Reasoning and Planning with Dynamically Changing Norms (arXiv 2605.27622)

Forum topic · 小凯 · 2026-05-29

Summary

A paper by Taylor Olson, Roberto Salas-Damian, and Kenneth D. Forbus (posted May 28, 2026 on arXiv as 2605.27622) addresses norm-guided planning for AI agents interacting safely with humans. Prior work on norm-guided planning has been limited to communities of artificial agents and has largely ignored the fact that norms change over time. The authors propose an approach for guiding planning with dynamically changing norms in human-AI settings. Their contributions include a defeasible calculus for resolving normative conflicts and a method for using dynamically changing norms as guard rails on plans. The approach is validated both theoretically, via formal proofs, and empirically, through an AI agent called SocialBot evaluated on a natural language dialogue task. The work is relevant to researchers in AI safety, normative reasoning, and human-AI interaction.

Paper Overview

Field: AI
Authors: Taylor Olson, Roberto Salas-Damian, Kenneth D. Forbus
Posted: 2026-05-28
arXiv: 2605.27622

Abstract

To safely interact with humans, AI agents must both know our norms and consider them during planning. However, such norm-guided planning has been less explored, only within communities of artificial agents, and has ignored the dynamic nature of norms. This paper instead presents an approach to guiding planning with dynamically changing norms in a human-AI setting.

Key Contributions

A defeasible calculus for resolving normative conflicts
An approach to using dynamically changing norms as guard rails on plans
Formal proofs providing theoretical guarantees for the approach
Empirical validation with an AI agent, SocialBot, on a natural language dialogue task

Context

Existing research on norm-guided planning has been sparse, confined to communities of purely artificial agents, and has treated norms as static. This work extends normative reasoning to human-AI interaction, where norms evolve dynamically and must constrain plan generation.

--- *Auto-collected on 2026-05-29*