A division-of-labor design assigns different jobs to a language model and a reinforcement-learning agent: the model turns messy inputs into structured representations, and the agent makes decisions from them. A small, validated, versioned contract connects the two. section 12.10
For real-time or large-scale decision loops, a frontier model may be too slow or costly to call at every step. One approach is to use it offline as an oracle to teach a smaller model, then run the student as the encoder. Keep the model’s role descriptive: for clinical support, it can encode patient notes without prescribing treatment. section 12.10
The same pattern can support different operational decisions:
- Turn a camera frame and instruction into objects, positions, and a goal for a control policy. section 12.10
- Extract supply-chain disruptions from supplier messages, weather advisories, and port-closure news. section 12.10
- Encode clinical notes and discharge summaries into a patient state for a treatment policy. section 12.10
