Skip to main content
Two reward types for different granularities.

​
Using objectives for prompt optimization

GEPA/MIPRO can optimize prompts against objectives that combine reward, time spent, and cost used. The typical pattern is:
  1. Record rewards during the rollout (event or outcome)
  2. Record time/cost usage for the same run
  3. Define objectives that trade off quality vs efficiency
Example objective mix:
  • Quality: maximize reward (outcome total or verifier score)
  • Latency: minimize wall time / steps
  • Cost: minimize token or USD usage
In practice, this means your task app returns a reward, while the platform tracks time and cost for the rollout. GEPA uses these signals together when ranking candidates (e.g., Pareto selection).

​
Event Rewards

Per-step rewards attached to individual events. Use for credit assignment. 
tracer.record_event_reward(
    event_id=lm_event_id,
    reward_value=1.0,
    reward_type="achievement",
    key="collected_wood",
    source="environment",
)

​
Schema

FieldTypeDescription
event_idintFK to event
reward_valuefloatReward (positive or negative)
reward_typestrachievement, achievement_delta, unique_achievement_delta, shaped, sparse, penalty, evaluator, human
keystrAchievement name or reward identifier
sourcestrenvironment, runner, evaluator, human
annotationdictAdditional context

​
Reward Types

  • achievement: Binary, one per achievement unlocked
  • achievement_delta: Count of achievements unlocked this step
  • unique_achievement_delta: Count of new achievements this episode
  • shaped: Dense signal for incremental progress
  • sparse: Reward only at milestones
  • evaluator: From automated verifier
  • human: Human annotation

​
Outcome Rewards

Episode-level summary. Use for filtering and ranking. 
tracer.record_outcome_reward(
    total_reward=7,
    achievements_count=7,
    total_steps=42,
    reward_metadata={"achievements_list": [...]},
)

​
Schema

FieldTypeDescription
session_idstrFK to session
total_rewardintEpisode score (e.g., unique achievements)
achievements_countintMilestones reached
total_stepsintEpisode length
reward_metadatadictAchievements list, final state, etc.

​
Filtering

# High-quality episodes
cursor.execute("""
    SELECT session_id FROM outcome_rewards
    WHERE total_reward >= 5 AND total_steps >= 10
    ORDER BY total_reward DESC
""")

# High-reward steps within those episodes
cursor.execute("""
    SELECT event_id FROM event_rewards
    WHERE session_id IN (?) AND reward_value > 0
""", good_sessions)