Lion-ER Aims to Teach Physical Reasoning Without Big Data
China Merchants’ Lion Hill laboratory released Lion-ER on Wednesday, 4 November, an embodied-reasoning model whose published identifier indicates 35 billion parameters with 3 billion active. The pitch is that robots can understand physical scenes without a very large data set. Low-data claims are cheap to make and expensive to verify; the test that matters is what the model does on hardware it has never seen, and no such evaluation was published.
