[Feature] Adding Test-Time diffusion opt #8670
Description
What feature would you like to see?
I recently read this paper and implemented a simple version of it locally using DSPY. I think it could be worthwhile addition to list of available optimizers. TTD use a draft to generate downstream tasks then each task generate multiple questions, answers them, llm-as-jduge refines it and at last gets aggregated; and this part can be done in a loop as "self-evolving" mechanism. Then all taks feedback goes back to the original draft, edits are made and the same thing happens again for refinement. At the end all the different tasks and outline are gathered into a report. I realize anyone who wants to implement DeepResearch may go about it differently but some of the components here can be used across different implementations to come.
Would you like to contribute?
- Yes, I'd like to help implement this.
- No, I just want to request it.
Additional Context
