A robot relies on more than a trained neural network. It also uses a software harness^[1]: code that calls tools, remembers mistakes and decides when to retry. Kehong Gong and colleagues introduced TinyAstral^[2] to improve this surrounding software while leaving the larger model's parameters unchanged. They then trained a smaller model to work inside the improved system.

Coding agents examined records of previous attempts and proposed harness changes. Each change had to pass a paired statistical comparison before it was accepted. The researchers then collected examples of the larger model's behaviour, including recovery from failures, and used them to train a compact student. This transfer of knowledge from a large model to a smaller one is called distillation^[3]. It does not mean the robot acquired new motors or sensors.

On the simulated RoboTwin benchmark, teacher success rose from 33.4% to 63.6% after harness improvements. A student with four billion parameters reached 62.8% using the same software. During measured episodes, the student consumed about one fifth of the teacher's accelerator time. That measures computing effort, not the price of a complete robot or its maintenance.

The striking detail is that better organization around a model can sometimes matter more than making the neural network bigger. Yet success in simulated tasks does not guarantee that automatically edited software will be safe around people. Reproduction requires the attempt logs, comparison conditions and exact harness versions.