Dear community,
In the M2 graded lab, we reflected on the generated draft in one step and then applied the feedback from that reflection in a subsequent step.
In the M3 graded lab, the reflection and the application of its feedback were performed within the same step.
I have a few questions about these approaches:
-
Which pattern is generally more advisable when applying reflection:
-
separating the generation of the reflection and the application of that reflection into two steps (possibly using different LLMs), or
-
performing both within a single step using the same model to reduce cost and latency?
-
-
Does it make sense to have the same LLM both perform a task (e.g., generate a report) and review the result (e.g., reflect on the generated report and optionally produce an improved version), assuming different prompts are used? Could the use of the same LLM introduce bias in the reflection process? For example, might a model be more inclined to favor its own outputs and be more critical of outputs produced by other models?
Curious to hear how others approach this.
Thanks!