Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

An Identity Based Agent Model for Value Alignment

Karthik Sama, Janvi Chhabra, Arpitha Srivatsha Malavalli, Jayati Deshmukh, Srinath Srinivasa

TL;DR

This work tackles AI value alignment in multi-agent settings by extending the Computational Transcendence (CT) framework to model agents that identify with abstract human values. It introduces schemas to bridge identity objects with contextual observables and enables adaptive identity association updates, while incorporating conformity as an external social factor; these extensions are applied to urban transit choices between taxi and bus. The study reveals that values like Frugalism and Individualism can strongly influence decisions, that initial belief distributions significantly shape final outcomes, and that conformity can drive polarization and reduce behavioral diversity, thereby highlighting the social dynamics of value-driven AI behavior. The approach yields an interpretable, domain-adaptable framework for designing value-aligned agents in socio-technical contexts, with potential policy implications for shaping collective outcomes through value encodings and social influence mechanisms.

Abstract

Social identities play an important role in the dynamics of human societies, and it can be argued that some sense of identification with a larger cause or idea plays a critical role in making humans act responsibly. Often social activists strive to get populations to identify with some cause or notion -- like green energy, diversity, etc. in order to bring about desired social changes. We explore the problem of designing computational models for social identities in the context of autonomous AI agents. For this, we propose an agent model that enables agents to identify with certain notions and show how this affects collective outcomes. We also contrast between associations of identity with rational preferences. The proposed model is simulated in an application context of urban mobility, where we show how changes in social identity affect mobility patterns and collective outcomes.

An Identity Based Agent Model for Value Alignment

TL;DR

This work tackles AI value alignment in multi-agent settings by extending the Computational Transcendence (CT) framework to model agents that identify with abstract human values. It introduces schemas to bridge identity objects with contextual observables and enables adaptive identity association updates, while incorporating conformity as an external social factor; these extensions are applied to urban transit choices between taxi and bus. The study reveals that values like Frugalism and Individualism can strongly influence decisions, that initial belief distributions significantly shape final outcomes, and that conformity can drive polarization and reduce behavioral diversity, thereby highlighting the social dynamics of value-driven AI behavior. The approach yields an interpretable, domain-adaptable framework for designing value-aligned agents in socio-technical contexts, with potential policy implications for shaping collective outcomes through value encodings and social influence mechanisms.

Abstract

Social identities play an important role in the dynamics of human societies, and it can be argued that some sense of identification with a larger cause or idea plays a critical role in making humans act responsibly. Often social activists strive to get populations to identify with some cause or notion -- like green energy, diversity, etc. in order to bring about desired social changes. We explore the problem of designing computational models for social identities in the context of autonomous AI agents. For this, we propose an agent model that enables agents to identify with certain notions and show how this affects collective outcomes. We also contrast between associations of identity with rational preferences. The proposed model is simulated in an application context of urban mobility, where we show how changes in social identity affect mobility patterns and collective outcomes.
Paper Structure (18 sections, 4 equations, 1 figure, 3 tables)

This paper contains 18 sections, 4 equations, 1 figure, 3 tables.

Table of Contents

  1. Introduction
  2. AI Value Alignment
  3. Identity based Decision making frameworks
  4. Modelling Agents that Identify with Values
  5. Computational Transcendence
  6. Schemas to Bridge Identity Objects with Observables
  7. Extended CT framework
  8. Adaptive Agents via Identity Association Updates
  9. Modeling Transit Choices Using the Extended CT Framework
  10. Contextual Observables underlying Transit Choices
  11. Perceived Utilities of Observables
  12. Value Considerations in Transit Choices
  13. Incorporating Conformity as an External Social Factor
  14. Experiments and Results
  15. Impact of Values on populations' choices
  16. ...and 3 more sections

Figures (1)

  • Figure 1: Heatmap of the effect of conformity and initialization of semantic distance to values on the proportion of stabilized population choosing public transport