DataClaw

LLM-judge semantic match between predicted answer â and gold answer a; multi-subquestion tasks return the normalized mean over L subquestions.

Ratio of gold reference step count N to the agent's actual step count T. EE > 1 means the agent solved it in fewer steps than the gold trajectory; EE < 1 means it took more.

Fraction of M annotated milestones that the agent achieves anywhere in its trajectory. Captures partial process credit when the final answer is wrong.

Averages an exponential temporal-decay factor γ^max(tⱼ−N,0) over the milestones the agent achieved. Milestones reached by step N contribute 1; later ones decay (γ = 0.9). TPE = 1 means all achieved milestones were on time; lower values mean later concentration.