Training Numerical Intelligence via Auto-Diagnosis and Skill Discovery
ADSD makes solver improvement diagnosis-led, turning execution failures into reusable numerical-method skills.
The paper says the framework first explains why a numerical solver is performing poorly, then uses that diagnosis to find better methods and implement them. It reports gains across power flow equations, AC optimal power flow control, stiff ODEs, and heterogeneous diffusion PDEs. On GOC-500 power flow, the authors say mean solver error fell by nearly 71x, with transfer to unseen grid topologies and operating regimes. Source: HF Daily Papers' note
The paper says the framework first explains why a numerical solver is performing poorly, then uses that diagnosis to find better methods and implement them. It reports gains across power flow equations, AC optimal power flow control, stiff ODEs, and heterogeneous diffusion PDEs. On GOC-500 power flow, the authors say mean solver error fell by nearly 71x, with transfer to unseen grid topologies and operating regimes. Source: HF Daily Papers' note
score 4