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

The study investigates how people in freely interacting groups balance personal and social information during collective search, and how this balance shapes individual and collective performance under realistic perceptual and spatial constraints. A key theoretical trade-off is that overreliance on personal information can prevent information diffusion, while overreliance on social information can reduce exploration and induce over-exploitation and herding. Prior laboratory paradigms often use static or simplified social information and discrete choices, limiting ecological validity. The authors introduce a naturalistic 3D immersive-reality foraging task to examine how resource distribution (concentrated vs. distributed) and incentive structures (individual vs. group) influence social information use. They develop a Social Hidden Markov Decision model to infer latent decisions to use social information from movement and visibility data, aiming to link cognitive mechanisms to emergent collective dynamics.

The work builds on producer–scrounger theory and models of social learning that predict increased scrounging when resources are hard to find but rich and potential misalignment between individual incentives and collective welfare. Prior experiments often used static or limited social information sources and constrained choice sets. Research on social learning strategies emphasizes selective, strategic use over indiscriminate copying. Visual and spatial constraints are known to shape information transfer in animal groups, and ecological models predict environment-dependent exploration–exploitation trade-offs. Institutional interventions (e.g., majority-rule or collective incentives) have been shown to reduce information cascades and conformity, potentially improving group outcomes. The present study extends these insights to unconstrained, spatially explicit human interactions using dynamic, first-person visual information and continuous movement.

Participants: 160 participants (63 male, 97 female; mean age 28.5 years, SD 6.4), grouped in fours, plus 40 additional solitary foragers (14 male, 26 female; mean age 29.8, SD 5.7). All were proficient in German; most from WEIRD populations. Ethics approval obtained; informed consent provided. Compensation: €18 base plus €0.01 per coin (individual vs. group payout depending on condition). Design and task: 2 × 2 design with between-subject incentives (Individual vs. Group) and within-subject resource distribution (Concentrated vs. Distributed). Each group completed four 12-minute rounds in a 90 m × 90 m 3D virtual courtyard. Resource distributions: Concentrated (5 patches × 48 coins) vs. Distributed (15 patches × 16 coins), same total coins per round; patches non-overlapping, radius 3 m. Upon depletion, a new patch of the same size and coins spawned at a random, non-overlapping location, keeping task structure constant. Players navigated with forward/turn controls (W/A/D), max speed 2 m/s, first-person view (108° × 76°). When on a patch, a metal detector signaled discovery and players entered a mini-game clicking coins appearing every 2 s; while exploiting, players could not move. Multiple players on a patch extracted at the same per-player rate (one coin every 2 s), creating exploitative competition without interference. Avatars displayed a digging animation when exploiting, providing real-time social cues. Feedback on total coins was given only after each round (individual or group total depending on condition). Solitary foragers completed the same task without other players or competition and were paid by own coins. Data collection: Logged at 25 Hz: positions (X,Y), orientation, velocity, coin count, exploit status; events for patch arrivals, coin extractions, patch depletion, and new patch generation. From trajectories, line-of-sight visibility and field-of-view social information were reconstructed via triangulation. Two of 160 rounds were omitted for fine-grained visibility/movement analyses due to technical errors. Behavioral analyses: Hierarchical Bayesian models assessed performance dynamics (Bernoulli likelihood for exploitation at 1 s, monotonic time effects), scrounging rates (binomial model for conditional probability to join a visible exploited patch), and relationships between social behavior (scrounging, proximity) and individual/group coin outcomes (multilevel Poisson/lognormal regressions). Frequentist mixed-model replications yielded similar conclusions. Computational model: Social Hidden Markov Decision (HMD) model inferred latent states at 1 s resolution: Individual Exploration and Social Relocation (Exploitation was observed). State-dependent variables: turning angles (von Mises), change in distance to visible exploiting players (normal), and relative bearing to others (log-normal). State predictors of switching from exploration to social relocation included: visibility of exploiters, distance to nearest visible exploiter, number of exploiters at nearest visible patch, and time since last personal success, estimated hierarchically by incentive condition and environment; also time-varying (monotonic) effects across round time. The forward algorithm computed marginal likelihood; Viterbi decoded most likely state sequences post hoc. Collective dynamics: Time-lagged Gaussian-process regressions (binomial likelihood) linked group-level distance and visibility to collective success (number/proportion exploiting) across lags from 1 to 180 s, per incentive and environment. Implementation: Unity (2020.3.21) with Netcode, server tick 25 Hz; analyses in R with Stan (HMC), weakly informative priors, LKJ priors for correlations, non-centered parameterization; convergence R ≤ 1.01. Data and code available on GitHub/Zenodo.

• Performance by incentives and environment: Group-incentivized participants showed no difference in coins between concentrated and distributed environments (−0.05 [−8.9, 8.1], ER=1.05). Individually incentivized participants performed worse in concentrated than distributed environments (−8.4 [−16.7, −0.5], ER=20.7). Over time, in concentrated environments, group-incentivized participants outperformed individually incentivized ones after ~4 minutes; in distributed environments, individually incentivized participants started better but converged to similar success.
• Order effects: Group-incentivized participants performed better than individually incentivized ones when foraging in concentrated environments for a second consecutive time (15.0 [4.3, 26.2], ER=76.7), but not in the first round (3.1 [−8.6, 14.3], ER=2.1) or when previously foraging in distributed environments (0.01 [−11.3, 11.4], ER=0.99).
• Behavioral shifts by environment: In concentrated environments, participants discovered fewer new patches (group: −5.2 [−5.7, −4.7], ER>100; individual: −5.6 [−6.1, −5.1], ER>100) but joined more patches discovered by others (group: 1.4 [0.7, 2], ER>100; individual: 1.1 [0.4, 1.8], ER>100). They stayed closer to others (group: −6.4 [−7.6, −5.3], ER>100; individual: −8.8 [−10.1, −7.6], ER>100) and looked more at others (group: 0.08 [0.05, 0.1], ER>100; individual: 0.14 [0.11, 0.17], ER>100).
• Scrounging rates (conditional on visibility): Higher in concentrated vs. distributed environments (group: 0.44 [0.37, 0.50], ER>100; individual: 0.48 [0.42, 0.54], ER>100). Tended to be higher under individual vs. group incentives in concentrated environments (difference 0.06 [−0.03, 0.16], ER=5.7).
• Determinants of success: In concentrated environments, individuals gained from higher scrounging rates and closer proximity, but groups earned more when fewer players exploited a patch (lower forager density) and when members kept greater distances. In distributed environments, both individuals and groups performed better when participants joined fewer others’ patches and maintained larger inter-individual distances. Independent patch discovery increased coins in both environments (concentrated: 0.14 [0.12, 0.15], ER>100; distributed: 0.09 [0.08, 0.09], ER>100). More directed/regular movement led to more discoveries.
• Solitary vs. group foragers: Solitary foragers collected more coins than group participants, performed similarly across environments, and discovered more patches in distributed environments.
• HMD model—baseline switching: When at least one exploiter was visible, switching to social relocation was more likely in concentrated vs. distributed environments under both group (0.03 [0.004, 0.06], ER=26.4) and individual incentives (0.05 [0.02, 0.08], ER>100). Individual incentives increased switching propensity in concentrated (0.04 [0.002, 0.07], ER=21.7) more than in distributed environments (0.01 [−0.01, 0.04], ER=5.8). Individual-level decision weights for switching predicted higher personal success in both environments.
• Selectivity: Participants were more likely to switch when successful others were closer, indicating distance-based selectivity; this selectivity was adaptive in distributed environments. Selectivity w.r.t. the number of exploiters at a patch appeared mainly under individual incentives; its impact on success was neutral, likely due to rarity of multi-exploiter patches (≥2 exploiters observed in 17% of cases). Time since last success did not increase switching nor predict success.
• Collective impact of latent decisions: Higher group-level baseline switching propensity was negatively related to collective success in concentrated environments; other decision weights were not clearly related to collective outcomes.
• Temporal dynamics of decision weights: Under group incentives, baseline social information use increased over time in concentrated (0.20 [−0.01, 0.42], ER=12.8) and decreased in distributed environments (−0.11 [−0.27, 0.05], ER=6.2). Selectivity increased over time in concentrated environments, with stronger negative weights for distance (−0.45 [−0.64, −0.26], ER>100) and number of exploiters (−0.55 [−0.88, −0.24], ER>100), suggesting more selective social use. Under individual incentives, baseline switching started higher and increased further in concentrated environments (0.18 [−0.03, 0.40], ER=10.4) without clear increases in selectivity.
• Time-lagged collective dynamics: Distributed environments: greater inter-individual distance consistently predicted higher collective success across lags and incentives. Concentrated environments: at short lags (≤15 s group incentives; ≤8 s individual incentives), smaller distances predicted higher success (facilitating rapid joint exploitation); at longer lags, larger distances predicted higher success (facilitating exploration and future discoveries). Visibility: negative association with success at short lags (≤20 s) across conditions (many players currently exploiting), but positive at longer lags, indicating that paying attention to others improved future collective success.

The findings show that humans strategically integrate personal and social information in naturalistic, spatially constrained environments. As predicted by producer–scrounger theory, participants relied more on social information when resources were rare but rich (concentrated), increasing individual success via scrounging and proximity. However, heavy social reliance reduced collective performance by concentrating foragers on the same patches and limiting exploration. Group-level incentives mitigated these collective costs by lowering responsiveness to social cues and promoting time-dependent selectivity, particularly in concentrated environments, aligning individual behavior with group welfare. The computational Social HMD model revealed latent decisions to switch toward social relocation as the mechanism behind observed scrounging patterns and clarified why individuals can benefit from frequent responsiveness even when overt scrounging is sometimes unproductive (e.g., discovering new patches en route). Time-lagged analyses uncovered a dynamic exploration–exploitation trade-off: in concentrated environments, transient grouping benefits joint exploitation in the short term, whereas dispersal enhances future discoveries; in distributed environments, dispersion consistently benefits groups. These insights bridge idealized theory and real-world behavior by highlighting the roles of space, perception, incentives, and temporal dynamics in collective adaptation.

This study links latent individual decision strategies to emergent collective outcomes in an immersive, unconstrained group foraging task. It shows that collective incentives reduce individually beneficial but collectively costly exploitation of social information, improving group performance when resources are concentrated. The Social Hidden Markov Decision model provides a principled way to infer latent social decisions from movement and visibility data, advancing methods for studying dynamic, interactive behavior. Future research can explore non–zero-sum social benefits (e.g., tracking mobile resources), vary group size, introduce heterogeneity in patch quality and leaving decisions, and probe how individuals and groups update strategies over repeated interactions. The modeling framework can be adapted to other domains (e.g., crowd dynamics, sports analytics, animal leadership, GPS-based movement) to understand socio-ecological drivers of decision-making across scales.

• Participant sample: Predominantly WEIRD, German-speaking adults; generalizability to other populations is uncertain.
• Experimental ecology: Virtual 3D environment with fixed first-person field of view, specific movement constraints, and patch mechanics; while naturalistic, it remains a laboratory simulation.
• Group size and structure: Fixed groups of four; effects may differ with other group sizes or network structures.
• Resource structure: Zero-sum, stationary patches with immediate replenishment elsewhere; results may differ for mobile or heterogeneous-quality resources, or when strategic patch-leaving is required.
• Incentive feedback: Participants only received cumulative feedback after rounds; different feedback regimes might alter learning and adaptation.
• Measurement constraints: Two rounds excluded from fine-grained analyses due to technical issues; scrounging rates depend on others’ behavior and visibility, potentially adding noise relative to latent switching estimates.