Le chiffrement. A complete scientific fields where llm fails at the most basic tasks.
You want a poorly coded cubic sieve for solving discrete logarithms ? All models useless.
You want to code a 10 line algorithm from it’s mathematical description ?. A resulting 100% human task.
You give a llm this paper then ask why in the case of local p adic lifting moving to a formal group is required ? You can go at the University to find a professor willing to explain the mathematical reasoning about the requirement.
You want an estimation of algrothmic complexity of this paper ? You have to do it yourself as you’ll only get useless garbage as answers.
Even on the basic cryptocurrency side if you use Ethereum’s altbn128 and want to unpack a G2 point to G12. You’d better have to know how to use the Pari/gp syntax to peform the 5 line computation.
Basically, I can add 20 to 50 annoted question to your benchmark where llm would see theit ability to answer crushed with their utility. Though the hardest ones is if you find help.
Why about including such thème ?
Hi! We won't include such a theme in GAIA as it's unrelated to what we are evaluating, but you are welcome to create your own leaderboard on this topic if you feel it could be interesting for the community!