[2603.03242] Density-Guided Response Optimization: Community-Grounded Alignment via Implicit Acceptance Signals

Computer Science > Artificial Intelligence arXiv:2603.03242 (cs) [Submitted on 3 Mar 2026] Title:Density-Guided Response Optimization: Community-Grounded Alignment via Implicit Acceptance Signals Authors:Patrick Gerard, Svitlana Volkova View a PDF of the paper titled Density-Guided Response Optimization: Community-Grounded Alignment via Implicit Acceptance Signals, by Patrick Gerard and 1 other authors View PDF HTML (experimental) Abstract:Language models deployed in online communities must adapt to norms that vary across social, cultural, and domain-specific contexts. Prior alignment approaches rely on explicit preference supervision or predefined principles, which are effective for well-resourced settings but exclude most online communities -- particularly those without institutional backing, annotation infrastructure, or organized around sensitive topics -- where preference elicitation is costly, ethically fraught, or culturally misaligned. We observe that communities already express preferences implicitly through what content they accept, engage with, and allow to persist. We show that this acceptance behavior induces measurable geometric structure in representation space: accepted responses occupy coherent, high-density regions that reflect community-specific norms, while rejected content falls in sparser or misaligned areas. We operationalize this structure as an implicit preference signal for alignment and introduce density-guided response optimization (DGRO), a met...