Research

Large language model agents increasingly make decisions collectively, but collective performance depends on more than the capabilities of the individual models. This study examines how two features of the communication system, what agents transmit and how much peer information they receive, affect whether a group preserves or loses an initially correct judgment. The experiment creates a deliberate conflict between numerical majority and evidentiary strength: most agents receive weaker evidence favoring the wrong answer, while a smaller group receives stronger evidence favoring the correct one.

The results show that the form of communication matters substantially. When agents share the evidence underlying their judgments, the collective largely preserves its initial accuracy because weak-majority agents can update toward stronger evidence while better-informed minority agents are less likely to capitulate. When agents exchange conclusions without the evidence behind them, the group can instead move toward the incorrect numerical majority. The effect varies across model families and becomes more severe with greater local exposure to peer conclusions, identifying communication content and local fan-in as system-level design variables rather than treating collective error solely as a property of individual models.

Multi-agent AI systems divide problems across interacting models, but the structure through which those models communicate is often treated as an implementation detail. This study instead makes communication architecture an experimental variable. Networks of stateless language-model agents are placed into six different architectures and asked to solve both convergent constraint-satisfaction problems and more open-ended synthesis tasks. Additional manipulations vary where a higher-capability model is placed and whether other agents are told which model has greater capability.

The results show that there is no generally superior communication network independent of the task. On simple constraint problems, centralized architectures use substantially fewer communication rounds and tokens, and the tiered hierarchy performs especially well. Full connectivity performs poorly under the study's centralized readout procedure, consistent with a possible premature-consensus mechanism that the current data do not directly identify. The architecture differences on complex synthesis tasks are less sharply resolved. Overall, the study shows that collective AI performance depends on the fit between task demands and communication structure, while also distinguishing robust findings from directional results concerning synthesis and capability placement.

Information passed through AI systems is often repeatedly summarized, reformulated, combined, or reconciled before a human sees the final output. This study asks how the architecture of that process affects what survives. Across 420 independent transmission chains and 5,880 node-level observations, identical source information is routed through six multi-agent communication architectures designed to represent sequential, centralized, hierarchical, and synthesis-based forms of information processing.

The architectures produce distinct patterns of semantic drift, divergence among separate transmission lineages, and unsupported fabrication. Hierarchical and centralized structures perform comparatively well on several dimensions, while long sequential chains and bidirectional synthesis structures generate different forms of information degradation. The study also finds that preserving factual markers such as names or numbers does not guarantee preservation of their evidentiary or relational meaning. Because architecture, path depth, node count, and synthesis demands vary together in the implemented systems, the findings concern the architectures as system-level bundles rather than isolating a single graph statistic as the cause of every difference.

Communication systems are often observed through their outputs rather than through the processes that produced them. A report, message, or propagated claim may be visible even when the dependencies behind it are not. This study treats that situation as an inverse problem: given a terminal communication artifact, what can be recovered about the hidden production structure that generated it? The experiment holds the underlying events constant while passing them through eight different production motifs, allowing differences in the resulting messages to be attributed to differences in the production process.

The results show that hidden structure can leave measurable residue, but recoverability depends strongly on both the production motif and the method used to detect it. A supervised classifier restricted to fourteen surface-form features recovers motif class substantially better than a prompted language-model reader given richer semantic and trace information. Some motifs leave highly distinctive formal signatures, while others conceal or combine their origins well enough to remain difficult to distinguish. The result is therefore not a claim that complete hidden networks can be reconstructed from text. It establishes conditional recoverability of production-motif class under known generating conditions and shows that the relevant signal may reside in how a message is formed rather than simply in what it says.

The controlled message-residue experiments establish that different communication processes can leave distinguishable traces. This companion study asks whether that result survives contact with naturally occurring adversarial communication. It applies the same general framework to the Twitter archive of accounts attributed to the Internet Research Agency, using documented account roles and production practices to map portions of the corpus onto several communication regimes. The resulting analysis covers nearly two million English-language tweets from more than one thousand attributed accounts.

The real-world results reproduce the broader asymmetry between form-sensitive measurement and semantic reading, while also showing that the specific recoverability of different production regimes changes outside the controlled experiment. Highly templated production processes leave especially strong formal signatures, while more heterogeneous forms of coordination are harder to recover. Additional analyses examine stability over time, transfer to an Iranian state-backed operation, and the effects of removing explicit coordination cues. The paper therefore functions as an external-validity test of the residue framework while retaining an important limitation: production regimes in observational data must be mapped from documented operational roles rather than experimentally assigned as ground truth.

Diffusion is usually measured by observing when people adopt a behavior, practice, or innovation. This model begins from the possibility that the material producing adoption may already be spreading before adoption itself becomes visible. Components of a larger practice can be acquired, retained, and transmitted independently. Someone may therefore participate in the diffusion process without ever adopting the larger object whose components they are carrying.

The simulations show that component transmission can pass extensively through nonadopters and that a substantial latent carrier population can emerge before visible adoption expands. Adoption depends not simply on the amount of information received, but on whether the particular components required for the larger object coexist long enough to be assembled. Controlled comparisons distinguish component identity from the number of inputs and the number of distinct senders. The model is intended as a mechanism-isolation device rather than an estimate of how frequently latent diffusion occurs empirically. Its central implication is that adoption records can omit a consequential diffusion process already underway beneath the observed adoption curve.

Internet memes vary substantially in their surface form while remaining recognizable as members of recurring families. This creates a measurement problem: image similarity, repeated text, and near-duplicate detection can identify related artifacts without necessarily capturing the structure that distinguishes one meme family from another. This study develops a contrastive sparse-feature signature that represents the features recurring within a documented family while emphasizing those that distinguish it from other families.

The measure is evaluated on 680 variants from 31 documented meme families and then applied to 4,263 Reddit image posts. It outperforms several simpler similarity and prototype measures in recovering documented family membership, although a supervised classifier using the same sparse representation performs better on the classification task. In the Reddit application, raw resemblance to canonical meme structure contributes a modest amount of additional information about engagement. Follow-up analyses show that this relationship is driven more by general resemblance to canonical meme structure than by calibrated fit to the correct family. The study therefore separates successful measurement from broader substantive interpretation rather than treating predictive association as automatic construct validation.

Schelling's classic segregation model is usually interpreted as showing how relatively modest preferences can generate substantial residential sorting. This paper examines a less visible assumption built into that mechanism: agents directly observe the demographic composition of their neighborhoods, classify what they observe in the same way, and base movement decisions on that common signal. The model removes those assumptions. Agents instead infer latent neighborhood types from noisy, non-demographic cues interpreted through schema-specific mappings.

Across a large simulation sweep, direct Schelling-style sorting is not the dominant result. The most common regime is continued mobility without stable residential separation. Stable sorting becomes concentrated where cues are highly informative and agents share similar mappings from cues to latent categories. The model also shows that disagreement about neighborhood type and residential separation are analytically separable outcomes rather than two measures of the same process. The contribution is not a calibrated model of any particular city. It identifies the informational conditions required for the classic preference-to-segregation mechanism to survive once perception itself becomes part of the model.

Error correction is usually treated as a mechanism for preserving information by removing mistakes. This study instead asks how correction changes adaptive search. Textual replicators evolve toward several possible target sentences while mutation introduces variation and a dictionary-based correction process constrains malformed variants toward an admissible set of valid forms. A multi-peak version of the model allows several target-specific lineages to remain active simultaneously rather than forcing the population to converge immediately on whichever target is easiest to reach.

Correction substantially increases the number of targets reached within the finite simulation horizon and reduces the frequency with which runs fail to reach any target. Extended runs indicate that much of the advantage reflects faster attainment rather than an established difference in eventual long-run coverage. Critically, the effect depends on the structure imposed by the correcting codebook. Removing target-relevant components from that admissible set removes the advantage. Correction therefore helps not because constraint is intrinsically adaptive, but because compatible constraints preserve useful partial structure while reducing destructive variation.

Human social networks are not external structures that people simply inhabit. They are partly products of the cognitive capacities humans evolved to manage social life. This chapter reviews the relationship between cognition and network structure, beginning with the social-brain argument that human intelligence developed in response to the demands of increasingly complex social environments. It then examines how people mentally represent networks, using memory, schemas, and compression heuristics to reconstruct relational structure that is too complex to retain in complete detail. Those cognitive processes shape which relationships become salient, which contacts people mobilize, and ultimately which social networks are realized.

The chapter then turns from cognition as an evolved human capacity to cognition increasingly shared with computational systems. Digital communication technologies already extend memory, reduce the costs of maintaining distant relationships, and shape which people and information become accessible. Contemporary AI goes further by allowing people to offload parts of judgment, recall, recommendation, and social-network management to algorithmic systems whose operations may be only partially visible to them. The chapter argues that understanding future social networks therefore requires treating human cognition, computational mediation, and network structure as interacting parts of the same system rather than treating online and offline sociality, or human and artificial cognition, as separate domains.

Traditional memetic accounts often borrow the language of biological replication, treating cultural persistence as a problem of copying, mutation, and selection. This paper argues that meaning-bearing cultural forms require a different emphasis. A representation can circulate widely without producing the same interpretation across receivers because its effects depend on the interpretive structures it encounters. The paper therefore reframes cultural persistence around compatibility between features of a sign and the interpretive environment of the receiver.

The model draws on pharmacophore reasoning from computational chemistry and Peircean semiotics. A meme is treated not as an indivisible replicator but as a structured configuration containing features capable of binding with compatible interpretive structures. The larger cultural environment becomes a “memetic reactome” in which representations can activate, inhibit, recombine, or fail to bind depending on context. The framework is primarily theoretical. Its contribution is to make interpretive compatibility, rather than copying fidelity alone, a candidate mechanism for explaining why structurally related cultural forms can produce different effects across populations and contexts.