Learning to Collaborate: A Capability Vectors-based Architecture for Adaptive Human-AI Decision Making

TL;DR

This work tackles adaptive human-AI decision making by introducing learnable capability vectors that uniformly encode the decision-making proficiencies of both humans and AI models. A transformer-based weight generator computes instance-specific aggregation weights, enabling a principled, end-to-end fusion of outputs from heterogeneous agents via a final score vector with $s_f = w S$. The authors also present a learning-free global baseline for case studies and demonstrate superior performance across image classification and hate speech detection, including real human labels (CIFAR-10H, GalaxyZoo) and synthetic expert scenarios. The approach demonstrates strong robustness, scalability, and practical potential for crowdsourcing, expert selection, and large-scale multi-task settings. Overall, capability vectors offer a unified, extensible framework for multi-agent collaboration with tangible gains in decision accuracy and applicability to real-world decision-making pipelines.

Abstract

Effective human-AI collaboration hinges on the ability to dynamically integrate the complementary strengths of human experts and AI models across diverse decision contexts. Context-aware weighted combination of human and AI outputs is a promising technique, which involves the optimization of combination weights based on capabilities of decision agents on a given task. However, existing approaches treat humans and AI as isolated entities, lacking a unified representation to model the heterogeneous capabilities of multiple decision agents. To address this gap, we propose a novel capability-aware architecture that models both human and AI decision-makers using learnable capability vectors. These vectors encode task-relevant competencies in a shared latent space and are used by a transformer-based weight generation module to produce instance-specific aggregation weights. Our framework supports flexible integration of confidence scores or one-hot decisions from a variable number of agents. We further introduce a learning-free baseline using optimized global weights for human-AI collaboration. Extensive experiments on image classification and hate speech detection tasks demonstrate that our approach outperforms state-of-the-art methods under various collaboration settings with both simulated and real human labels. The results highlight the robustness, scalability, and superior accuracy of our method, underscoring its potential for real-world applications.

Figures (6)

• Figure 1: The diagram of the proposed architecture. The candidate decision options may be either pre-defined or generated by models or human experts. We employ a weighted aggregation of the scores assigned by each decision agent to each option in order to derive the final score for each choice. The weights assigned to each agent are determined by a weight generation model, which takes into account the capability vectors and the embedding of the decision task.
• Figure 2: The diagram of the dimension alignment and the learning of capability vectors. At each iteration, we sample a subset of decision makers along with their corresponding one-hot vectors, to predict their respective weights. The final decision is derived from the weighted combination from all their scores. Furthermore, the capability matrix is subject to learning through back-propagation.
• Figure 3: Estimated accurate probability under different collaborative settings. In each subplot, x-axis corresponds to different value of $\alpha$, and we utilize different colors for different values of $p$s.
• Figure 4: Diagram of accuracy under different non-expertise capability levels.
• Figure 5: Diagram of accuracy under collaborative decision by AI models and multiple human experts.