolution not only reflects a technological leap but also indicates that virtual worlds are transforming into vibrant digital ecosystems.
Contemporary game agents, through cutting-edge technologies like reinforcement learning, large language models, and multimodal perception, are now capable of understanding complex game environments, learning player behavior patterns, generating personalized interactive content, and even exhibiting human-like emotional responses and social abilities.This immersive evolution makes AI no longer a supporting role in virtual worlds but a core driving force behind game narratives, emergent gameplay, and dynamic social networks.
With continuous technological breakthroughs, game agents are laying a solid foundation for the arrival of the metaverse era, making virtual worlds a true second space for human life, work, and entertainment.
AI Moving into the Physical World
Trend 08
Embodied Intelligence’s “GPT-2 Moment”: Collaborative Evolution of Foundational Models, Data Engineering, and Software Platforms
From the trajectory exhibited by generative artificial intelligence, particularly GPT, it appears that the field of embodied intelligence is accumulating leap momentum through scale effects: 2025 is likely to become the “GPT-2 moment” for embodied intelligence.
Many advancements, represented by breakthroughs in VLA (Vision-Language-Action) multimodal large models, mark a critical step for embodied intelligence from specialized scenarios and single tasks towards more general, intelligent, and autonomous machine intelligence, further activating the application potential of robots in human living environments. Specifically:
-
First, powerful end-to-end multimodal foundational models are being constructed, which will empower robots’ cognitive and execution capabilities to leap to higher levels and exhibit a degree of generalization;
-
Second, large-scale real and synthetic data are being generated and utilized at an unprecedented scale, providing support for model training;
-
Third, cross-modal and cross-ontology software platforms are unifying development processes; internet companies like Tencent and NVIDIA are accelerating technology implementation by building robot simulation and training platforms, activating a trillion-level upstream and downstream ecosystem.
Trend 09
Spatial Intelligence: From Conversational to Truly Understanding the World
As AI technology continues to break through, the application of intelligence is transitioning from linguistic intelligence to spatial intelligence.
The rise of spatial intelligence signifies that AI is evolving from processing tokens to understanding systems (Voxels), acquiring the core capabilities to understand and process the three-dimensional world, including perception, reasoning, interaction, and generation in 3D environments.This technological breakthrough enables AI to predict three-dimensional space and emerge just as it predicts the next line of text.
Spatial intelligence is reshaping the ways of working in fields such as autonomous driving, robotics manufacturing, XR mixed reality, medical surgery, architectural design, and smart cities, providing AI with critical physical common sense and causal reasoning capabilities to propel it from “being conversational” to truly “understanding the world” in a leap forward.
Trend 10
From Testing to Mass Production: Applications Accelerating the Maturity of Embodied Intelligence
The “2025 State Council Government Work Report,” released in March 2025, first positioned intelligent robots as “next-generation intelligent terminals and intelligent manufacturing equipment,” incorporating them into the “AI+” action plan. At the same time, the report clearly states the need to “cultivate future industries such as embodied intelligence,” marking that embodied intelligence has risen to a national strategic height.
With continuous technological breakthroughs and the expansion of application scenarios, embodied intelligence entities – robots are transitioning from laboratories to industrialization, reaching a critical turning point from testing to mass production, with breakthroughs in hardware configuration, capability enhancement, and industrial impact across three dimensions:
-
In terms of hardware configuration, the configuration of the three systems of “motion, perception, and foundation” is basically finalized, with large-scale mass production and application imminent;
-
In terms of capability enhancement, applications are driving robots to have stronger coordination and collaboration capabilities, with overall quality reaching the level of “elementary school students”;
-
In terms of industrial impact, multi-form and multi-collaborative embodied intelligent robots will become a strong supplement to the future labor market.
Source: Internet
Copyright Statement
Disclaimer: The copyright of the article belongs to the original author. If there are any issues regarding the content, copyright, or other matters, please contact us for deletion! The content of the article represents the author’s personal views and does not reflect the views or support of this public account. This public account reserves the final interpretation rights of this statement.