Position: General Alignment Has Hit a Ceiling; Edge Alignment Must Be Taken Seriously

TL;DR

The paper argues that General Alignment, which compresses diverse human values into a single scalar objective, hits a theoretical and practical ceiling at decision boundaries where values conflict. It proposes Edge Alignment, a three-phase framework with seven pillars, to preserve multi-dimensional value structure, enable plural normative governance, and empower dynamic cognitive arbitration through vector-valued optimization, governance-influenced data design, and interactive clarification. Key contributions include formalizing Multi-Objective Alignment and lexicographic constraints, introducing Pluralistic, Contextual, and Collective Alignment, and detailing uncertainty-aware and interactive mechanisms for resolving edge cases. The work reframes alignment as a lifecycle problem requiring governance, domain conditioning, and ongoing negotiation rather than a one-shot optimization, with practical guidance on data, objectives, evaluation, and community participation. This approach aims to produce AI systems that remain aligned amid value conflicts, uncertainty, and diverse stakeholder inputs, improving safety, legitimacy, and adaptability in real-world deployments.

Abstract

Large language models are being deployed in complex socio-technical systems, which exposes limits in current alignment practice. We take the position that the dominant paradigm of General Alignment, which compresses diverse human values into a single scalar reward, reaches a structural ceiling in settings with conflicting values, plural stakeholders, and irreducible uncertainty. These failures follow from the mathematics and incentives of scalarization and lead to \textbf{structural} value flattening, \textbf{normative} representation loss, and \textbf{cognitive} uncertainty blindness. We introduce Edge Alignment as a distinct approach in which systems preserve multi dimensional value structure, support plural and democratic representation, and incorporate epistemic mechanisms for interaction and clarification. To make this approach practical, we propose seven interdependent pillars organized into three phases. We identify key challenges in data collection, training objectives, and evaluation, outlining complementary technical and governance directions. Taken together, these measures reframe alignment as a lifecycle problem of dynamic normative governance rather than as a single instance optimization task.

Figures (5)

• Figure 1: A geometric view of alignment under the HHH criteria. Thresholds $\tau_H$, $\tau_O$, and $\tau_S$ define the feasible region. General Alignment applies in the interior, whereas Edge Alignment concerns behavior near edges and vertices where HHH constraints conflict.
• Figure 2: Three failure modes of scalar alignment and corresponding remedies: value flattening (left), representation loss (right top), and uncertainty blindness (right bottom).
• Figure 3: Pillars of Edge Alignment. A three-phase framework spanning objective structure (\ref{['sec:phase_1']}), normative governance (\ref{['sec:phase_2']}), and dynamic cognition (\ref{['sec:phase_3']}).
• Figure 4: Lifecycle of Edge Alignment .
• Figure 5: Macro-level evidence of the paradigm imbalance in alignment research. The literature remains dominated by general alignment, while edge-alignment work has only recently begun to emerge.