Doesn’t seem like a “wall” to me.
This is a dataset basically constructed by taking a bunch of DeepSeek R1 trajectories and filtering out the incredibly long rambling ones: https://huggingface.co/datasets/open-thoughts/OpenThoughts-1...
There's also a bunch of RL research around strategies to coax the student model towards more concise trajectories.
In general, a lot of people in industry are finding cost savings these days by tuning a smaller model over longer agent trajectories (not exactly the same as distilling the reasoning traces, since those aren't exposed by many frontier labs anymore) collected from their production agents. Typically works best when you have a pretty workflow-y agent whose tasks all share a similar shape.
Let say you have a model that score 100 in a benchmark and you distill it into a model half the size. It has very few guarantees that it will score close to 100 again.1
But I assume one of the hard problems is to know that the model might not perform well enough outside of the distillation part.
There is also another caveat that models do display some level of sub language control, by this I mean that innocuous text might contain information relevant for the own agent or for other agent with it shares tasks. (Like when your partner says there is nothing wrong) Because distillation uses very few resources, we don't know exactly what gets send forward or not.2
As Caleb said, many companies found that they can tune small open weight agents over procedural known tasks because that's easy to inspect and check and because there is little space dor ai-loitering, it tends to be less problematic. The key difference, at least in my understanding, is that you are not concerned with the weights or reducing data. You are basically tuning agents existing weights to emulate another agent answer.
1 - The process of distillation is super effective for specific domains though.
2 - I know I will be called out for supposed AI mystery machine. What I mean is that Large models rely on latent reasoning and nuanced context, often lost during condensation. Distilling models to use fewer tokens or parameters makes predicting retained details difficult. Some papers suggest that model preferences introduced during training, whether intentionally or inadvertently, may be inherited by distilled and fine-tuned models.
To do that you need a larger model to have many many branches to try.