Zehao Wang

Current VLA models fail at zero-shot robotic personalization, highlighting a critical gap in embodied AI's ability to adapt to user preferences.

Sparse human guidance, covering less than 20% of interactions, is sufficient to dramatically improve real-world mobile manipulation performance by reshaping the critic's value landscape.

Quality-aware pretraining strategies, particularly approach-aware episode conditioning, can significantly outperform plain behavioral cloning in VLA tasks, with optimal choices depending on label granularity and model architecture.

A single Transformer can learn both robotic control and geometric state prediction by treating scene dynamics as a language modeling problem.

The hardest AI tasks remain largely unsolved, with current models achieving only a 2.6% success rate on economically valuable workflows.

MLLMs can achieve 10% gains on multimodal reasoning benchmarks by using ground-truth anchored data curation and scaffold-stripping to avoid cognitive drift during self-evolution.

ML evaluation harnesses, the unsung heroes of model development, are plagued by surprisingly mundane software engineering issues like missing documentation and unimplemented features, hindering reliable assessment.

Cats are helping AI researchers: a Bayesian-inspired model that treats context as a prior significantly improves intent inference for non-speaking agents and avoids shortcut biases.