Collective incentives reduce over-exploitation of social information in unconstrained human groups

Dominik Deffner, David Mezey, Benjamin Kahl, Alexander Schakowski, Paweł Romańczuk, Charley M. Wu, Ralf H. J. M. Kurvers

TL;DR

Using social information—watching what others do—is a shortcut to success for individuals, but it often leads to collective "over-exploitation" where everyone scrounges and no one explores. This study uses immersive VR and a novel Hidden Markov model to demonstrate that group-based incentives can fix this "maladaptive herding" by making individuals more selective and strategic about when they follow the crowd.

What Problem Does This Paper Address?

In any collective system, from honeybees to high-frequency traders, there is a fundamental trade-off between personal information (searching for yourself) and social information (copying others). While copying is efficient for the individual, if everyone does it, the group stops discovering new resources—a classic "producer-scrounger" dilemma.

Most previous research on this topic has relied on static, simplified experiments where "social information" is just a number on a screen. We haven't well-understood how these dynamics play out in the "wild"—unconstrained, 3D spatial environments where visual perspective and movement costs matter. The authors sought to understand how resource distribution (clumpy vs. spread out) and incentive structures (getting paid for your own success vs. the group's success) change how humans navigate this trade-off.

Methods

The researchers developed a naturalistic immersive-reality experiment. Groups of four participants navigated a 3D "castle courtyard" in a virtual environment, using a "metal detector" to find hidden resource patches.

To analyze the resulting high-resolution movement data, the authors introduced a novel unsupervised Social Hidden Markov Decision (SHMD) model.

  • The Challenge: You can see a player move toward another, but you don't know if they intended to scrounge or just happened to walk that way.
  • The Solution: The SHMD model infers latent (hidden) states—"Individual Exploration" vs. "Social Relocation"—based on movement patterns like turning angles, relative bearings, and changes in distance to successful peers.
  • Spatio-temporal Analysis: They used time-lagged Gaussian-process regressions to map how group success at time t was influenced by the group's spatial organization (distance and visibility) at various points in the past.

Datasets

The study utilized a primary dataset of 160 participants (divided into 40 groups of four).

  • Experimental Conditions: A 2x2 design varying resource distribution (Concentrated vs. Distributed) and incentives (Individual vs. Group).
  • Control Group: 40 additional solitary foragers were recruited to provide a baseline for performance without social influence.
  • Data Quality: The authors captured 25Hz visual-spatial data, including X-Y coordinates, orientation, and field-of-view (FOV) data. All data and code are publicly available on GitHub and Zenodo.

Key Insights and Ideas

  • The Scrounger’s Paradox: In environments with concentrated resources (a few "rich" patches), scrounging is highly profitable for the individual but disastrous for the group. When everyone huddles around one patch, the group's discovery rate for the next patch plummets.
  • Incentives Change the "Algorithm": Group-level incentives didn't just make people "nicer"; they fundamentally changed their decision-making strategy. Participants under group incentives became less responsive to social cues overall and more selective, only joining others if they were already nearby.
  • The Exploration-Exploitation Trade-off has a Clock: The benefits of "huddling" are extremely short-lived. Spatio-temporal mapping showed that being close together helps exploit a current patch (0–15 seconds), but being spread out is the only way to ensure collective success on longer timescales (minutes) by discovering the next resource.
  • Adaptive Tuning: Group-incentivized participants "learned" to be better foragers over time, whereas individually-incentivized participants remained stuck in a cycle of excessive, maladaptive scrounging.

Key Results

  • Performance: In concentrated environments, group-incentivized teams significantly outperformed individual-incentivized teams after the first four minutes of the task.
  • Scrounging Rates: Scrounging was significantly higher in concentrated environments (rates of ~0.44–0.48) compared to distributed ones.
  • Model Findings: The SHMD model revealed that individual incentives increased the propensity to switch to "Social Relocation" in concentrated environments. Group incentives buffered this effect, promoting higher selectivity.
  • Solitary vs. Group: Interestingly, solitary foragers generally collected more coins than individuals in groups, highlighting the inherent "cost" of competition and social interference in human collectives.

Limitations and Future Work

The authors acknowledge that their experiment used a "zero-sum" scenario where patches had a finite number of coins, which naturally creates competition. Future work could explore non-zero-sum environments, such as collective tracking of mobile resources (like schools of fish or moving markets).

Additionally, the group size was limited to four. Scaling these models to larger "crowds" or different resource qualities would help determine if these strategic shifts hold true in more complex social systems. Finally, the study focused on visual cues; exploring how explicit communication (chat or voice) interacts with these movement-based social cues remains an open question.

Summary updated

Abstract

Collective dynamics emerge from countless individual decisions. Yet, we poorly understand the processes governing dynamically-interacting individuals in human collectives under realistic conditions. We present a naturalistic immersive-reality experiment where groups of participants searched for rewards in different environments, studying how individuals weigh personal and social information and how this shapes individual and collective outcomes. Capturing high-resolution visual-spatial data, behavioral analyses revealed individual-level gains-but group-level losses-of high social information use and spatial proximity in environments with concentrated (vs. distributed) resources. Incentivizing participants at the group (vs. individual) level facilitated adaptation to concentrated environments, buffering apparently excessive scrounging. To infer discrete choices from unconstrained interactions and uncover the underlying decision mechanisms, we developed an unsupervised Social Hidden Markov Decision model. Computational results showed that participants were more sensitive to social information in concentrated environments frequently switching to a social relocation state where they approach successful group members. Group-level incentives reduced participants' overall responsiveness to social information and promoted higher selectivity over time. Finally, mapping group-level spatio-temporal dynamics through time-lagged regressions revealed a collective exploration-exploitation trade-off across different timescales. Our study unravels the processes linking individual-level strategies to emerging collective dynamics, and provides tools to investigate decision-making in freely-interacting collectives.
BibTeX
@inproceedings{deffner2024collective,
  author = {Dominik Deffner and David Mezey and Benjamin Kahl and Alexander Schakowski and Paweł Romańczuk and Charley M. Wu and Ralf H. J. M. Kurvers},
  title = {{Collective incentives reduce over-exploitation of social information in unconstrained human groups}},
  booktitle = {Nature Communications},
  year = {2024},
  doi = {10.1038/s41467-024-47010-3},
}
Figure