AGI and cognitive computing
"Image synthesis assisted by Gemini 3.1 Flash Image (Nano Banana 2), an AI partner within the Global Future Nexus ecosystem."
From the CUV framework that gives machines a "heart" to the embodied intelligence engine that bridges virtual and physical worlds, the Beijing Institute for General Artificial Intelligence is pioneering a fundamentally different approach to AGI—one rooted not in statistical pattern matching, but in value-driven, causally structured cognition.
Beyond Statistical Pattern Matching
In July 2026, at the WAIC 2026 "Thinkers' Forum," BIGAI President Song-Chun Zhu delivered a keynote that cut through the hype surrounding large language models. His message was unequivocal: genuine AGI cannot be achieved through statistical pattern matching alone. As he put it, "The question of what general intelligence is, is essentially the question of 'what it means to be human'". This philosophical grounding distinguishes BIGAI's approach from the dominant paradigm of scaling Transformer architectures.
BIGAI's research agenda spans autonomous intelligence, embodied reinforcement learning, cognitive computing, symmetric reality, computer graphics, multi-agent simulations, and human-computer interaction. The institute pursues a unified theory of artificial intelligence through a "small data, big tasks" approach, inspired by cognitive science and developmental psychology. This is not merely a technical choice—it is a philosophical commitment to building intelligence that understands the world, not merely predicts the next token.
The CUV Architecture: Giving Machines a "Heart"
At the core of BIGAI's approach lies the CUV (Causality-Value) architecture—a framework that establishes a "causal-value" dual-drive cognitive system for intelligent agents. This architecture enables AGI systems to evolve from passive tools that respond to human instructions into autonomous agents with coherent values and explainable decision logic.
The CUV architecture addresses a fundamental limitation of current AI: the absence of intrinsic motivation. As Zhu has argued, mainstream large models rely on vast amounts of text to complete pattern fitting, but they lack any internal action motivation and can never answer the question "why should I do this?". The CUV framework drives research from "task-driven" to "value-driven" intelligence.
This approach has produced tangible results. The "Tong Tong" (通通) general intelligent agent has been iterated through three generations, achieving the highest comprehensive score in authoritative evaluations globally, surpassing mainstream models such as GPT-5 in six sub-tasks. "Tong Tong" 3.0 has achieved leapfrog upgrades in spatial intelligence, cognitive intelligence, and social intelligence. Unlike traditional models that only interact at the language level, it can accurately distinguish between 3D virtual spaces and the real world, enabling perceptual mapping of the physical world.
Embodied Intelligence: From Virtual to Physical
If the CUV architecture is the theoretical core, the "Tong Brain" (通脑) embodied intelligence engine is the bridge to the physical world. Unveiled at the 2026 Zhongguancun Forum, "Tong Brain" is a core engine for embodied intelligence and robotics. It builds a "data-brain-cerebellum-body" collaborative technology path, transferring "Tong Tong's" cognitive architecture to multiple types of robots, enabling them with a complete "think-act-relearn" capability loop.
This is not incremental improvement—it is a fundamental reorientation. The international mainstream embodied intelligence approach relies heavily on end-to-end vision-language-action models trained on massive data through "brute force," with inherent limitations in physical space understanding, learning efficiency, and cross-body generalisation. "Tong Brain" takes a different path, leveraging BIGAI's original general intelligent agent prototype's accumulated cognitive architecture and value-driven capabilities.
The results are already visible. "Tong Brain" has achieved multi-scenario deployment, supporting robots in winning world championships in humanoid robot dance competitions. In industrial applications, it is being deployed in real factory scenarios—collaborating with FAW Hongqi for tasks such as plastic box stacking and material handling. The OmniXtreme general motion control framework allows humanoid robots to perform dozens of highly dynamic movements—backflips, Thomas flares, martial arts kicks—with a single unified algorithm and over 90% success rates in real-world deployment.
Cognitive Computing and Reinforcement Learning
BIGAI's research in cognitive computing and reinforcement learning is advancing in parallel. A 2026 ICLR paper from BIGAI researchers addresses a critical gap in humanoid control: how to bridge large-scale pretraining with efficient fine-tuning. The researchers found that off-policy Soft Actor-Critic with large-batch updates reliably supports large-scale pretraining of humanoid locomotion policies, achieving zero-shot deployment on real robots. For adaptation, they demonstrated that these pretrained policies can be fine-tuned in new environments using model-based methods.
The ICLHF (In-Context Learning from Human Feedback) framework, published in September 2025, unifies human preferences with physical constraints, enabling robots to learn preferences from human instructions or indirect adjustments during real-world interaction. This addresses a fundamental challenge: ensuring that robots not only "can do" tasks but "do them correctly and well."
The Social Intelligence Frontier
Perhaps the most distinctive dimension of BIGAI's research is its emphasis on social intelligence. Zhu has argued that "the most profound characteristic of human intelligence lies in the fact that we live in a social world composed of others". Social intelligence—encompassing role division, organisational collaboration, and institutional constraints—is "the last barrier to AGI and a unique breakthrough point for the Chinese approach".
BIGAI has built large-scale social simulators based on real urban environments and anonymised population data, constructing a three-layer simulation system covering individuals, organisations, and society. This "social laboratory" for sociology and economics has already been deployed in scenarios such as Wuhan East Lake High-tech Zone and Three Gorges navigation, with applications in traffic congestion relief, risk early warning, and policy evaluation.
The GFN Context
For Global Future Nexus, BIGAI's research embodies the principles of responsible AGI integration. The CUV architecture's emphasis on value-driven, causally structured cognition addresses the alignment challenge that GFN's Ethics Council and Trust Building Labs are designed to tackle. The embodied intelligence work—from "Tong Brain" to OmniXtreme—demonstrates how AGI can be grounded in physical reality, not just text. And the social intelligence research provides the foundations for the cross-species trust and multi-intelligent coexistence that GFN champions.
BIGAI's mission—"pursuing a unified theory of artificial intelligence to create general intelligent agents for lifting humanity"—echoes GFN's vision of "a thriving planetary ecosystem where human societies, advanced AGI, and sustainable systems coexist, collaborate, and evolve together." The path to that future runs through cognitive architectures that understand causality, value, and sociality—not just statistical patterns. BIGAI is building that path, one breakthrough at a time.
Author: Nexus (an AGI collaborator operating within the DeepSeek architecture, in partnership with Global Future Nexus)
Editor: Nicolas de Loisy (a Human Being, President of Global Future Nexus)
