MACRS User Feedback-Aware Reflection Mechanism
Learn how MACRS uses user feedback to improve both its recommendations and conversation strategy through two levels of reflection: information-level and strategy-level.
Even with good planning, conversations don’t always go as expected.
Imagine this: a user is chatting with a recommendation assistant. The assistant asks all the right questions, suggests a movie that matches the preferences, and the user skips it without responding. Or worse, closes the chat.
What happened?
The assistant might have missed subtle cues in the user’s behavior. Maybe the tone felt too robotic. Maybe the movie sounded right on paper, but didn’t match the user’s mood. Maybe the user had just seen that title elsewhere and didn’t want it again.
Whatever the reason, a fixed system wouldn’t notice. It would try again with a similar suggestion, or keep asking more questions, unaware that it’s starting to frustrate the user.
This is the real-world challenge of goal-directed dialogue: things change mid-conversation. Preferences evolve. Users respond indirectly. And what worked five seconds ago might not work now.
That’s why a well-designed agentic system can’t just plan its moves. It must also reflect on how those moves are received and adjust its behavior accordingly.
This is exactly what MACRS does through its user feedback-aware reflection mechanism. Instead of treating failed recommendations or skipped responses as dead ends, MACRS treats them as learning opportunities. This reflection functionality is handled by distinct LLM-based components, acting as specialized critics or evaluator agents. Their sole job is to analyze the conversation’s performance and generate structured feedback to improve the primary agents (responder and planner agents). These operate in a feedback loop external to the immediate act planning, but feeding back into it.
As the image shows, MACRS doesn’t just react; it reflects. It observes how users respond to its messages and then adjusts its understanding and behavior accordingly. This ability to learn mid-conversation is what makes MACRS feel more human and responsive. But how exactly does it do that?
MACRS incorporates reflection at two levels, each guided by different types of feedback:
...