Human-computer interaction extends beyond graphical interfaces to everyday negotiations between people and autonomous systems, such as a human and a self-driving taxi agreeing on where to meet. In such settings, each party holds a private preference over candidate meeting points and would rather not reveal it: sharing a preference score up front hands the other side leverage to demand a worse deal. Yet even when parties exchange only proposed points and never their underlying scores, the negotiation can still leak those preferences, through which points are offered, in what order, and the route ultimately taken to reach the agreed location.
We present Spacer, a three-stage human-autonomous-vehicle coordination pipeline (Evaluation, Negotiation, and Navigation) trained with multi-agent reinforcement learning (IQL, QMIX, and MAPPO) on a grid environment with candidate points of interest and private, per-agent utility scores. Negotiating agents exchange only proposed points, never scores, and are evaluated both on outcome quality (agreement rate, solution quality relative to a scripted baseline) and on how much an eavesdropper can infer about each agent's private preferences from the exchange alone.
Early results show learned agents reach agreement and solution quality close to a scripted baseline (91-100% agreement, 0.79-0.89 vs. 0.86-0.95 quality) while leaking substantially less: opening-offer preference leakage drops to 35-62%, against 100% for the scripted baseline (chance is 20%). In navigation, learned policies match or exceed the baseline on success and route straightness, though route-based leakage remains comparable between the two, rising from ~46-52% a quarter route in to ~79-85% by route completion, pointing to route shape itself as an open privacy channel worth addressing next.
We are excited to invite you to the ACM GreekCHI-CYCHI Symposium 2026, held on November 6-7, 2026 jointly in Athens, Thessaloniki, Nicosia and Online.