Apologies if this post is totally off kilter, this is my first time posting here and I'm still working everything out.
Are there any AI research examples where we:
identify some piece of knowledge a model has obtained at pre-training;
seek to remove/suppress that specific piece of knowledge without decreasing reasoning capability;
post-train with the goal of having the model obtain that knowledge with reasoning and tool use instead of relying on that existing knowledge, maintaining benchmark scores?
While this is not very practical with current architectures, it seems to me that in principle it should be possible to separate memorised knowledge from 'thinking' in this way as we as humans make that separation between memorisation and learning when we think about these things conceptually? And therefore it should be possible to have a very small model which is hopeless by itself but very capable when it has access to tools?
I went looking on research in this area and I did find a decent amount on unlearning for the sake of copyright, illegal material and such, but not with the express goal of trying to erase as much 'knowledge' as possible.
I remember reading about the continuum hypothesis at university, and how the proof that it was undecidable involved effectively creating two separate 'universes' where the axioms were different, and demonstrating that both were valid in some sense. You could plausibly test the results of the model above by having benchmarks similar to this, i.e. some number of imagined worlds with nonsense facts, where you can drop this model and see how it does at problem solving within them. I went looking for benchmarks like this and they do exist in sort of small forms with imagined facts but not really to this degree or oriented towards testing this sort of thing. Are there any benchmarks along these lines that I may have missed?
Apologies if this post is totally off kilter, this is my first time posting here and I'm still working everything out.
Are there any AI research examples where we:
While this is not very practical with current architectures, it seems to me that in principle it should be possible to separate memorised knowledge from 'thinking' in this way as we as humans make that separation between memorisation and learning when we think about these things conceptually? And therefore it should be possible to have a very small model which is hopeless by itself but very capable when it has access to tools?
I went looking on research in this area and I did find a decent amount on unlearning for the sake of copyright, illegal material and such, but not with the express goal of trying to erase as much 'knowledge' as possible.
I remember reading about the continuum hypothesis at university, and how the proof that it was undecidable involved effectively creating two separate 'universes' where the axioms were different, and demonstrating that both were valid in some sense. You could plausibly test the results of the model above by having benchmarks similar to this, i.e. some number of imagined worlds with nonsense facts, where you can drop this model and see how it does at problem solving within them. I went looking for benchmarks like this and they do exist in sort of small forms with imagined facts but not really to this degree or oriented towards testing this sort of thing. Are there any benchmarks along these lines that I may have missed?
I see SynthWorlds as the closest https://arxiv.org/abs/2510.24427v1
This research is tangential too https://arxiv.org/html/2506.15732v2
Perhaps a semi-practical way to implement this is to train on data up to date X only and have the different worlds be different points in history?
Grateful for any comments or suggested reading.