Multi-attribute negotiation is an important mechanism for distributed decision makers to reach agreements in real-world situations. It allows the possibility of reaching “win-win” solutions for both parties, who trade off different attributes in a solution. Existing research on multi-attribute negotiations has mainly focused on the situations when negotiation parties have complete information about each other's preference. This paper presents a model with incomplete information, while considering Pareto-efficiency and computational efficiency. A non-biased mediator, who applies query learning to maintain near Pareto-efficiency without heavy computation, is adopted in the model. In addition, the mediating mechanism proposed in the model overcomes the difficulty of preference elicitation which usually arises in the preliminary step of a multi-attribute negotiation. Our model also reduces the negotiation complexity by decomposing the original n-dimensional negotiation space into a sequence of negotiation base lines. Agents can negotiate upon a base line with rather simple strategies. The experimental results show that near Pareto-efficient agreements can be reached effectively.