Definition
Plain language
A reinforcement-learning recipe that lets a model aggressively cut back bad behaviors while limiting how fast it doubles down on good ones.
As stated in the literature
MiniMax's asymmetric-clipping policy gradient objective for agent RL, permitting strong down-weighting of negative-advantage actions while constraining upward updates on positive-advantage actions to stabilize long-horizon training.
Why it matters: This asymmetry helps long-horizon agent training avoid the runaway exploitation of lucky trajectories that destabilizes standard policy gradients.
For example, when an agent takes a clearly bad action, CISPO allows a large negative update, but when it stumbles into a good outcome, the positive update is kept modest.
Heard on the show
“There's one more thing worth flagging in their RL recipe — their policy gradient objective, which they call CISPO.”Episode 090 — How MiniMax-M2 Bets That Sparsity Plus Verifiable Rewards Can Match Frontier Agents