TRIBE: Predicting Team Performance via Communication Behavior Ensembles

# TRIBE: Predicting Team Performance via Communication Behavior Ensembles
Source: [https://arxiv.org/html/2608.06926](https://arxiv.org/html/2608.06926)
###### Abstract

Designing autonomous agents that effectively assist human teams hinges on understanding team dynamics, often without task specific knowledge\. We present TRIBE, a domain independent approach that reveals team behavioral dynamics invisible to traditional performance metrics\. We show that communication patterns can categorize teams into performance predictive behavioral tribes, as early as 10% into the task, enabling timely interventions\. We test TRIBE on four diverse datasets and demonstrate that communication patterns predict team performance while the prediction strength varies by the degree a task structure allows for behavioral freedom\. Our temporal analysis reveals that AI agents significantly alter team behavioral trajectories while human advisors align with natural dynamics, and that teams maintain behavioral flexibility throughout collaboration\. Further, we compare TRIBE to Llama and optimize the pipeline, achieving significant speedup with performance improvement\.

## Introduction

Autonomous agents increasingly support human teams in complex, real time tasks through automated assessment and assistance during task performance\(Sukthankaret al\.[2007](https://arxiv.org/html/2608.06926#bib.bib2); Webberet al\.[2019](https://arxiv.org/html/2608.06926#bib.bib3); Seoet al\.[2021](https://arxiv.org/html/2608.06926#bib.bib1)\)\. However, their ability to understand the people they help often limits the usefulness of agents\. The problem is even harder when interacting with teams, since the agent needs to model the communication and behavioral patterns among them\.

Research in human agent interaction has long recognized the importance and difficulty of maintaining accurate models of individuals in collaboration\(Albrecht and Stone[2018](https://arxiv.org/html/2608.06926#bib.bib4)\)\. When multiple people work together, communication becomes an integral part of their behavior, a reality documented in diverse domains\(Emmitt and Gorse[2006](https://arxiv.org/html/2608.06926#bib.bib5); Marlowet al\.[2017](https://arxiv.org/html/2608.06926#bib.bib6); Stempfle and Badke\-Schaub[2002](https://arxiv.org/html/2608.06926#bib.bib7)\)\. This is advantageous for autonomous agents because communication has proven to be a valuable source of information about the characteristics and processes that underlie team performance\(Tiferes and Bisantz[2018](https://arxiv.org/html/2608.06926#bib.bib8)\)\. For an agent seeking to improve team performance, the question is: can we predict which teams struggle, early enough to intervene effectively?

We present TRIBE \(Team based Raw Interactions to Behavioral Ensembles\), a pipeline that addresses this challenge\. We demonstrate that topic modeling combined with clustering extracts meaningful patterns from team communication that are strongly correlated with performance outcomes, enabling early interventions\. Through temporal analysis, we show how pre\-existing AI agent interventions impact team behavioral dynamics\. Lastly, we validate TRIBE across multiple domains and systematically optimize the pipeline for practical deployment\.

## Related Work

Traditional team assessment relies on outcome metrics such as task completion time and success rates, which provide clear signals but only after events conclude, limiting utility in ad hoc, high stakes contexts requiring real time intervention\. Recent work has shifted toward process oriented approaches using communication patterns\(Marlowet al\.[2017](https://arxiv.org/html/2608.06926#bib.bib6); Tiferes and Bisantz[2018](https://arxiv.org/html/2608.06926#bib.bib8)\)\.\(O’Neillet al\.[2022](https://arxiv.org/html/2608.06926#bib.bib9)\)provide a comprehensive review of the human autonomy teaming literature, highlighting the importance of mutual understanding between humans and AI agents\.\(Zhanget al\.[2021](https://arxiv.org/html/2608.06926#bib.bib10)\)explore expectations of AI teammates in human AI collaboration, emphasizing that effective teaming requires agents to adapt to human behavioral patterns rather than imposing rigid protocols\.\(Zijlstraet al\.[2012](https://arxiv.org/html/2608.06926#bib.bib39)\)identified early interaction patterns in swift starting teams, finding that specific communication patterns emerge very quickly and predict effectiveness in ad hoc teams like aviation crews\. More recently,\(Bisberget al\.[2025](https://arxiv.org/html/2608.06926#bib.bib40)\)demonstrated that communication patterns in multiplayer online games achieve meaningful predictive power for team skill levels\. Entropy based approaches\(Engome Tchupo and Macht[2023](https://arxiv.org/html/2608.06926#bib.bib41)\)use sliding window entropy to analyze team coordination dynamics, showing that communication entropy can differentiate stable from unstable team communications\.

Successful agent intervention hinges on identifying when and how to intervene in human teamwork\. Examples from the literature include:\(Seoet al\.[2021](https://arxiv.org/html/2608.06926#bib.bib1)\)propose AI coaches to infer team mental model alignment in healthcare settings\.\(Myerset al\.[2018](https://arxiv.org/html/2608.06926#bib.bib11)\)develop autonomous intelligent agents for team training that must understand team state to provide appropriate guidance\. Recent work by\(Liet al\.[2023](https://arxiv.org/html/2608.06926#bib.bib12)\)introduces Markovian approaches to modeling human trust and reliance in AI assisted decision making that can capture dynamic human AI interaction patterns\. Further work shows effective human\-AI teaming requires behavior\-aware design\(Mahmoodet al\.[2024](https://arxiv.org/html/2608.06926#bib.bib13)\), learned interaction rules\(Mozannaret al\.[2023](https://arxiv.org/html/2608.06926#bib.bib14)\), and explanation\(Schleibaumet al\.[2024](https://arxiv.org/html/2608.06926#bib.bib15)\)\. These works collectively suggest that agent intervention requires understanding team behavioral states\.

The approaches outlined have four key limitations\. First, they rely on retrospective analysis where the patterns are identified after performance concludes\. This limits their applicability in early real time interventions\. Second, they are often domain specific: metrics calibrated for gaming communities do not generalize to planning or reasoning tasks\. Third, they identify coarse grained patterns without partitioning teams into behavioral categories with clear performance implications\. Fourth, agent monitoring frameworks track agent reliability but cannot show how various agents with different logics affect the team behavioral trajectories\.

TRIBE addresses all four limitations mentioned above: it enables early prediction, operates domain independently across diverse collaborative contexts, partitions teams into performance predictive behavioral ensembles \(tribes\), and directly quantifies intervention impact through temporal analysis, revealing how different agents impact the team dynamics independent of proprietary intervention logic\.

## Data and Task Environments

We analyze team communication from the publicly available \(third party\) datasets of ASIST program \(Artificial Social Intelligence for Successful Teams\)\(Huanget al\.[2021](https://arxiv.org/html/2608.06926#bib.bib16)\), which developed Minecraft task environments to study human teamwork and AI agent intervention\. The data was collected at Arizona State University with full IRB approval\.

### Study 3: Urban Search and Rescue

ASIST Study 3\(Huanget al\.[2021](https://arxiv.org/html/2608.06926#bib.bib16)\)involved teams of three participants performing collaborative urban search and rescue missions in a Minecraft based environment\(Corralet al\.[2021](https://arxiv.org/html/2608.06926#bib.bib17); Freemanet al\.[2022](https://arxiv.org/html/2608.06926#bib.bib18)\)\. Each team member had a unique role: Engineer, Medic, and Transporter\. Teams underwent hands\-on training followed by two consecutive 17 minute trials\. Each trial began with a mandatory 2 minute planning phase where teams remained at the entrance, followed by 15 minutes of gameplay where communication continued\. The score is based on the number of rescues in each trial\.

Study 3 involved multiple research teams developing different AI agent systems with proprietary intervention logic and strategies\. The teams were distributed across these AI agents: Univ\. of Southern California: USC agent\(Pynadathet al\.[2023](https://arxiv.org/html/2608.06926#bib.bib21)\), CMU Robotics Institute: CMURI agent\(Sycaraet al\.[2021](https://arxiv.org/html/2608.06926#bib.bib22)\), Charles River Analytics: CRA agent\(Charles River Analytics[2024](https://arxiv.org/html/2608.06926#bib.bib23)\), Smart Information Flow Technologies: SIFT agent\(Kuter and Miller[2022](https://arxiv.org/html/2608.06926#bib.bib24)\), Univ\. of Arizona: UAZ agent\(Pyarelalet al\.[2023](https://arxiv.org/html/2608.06926#bib.bib26)\), MIT: DOLL agent\(Simon\-Dempseyet al\.[2023](https://arxiv.org/html/2608.06926#bib.bib25)\); plus two categories for a Human\-Advisor and No\-Advisor, for a total of 8 intervention categories\. It is worth mentioning that TRIBE has no information about or access to the inner logic of any AI agent, and it only observes the interactions\.

To inform TRIBE, we analyze only the natural language transcripts of team communications, excluding other telemetry data from the testbed logs\.

## The TRIBE Pipeline

### Preprocessing of Transcripts

We implemented a comprehensive preprocessing pipeline to prepare transcripts for analysis\. We removed all annotations, encodings, and administrator messages, retaining only team communication transcripts\. Since each team performed two trials, we split each team file into two separate trials\. After removing duplicates, we obtained 222 trials from 111 teams, with trial transcripts with 80 to 370 lines\.

We created Document\-Term Matrix \(DTM\) representations for each trial using the textmineR package\(Jones[2021](https://arxiv.org/html/2608.06926#bib.bib19)\)\. This process involved converting text to lowercase, removing punctuation and numbers, and filtering standard stopwords\. We used unigram models throughout our analysis\.

### Topic Modeling

From an agent’s perspective, determining which teams may need assistance requires understanding how team communication patterns relate to performance\. TRIBE leverages topic modeling to extract latent themes from team conversations, then uses clustering to identify behavioral ensembles \(tribes\) that correlate with performance outcomes\.

#### Latent Dirichlet Allocation

After preprocessing, we apply topic modeling to discover the content structure of team communications\. Topic modeling is an unsupervised method that extracts latent topics from text, where topics represent recurring themes characterized by word distributions\. Latent Dirichlet Allocation \(LDA\)111Due to space constraints, for all statistical, AI, and machine learning methods, we provide citations for detailed descriptions\.\(Bleiet al\.[2003](https://arxiv.org/html/2608.06926#bib.bib20)\)is the standard method for topic modeling and we used the R package textmineR\(Jones[2021](https://arxiv.org/html/2608.06926#bib.bib19)\)\.

The first step was to determine the optimal number of topics\. To do so, we evaluated average probabilistic coherence\(Mimnoet al\.[2011](https://arxiv.org/html/2608.06926#bib.bib36)\), a metric of topic quality\. To ensure stability, we evaluated coherence per topic count \(2\-20\) by averaging 100 runs, yielding k=12 as optimal\.

### Behavioral Clustering

To identify potential behavioral categories among teams that indicate different performance levels, we perform clustering over the topic probability distributions\. To do so, each trial is represented by its topic distribution vector \(the theta matrix from LDA\), which captures the proportional probability of each topic in each trial’s transcript\.

We used gap statistics\(Tibshiraniet al\.[2001](https://arxiv.org/html/2608.06926#bib.bib32)\)with 500 iterations to determine the optimal cluster count, yielding 8 clusters\. We then applied k\-means clustering to assign each of the 222 trials to one of these 8 clusters, representing distinct behavioral patterns\.

### Performance Prediction

To evaluate whether TRIBE’s behavioral clusters relate to trial scores, linear regression revealed a significant relationship between cluster assignment and trial scores, indicating strong correlation between communication patterns captured by clusters and performance outcomes\. Table[1](https://arxiv.org/html/2608.06926#Sx4.T1)contains the cluster sorted by their mean score, how they are split between first and second trials, and the significance of that split\.

Table 1:Cluster trial separation, ranked by score mean\.As shown in the table, seven of eight clusters show a significant trial skew \(binomialp<p<0\.05\), meaning TRIBE differentiates trial order from communications alone, with no knowledge of trial order, semantic content, or scores\. This is not simply tracking the score increase between trials since scores alone poorly separate 1st and 2nd trials \(pseudo\-R2R^\{2\}=0\.08\), while TRIBE clusters do so far more strongly \(pseudo\-R2R^\{2\}=0\.31\)\. Only the worst performing cluster shows no skew \(53%, 47%;p=p=1\.00\) indicating poor team behavior is trial agnostic, which is a strong signal for intervention if we can predict it early on\.

### Early Prediction: Real Time Viability

A critical capability for any intervention system is the ability to assess team status early enough to enable timely assistance\. To test TRIBE’s viability for real time monitoring, we evaluated prediction accuracy when classifying teams based on progressively larger transcript segments\.

The methodology: for each trial, we extracted cumulative 10% increments of the transcript and inferred each segment’s topic distribution via fold\-in Gibbs sampling against the original model’s fixed topic\-word distribution, rather than refitting a new topic model per segment\. This way we ensured that the segment and the full trial theta vectors share an identical topic space\. Each segment was then assigned to its nearest of the 8 original cluster centroids and compared against the trial’s full transcript cluster assignment\.

Results:TRIBE achieves 47% prediction accuracy at 10% of the transcript, a 3\.0x improvement over the majority cluster baseline of 15\.8% \(and 3\.76x over uniform random of 12\.5%\), rising to 76% at 30%, and 90% at 50%\.

These results establish proof of concept that early communication patterns can predict behavioral cluster membership with substantial accuracy\. Given that a team starting in a high performing cluster at 10% may migrate to lower performance by 50%, such checkpoints represent potential intervention opportunities where an external agent could detect and respond to behavioral degradation\.

## Behavioral Dynamics Under Various Intervention Strategies

The previous section demonstrated the proof of concept: TRIBE captures meaningful performance clusters and enables early prediction\. We now investigate deeper questions: What happens when the agents intervene? Do different agent designs produce different impacts? Can we characterize the temporal dynamics of behavioral change?

To answer these questions, we analyze the \(pre\-existing\) intervention metadata embedded in Study 3: teams were assigned to 8 different intervention conditions, the 6 different AI agents, Human\-Advisor, and No\-Advisor\. We apply temporal analyses namely survival analysis, intervention effectiveness quantification, and Markov chain modeling to reveal how each agent shapes team behavioral trajectory\. A critical emphasis of this work is that TRIBE assesses intervention impact regardless of each agent’s underlying intervention logic\. We do not validate or compare the proprietary mechanisms by which these agents decide to intervene\. Rather, we demonstrate that TRIBE can reveal the behavioral consequences of interventions across different agent systems, enabling a domain independent approach to understanding how various AI architectures affect team dynamics\.

### Survival Analysis: Behavioral Stability

Survival analysis\(Kalbfleisch and Prentice[2002](https://arxiv.org/html/2608.06926#bib.bib37)\)provides a framework for understanding team behavioral stability over time\. To perform survival analysis, we split trials into 10% segments and model cluster persistence as a survival problem, where the event is the transition from initial cluster\.

Of the 8 intervention categories, No\-Advisor, SIFT agent, and CRA agent resulted in 100% of teams changing clusters, and for the remaining 5, most teams changed cluster in the 2 first segments\. The universal median survival time of 1\.0 window revealed that most teams abandon initial communication patterns between 10% and 20% marks of trials\. This timing aligns with Study 3’s task structure: each 17 minute trial begins with a 2 minute planning phase followed by 15 minutes of field execution\. The cluster change at 10\-20% indicates that TRIBE accurately captures the shift from planning to action which is the real communication behavior shift rather than just being noise in communication patterns\.

### Intervention Effectiveness Analysis

Beyond understanding when interventions cause teams to change clusters, we evaluate whether these impacts improve or degrade team performance\. We classified all cluster transitions into four categories: Positive Change \(transition to higher ranked cluster\), Negative Change \(transition to lower ranked cluster\), Good Stable \(remaining in the same high performing cluster\), and Bad Stable \(remaining in the same low performing cluster\)\. We calculate Net Change as the difference between Positive and Negative changes, providing a single metric for intervention impact\. Table[2](https://arxiv.org/html/2608.06926#Sx5.T2)presents the percentage breakdown\.

Table 2:Intervention outcome percentages by agent\.AgentBadGoodNeg\.Pos\.NetStableStableChangeChangeChangeUSC22\.5%25\.0%20\.0%32\.5%12\.5Human15\.7%28\.3%25\.2%30\.7%5\.5CRA15\.1%20\.2%32\.5%32\.1%\-0\.4UAZ13\.8%28\.9%28\.0%29\.4%\-1\.4DOLL26\.5%13\.7%30\.8%29\.1%\-1\.7CMURI15\.2%37\.8%24\.4%22\.6%\-1\.8SIFT13\.1%18\.8%35\.0%33\.1%\-1\.9TRIBE exposes two distinctive patterns in interventions:

- •AlthoughCMURIagent interventions have the largest impact \(37\.8%\) keeping good teams stable in their cluster, when CMURI interventions do change team behaviors, they have a negative net effect \(\-1\.8%\), suggesting successful teams may dismiss interventions\.
- •USCagent interventions, while less frequent \(80 total vs\. CMURI’s 217\), demonstrate positive impact when changing team behavior \(12\.5% net positive\), more than twice the positive impact of Human\-Advisor \(5\.5%\)\.

To quantify intervention impact depth, we examined cluster rank change post intervention\. While two interventions could both have positive effects, one may improve cluster rank more substantially\. Figure[1](https://arxiv.org/html/2608.06926#Sx5.F1)shows that USC rank improvement is even more impactful, 3x the Human\-Advisor impact and more than 4x the UAZ agent\.

![Refer to caption](https://arxiv.org/html/2608.06926v1/figures/sectional_survival_avg_rank_improvement_actual_changes_only.png)Figure 1:Average rank improvement\.nnis the number of interventions that triggered changes, not the total count\.To determine statistical significance of observed differences, we compared each agent’s transition patterns to No\-Advisor baseline using chi square tests, see Table[3](https://arxiv.org/html/2608.06926#Sx5.T3)\.

Table 3:Statistical comparison to No\-Advisor baseline\. Tests compare flattened8×88\{\\times\}8transition matrices\. df ranges 53–59 among tests with exactpp, varying because cluster\-pair cells unobserved in both conditions are excluded\.Notably, Human\-Advisor shows no significant difference from No\-Advisor \(χ2\\chi^\{2\}=57\.53, p=0\.31\), although a larger sample could show a real difference\. By contrast, all AI agents show significant divergence \(p<<0\.05\)\.

### Markov Chain Analysis

A Markov chain\(Norris[1998](https://arxiv.org/html/2608.06926#bib.bib42)\)models a system transitioning between discrete states according to fixed probabilities\. In our context, states are the 8 communication clusters identified by TRIBE, transitions represent movement between clusters at each 10% segment, and the Markov property assumes future cluster depends only on current cluster, not history\.

Each agent’s influence on team dynamics is represented as an 8×8 transition probability matrix, where P\[i,j\] is the probability of transitioning from cluster i to cluster j\.

Analysis of 252 transitions \(28 trials per agent×\\times9 transitions\) per agent reveals several patterns\. Surprisingly, no clusters showed absorption behavior \(\>\>90% self transition probability\), indicating teams remain behaviorally dynamic throughout trials, no communication pattern becomes stuck, even poor performing clusters\.

![Refer to caption](https://arxiv.org/html/2608.06926v1/figures/sectional_markov_good_cluster_evolution.png)Figure 2:Good cluster evolution over 10 time slots\.Every cluster can reach every other cluster, with varying probabilities, revealing that no behavioral states are isolated and improvement is always possible from any starting point\. Figure[2](https://arxiv.org/html/2608.06926#Sx5.F2)shows how the probability of being in good clusters \(3 best performing clusters: 1, 2, 3\) evolves\.

No\-Advisor starts at 77% and stabilizes around 78%, showing natural dynamics favor good clusters\. CMURI, UAZ, and Human\-Advisor start around 66\-67% and improve to 70%\. USC, SIFT, and DOLL start and remain in 63\-66% range\. CRA starts lowest at approximately 58% with minimal improvement\. Most changes occur by step 2\-3, confirming that the teams quickly establish behavioral patterns persisting throughout trials\. We calculate Mean First Passage Time \(MFPT\)\(Feller[1968](https://arxiv.org/html/2608.06926#bib.bib35)\)to quantify the average steps needed to transition from worst to best cluster; as shown in Table[4](https://arxiv.org/html/2608.06926#Sx5.T4)\.

Table 4:Average steps to reach best cluster from worst\.The fact that No\-Advisor provides the fastest path from worst to best performance is striking, yet, it does raise questions about the fairness of team quality distribution over the agent cases vs\. the No\-Advisor case\. We checked whether team quality was fairly distributed across the 8 categories\. Table[5](https://arxiv.org/html/2608.06926#Sx5.T5)shows the distribution of trial quality per category and reveals that No\-Advisor received disproportionately high quality teams while USC received low quality teams, explaining why No\-Advisor achieved faster MFPT\.

Table 5:Trial count by agent, based on cluster ranking\.ClusterCMURICRADOLLSIFTUAZUSCHumanNo\-Adv155522547246293343344516116431033233551624464606583432753122543822415410No\-Advisor has 0 trials in worst performing cluster and largest number in best cluster\. Aggregating clusters into three performance tiers \(top: 1,2,3; middle: 4,5; low: 6,7,8\) we arrive at the distributions in Table[6](https://arxiv.org/html/2608.06926#Sx5.T6)which indicate that No\-Advisor and CMURI were assigned disproportionately high quality teams, whereas USC received predominantly low performing teams\. Such allocation bias in Study 3 dataset, rather than inherent differences in natural team dynamics, can explain the counterintuitive finding that unguided teams outperformed those with AI assistance in MFPT\.

Table 6:Trial count by agent and performance group\.Starting tier distributions differ across agents, so cross agent comparisons should be read jointly with the rank improvement analysis\. Critically, because clusters derive purely from communication and are independent of trial score, detecting this imbalance by score alone would be circular; meaning labeling high scoring teams as better and citing those same scores as evidence of bias explains nothing about*why*they scored well\. Was it that stronger teams were assigned to No\-Advisor; or was it that withholding intervention improving performance\. TRIBE resolves this: the imbalance is attributable to team composition, not intervention effect\.

## Method Comparison and Optimization

Having established that TRIBE works for performance prediction and intervention analysis, we now ask two complementary questions: \(1\) Does a general purpose LLM, deployable locally within an agent, extract performance\-predictive behavioral patterns without task specific fitting? We test whether a modern Large Language Model with vast semantic knowledge can extract meaningful behavioral patterns from team communication, potentially improving on our approach\. \(2\) Can we optimize the pipeline beyond the baseline? The baseline TRIBE uses LDA for topic modeling and k\-means for clustering\. We systematically test alternative methods for each component to determine whether different combinations might achieve better performance prediction\. \(All temporal analysis in previous section employed the LDA\+k\-means configuration to ensure fair intervention assessment\.\)

### Llama Behavioral Extraction

We tested whether Llama 3\.3 70B via Ollama\(Dubeyet al\.[2024](https://arxiv.org/html/2608.06926#bib.bib31)\), a modern LLM with extensive training on collaborative task data, could extract behavioral patterns from team communication\. At Stage 0, we provided the LLM with the dataset, without any mention of clusters and asked for behavioral patterns that could predict the performance, but the outcomes were uninformative\. Then, we proceeded in three stages with progressively more specific prompts\.

Stage 1 \- Free labeling prompt: We let the LLM label freely with the prompt:"Based on team communication patterns, create a 2 word label for each cluster that reflects its behavioral patterns\."The labels were too generic and all clusters received positive labels regardless of performance\.

Stage 2 \- Domain informed prompt: We gave Llama domain information and made it clear that we were looking for labels based on behaviors and performances, asking for differentiation and uniqueness\. The labels were still extremely generic and not indicating anything about performance as the worst cluster was labeled "Good Communication\."

Stage 3 \- Performance informed prompt: We provided all detailed information about the task and data, crucially provided thecluster score meansalong withcluster performance rankings, and asked for behavior and performance based labeling\. Only the explicit ranking of the cluster made the LLM come up with somewhat relevant labels, and yet, Cluster 4 was labeled "Rank\-4 Team" \(Table[7](https://arxiv.org/html/2608.06926#Sx6.T7)\), meaning it failed to extract meaningful behavioral information\.

Table 7:Llama behavioral labeling based on cluster ranking\.Even with explicit performance rankings and detailed task information, Llama did not identify meaningful behavioral patterns, despite its vast training on collaborative language\. This comparison is deliberately scoped as one open weight model under three prompting regimes\. Frontier models, fine tuning, or embedding based pipelines may perform better\. Our claim is therefore narrower than "TRIBE outperforms LLMs": for real time, on device deployment within agents that must analyze team communications and deliver timely interventions, our lightweight statistical pipeline extracted performance\-predictive structure that this LLM configuration did not, despite its vastly greater scale and compute\.

### Topic Modeling Alternative Methods

LDA’s main limitation is the computational cost \(\>\>2 hours for 100 runs with reproducibility challenges\)\. We evaluated NMF\(Lee and Seung[1999](https://arxiv.org/html/2608.06926#bib.bib33),[2000](https://arxiv.org/html/2608.06926#bib.bib43)\), Top2Vec\(Angelov[2020](https://arxiv.org/html/2608.06926#bib.bib44)\), DTM\(Blei and Lafferty[2006](https://arxiv.org/html/2608.06926#bib.bib45)\), and STM\(Robertset al\.[2014](https://arxiv.org/html/2608.06926#bib.bib46)\)as alternatives\. \(LSA/LSI lacks topic structure; HDP over fragments small corpora; BERTopic’s density based clustering is unreliable below 1k documents\.\) NMF emerged as the most promising alternative\. With optimal configuration \(6 topics\), NMF finished in just 5 minutes achieving a 24x speedup over LDA which took 120 minutes\. Top2Vec consistently discovered only 2\-3 topics \(insufficient for communication granularity\)\. DTM and STM showed promise theoretically but didn’t improve the performance prediction\.

Unigram based NMF with k=6 topics, also outperformed bigram variants, becoming optimal topic modeling method for the combined pipeline optimization\.

### Clustering Alternative Methods

All clustering methods in this comparison operate on the NMF \(k=6\) topic\-distribution vectors identified as optimal in the previous subsection\. Beyond k\-means, we tested: Gaussian Mixture Models \(GMM\)\(Reynolds[2015](https://arxiv.org/html/2608.06926#bib.bib34)\)with 4 covariance types, Bayesian GMM \(BGMM\)\(Bishop[2006](https://arxiv.org/html/2608.06926#bib.bib47)\), hierarchical clustering \(4 linkage methods\)\(Ward Jr[1963](https://arxiv.org/html/2608.06926#bib.bib49)\), and Dynamic Time Warping \(DTW\)\(Sakoe and Chiba[1978](https://arxiv.org/html/2608.06926#bib.bib48)\)\.

Internal vs\. external validation disconnect: Standard clustering metrics \(BIC, silhouette, gap statistics\) suggested k=3\-4 as optimal cluster counts\. However, external validation against team performance revealed the opposite that more granular clustering \(k=9\-16\) actually predicted performance better \(Table[8](https://arxiv.org/html/2608.06926#Sx6.T8)\)\. SinceR2R^\{2\}increases with the number of clusters and each additional cluster adds a regression degree of freedom, to have a fair comparison across configurations with different k, we reportR¯2\\bar\{R\}^\{2\}\(adjustedR2R^\{2\}\) instead ofR2R^\{2\}\.

To ensure reported values reflect stable estimates rather than a single favorable initialization, each stochastic clustering method \(K\-means, GMM, BGMM\) was refit across 100 random seeds, with hyperparameters selected per fit via unsupervised criteria \(BIC or Calinski\-Harabasz score\), not against trial score\. We report the meanR¯2\\bar\{R\}^\{2\}across 100 seeds\.

Table 8:Clustering method comparison\.GMM with k=16 and BGMM with k=9 were the top performers\. GMM\(16\) achieved 25\.4%R¯2\\bar\{R\}^\{2\}and became our optimal clustering component\. To confirm the clustering–performance relationship is not an artifact of chance, we ran a permutation test at our selected configuration \(GMM, k=16\): shuffling trial scores 1,000 times and refitting on each which yielded aR¯2\\bar\{R\}^\{2\}that never exceeded 0\.152 \(mean 0\.028\), far below the observed 0\.254 \(p<0\.001p<0\.001\)\.

### The Optimal Pipeline

The first thing we tested, before the baseline TRIBE \(LDA 12\+K\-means 8\), was TF\-IDF \(SVD\-reduced to 12 dims, k=8\) which producedR¯2\\bar\{R\}^\{2\}=0\.067\. The baseline TRIBE achievedR¯2\\bar\{R\}^\{2\}= 0\.185, confirming LDA’s topic structure captures signals beyond raw lexical features\. Our optimization analysis showed that NMF achieves the best performance amongst topic modeling methods and GMM amongst clustering methods, meaning NMF\(6\)\+GMM\(16\) is the overall optimal TRIBE pipeline withR¯2\\bar\{R\}^\{2\}=0\.254\. Comparing the baseline TRIBE to the optimal TRIBE, we see a 37\.3% improvement in performance, along with 24x speedup \(120m vs\. 5m\)\.

## Domain Independence

A critical test for TRIBE is whether insights generalize beyond Study 3\. This section examines domain independence by applying TRIBE to four diverse datasets\. To ensure fair cross domain comparison, all analyses in this section use the LDA\+K\-means pipeline\.

### Cross Domain Methodology

Each dataset underwent identical processing to ensure fair comparison\. The standardized pipeline involved: preprocessing data by cleaning utterances, setting minimum document length thresholds, and matching team data across trials; then performing topic model evaluation with topic counts ranging from 2 to 25 to find optimal configurations and similarly for clustering configurations\.

We selected 4 datasets representing various collaborative tasks with team conversations and performance measures\.

Study 3\(Huanget al\.[2022](https://arxiv.org/html/2608.06926#bib.bib27)\): detailed in Data section\.

Study 4\(Huanget al\.[2024](https://arxiv.org/html/2608.06926#bib.bib28)\): involves teams of 3 for a virtual search and defusal task in an environment where the team score goes down if an explosive explodes\. Unlike Study 3, teams in Study 4 can only communicate during planning phases when gathered in a shop\. Once they enter the field to defuse explosives, verbal communication is disabled\. Teams can return to the shop multiple times to replan as needed, allowing for extended discussions\. We chose minimum of 40 lines per trial, leading to 231 trials from 30 teams\.

DELIdataset\(Karadzhovet al\.[2023](https://arxiv.org/html/2608.06926#bib.bib29)\): involves teams solving logical reasoning tasks \(Wason selection task\) through discussion\. In this task design, there is one correct answer that can be logically deduced\. Teams of 3\-5 people deliberate to identify which cards to check to test logical rules\. Dataset contains 333 team deliberations\.

GAPdataset\(Braley and Murray[2018](https://arxiv.org/html/2608.06926#bib.bib30)\): involves teams ranking 15 survival items by importance after a hypothetical plane crash\. The items are then ranked by domain experts for score allocation\. The dataset contains 28 team discussions\.

### Performance Prediction

The key test of TRIBE is how it performs across these 4 different domains\. Table[9](https://arxiv.org/html/2608.06926#Sx7.T9)presents the results:

Table 9:Statistical performance of clustering\.ANOVA F\-tests assess variance across clusters; Cohen’s d\(Cohen[1988](https://arxiv.org/html/2608.06926#bib.bib38)\)quantifies effect size\. Study 3 achievedR¯2\\bar\{R\}^\{2\}= 18\.5% with clear cluster differentiation\. Cohen’s d = 0\.94 indicates a large effect size, with 82% of worst cluster teams scoring below average of best cluster\. Study 4’s setting yielded more dramatic results withR¯2\\bar\{R\}^\{2\}= 44\.3%\. Cohen’s d = 1\.82 represents a very large effect size, with 97% of worst cluster teams scoring below average of best cluster\.

Study 4’s "unlimited" time communication structure allowed for more re\-strategizing and behavioral expressions to surface, whereas Study 3’s 17 minute time limit for planning and execution combined might have created the urgency for taking quick actions to increase the score, limiting the opportunity for behavioral expression\.

DELI achieved a modestR¯2\\bar\{R\}^\{2\}= 14\.8% with limited differentiation\. Cohen’s d = 0\.51 represents a medium effect size, with approximately 69% of the worst cluster teams scoring below the average of the best cluster\. DELI’s finite answer structure limits behavioral variation, explaining the lowR¯2\\bar\{R\}^\{2\}\.

While GAP dataset does allow for more behavioral expressions withR¯2\\bar\{R\}^\{2\}= 28\.6%, it failed the significance test \(p=0\.253p=0\.253\), reflecting small sample inflation\. However, this is promising for a similar task with larger data collections\.

The cross domain analysis, not only showcases the domain independence capabilities of TRIBE, but also reveals that the task design determines the capacity for behavioral expressions\. Consequently, TRIBE’s effectiveness depends on the degree that a task allows for behavioral variation\.

### Principal Component Analysis of Study 4

To understand Study 4’s superior predictive power, we performed PCA on its topic distributions\. Study 4’s LDA model discovered 13 topics from planning phase communications\. Each team’s communication pattern can be represented as a 13 dimensional vector showing emphasis on each topic\. PCA transforms this 13 dimensional space into a new coordinate system where the first principal component \(PC1\) captures the maximum variance\.

Table 10:PC1 vs\. Cluster PerformanceThe analysis revealed a striking finding: At the cluster level, PC1 explainedR2=0\.927R^\{2\}=0\.927of the variance in mean cluster performance \(p=0\.002p=0\.002\) and at the trial level, PC1 explainedR2=0\.49R^\{2\}=0\.49\(p<0\.001p<0\.001\)\. This result reveals that Study 4’s performance is substantially explained by a single behavioral dimension, where teams can be positioned on a performance continuum of communication patterns\.

PC1 represents a weighted combination of all 13 topics, with topics 9 and 7 having strong negative loadings \(−0\.427\-0\.427and−0\.393\-0\.393respectively\) and topic 11 having strong positive loading \(\+0\.481\+0\.481\), at the cluster level\. This means teams emphasizing topics 9 and 7 receive negative PC1 values \(associated with higher performance\), while teams emphasizing topic 11 receive positive PC1 values \(associated with lower performance\)\. In other words, PC1 provides a continuous spectrum from high to low performance communication patterns, enabling real time monitoring for intervention design\.

## Conclusion

This work introduced TRIBE, a domain independent approach for understanding and predicting team performance through communication patterns\. By applying topic modeling and clustering to team conversations, we demonstrated that behavioral dynamics invisible to traditional performance metrics can be extracted for early intervention\. Teams can be classified into performance predictive behavioral clusters as early as 10% into the task, enabling timely support\.

Our analysis revealed several counterintuitive findings\. While teams without advisors outperformed those with AI guidance due to bias in team quality assignment, TRIBE’s ability to detect this bias validates its capacity to identify meaningful performance related patterns\. Human advisors produced patterns statistically indistinguishable from natural team evolution, whereas all AI agents significantly altered behavioral trajectories, showing how TRIBE can evaluate the impact of interventions\. TRIBE’s survival and Markov analyses revealed behavioral flexibility throughout trials\.

Cross domain validation established TRIBE’s generalizability, while task characteristics determine how much behavioral patterns emerge\. Study 4’s planning structure achieved 44\.3% predictive power, with PC1 alone explaining 49% of trial level performance variance, a continuum enabling real time intervention triggers; whereas DELI’s finite answer task reached only 14\.8%\.

Method optimization improved predictive power by 37\.3% with a 24x speedup\. TRIBE’s comparison to a contemporary LLM showed that, despite the LLM’s semantic knowledge, it could not identify meaningful behavioral patterns, suggesting the statistical structure of interaction patterns reveals team behavioral signatures\. Overall, TRIBE offers a scalable, domain independent approach for team performance prediction where task design allows sufficient behavioral expression, providing actionable insights for autonomous agents supporting teams\.

## References

- S\. V\. Albrecht and P\. Stone \(2018\)Autonomous agents modelling other agents: a comprehensive survey and open problems\.Artificial Intelligence258,pp\. 66–95\.Cited by:[Introduction](https://arxiv.org/html/2608.06926#Sx1.p2.1)\.
- D\. Angelov \(2020\)Top2Vec: distributed representations of topics\.arXiv preprint arXiv:2008\.09470\.Cited by:[Topic Modeling Alternative Methods](https://arxiv.org/html/2608.06926#Sx6.SSx2.p1.1)\.
- A\. J\. Bisberg, S\. J\. Shaikh, Y\. Zeng, F\. Morstatter, E\. Chen, E\. Ferrara, and D\. Williams \(2025\)Communication patterns predict team skill in multiplayer online games\.Proceedings of the ACM on Human\-Computer Interaction9\(7\)\.Note:CSCW388External Links:[Document](https://dx.doi.org/10.1145/3757569)Cited by:[Related Work](https://arxiv.org/html/2608.06926#Sx2.p1.1)\.
- C\. M\. Bishop \(2006\)Pattern recognition and machine learning\.Vol\.4,Springer,New York\.Cited by:[Clustering Alternative Methods](https://arxiv.org/html/2608.06926#Sx6.SSx3.p1.1)\.
- D\. M\. Blei and J\. D\. Lafferty \(2006\)Dynamic topic models\.InProceedings of the 23rd international conference on machine learning,pp\. 113–120\.Cited by:[Topic Modeling Alternative Methods](https://arxiv.org/html/2608.06926#Sx6.SSx2.p1.1)\.
- D\. M\. Blei, A\. Y\. Ng, and M\. I\. Jordan \(2003\)Latent Dirichlet allocation\.Journal of Machine Learning Research3,pp\. 993–1022\.Cited by:[Latent Dirichlet Allocation](https://arxiv.org/html/2608.06926#Sx4.SSx2.SSSx1.p1.1)\.
- M\. Braley and G\. Murray \(2018\)The group affect and performance \(gap\) corpus\.InProceedings of the Group Interaction Frontiers in Technology,External Links:[Document](https://dx.doi.org/10.1145/3279981.3279985)Cited by:[Cross Domain Methodology](https://arxiv.org/html/2608.06926#Sx7.SSx1.p6.1)\.
- Charles River Analytics \(2024\)Using ai to understand and assist human teams: charles river analytics develops foundational ai technology to understand human behavior and improve team performance\.Note:https://cra\.com/using\-ai\-to\-understand\-and\-assist\-human\-teams\-charles\-river\-analytics\-develops\-foundational\-ai\-technology\-to\-understand\-human\-behavior\-and\-improve\-team\-performance/Cited by:[Study 3: Urban Search and Rescue](https://arxiv.org/html/2608.06926#Sx3.SSx1.p2.1)\.
- J\. Cohen \(1988\)Statistical power analysis for the behavioral sciences\.2nd edition,Lawrence Erlbaum Associates\.Cited by:[Performance Prediction](https://arxiv.org/html/2608.06926#Sx7.SSx2.p2.2)\.
- C\. C\. Corral, K\. S\. Tatapudi, V\. Buchanan, L\. Huang, and N\. J\. Cooke \(2021\)Building a synthetic task environment to support artificial social intelligence research\.InProceedings of the Human Factors and Ergonomics Society Annual Meeting,Vol\.65,pp\. 660–664\.Cited by:[Study 3: Urban Search and Rescue](https://arxiv.org/html/2608.06926#Sx3.SSx1.p1.1)\.
- A\. Dubey, A\. Jauhri, A\. Pandey,et al\.\(2024\)The llama 3 herd of models\.External Links:2407\.21783Cited by:[Llama Behavioral Extraction](https://arxiv.org/html/2608.06926#Sx6.SSx1.p1.1)\.
- S\. Emmitt and C\. Gorse \(2006\)Communication in construction teams\.Routledge\.Cited by:[Introduction](https://arxiv.org/html/2608.06926#Sx1.p2.1)\.
- D\. Engome Tchupo and G\. A\. Macht \(2023\)Entropy for team communication pattern recognition\.Applied Ergonomics111\.Note:104038External Links:[Document](https://dx.doi.org/10.1016/j.apergo.2023.104038)Cited by:[Related Work](https://arxiv.org/html/2608.06926#Sx2.p1.1)\.
- W\. Feller \(1968\)An introduction to probability theory and its applications\.3rd edition,John Wiley & Sons\.Cited by:[Markov Chain Analysis](https://arxiv.org/html/2608.06926#Sx5.SSx3.p5.1)\.
- J\. T\. Freeman, L\. Huang, M\. Wood, and S\. J\. Cauffman \(2022\)Evaluating artificial social intelligence in an urban search and rescue task environment\.InComputational Theory of Mind for Human\-Machine Teams,N\. Gurney and G\. Sukthankar \(Eds\.\),pp\. 72–84\.External Links:[Document](https://dx.doi.org/10.1007/978-3-031-21671-8%5F5)Cited by:[Study 3: Urban Search and Rescue](https://arxiv.org/html/2608.06926#Sx3.SSx1.p1.1)\.
- L\. Huang, N\. J\. Cooke, R\. S\. Gutzwiller, S\. Berman, E\. K\. Chiou, M\. Demir, and W\. Zhang \(2021\)Distributed dynamic team trust in human, artificial intelligence, and robot teaming\.InTrust in Human\-Robot Interaction,C\. S\. Nam and J\. B\. Lyons \(Eds\.\),pp\. 301–319\.External Links:[Document](https://dx.doi.org/10.1016/B978-0-12-819472-0.00013-7)Cited by:[Study 3: Urban Search and Rescue](https://arxiv.org/html/2608.06926#Sx3.SSx1.p1.1),[Data and Task Environments](https://arxiv.org/html/2608.06926#Sx3.p1.1)\.
- L\. Huang, A\. Fouse, N\. Cooke, and E\. Weiss \(2024\)Artificial Social Intelligence for Successful Teams \(ASIST\) Study 4 Dragon Testbed Dataset\.ASU Library Research Data Repository\.External Links:[Document](https://dx.doi.org/10.48349/ASU/ZO6XVR),[Link](https://doi.org/10.48349/ASU/ZO6XVR)Cited by:[Cross Domain Methodology](https://arxiv.org/html/2608.06926#Sx7.SSx1.p4.1)\.
- L\. Huang, J\. Freeman, N\. Cooke, J\. "\. Colonna\-Romano, M\. Wood, V\. Buchanan, and S\. Caufman \(2022\)Artificial Social Intelligence for Successful Teams \(ASIST\) Study 3\.ASU Library Research Data Repository\.External Links:[Document](https://dx.doi.org/10.48349/ASU/QDQ4MH),[Link](https://doi.org/10.48349/ASU/QDQ4MH)Cited by:[Cross Domain Methodology](https://arxiv.org/html/2608.06926#Sx7.SSx1.p3.1)\.
- T\. Jones \(2021\)TextmineR: functions for text mining and topic modeling\.Note:R package version 3\.0\.5External Links:[Link](https://cran.r-project.org/package=textmineR)Cited by:[Preprocessing of Transcripts](https://arxiv.org/html/2608.06926#Sx4.SSx1.p2.1),[Latent Dirichlet Allocation](https://arxiv.org/html/2608.06926#Sx4.SSx2.SSSx1.p1.1)\.
- J\. D\. Kalbfleisch and R\. L\. Prentice \(2002\)The statistical analysis of failure time data\.2nd edition,Wiley\.Cited by:[Survival Analysis: Behavioral Stability](https://arxiv.org/html/2608.06926#Sx5.SSx1.p1.1)\.
- G\. Karadzhov, T\. Stafford, and A\. Vlachos \(2023\)DeliData: a dataset for deliberation in multi\-party problem solving\.InProceedings of the ACM on Human\-Computer Interaction,Vol\.7,pp\. 1–25\.External Links:[Document](https://dx.doi.org/10.1145/3610056)Cited by:[Cross Domain Methodology](https://arxiv.org/html/2608.06926#Sx7.SSx1.p5.1)\.
- U\. Kuter and C\. Miller \(2022\)Artificial social intelligence for successful teams \(asist\): sift team project\.Note:https://www\.sift\.net/research/artificial\-intelligence/asistCited by:[Study 3: Urban Search and Rescue](https://arxiv.org/html/2608.06926#Sx3.SSx1.p2.1)\.
- D\. D\. Lee and H\. S\. Seung \(1999\)Learning the parts of objects by non\-negative matrix factorization\.Nature401\(6755\),pp\. 788–791\.Cited by:[Topic Modeling Alternative Methods](https://arxiv.org/html/2608.06926#Sx6.SSx2.p1.1)\.
- D\. D\. Lee and H\. S\. Seung \(2000\)Algorithms for non\-negative matrix factorization\.InAdvances in neural information processing systems,Vol\.13,pp\. 556–562\.Cited by:[Topic Modeling Alternative Methods](https://arxiv.org/html/2608.06926#Sx6.SSx2.p1.1)\.
- Z\. Li, Z\. Lu, and M\. Yin \(2023\)Modeling human trust and reliance in ai\-assisted decision making: a markovian approach\.InProceedings of the AAAI Conference on Artificial Intelligence,Vol\.37,pp\. 6056–6064\.External Links:[Document](https://dx.doi.org/10.1609/aaai.v37i5.25748)Cited by:[Related Work](https://arxiv.org/html/2608.06926#Sx2.p2.1)\.
- S\. H\. A\. Mahmood, Z\. Lu, and M\. Yin \(2024\)Designing behavior\-aware ai to improve the human\-ai team performance in ai\-assisted decision making\.InProceedings of the Thirty\-Third International Joint Conference on Artificial Intelligence \(IJCAI\),pp\. 3106–3114\.External Links:[Document](https://dx.doi.org/10.24963/ijcai.2024/344)Cited by:[Related Work](https://arxiv.org/html/2608.06926#Sx2.p2.1)\.
- S\. L\. Marlow, C\. N\. Lacerenza, and E\. Salas \(2017\)Communication in virtual teams: a conceptual framework and research agenda\.Human Resource Management Review27\(4\),pp\. 575–589\.Cited by:[Introduction](https://arxiv.org/html/2608.06926#Sx1.p2.1),[Related Work](https://arxiv.org/html/2608.06926#Sx2.p1.1)\.
- D\. Mimno, H\. Wallach, E\. Talley, and M\. Leenders \(2011\)Optimizing semantic coherence in topic models\.InProceedings of the Conference on Empirical Methods in Natural Language Processing,pp\. 262–272\.Cited by:[Latent Dirichlet Allocation](https://arxiv.org/html/2608.06926#Sx4.SSx2.SSSx1.p2.1)\.
- H\. Mozannar, J\. J\. Lee, D\. Wei, P\. Sattigeri, S\. Das, and D\. Sontag \(2023\)Effective human\-ai teams via learned natural language rules and onboarding\.InAdvances in Neural Information Processing Systems 36 \(NeurIPS 2023\),Note:arXiv:2311\.01007Cited by:[Related Work](https://arxiv.org/html/2608.06926#Sx2.p2.1)\.
- C\. Myers, J\. Ball, N\. Cooke, M\. Freiman, M\. Caisse, S\. Rodgers, M\. Demir, and N\. McNeese \(2018\)Autonomous intelligent agents for team training\.IEEE Intelligent Systems34\(2\),pp\. 3–14\.Cited by:[Related Work](https://arxiv.org/html/2608.06926#Sx2.p2.1)\.
- J\. R\. Norris \(1998\)Markov chains\.Cambridge University Press\.Cited by:[Markov Chain Analysis](https://arxiv.org/html/2608.06926#Sx5.SSx3.p1.1)\.
- T\. O’Neill, N\. McNeese, A\. Barron, and B\. Schelble \(2022\)Human–autonomy teaming: a review and analysis of the empirical literature\.Human Factors64\(5\),pp\. 904–938\.Cited by:[Related Work](https://arxiv.org/html/2608.06926#Sx2.p1.1)\.
- A\. Pyarelal, S\. Ashton, J\. Astier, L\. Champlin, K\. Barnard, E\. Butler, R\. Huang, C\. Jeong, S\. Kim, C\. Morrison,et al\.\(2023\)UAZ study 3 results\.OSF\.External Links:[Link](https://osf.io/t2rjm/)Cited by:[Study 3: Urban Search and Rescue](https://arxiv.org/html/2608.06926#Sx3.SSx1.p2.1)\.
- D\. V\. Pynadath, N\. Gurney, S\. Kenny, R\. Kumar, S\. C\. Marsella, H\. Matuszak, H\. Mostafa, P\. Sequeira, V\. Ustun, and P\. Wu \(2023\)Effectiveness of teamwork\-level interventions through decision\-theoretic reasoning in a minecraft search\-and\-rescue task\.InProceedings of the 2023 International Conference on Autonomous Agents and Multiagent Systems \(AAMAS\),pp\. 2334–2336\.Cited by:[Study 3: Urban Search and Rescue](https://arxiv.org/html/2608.06926#Sx3.SSx1.p2.1)\.
- D\. A\. Reynolds \(2015\)Gaussian mixture models\.Encyclopedia of Biometrics741,pp\. 659–663\.Cited by:[Clustering Alternative Methods](https://arxiv.org/html/2608.06926#Sx6.SSx3.p1.1)\.
- M\. E\. Roberts, B\. M\. Stewart, D\. Tingley, C\. Lucas, J\. Leder\-Luis, S\. K\. Gadarian, B\. Albertson, and D\. G\. Rand \(2014\)Structural topic models for open\-ended survey responses\.American Journal of Political Science58\(4\),pp\. 1064–1082\.Cited by:[Topic Modeling Alternative Methods](https://arxiv.org/html/2608.06926#Sx6.SSx2.p1.1)\.
- H\. Sakoe and S\. Chiba \(1978\)Dynamic programming algorithm optimization for spoken word recognition\.IEEE Transactions on Acoustics, Speech and Signal Processing26\(1\),pp\. 43–49\.Cited by:[Clustering Alternative Methods](https://arxiv.org/html/2608.06926#Sx6.SSx3.p1.1)\.
- S\. Schleibaum, L\. Feng, S\. Kraus, and J\. P\. Müller \(2024\)ADESSE: advice explanations in complex repeated decision\-making environments\.InProceedings of the Thirty\-Third International Joint Conference on Artificial Intelligence \(IJCAI\),pp\. 7904–7912\.External Links:[Document](https://dx.doi.org/10.24963/ijcai.2024/875)Cited by:[Related Work](https://arxiv.org/html/2608.06926#Sx2.p2.1)\.
- S\. Seo, L\. R\. Kennedy\-Metz, M\. A\. Zenati, J\. A\. Shah, R\. D\. Dias, and V\. V\. Unhelkar \(2021\)Towards an ai coach to infer team mental model alignment in healthcare\.InProceedings of the IEEE Conference on Cognitive and Computational Aspects of Situation Management,pp\. 39–44\.Cited by:[Introduction](https://arxiv.org/html/2608.06926#Sx1.p1.1),[Related Work](https://arxiv.org/html/2608.06926#Sx2.p2.1)\.
- H\. Simon\-Dempsey, R\. Huang,et al\.\(2023\)DOLL agent for asist study 3: team intervention through dynamic observation and learning\.Open Science Framework\.Note:https://osf\.io/mtqkbCited by:[Study 3: Urban Search and Rescue](https://arxiv.org/html/2608.06926#Sx3.SSx1.p2.1)\.
- J\. Stempfle and P\. Badke\-Schaub \(2002\)Thinking in design teams\-an analysis of team communication\.Design Studies23\(5\),pp\. 473–496\.Cited by:[Introduction](https://arxiv.org/html/2608.06926#Sx1.p2.1)\.
- G\. Sukthankar, K\. Sycara, J\. A\. Giampapa, C\. Burnett, and A\. Preece \(2007\)Towards a model of agent\-assisted team search\.InProceedings of the First Annual Conference of the International Technology Alliance in Network and Information Science,Cited by:[Introduction](https://arxiv.org/html/2608.06926#Sx1.p1.1)\.
- K\. Sycara, M\. Lewis, and D\. Hughes \(2021\)DARPA asist study 3 preregistration: cmu\-ri team\.Open Science Framework\.Note:https://osf\.io/yj52e/Cited by:[Study 3: Urban Search and Rescue](https://arxiv.org/html/2608.06926#Sx3.SSx1.p2.1)\.
- R\. Tibshirani, G\. Walther, and T\. Hastie \(2001\)Estimating the number of clusters in a data set via the gap statistic\.Journal of the Royal Statistical Society: Series B \(Statistical Methodology\)63\(2\),pp\. 411–423\.Cited by:[Behavioral Clustering](https://arxiv.org/html/2608.06926#Sx4.SSx3.p2.1)\.
- J\. Tiferes and A\. M\. Bisantz \(2018\)The impact of team characteristics and context on team communication: an integrative literature review\.Applied Ergonomics68,pp\. 146–159\.Cited by:[Introduction](https://arxiv.org/html/2608.06926#Sx1.p2.1),[Related Work](https://arxiv.org/html/2608.06926#Sx2.p1.1)\.
- J\. H\. Ward Jr \(1963\)Hierarchical grouping to optimize an objective function\.Journal of the American statistical association58\(301\),pp\. 236–244\.Cited by:[Clustering Alternative Methods](https://arxiv.org/html/2608.06926#Sx6.SSx3.p1.1)\.
- S\. S\. Webber, J\. Detjen, T\. L\. MacLean, and D\. Thomas \(2019\)Team challenges: is artificial intelligence the solution?\.Business Horizons62\(6\),pp\. 741–750\.Cited by:[Introduction](https://arxiv.org/html/2608.06926#Sx1.p1.1)\.
- R\. Zhang, N\. J\. McNeese, G\. Freeman, and G\. Musick \(2021\)"An ideal human" expectations of ai teammates in human\-ai teaming\.InProceedings of the ACM on Human\-Computer Interaction,Vol\.4,pp\. 1–25\.Cited by:[Related Work](https://arxiv.org/html/2608.06926#Sx2.p1.1)\.
- F\. R\. H\. Zijlstra, M\. J\. Waller, and S\. I\. Phillips \(2012\)Setting the tone: early interaction patterns in swift\-starting teams as a predictor of effectiveness\.European Journal of Work and Organizational Psychology21,pp\. 749–777\.External Links:[Document](https://dx.doi.org/10.1080/1359432X.2012.690399)Cited by:[Related Work](https://arxiv.org/html/2608.06926#Sx2.p1.1)\.