Teaching a language model one skill can quietly make it worse at another, and the damage doesn't run both ways.
Researchers tested instruction-tuning mixtures on Qwen3 and Mistral models ranging from 0.6B to 32B parameters, running hundreds of fine-tuning jobs to measure how each training task affected every other held-out task. They built what they call a transfer map, a signed score estimating whether a given source task helps or hurts a given target task. The results undercut two common shortcuts: that piling on more training tasks only helps, and that if task A helps task B, task B should help task A back. Instead the team found genuinely asymmetric pairs, where A boosts B's accuracy while B drags A's down.
Teams fine-tuning models for a specific domain usually guess at which tasks to mix in, then burn a full training run to find out if the guess worked. The transfer map, estimated once per corpus, predicted accuracy on unseen mixtures with less than half the error of guessing without one, and it held up across model sizes from 0.6B to 32B. Using it to pick helpful tasks and drop harmful ones lifted accuracy on reasoning benchmarks, including causal explanation, multi-hop questions, and methodological critique, by up to 14 percentage points over training on every available task.
It's a corpus-specific lookup table, not a universal law of which skills play well together, so it won't replace fine-tuning judgment so much as make the guessing less expensive.