open models took the big hit from the new benchmarks in aa index v5: tb-sci hle-diamond and frontiercode.
DeepSeek 1017 ckpt is now on FrontierCS 2 with a small jump.
we brought PPO back just for one reason: get way more out of each rollout
GRPO is hard to scale, you can't just pay for as many rollouts as you want on every task. for math/coding an extra rollout is a couple more compiles and checks, whatever. for RSI it's a training run that costs thousands of dollars. and in bio we literally don't have multiple identical embryos to run in parallel
we did some early stuff on keeping PPO training stable, hopefully some of it's useful to people
Show more
MiMO just release an agentic critic. My feeling is that it can effectively speedup in early step but struggle with new introduced bias at long-term. It'll very good if we can not afford tons of rollout like research or science.
great one
Show more
