Hacker News new | ask | show | jobs
by OsrsNeedsf2P 9 days ago
So it's trained on the SWE Bench Pro evalset
2 comments
topsycatt 9 days ago
That's not accurate. Take a look at the paper to see what it is trained on! And specifically decontamination is called out in A.4

https://microsoft.ai/wp-content/uploads/2026/06/main_2026060...

link
lemonish97 9 days ago
What is your evidence for this claim?
link
fooker 9 days ago
They say hill climbing

https://microsoft.ai/news/building-a-hillclimbing-machine-la...

Unless they specifically clarify that the testing and training benchmarks are completely separate, we have to assume they test on the same 'hill' the model climbs.

link
artemisart 9 days ago
Hill climbing doesn't mean much but absolutely doesn't imply they cheat on benchmarks. They have more details here https://microsoft.ai/news/introducing-mai-thinking-1/ it seems to be "RL on everything".
link