reward = reward_correctness - α * |target_length - actual_length|