Post-training platforms that serve many business lines from one base model face a resource constraint: they can maintain only a limited number of experts. When lines outnumber experts, some lines must share. The problem is that business lines rarely come with a taxonomy indicating which lines are similar enough to share effectively.
The paper, posted on arXiv as 2610.14437, asks whether transfer probes can fill that gap. Transfer probes are a way to measure how well knowledge transfers from one task or domain to another, potentially revealing which lines would benefit from a shared expert. The authors frame this as an expert-grouping problem rather than assuming a pre-existing grouping.
The released abstract stops before presenting the method's outcome, so the article can only describe the motivation and open question. If transfer probes prove reliable, they could remove the need for manual taxonomies when allocating experts across business lines.