I understand the idea of a distillation challenge but it really feels like the poor man's solution to a problem that could be better solved by training a LLM and analysing the layers. In the end, as I value a condensed "cheat sheet", if the goal is to improve open-source model. A better approach seems to recreate the AlphaProof system, longer to do, but more efficient. The path taken by mathematicians now is agentic system with general LLM.
- Masto: https://mathstodon.xyz/@tao/116225525978210807
- AlphaProof: https://deepmind.google/blog/ai-solves-imo-problems-at-silve...
- AlphaProof description: https://www.nature.com/articles/s41586-025-09833-y
[1] https://mathoverflow.net/questions/64071/what-does-the-term-...
Presumably if it's written in plain text and useful to the AI, there may be some relevant information in there that will be interesting for humans too.
It's explicitly stated that the goal is to improve performance of cheap models but I assume, like you did, that they are hopping that the plain text may be useful to humans too.
