Team: Concepts, Formation & Coordination

• Team is a collective unit defined across various domains, including scientific collaboration, multi-agent systems, and game theory, characterized by coordinated decision-making and distributed information.
• It encompasses methodologies for optimal team formation and role allocation using network analysis, optimization techniques, and social dynamics to balance skills and minimize communication costs.
• The abstraction of a team spans practical implementations in robotics and reinforcement learning as well as theoretical frameworks in stochastic control and team semantics, offering insights into performance, robustness, and adversarial interactions.

Team denotes a collective unit of action, inference, or evaluation, but contemporary research uses the term in several distinct technical senses. A team may be the byline of a scientific paper, a temporary group of players in a competitive game, a collection of agents in a multi-agent system or robot-soccer controller, a decentralized set of decision makers in stochastic control and game theory, or a set of valuations in team semantics. Across these literatures, the recurring problems are team formation, role allocation, coordination, equilibrium computation, performance measurement, and robustness under uncertainty or adversarial pressure (Zhao et al., 25 Jun 2026, Elbert et al., 4 Jun 2025, Lau et al., 2023, Saldı, 2017, Yang et al., 2016).

1. Conceptual scope and formal definitions

In bibliometric research, a “scientific team” is identified with the byline of a published paper: any article with two or more coauthors constitutes a team, and the central independent variable may be the ratio of thought leaders, defined as $$\mathrm{RTL}\equiv(\#\text{ of thought leaders})/(\text{total \# of authors})$$ (Zhao et al., 25 Jun 2026). In social-network formulations of team formation, a team is a subset of experts chosen to cover a task’s required skills while minimizing a communication-cost objective such as diameter, leader distance, Steiner cost, or bottleneck cost (Datta et al., 2012, Addanki et al., 2020).

In stochastic team theory, a team consists of $N$ agents with observation kernels $W_i(\cdot|x)$, local policies $\gamma_i$, and a common nonnegative cost

$$J(\gamma)=\int c(x,y,u)\,\Bigl[\prod_i W_i(dy_i|x)\Bigr]\Bigl[\prod_i \gamma_i(du_i|y_i)\Bigr]P(dx),$$

with optimality defined by minimizing $J(\gamma)$ over the product policy space (Saldı, 2017). In two-team zero-sum stochastic games, the object is a team policy or a distribution over joint policies, and robustness is expressed through correlated-team maxmin equilibrium, where one team optimizes against an opponent team allowed arbitrary joint deviations (Liu et al., 2024).

In extensive-form team games, the same issue appears as the need to represent correlation plans for a team whose members have perfect recall individually but not as a single meta-player; this is the setting in which the Team Belief DAG was introduced as a convex representation of team strategies (Zhang et al., 2022). In propositional team logics, a team is not an organization at all but a set of valuations $X\subseteq 2^N$, and a formula is evaluated on $X$ rather than on a single valuation, making dependence, independence, inclusion, and non-emptiness semantically expressible (Yang et al., 2016).

This range of definitions suggests that “team” is not a single mathematical primitive. It is a family of formal abstractions tied together by shared concerns: distributed information, partial observability, correlated choice, and collective outcomes.

2. Team formation, capacity, and composition

Research on team formation in networks emphasizes the joint role of skill coverage and social proximity. In the capacitated team-formation problem, the social network is an undirected weighted graph $G=(V,E,w)$; each user $u$ has a skill set $N$0 and a capacity $N$1; and feasibility requires both covering the task items and respecting the packing constraint $N$2 (Datta et al., 2012). The resulting optimization problems include MINDIAMTEAM, MINAGGRTEAM, and MINMAXTEAM. The first two are NP-hard, whereas MINMAXTEAM is solvable in polynomial time via thresholding. Approximation algorithms are given for minimizing diameter and Steiner cost, and exact optimization is given for the bottleneck objective (Datta et al., 2012).

A distinct line of work exploits community structure rather than global skill search. The community-based strategy TFC chooses a leader and neighbors from within a community, rather than starting from the rarest required skill. TFC-R and TFC-N form teams on desirable communities, where a community $N$3 is desirable if $N$4, and use high-degree experts as candidate leaders. On DBLP, these methods yield much better communication cost than Rarestfirst and are faster than MinLD and MinSD; when run on communities, the runtime is several orders faster than on the larger network without compromising too much on the communication cost (Addanki et al., 2020).

Composition effects are also studied in temporary game teams. In Honor of Kings, team effectiveness was decomposed into performance, tenacity, and rapport. Across 96 million ranked matches, teams with more distinct roles win significantly more often and surrender much less in early games, but diversity is double-edged for abusive language: when a diverse team wins, it abuses less, and when it loses, it abuses more (Cheng et al., 2019). The same study reports that “experienced assassin” players have significantly higher historical abuse rates regardless of which role they play, and that the same individual does not become more abusive when playing assassin versus other roles, supporting a self-selection rather than role-induced account (Cheng et al., 2019).

Team selection by individual scores has also been examined in a probabilistic setting. For contest-style objectives based on the expected top-$N$5 order statistics,

$N$6

the canonical test $N$7 can be a factor of $N$8 worse than the optimal team, whereas a potential-based test yields a constant-factor approximation with a universal constant $N$9 (Kleinberg et al., 2015). This directly contradicts the common assumption that ranking individuals by isolated average performance is sufficient for team assembly under non-modular performance objectives.

3. Coordination architectures and communication protocols

In embodied multi-agent systems, team coordination is often layered. The FC Portugal 3D Simulation Team implements each agent in Java with seven main packages: WorldState, AgentModel, Geometry, Optimization, Skills, Utils, and Strategy. High-level coordination combines formations, tactics, set-plays, Situation Based Strategic Positioning, and Dynamic Positioning and Role Exchange, while low-level control uses model-based methods such as Zero-Moment-Point on an inverted-pendulum model, rigid-body dynamics of the form

$W_i(\cdot|x)$0

and optimizer-tuned cost functions over forward speed, energy consumption, foot-swing clearance, torso stability, and fall penalties (Lau et al., 2023). Agents also broadcast a Communicated World State, and a planner samples possible team formations $W_i(\cdot|x)$1 and evaluates a team utility $W_i(\cdot|x)$2 based on coverage and ball proximity (Lau et al., 2023).

The same system couples model-based control to reinforcement learning. Its generic framework is built atop OpenAI Gym with a SyncMode SimSpark interface, exposes the standard reset() and step(a)\to (s',r,done,info) methods, and hides inverse kinematics, ZMP solvers, optimizers, and referee protocols behind a binary library. The main algorithms are PPO for running and sprinting, A3C³ for mid-level tests, and CREPS-CMA for distance-controlled kicks (Lau et al., 2023).

In open-ended collaborative domains, TeamFusion replaces one-shot answer aggregation with modeled interaction. For each team member $W_i(\cdot|x)$3, it instantiates a proxy agent $W_i(\cdot|x)$4 conditioned in-context on participant-specific preference evidence $W_i(\cdot|x)$5. These agents participate in a round-robin discussion with shared history $W_i(\cdot|x)$6, surface agreements and disagreements explicitly in the transcript, and then a remix agent synthesizes a deliverable

$W_i(\cdot|x)$7

The method can iterate up to $W_i(\cdot|x)$8 rounds, with the paper reporting up to 3 iterations (Liu et al., 21 Apr 2026). A central claim of this design is that direct aggregation baselines commonly used in closed domains are ill-suited to open-ended settings because they tend to suppress minority perspectives rather than resolve disagreements (Liu et al., 21 Apr 2026).

Tempo-relational team modeling addresses a different coordination problem: how to represent interaction dynamics and relations jointly. TRENN constructs a temporal graph from diarized speaker-listener relations, applies a GCN at each slice,

$W_i(\cdot|x)$9

and then applies multi-head self-attention over time to obtain “social embeddings.” MT-TRENN replaces the single-task decoder with a multi-task head for Emergent Leadership, Leadership Style, and Teamwork components, and adds factual saliency and counterfactual CoDy explanations (Luca et al., 17 Jul 2025).

4. Performance, leadership, and automated support

Performance research increasingly separates technical proficiency from coordination-related contributions. In Age of Empires II, the probability that team $\gamma_i$0 wins match $\gamma_i$1 is modeled as a logistic function of differences in task proficiency, team player effect, and team familiarity. The key estimate is $\gamma_i$2 with standard error $\gamma_i$3 and $\gamma_i$4, while the marginal effect at the mean is $\gamma_i$5 points of win probability per one-unit $\gamma_i$6 (Elbert et al., 4 Jun 2025). The positive interaction $\gamma_i$7 and the positive interaction $\gamma_i$8 show that the team player effect is amplified by familiarity and grows with team size (Elbert et al., 4 Jun 2025). The paper’s interpretation is explicit: social skills and prior shared experience multiply each other’s benefits, rather than adding independently (Elbert et al., 4 Jun 2025).

Leadership composition in scientific teams exhibits a different pattern. Using more than 140,000 PLOS papers, thought leaders are defined as authors tagged as having “conceived and designed the experiments,” and performance is measured both by five-year citation impact, $\gamma_i$9, and by percentile-normalized disruptiveness based on the disruption index

$$J(\gamma)=\int c(x,y,u)\,\Bigl[\prod_i W_i(dy_i|x)\Bigr]\Bigl[\prod_i \gamma_i(du_i|y_i)\Bigr]P(dx),$$0

The estimated relationship between the ratio of thought leaders and team impact is inverted-U shaped, with $$J(\gamma)=\int c(x,y,u)\,\Bigl[\prod_i W_i(dy_i|x)\Bigr]\Bigl[\prod_i \gamma_i(du_i|y_i)\Bigr]P(dx),$$1 and $$J(\gamma)=\int c(x,y,u)\,\Bigl[\prod_i W_i(dy_i|x)\Bigr]\Bigl[\prod_i \gamma_i(du_i|y_i)\Bigr]P(dx),$$2, while disruptiveness declines with higher thought-leader ratios, with $$J(\gamma)=\int c(x,y,u)\,\Bigl[\prod_i W_i(dy_i|x)\Bigr]\Bigl[\prod_i \gamma_i(du_i|y_i)\Bigr]P(dx),$$3 and $$J(\gamma)=\int c(x,y,u)\,\Bigl[\prod_i W_i(dy_i|x)\Bigr]\Bigl[\prod_i \gamma_i(du_i|y_i)\Bigr]P(dx),$$4 and the turning point outside the feasible range (Zhao et al., 25 Jun 2026). This suggests that “more heads” are not unambiguously better: teams with more thought leaders tend to be less disruptive even when they may gain citation impact up to an interior optimum.

Automated feedback systems target team process rather than team composition. tAIfa is a Slack-integrated agent that ingests channel messages, computes deterministic metrics such as Language Style Matching, Sentiment, and Team Engagement, asks an LLM to evaluate contextual constructs such as Transactive Memory System, Collective Pronouns, Communication Flow, and Topic Coherence, and then delivers private individual feedback and public team feedback in a four-part format of summary, strengths, areas for improvement, and actionable steps (Almutairi et al., 19 Apr 2025). In a preregistered between-subjects study with 18 three-person teams, repeated-measures ANOVA showed a significant main effect of condition on conversation duration, $$J(\gamma)=\int c(x,y,u)\,\Bigl[\prod_i W_i(dy_i|x)\Bigr]\Bigl[\prod_i \gamma_i(du_i|y_i)\Bigr]P(dx),$$5, and turn frequency, $$J(\gamma)=\int c(x,y,u)\,\Bigl[\prod_i W_i(dy_i|x)\Bigr]\Bigl[\prod_i \gamma_i(du_i|y_i)\Bigr]P(dx),$$6, but not on task score (Almutairi et al., 19 Apr 2025).

TeamFusion also reports direct performance evidence. On civic comment synthesis with 500 random team samplings from DeliberationBank, TeamFusion reached representativeness scores of $$J(\gamma)=\int c(x,y,u)\,\Bigl[\prod_i W_i(dy_i|x)\Bigr]\Bigl[\prod_i \gamma_i(du_i|y_i)\Bigr]P(dx),$$7 on Llama-3.3-70B, $$J(\gamma)=\int c(x,y,u)\,\Bigl[\prod_i W_i(dy_i|x)\Bigr]\Bigl[\prod_i \gamma_i(du_i|y_i)\Bigr]P(dx),$$8 on GPT-4.1-mini, and $$J(\gamma)=\int c(x,y,u)\,\Bigl[\prod_i W_i(dy_i|x)\Bigr]\Bigl[\prod_i \gamma_i(du_i|y_i)\Bigr]P(dx),$$9 on GPT-4.1, outperforming Direct, CoT, Self-Refine, and MAD; across all four metrics it led by $J(\gamma)$0–$J(\gamma)$1 absolute improvements, with neutrality on par with baselines (Liu et al., 21 Apr 2026). In visual design convergence, Kendall’s $J(\gamma)$2 rose from $J(\gamma)$3 to $J(\gamma)$4, and in a pilot live within-team study decision time fell from $J(\gamma)$5 minutes under free discussion to $J(\gamma)$6 minutes under TeamFusion (Liu et al., 21 Apr 2026).

5. Competition, robustness, and adversarial interaction

In competitive settings, teams are often studied through worst-case guarantees. In two-team zero-sum stochastic games, a correlated-team maxmin equilibrium chooses a distribution $J(\gamma)$7 over joint policies to solve

$J(\gamma)$8

Its significance is that the inner minimization ranges over every joint opponent policy, so the equilibrium is unexploitable even when the opponent team can coordinate arbitrarily (Liu et al., 2024). Because evaluating all joint policies is intractable in large games, the paper introduces restricted correlated-team maxmin equilibrium, sampled deviation spaces, a sequential correlation mechanism, and the S-PSRO algorithm. On Google Research Football, S-PSRO obtains approximately $J(\gamma)$9 Elo points over Team-PSRO and $X\subseteq 2^N$0 goals per match over the built-in AI (Liu et al., 2024).

The Team Belief DAG addresses the same coordination problem at the representation level. It replaces the unavailable team-level sequence form with a perfect-information DAG whose active nodes are beliefs and inactive nodes are observations, and whose convex flow constraints characterize the same set of correlation plans as the original team decision problem. The paper states that TB-DAG can be exponentially smaller and can be computed exponentially faster than all other known representations, and that the converse is never true (Zhang et al., 2022). Paired with regret minimization, it yields state-of-the-art results on benchmark team games (Zhang et al., 2022).

Adversarial robustness introduces a different meaning of temporal interaction. The attack model TEAM, “Temporal Adversarial Examples Attack Model,” targets RNN-based network intrusion detection systems by reconstructing only non-functional features via an autoencoder and optimizing a loss that simultaneously induces current adversarial examples and “next-moment” misclassification on subsequent original examples. It augments the attacker’s guide model with a Time Dilation mechanism so that past hidden states are weighted more strongly, aligning temporal retention between the proxy and attacked RNNs (Liu et al., 2024). On CIC-IDS2017, TEAM achieves white-box attack success rates of at least $X\subseteq 2^N$1, with a best result of $X\subseteq 2^N$2, and on CIC-IDS2017 DDoS the misclassification of subsequent original samples rises from $X\subseteq 2^N$3 to $X\subseteq 2^N$4 (Liu et al., 2024). The technical lesson is that temporal dependence can itself become an attack surface.

6. Team as a mathematical and logical object

The most abstract uses of “team” appear in stochastic control and logic. In stochastic team theory, existence results are not automatic because decentralized policies interact with observation channels and topological properties of policy spaces. By introducing an $X\subseteq 2^N$5–$X\subseteq 2^N$6 topology on randomized policy kernels and imposing lower semicontinuity, local compactness, and total-variation continuity assumptions on the observation channels, existence of optimal policies is established for static teams and for a class of sequential dynamic teams via static reduction (Saldı, 2017). The paper further shows that deterministic optimal policies can be chosen and applies the framework to Witsenhausen’s counterexample and the Gaussian relay channel (Saldı, 2017).

In propositional team logics, the semantic shift is more radical: truth is evaluated on a team $X\subseteq 2^N$7 of valuations rather than on a single valuation. This permits atoms for functional dependence,

$X\subseteq 2^N$8

independence, inclusion, and non-emptiness, along with multiple disjunctions adapted to team semantics (Yang et al., 2016). The hierarchy culminates in full propositional team logic, which is expressively complete in the sense that for every finite $X\subseteq 2^N$9,

$X$0

Thus every family of teams on $X$1 is definable (Yang et al., 2016).

These theoretical results show that the study of teams is not confined to empirical collaboration or engineered coordination. It also includes foundational questions about topology, convex representation, correlated choice, and semantics. A plausible implication is that the term “team” has become a unifying bridge between social organization, multi-agent computation, equilibrium analysis, decentralized control, and logic, even though each field fixes a different formal object and a different criterion of success.