Owen L

What versions are you using?

Follow up, I actually just tried with a trivial implementation of Heun and it's also not working. Also, the mean/square stuff has no impact as well, tested without it. Follow...

Looking deeper, I thought it might be a situation like 4.1.2.4 of https://arxiv.org/abs/2406.09699 where AD is numerically wrong (see also https://github.com/ODINN-SciML/DiffEqSensitivity-Review/blob/main/code/SensitivityForwardAD/testgradient_python.py), since jacrev and jacfwd work. However, they argue this...

There was some good discussion in https://github.com/SciML/SciMLSensitivity.jl/issues/1094. Given that clearly isn't a fault of diffrax (or the Julia sciml ecosystem), the original points in my issue aren't as relevant. But...

I guess it seems to me that these are solvable problems in engineering (I hope), but there isn't any theoretical limitations (like if we want to just store every sample...

I think in general, having some concept of state makes sense for general controls. Even if all existing controls can be done without it, we are working on more flexible...

As we've explored more, I think stateful controls can make a lot of sense and would be useful. For the limited usage we currently have, just leeching off the solver...

Yea, that's basically the implementation we have currently. > Which is arguably the appropriate amount of fiddly for doing something questionable like this :))) I do agree, which is why...

Ahh the classic ``` ValueError: `terms` must be a PyTree of `AbstractTerms` (such as `ODETerm`), with structure ``` I do think there could be a more informative error message here...

> Thank you @lockwo for the help! > > By the way, I'd be very happy to take a pull request improving this error message, describing whatever you think most...