Stacking two unrelated LoRA adapters: most likely outcome

A **LoRA** adapter is a small set of extra weights you train on top of a frozen base model. It nudges the model toward one task, like writing code or answering medical questions. It is tempting to think you can train two adapters separately, then add them together to get a model that is good at both. More skills, same base, no extra training. Sadly that is not how it works. Each adapter was trained alone. It tuned the shared base weights as if no other adapter existed. When you merge them, both sets of changes land on the same weights and pull in conflicting directions. The result is often worse on BOTH original tasks than either adapter was by itself. The correct answer is C. The merged adapter usually underperforms each source adapter, because two independent updates collide. There are two safer ways to combine skills. Keep the adapters separate and route each request to the right one. Or train one new adapter on the combined data, so it learns both tasks together from the start.

The answer is C. Two LoRA adapters trained independently and then merged often perform worse on each task than either adapter alone. Why the merge fails. Each adapter learned a weight delta while assuming sole ownership of the base. Their updates were never co-optimised. When you add the deltas, overlapping directions interfere, and you can land on a third direction that serves neither task. This is the same destructive interference seen in task arithmetic and model merging. The distractors encode the classic traps. The capacity claim is wrong: independent fine-tunes never shared gradient information. The low-rank claim is wrong too. Two rank-r updates collide as easily as full-rank ones. The impossibility claim is wrong because the LoRA update is linear in its matrix product. Merging is always mathematically valid. Safer compositions keep the adapters apart. Route each request to the right adapter with multi-LoRA serving like S-LoRA or vLLM. Or train one joint adapter on the union of both datasets so the tasks share gradients. When you must merge, weighted or sign-aligned methods reduce, but do not eliminate, the conflict.