Influence Through Interaction

Updated 5 June 2026

Influence through Interaction is the process by which agents’ actions and communications causally shape behaviors and beliefs, modeled via network diffusion, transfer entropy, and latent space methods.
It applies across diverse domains—from social networks and human-robot systems to astrophysical phenomena—using quantitative tools like Laplacian models and second-order expansions.
Emerging research offers design guidelines and algorithmic strategies to sustain dynamic influence, balancing consensus and fragmentation in evolving interaction systems.

Influence through Interaction denotes the process by which the actions, language, or presence of individuals or agents causally shape the states, behaviors, beliefs, or properties of other entities through mutual or sequential engagement. This causal relationship is not restricted to direct persuasion but encompasses both explicit and implicit forms, ranging from the adoption of innovations in social collectives to intention shaping in human–robot systems, and even to physical effects in astrophysical settings. Contemporary research operationalizes and quantifies such influence by leveraging a spectrum of mathematical apparatus, spanning network diffusion, probabilistic reward modulation, variational inference, and group-level attribution methods.

1. Formal Definitions and Modeling Frameworks

Research across domains has proposed precise formalizations for influence through interaction, often tailored to the structure and observability of the underlying processes:

Network Diffusion and Laplacian Models: The diffusion of innovations is modeled via continuous-time differential equations, where the influence from agents at path-length $d$ contributes as a $d$ -path Laplacian $L_d$ , weighted by a decaying coefficient $c_d$ (e.g., $c_2\approx2/3$ , $c_3\approx1/3$ of direct-neighbor strength) (Miranda et al., 2024). The classic DeGroot model and its variants capture opinion updating as linear averaging, while extensions treat temporary/external influencers as scheduled, high-impact participants whose optimal strategies can be analytically characterized (Zhuang et al., 2022).
Influence Matrices in Time Series: The influence model represents the dynamics as a Markov process where each actor $c'$ ’s state distribution at $t$ is a row-stochastic convex mixture of pairwise templates:

$P(h_t^{(c')} \mid h_{t-1}^{(1)},...,h_{t-1}^{(C)}) = \sum_{c=1}^C R_{c',c} M^{c,c'}_{h_{t-1}^{(c)},h_t^{(c')}}$

Here, $R$ encodes the directed, weighted influence network, which can itself be dynamic (mode-dependent) (Pan et al., 2010).

Information-Theoretic Influence: In multi-agent and human–robot interaction, influence is operationalized as transfer entropy (TE), quantifying the reduction in uncertainty in the target agent’s action distribution given the history of another agent’s behavior. In a POMDP context, modulating the agent’s reward by a TE bonus yields emergent, legible and human-perceptible behavioral adaptation (Jiang et al., 2024).
Latent Space Models: Social influence on behavior is disentangled by embedding both network and item response data into a shared latent Euclidean space, where homophily and proximity directly modulate response tendencies. The overall effect is parameterized by a global influence weight $d$ 0, with behavioral heterogeneity visualized in “interaction maps” (Park et al., 2021).
Interaction-Aware Influence Functions (Machine Learning): Standard influence estimation for group attribution commonly sums first-order effects (single-point parameter gradients). Second-order expansions correct this by including pairwise interaction terms, revealing redundancy (positive alignment) or complementarity (negative alignment) among examples, significantly improving faithfulness to leave-group-out retraining (Heo et al., 15 May 2026).

2. Mechanisms and Phenomenology of Influence

Direct and Indirect Influence: Both experimental and theoretical evidence confirm that individuals are responsive not only to direct neighbors but also to second- and third-degree alters, with empirically determined decay weights (e.g., $d$ 1, $d$ 2 relative to direct) (Miranda et al., 2024).
Principles of Social Psychology: The canonical “principles” (liking, social validation, authority, reciprocation, consistency, scarcity) mediate uptake in informal, non-persuasive interactions. Linguistic markers—enthusiasm, qualifiers, modifications—act as unobtrusive but measurable facilitators of peer adoption, consistently raising classification accuracy of “influential” acts (Prabhumoye et al., 2017).
Temporal Adaptation and Feedback: Influence is inherently dynamic: humans and artificial agents anticipate, adapt, and update their strategies over repeated interactions. Models where robots maximize short-term influence (e.g., Stackelberg games) suffer attrition over repeated trials, requiring entropy-maximizing, unpredictability-inducing modifications to sustain long-term influence (Sagheb et al., 2022, Sagheb et al., 18 Mar 2025).
Interaction Streams vs. Static Networks: Modeling influence as emergent from temporally ordered interaction streams—which are processed via decaying kernels, synopses, and recency-weighted aggregators—more faithfully tracks real behavioral updating than traditional static or snapshot network approaches (Serwata et al., 28 Mar 2025).

3. Quantitative and Algorithmic Approaches

A selection of algorithmic instantiations and empirical results:

Modeling Paradigm	Influence Signal	Core Mechanism	Quantitative Effect
Multi-path Laplacian (Miranda et al., 2024)	Adoption propensity	Path-length Laplacian + $d$ 3	$d$ 4 ≈ 0.65, $d$ 5 ≈ 0.37
Transfer entropy reward (Jiang et al., 2024)	Action predictivity	TE bonus in RL/POMDP	+28% collab SRCL, significant
Dynamic influence matrix (Pan et al., 2010)	State sequence	Switching Markov, variational EM	Mode recovery, turn-taking
Linguistic marker classifier (Prabhumoye et al., 2017)	Uptake (percent)	Enthusiasm, qualifier, modif.	+3.15 points accuracy
Interaction-aware IF (Heo et al., 15 May 2026)	Val.loss change	2nd-order Taylor expansion	Up to +0.67 Spearman, LLM success
Latent space map (Park et al., 2021)	Item engagement	LSM+item response, $d$ 6	$d$ 7 median 0.92 if blocks matched

Notably, in human-robot interaction, modulating TE as a reward bonus with φ = +10 substantially increased collaborative success (from ≈63% to ≈92%) in grid navigation tasks, and human participants reported higher mutual legibility (Jiang et al., 2024). In a reinforcement learning context, LILI’s explicit modeling of latent partner strategies enabled a blocking robot to shape a real robot or human adversary into less optimal counter-strategies—raising block rates from 44–50% (baselines) to 91% (Xie et al., 2020).

4. Applications Across Domains

Social Innovation Diffusion: Precise quantification of direct and indirect influence has significant impact for predicting and engineering consensus formation and innovation adoption rates in complex networks (Miranda et al., 2024).
Human–Robot Interaction (HRI): Transfer entropy–driven and latent-strategy-aware policy optimization frameworks enable robots to produce both “legible” and unpredictable behavior, sustaining task-relevant influence over humans in collaborative and competitive settings, while accounting for human adaptation (Jiang et al., 2024, Sagheb et al., 2022, Sagheb et al., 18 Mar 2025).
Group Attribution in ML: Interaction-aware selection of data for LLM training yields improved generalization in resource-constrained regimes, outperforming both additive influence-based and representation-similarity baselines on benchmark instruction-tuning tasks (Heo et al., 15 May 2026).
Astrophysics: In planetary system formation, gravitational back-reaction between dense discs and self-gravitating protoplanets demonstrates that physical “influence through interaction” can stochastically hinder or foster planetary growth, depending on disc eccentricity and feedback dynamics (Gyergyovits et al., 2014).

5. Theoretical Insights and Limitations

Consensus vs. Fragmentation: In both abstract models and empirical data, influence through interaction induces a tension between homogeneous consensus (due to global averaging or expert dominance) and stable fragmentation (due to homophily and bounded-confidence)—with equilibrium characterized as independent sets in a similarity graph (Kempe et al., 2013, Cooper et al., 2023).
Role of Stubborn/Dictatorial Agents: The presence of non-susceptible vertices (stubborn influencers) can sharply alter opinion convergence, allowing persistent minorities up to a threshold of network density or stubborn agent fraction (Cooper et al., 2023).
Confounding and Identifiability: Causal identification of influence from observational data is nontrivial, as apparent information flow may reflect unmeasured homophily or environmental confounders. Dynamic models must be sufficiently expressive and properly regularized to avoid attributing spurious causality (Pan et al., 2010, Serwata et al., 28 Mar 2025).
Long-term Influence Sustainability: Without explicit unpredictability mechanisms or memory/personality-driven retrieval models (in robots), initial influence deteriorates as humans adapt, highlighting the need for entropic or personality-driven variation (Sagheb et al., 2022, Matcovich et al., 2024).

6. Design Guidelines and Emerging Directions

Emergent design principles and open challenges include:

Reward Alignment: In learning agents, calibrate influence-relevant reward bonuses (e.g., TE coefficients) to match task-reward scales; monitor TE during training to avoid overfitting or neglect (Jiang et al., 2024).
History and Window Size: Employ short history windows and recency weighting to capture most relevant influence signals while ensuring computational tractability and relevance to human cognitive constraints (Serwata et al., 28 Mar 2025).
Group-level Attribution: Utilize second-order corrections in group influence estimation to counteract redundancy and harness the complementarity of data points, especially in high-stakes selection or attribution tasks (Heo et al., 15 May 2026).
Memory and Personality Modeling: Leverage personality-conditioned memory retrieval strategies in social robots to render interaction influence more emotionally resonant and adaptive, mapping the OCEAN trait-space to structural memory policies (Matcovich et al., 2024).
Temporal and Contextual Adaptation: Design policies that manage the tradeoff between legibility and unpredictability, inject carefully regulated randomness/entropy, or directly model future strategy uncertainty in order to sustain influence as humans dynamically change their response rules (Sagheb et al., 2022, Sagheb et al., 18 Mar 2025, Xie et al., 2020).
Stream-based Analysis: Prefer streaming models for influence quantification in settings with high-frequency, temporally-resolved data to enhance fidelity and robustness to aggregation artifacts (Serwata et al., 28 Mar 2025).

Open areas involve (i) robust disentanglement of influence from homophily, (ii) adaptive, longitudinal models of memory and adaptation in both humans and artificial agents, and (iii) algorithmic frameworks for influence optimization in multi-agent, multi-modal, or high-dimensional environments.

Influence through interaction is thus a quantitatively tractable, computationally actionable, and theoretically rich construct, with validated impact across domains ranging from social dynamics and linguistics to human-robot interaction and data-driven machine learning attribution. The most effective contemporary models encode not just direct transmission, but also the temporal, contextual, and second-order structure of influence, and enable robust, principled design of interaction systems that shape collective and individual behavior in predictable, sustainable, and adaptive ways.

Key references: (Miranda et al., 2024, Pan et al., 2010, Prabhumoye et al., 2017, Jiang et al., 2024, Sagheb et al., 2022, Sagheb et al., 18 Mar 2025, Heo et al., 15 May 2026, Serwata et al., 28 Mar 2025, Cooper et al., 2023, Park et al., 2021, Matcovich et al., 2024, Xie et al., 2020, Kempe et al., 2013, Gyergyovits et al., 2014, Zhuang et al., 2022, Moussaid et al., 2013).