Fudan’s Jia Xiaosong argues training, not VLA or WAM architecture, is the bottleneck

What happened

In a Leiphone interview, Fudan University’s Jia Xiaosong argues that training methods may constrain autonomous driving more than model architecture on some academic benchmarks. Drawing on his research, he advocates imitation learning followed by reinforcement learning and evaluates world models by their usefulness for training, evaluation, and decisions. He sees similar challenges emerging in robotics, where complex contact physics makes simulation harder.

The story so far

Fudan University’s Jia Xiaosong argued in a Leiphone interview that training methods may now constrain end-to-end autonomous driving more than model architecture, at least on some academic benchmarks. He advocated imitation learning followed by reinforcement learning, framing how models are trained as a potentially greater bottleneck than their perception or cognition capabilities. Jia said world models should be judged by their usefulness for training, evaluation, and decision-making. He also sees similar challenges emerging in robotics, where complex contact physics makes simulation harder.

Keep reading

Trending now

1. 1DeepSeek Releases V4.1-Flash-1017, Now Available on Ollama Cloud With US Hosting
2. 2Anthropic Pledges 5% of Its Shares to New Public Benefit Fund
3. 3Meta Removes Business Terms Language That Raised Website Scraping and AI Training Concerns
4. 4Anthropic Says Claude Reconstructed Five NIST Elliptic-Curve Seeds Using a Timestamp-Initialized Generator