Karouzos and Aletras introduce LambdaMix, a parameter-efficient method for learning a preference-conditioned family of policies via group-relative online reinforcement learning. LambdaMix jointly trains one low-rank adapter per objective over a frozen base model; a preference vector selects the adapter mixture for response generation and weights the separately normalized rewards used to update each adapter, enabling a single training run to support arbitrary preference changes at inference without further optimization. Experiments on four models across helpfulness/harmlessness trade-offs and math problem solving with instruction following, plus evaluation of the two largest models on comment rewriting to reduce personal-attribute clues while preserving wording, show LambdaMix achieves higher hypervolume than all conditioned and policy-merging baselines on the first two tasks. LambdaMix uses fewer training completions and stores fewer adapter parameters than a set of GRPO specialist policies. With architecture and initialization fixed, jointly varying adapter mixtures and objective weights yields wider score ranges and higher coverage than varying either factor alone.