Modeling human reputation-seeking behavior in a spatio-temporally complex public good provision game

"Groups of agents in the computational model appear to coordinate their efforts with a temporal rotation scheme. After measuring the extent to which group members’ take “turns” entering and cleaning the river, we observe significantly greater turn taking under identifiable conditions than under anonymous conditions, 𝑝 < 0.0001 (repeated-measures ANOVA, Figure 3c). Further, results from the model confirm that this turn-taking strategy is associated with higher group performance."

Results, p. 9

This demonstrates temporal coordination (turn-taking) as a mechanism that segments and binds cooperative behavior over time in AI agents, aligning with temporal coordination markers relevant to consciousness-inspired organization of information flow in complex systems .

Figures

Figure 3 (p. 9) : Figure 3c quantifies increased turn-taking under identifiability, evidencing structured temporal coordination that supports higher collective performance in AI agents .

"Furthermore, the model also captured the coordination mode used by the human participants. Both humans and artificial agents used a turn-taking strategy, not a territorial strategy."

Results, p. 11

This links human temporal coordination to the same structured timing strategy observed in AI agents, supporting cross-system relevance of temporal binding/segmentation to cooperative performance and potential consciousness-related timing mechanisms .

Figures

Figure 4 (p. 12) : Figure 4c shows that identifiability boosts human turn-taking, evidencing temporally coordinated group behavior correlated with better outcomes .

"MARL allows agents to explore their environment and potentially discover strategies the designer did not even know were possible (Hassabis, 2017; Leibo, Hughes, et al., 2019; Baker et al., 2019)."

Constructing MARL models of social behavior, p. 5

This highlights emergent dynamics—novel strategies arising from agent-environment interactions—mirroring higher-order phenomena like self-organization that are important to consciousness research in both AI and neuroscience .

"It’s important to note that the intrinsic reward is always negative (since 𝛼 > 0 and 𝛽 > 0)... at convergence the magnitude of the intrinsic reward is always small compared to the extrinsic reward (|𝑟𝑖| ≪ |𝑟𝑒|), see Fig. S1 and Fig. S2."

Results, p. 8

A distinct negative reward channel implements an aversive-like cost signal, aligning with AI valence markers and informing welfare-relevant considerations about persistent negative states in artificial agents .

Figures

Figure S1 (p. 16) : Shows the form of the intrinsic signal across agents, clarifying its negative structure and variability in the population .

Figure S2 (p. 16) : Quantifies that intrinsic (negative) reward contributes less than extrinsic reward, supporting the claim about relative magnitudes and valence channeling .

"The environment provides contribution information for each group member currently in view as an input to the agent. In the anonymous condition the contribution information is corrupted by substantial noise."

Supplementary Information, p. 16

This specifies the intervention on information access (routing of contribution signals) that changes downstream behavior, aligning with causal-control markers like gating/masking in AI .

"As expected, identifiability produced a significant increase in group contribution levels, 𝑝 < 0.0001 (repeated-measures ANOVA, Figure 3a). This increase in contribution levels led to significantly higher collective returns, 𝑝 < 0.0001 (repeated-measures ANOVA, Figure 3a)."

Results, p. 8

The manipulated access condition (identifiability vs. anonymity) causally increased cooperation and welfare, demonstrating controllable changes in computation and behavior consistent with causal-control criteria .

Figures

Figure 3 (p. 9) : Shows outcome changes due to the identifiability manipulation, evidencing a causal link from information access to cooperative behavior .