TUM Li Ziyue: Models Are So Big and Powerful, Why Are They Still Unable to 'Get on the Road'? | IJCAI 2026
At IJCAI 2026, TUM's Li Ziyue pointed out that large models and agents are difficult to deploy in real-world scenarios such as transportation, and that the differences between academic data and real data, along with insufficient explainability, are key bottlenecks.
Key Takeaways
- Key Highlight：At IJCAI 2026, TUM's Li Ziyue pointed out that large models and agents are difficult to deploy in real-world scenarios such as transportation, and that the differences between academic data and real data, along with insufficient explainability, are key bottlenecks.
- Innovation & Tech：Highlights advancements in TUM, Li, Ziyue, demonstrating rapid progress in model capabilities.
- Industry Impact：Reported via 雷峰网 (CN), offering actionable signals for developers and technology leaders.
At IJCAI 2026, TUM's Li Ziyue discussed the challenges of deploying large models and agents. Her core argument is that there are fundamental differences between academic data and real-world data, causing models to perform excellently in the lab but struggle to gain trust once applied to actual scenarios. This directly addresses the core pain point of AI moving from research to industrial application.
The article cites an earlier McKinsey study, stating that achieving true intelligence in transportation and industry will still take 10 to 20 years. Despite the continuous expansion of large model parameter scales, the surge in the number of agents, and an endless stream of spatiotemporal data mining papers, no city worldwide has yet deployed a reinforcement learning-based traffic signal control system over the long term, indicating a significant gap between technological breakthroughs and engineering deployment.
Explainability is regarded as the 'lifeline' for deployment. When traffic management departments encounter prediction models that 'surpass SOTA,' they are reluctant to adopt them because they cannot understand the decision-making logic. This reflects that AI systems in critical infrastructure not only require performance, but also transparency and verifiability to build human trust.
This phenomenon serves as a warning to the AI industry: simply pursuing model scale and leaderboard results cannot automatically translate into real-world productivity. Future research needs to pay more attention to data distribution differences, explainability frameworks, and deep integration with industry scenarios in order to push AI from papers to real deployment.
Li Ziyue's remarks represent frontline researchers' reflection on the bottlenecks of AI deployment, and also provide a reference for future research directions: while model capabilities continue to improve, engineering issues such as reliability, safety, and human-machine collaboration must be addressed in parallel; otherwise, no matter how powerful the model is, it will be difficult to hit the road.
This page provides an editorial summary based on publicly available information. It is not a republished article. Use the source link below for the original report.
Industry Insights & Analysis
As artificial intelligence rapidly evolves, breakthroughs surrounding TUM, Li, Ziyue, Models are shifting toward scalable, robust real-world implementations.
Driven by both open-source ecosystems and proprietary model architectures, the integration between compute optimization, data engineering, and agentic workflows is accelerating. This development provides a strategic benchmark for upcoming AI tooling and developer workflows.