Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Co-Alignment: Rethinking Alignment as Bidirectional Human-AI Cognitive Adaptation

Yubo Li, Weiyi Song

TL;DR

The paper addresses the limitations of unidirectional AI alignment by introducing Bidirectional Cognitive Alignment (BiCA), a framework in which humans and AI mutually adapt through learnable protocols, representation mapping, and KL-budget constraints. BiCA combines five components—AI Policy Network, Human Surrogate Network, Protocol Generator, Representation Mapper, and Instructor Network—within a partially observable multi-agent setting and optimizes a composite objective that balances task performance with cognitive alignment. Empirical results on MapTalk and Latent Navigator demonstrate substantial gains over single-directional baselines, including improved success rates, protocol convergence, and representation alignment, as well as enhanced safety under distribution shifts. The findings suggest that optimal collaboration emerges at the intersection of human and AI capabilities, with emergent protocols outperforming handcrafted ones and bidirectional adaptation yielding meaningful practical benefits for cooperative AI systems.

Abstract

Current AI alignment through RLHF follows a single directional paradigm that AI conforms to human preferences while treating human cognition as fixed. We propose a shift to co-alignment through Bidirectional Cognitive Alignment (BiCA), where humans and AI mutually adapt. BiCA uses learnable protocols, representation mapping, and KL-budget constraints for controlled co-evolution. In collaborative navigation, BiCA achieved 85.5% success versus 70.3% baseline, with 230% better mutual adaptation and 332% better protocol convergence. Emergent protocols outperformed handcrafted ones by 84%, while bidirectional adaptation unexpectedly improved safety (+23% out-of-distribution robustness). The 46% synergy improvement demonstrates optimal collaboration exists at the intersection, not union, of human and AI capabilities, validating the shift from single-directional to co-alignment paradigms.

Co-Alignment: Rethinking Alignment as Bidirectional Human-AI Cognitive Adaptation

TL;DR

The paper addresses the limitations of unidirectional AI alignment by introducing Bidirectional Cognitive Alignment (BiCA), a framework in which humans and AI mutually adapt through learnable protocols, representation mapping, and KL-budget constraints. BiCA combines five components—AI Policy Network, Human Surrogate Network, Protocol Generator, Representation Mapper, and Instructor Network—within a partially observable multi-agent setting and optimizes a composite objective that balances task performance with cognitive alignment. Empirical results on MapTalk and Latent Navigator demonstrate substantial gains over single-directional baselines, including improved success rates, protocol convergence, and representation alignment, as well as enhanced safety under distribution shifts. The findings suggest that optimal collaboration emerges at the intersection of human and AI capabilities, with emergent protocols outperforming handcrafted ones and bidirectional adaptation yielding meaningful practical benefits for cooperative AI systems.

Abstract

Current AI alignment through RLHF follows a single directional paradigm that AI conforms to human preferences while treating human cognition as fixed. We propose a shift to co-alignment through Bidirectional Cognitive Alignment (BiCA), where humans and AI mutually adapt. BiCA uses learnable protocols, representation mapping, and KL-budget constraints for controlled co-evolution. In collaborative navigation, BiCA achieved 85.5% success versus 70.3% baseline, with 230% better mutual adaptation and 332% better protocol convergence. Emergent protocols outperformed handcrafted ones by 84%, while bidirectional adaptation unexpectedly improved safety (+23% out-of-distribution robustness). The 46% synergy improvement demonstrates optimal collaboration exists at the intersection, not union, of human and AI capabilities, validating the shift from single-directional to co-alignment paradigms.

Paper Structure

This paper contains 52 sections, 21 equations, 2 figures, 6 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. AI Alignment and Human-AI Collaboration
  4. Multi-Agent Learning and Emergent Communication
  5. Cognitive Foundations of Collaboration
  6. Methods
  7. Problem Formulation
  8. BiCA Framework
  9. AI Policy Network
  10. Human Surrogate Network
  11. Protocol Generator
  12. Representation Mapper
  13. Instructor Network
  14. BiCA Objective
  15. Experiments
  16. ...and 37 more sections

Figures (2)

  • Figure 1: Environment screenshots for our two tasks. (a) MapTalk: collaborative navigation with asymmetric observations and discrete protocol. (b) Latent Navigator: human-in-the-loop exploration of latent space with VAE decoding.
  • Figure 2: Ablation study overview: normalized colors (per metric) with raw values annotated. Metrics shown: success rate, BAS score, CCM score, and average steps. Variants are ordered by success rate.