Frameworks

  • LLM (In-context learning)
  • AlphaGeometry (hybrid system)
  • ? New Architecture
  • ? (Finally) Neural Engine

2 - Auto-formalization

translate NL into FL less received than AI 4 theorem proving Argument: for semantic alignment, a better way to validate is just to do the proof Suggest: more unit tests(i..e examples) to improve redability of formalizaed artifact

3 - Scientific Verification

The ease of training AI to solve a task is proportional to how verifiable the task is

It might also be tempting to promote LLM autoformalization to other scientific fields, such as physics, chemistry, etc.

Q & A

  • What is the middle language KELPS like
    • it is a general syntax having a triple: “declaration, fact, query” for representing knowowledge
    • with extensible libraries to deal with various theories
  • What about autonomous theory discovery
    • we need to first have a criteria of what is a good extension (of formal definitions …)
    • this echos with the argument that definitions do not come alone, but need to be tested by whether theorems can be proved