dots3-note preview just dropped and there's a wild capability demo here.
Knight placement game, 64 rounds, identical reward signals on two separate runs. One agent actually learned the correct rule. The other was optimizing for the wrong objective entirely.
The critic model scored them 3.8 vs 2.29 - it could distinguish between "solving the right problem" and "getting lucky on the wrong problem."
This is huge because most LLMs can't tell when they're fundamentally confused. They'll confidently optimize toward the wrong goal and never flag it. This critic can apparently detect misalignment between what the model thinks it's doing vs what it's actually doing.
That's the kind of meta-reasoning we need for agents that don't just hallucinate their way into the wrong solution space.
Knight placement game, 64 rounds, identical reward signals on two separate runs. One agent actually learned the correct rule. The other was optimizing for the wrong objective entirely.
The critic model scored them 3.8 vs 2.29 - it could distinguish between "solving the right problem" and "getting lucky on the wrong problem."
This is huge because most LLMs can't tell when they're fundamentally confused. They'll confidently optimize toward the wrong goal and never flag it. This critic can apparently detect misalignment between what the model thinks it's doing vs what it's actually doing.
That's the kind of meta-reasoning we need for agents that don't just hallucinate their way into the wrong solution space.