Learning Whom to Trust : Decision-Generated Credibility in Social Learning
Introduces a reinforcement learning-based decision confidence model for social credibility, revealing non-monotonic effects on group consensus and error propagation.
Key Findings
Methodology
The study integrates a drift-diffusion model (DDM) with reinforcement learning to simulate binary choices, decision times, and confidence levels. Social credibility is dynamically generated by weighting the sender’s confidence, which influences social influence via a community network represented by a community-coupling matrix. Linearization yields an analytical community-level feedback mechanism, with the Jacobian indicating how confidence modulates information amplification. Monte Carlo simulations explore the effects of transmission strength and community permeability on collective accuracy and polarization, revealing non-monotonic performance patterns. The model incorporates two information channels: anticipatory influence and retrospective learning, both modulated by process-generated confidence, capturing the dual role of confidence in social influence and private learning.
Key Results
- Simulations show moderate transmission speeds (around 0.3) accelerate correction, raising correct consensus to 85%, while strong transmission (>0.6) causes persistent wrong consensus at 30%. High community permeability spreads confident errors across communities, increasing misinformed polarization. Confidence-dependent private learning stabilizes decisions, reducing error spread. Early high-confidence errors tend to persist longer, amplifying initial mistakes. The model predicts that sender confidence correlates with receiver behavior, especially when sender accuracy is high, providing testable behavioral hypotheses.
- Ablation studies confirm that decision-generated confidence influences both social amplification and private learning. High confidence increases the learning rate for negative prediction errors, leading to faster correction of mistakes but also higher risk of error reinforcement. Conversely, confidence-dependent private learning acts as a stabilizer, damping error propagation. The community feedback mechanism’s analytical threshold determines when social influence shifts from correction to amplification, depending on network permeability and confidence levels.
- Results demonstrate that the interplay of transmission strength, community structure, and confidence dynamics critically shapes collective decision outcomes. The model’s non-monotonic performance pattern aligns with empirical observations of social polarization and misinformation spread, offering a nuanced understanding of how confidence modulates social influence, error correction, and polarization in networked groups.
Significance
This work advances the understanding of social influence by embedding decision-generated confidence into a formal reinforcement learning framework, bridging cognitive decision processes with social dynamics. It explains how confidence fluctuations can lead to both efficient consensus and persistent errors, providing insights into phenomena like polarization and misinformation. The model’s analytical tractability and simulation validation offer a robust tool for designing interventions to mitigate social errors. Its implications extend to social media, political discourse, and collective decision-making, addressing long-standing challenges in understanding and controlling social influence effects in complex networks.
Technical Contribution
The paper introduces a novel integration of drift-diffusion decision models with reinforcement learning to generate dynamic social credibility signals. It derives an exact quotient representation of community-level social influence, with a linearized Jacobian capturing how confidence modulates amplification thresholds. The framework separates anticipatory influence from retrospective learning, providing a detailed mathematical analysis of feedback mechanisms. It offers a new theoretical foundation for understanding how confidence-driven social influence can both accelerate correction and cause persistent errors, supported by rigorous simulation experiments. This approach opens pathways for more realistic modeling of social cognition and influence dynamics.
Novelty
This is the first model to embed process-generated confidence from drift-diffusion decision mechanisms into a social learning framework, explicitly linking individual decision uncertainty with social influence. Unlike traditional fixed-weight models, it captures the dynamic, feedback-driven nature of trust formation. The analytical derivation of community feedback thresholds and the demonstration of non-monotonic group performance patterns represent significant innovations, providing a comprehensive, mechanistic understanding of how confidence influences social contagion, polarization, and error correction.
Limitations
- The model assumes static community structures, neglecting dynamic network evolution which could influence information flow and trust dynamics in real-world settings.
- Confidence is modeled based on a simplified drift-diffusion process for binary choices, which may not generalize well to multi-dimensional or multi-stage decision contexts.
- Simulations are conducted under idealized conditions with limited noise sources; real social environments involve complex, non-rational behaviors that could alter the predicted effects.
Future Work
Future research will extend the model to incorporate dynamic network topologies and multi-dimensional decision spaces, enhancing realism. Empirical validation using neuroimaging and behavioral data will test the neural correlates of decision-generated confidence. Additionally, designing intervention strategies to mitigate harmful polarization driven by high-confidence errors will be explored, aiming to inform social media moderation and policy-making. Cross-disciplinary efforts will integrate insights from neuroscience, sociology, and AI to refine the framework and expand its applicability.
AI Executive Summary
In an era dominated by rapid information exchange, understanding how social influence shapes collective decision-making is crucial. Traditional models often assume fixed influence weights, overlooking the dynamic nature of trust and confidence that evolve during decision processes. This study introduces a novel framework combining reinforcement learning with drift-diffusion models to simulate how individuals generate and transmit confidence-based credibility. By embedding decision-generated confidence into social influence, the model captures the complex feedback mechanisms that can either accelerate correction or entrench errors within groups.
The core innovation lies in analytically deriving community-level feedback thresholds, revealing how the interplay of transmission strength and community permeability determines whether social influence leads to consensus or polarization. Simulations demonstrate a non-monotonic pattern: moderate information flow enhances accuracy, but excessive transmission amplifies early mistakes, causing persistent wrong consensus. High confidence errors, especially when propagated across communities, significantly contribute to polarization, aligning with real-world observations in social media and political debates.
This work offers a mechanistic understanding of the dual role of confidence in social learning, bridging cognitive decision processes with network dynamics. Its implications extend to designing interventions that mitigate misinformation and polarization, emphasizing the importance of controlling information flow and trust calibration. While the model provides valuable insights, future extensions will address dynamic networks and more complex decision scenarios, aiming for broader applicability and empirical validation. Overall, this research advances the theoretical foundation for understanding trust formation and influence in complex social systems, with profound implications for societal decision-making.
Deep Dive
Abstract
Social interaction can improve collective learning but also amplify early mistakes. We study this tension when the credibility of social information is generated by the sender's own decision process rather than fixed ex ante. Reinforcement-learning agents make binary choices through a drift--diffusion process that jointly determines choice, decision time, and confidence; decision confidence then becomes social credibility by weighting anticipatory influence and retrospective social learning. Under balanced community exposure, the anticipatory field admits an exact quotient representation. Its local Jacobian is a scalar decision-sensitivity term multiplying the community-coupling matrix, which yields a common-mode amplification threshold and an analytical role for cross-community permeability in damping relative community differences. Monte Carlo experiments show the corresponding non-monotone performance pattern: moderate transmission accelerates correction, whereas strong transmission can lock populations into wrong consensus; low permeability instead sustains disagreement. Ablations reveal a dual role for confidence: credibility-sensitive transmission amplifies social error, while confidence-dependent private learning stabilises it. The model yields testable predictions linking sender confidence to receiver behaviour conditional on accuracy.