Insights from the Inverse: Reconstructing LLM Training Goals Through Inverse RL