Same scenario, different trajectories
The policy attempts one task several ways. Tool calls and conversation turns stay inside each trajectory.
3D view unavailable.
The trajectory table and every learning control remain fully usable below.
ART by OpenPipe / simplified trajectory-group model
GRPO Rollout Lab
Predict a multi-turn strategy. Score the group. Watch experience reshape the next policy.
PlanToolVerifyAnswer
Live trajectory evidence
Awaiting rollout
What the group contributes
Reward becomes useful when it is relative.
A rollout is not just an answer. It is the whole sequence of planning, tool calls, observations, and revisions. GRPO compares several trajectories from the same scenario, turning their differences into a learning signal.
Group mean
μ = Σr / G
Relative advantage
Ai = (ri - μ) / σ
Simplified policy step
p'i ∝ pi exp(ηAi)
Above or below this group?
Normalization gives the strong rollouts positive advantage and weak rollouts negative advantage without needing a fixed class label.
Make useful behavior more likely
The learning step increases probability for above-group behavior. The lab shows a simplified strategy-level update, not full token-level GRPO.
Misconception recovery
The optimizer cannot rescue a bad reward.
Choose Speed only in the arena and run a group. The fast guess wins the proxy while skipping verification. Then repair the metric with verified correctness.
Next-policy distribution100%
The trap has not been triggered yet.
Transfer the idea
Should this task use ART?
Assess an expense-approval agent. ART's official guidance favors tasks the base model can sometimes complete, with a consistent quantifiable reward and an environment that can be repeated without real-world side effects.
Can the current model already succeed sometimes?
Can success be scored consistently?
Can attempts run without approving real expenses?
Increase safe repeatability by moving approvals into a simulator.
Keep the evidence
Export the reasoning, not a generic brief.
The trace contains your prediction, trajectory rewards, computed advantages, policy shift, misconception correction, mastery, and transfer guidance.