TRIBE analyzes team communications to predict performance outcomes. The system discovers hidden patterns of behaviors—and correlates these patterns with measurable results. Communication patterns serve as performance indicators, revealing early on to a task which teams could fail or succeed.
Using natural language processing and unsupervised machine learning, TRIBE identifies communication topics without predetermined categories, maps how teams distribute their attention across these topics, and finds which distributions correlate with their success or failure. The system requires only conversation data and performance metrics.
To showcase the system, we applied it to two collaborative team tasks. In controlled studies involving hundreds of teams, A single behavioral dimension recovered by dimension reduction explains 93% of the variance in mean cluster performance and 49% at the individual trial level. Teams at opposite ends of that dimension differ in score by 134%.
TRIBE transforms unstructured communication into quantifiable behavioral categories through sophisticated analysis techniques.
TRIBE operates through a systematic process that transforms raw communications into actionable performance predictions. The pipeline progresses through six stages, each building upon the previous to extract increasingly sophisticated behavioral insights.
Each stage uses established computational methods in novel combination. Topics emerge through various topic modeling techniques without supervision. Patterns self-organize through clustering algorithms. Performance correlations reveal themselves through dimensionality reduction. The innovation lies not in any single technique but in recognizing that team communication contains predictive signal that can be automatically extracted.
We demonstrate TRIBE's capabilities through analysis of two off-the-shelf datasets from collaborative team tasks: first, a search and defusal scenario, and then, a search and rescue scenario*.
*Artificial Social Intelligence for Successful Teams (ASIST) Program, DARPA, 2020-2024
To demonstrate TRIBE's capabilities, we analyzed data from a team task where team members collaborate in a Minecraft search and defusal scenario.
The study involved 231 three-person teams operating in a Minecraft-based simulation. Each team member was identified by color (red, blue, green) and had specialized capabilities. The task required two distinct phases:
Planning Phase: Teams gathered in a virtual shop where they could communicate verbally to strategize. They discussed approaches, assigned responsibilities, and coordinated intended actions. All verbal communication was recorded and transcribed for analysis. The timer would pause every time team entered a shop, allowing for unlimited number of team planning sessions.
Execution Phase: Teams deployed to the field where verbal communication was prohibited. They could only communicate through placing colored flags visible to teammates. Success depended entirely on the quality of their planning phase communication.
Performance was measured through objective scoring based on successful defusals, with scores ranging from -192 to 1080 points, meaning the team could take hit and lose score in unsuccessful defusals. This created ideal conditions for TRIBE analysis: natural team communication with clear, measurable outcomes.
TRIBE discovered 13 distinct communication topics in the planning conversations, and 6 behavioral pattern clusters. In order to make sense of what each cluster is focused on, we need to know what topic-combinations are forming the behavioral each cluster. The question is how can we reduce these 13 dimensions (topics) that describe the behaviors of the teams into a scale for change in communication patterns of each behavioral cluster?
Principal Component Analysis (PCA) is a statistical technique that simplifies complex data by finding the most important patterns of variation. Imagine you have 13 different measurements for each team (their emphasis on each of the 13 topics). PCA finds new dimensions—called principal components (PCs)—that capture the essential differences between teams while reducing this complexity. The first principal component (PC1) is the single dimension that explains the most variation in the data, the second (PC2) explains the next most, and so on. Each PC is a weighted combination of the original measurements, where the weights show how much each original variable contributes to that dimension.
In TRIBE's analysis, Principal Component Analysis revealed that despite 13 topics and 6 behavioral clusters, teams fundamentally varied along PC1 (the first principal component). This single axis captured nearly all performance-relevant variation. PC1 represents a specific blend of topics: teams with negative PC1 values emphasized topics 9 and 7 (completion-focused language), while teams with positive PC1 values emphasized topic 11 (role-focused language).
In this dataset, PC1 emerged as the dominant predictor, explaining 93% of the variance in mean cluster performance (p = 0.002) and 49% of variance at the individual trial level (p < 0.001). Other datasets may reveal multiple important components depending on their communication complexity. The key insight is that TRIBE automatically identifies whichever dimensions matter for performance.
Move the slider to explore how PC1 position correlates with team performance. Teams on the left (negative PC1) emphasize completion and execution topics, while teams on the right (positive PC1) focus more on role discussion. Below the sliding scale are the clusters with their corresponding average scores.
Note: The percentages show the emphasis level of each top 3 topic (out of 13 total) at the current PC1 position.
The six behavioral clusters showed dramatically different performance levels, perfectly ordered by their PC1 positions. The bold entries in the table below highlight the extreme performers—the top two highest-scoring clusters (941 and 876 points) and the lowest-scoring cluster (402 points). These extremes demonstrate the remarkable 539-point performance gap between teams with the best communication patterns versus those with the worst, representing more than a 2x difference in effectiveness.
The key to understanding PC1 lies in how different clusters emphasized different topics during planning.
Topic 9: Cluster 4 (best) devoted 42.1% of communication to this topic. Cluster 2 (worst) used it only 8%. A 5x difference.
Topic 7: Cluster 4 (best) shows 25.0% emphasis while Cluster 2 (worst) shows only 6.0%. A 4x difference.
Topic 11: Cluster 2 spent 32.1% on this topic. Cluster 4 used it only 5%. A 6x difference in opposite direction.
Teams with communication patterns mathematically similar to high performers achieved similar results. The correlation between pattern distance and performance reached r = -0.907 (p = 0.013).
This enables real-time assessment: measure a new team's topic distribution, calculate its distance from proven successful patterns, predict likely performance. The closer the match, the better the predicted outcome.
TRIBE was also applied to a Minecraft search and rescue scenario, demonstrating the system's versatility across different collaborative contexts.
This study involved 222 three-person teams performing urban search-and-rescue missions. Teams had to coordinate to locate and rescue victims while managing hazards. Unlike the defusal task, teams could communicate continuously throughout the 17-minute trials.
Communication Structure: Teams began with a 2-minute planning phase while locked in the entrance, then had 15 minutes of active gameplay with continuous voice communication.
Performance Metrics: Success was measured by victims saved and their severity levels, creating scores that reflected both speed and prioritization decisions.
While this task showed lower variance explanation, it revealed important behavioral dynamics:
TRIBE accurately found the changes in the behavior since most teams changed behavior after the first 2 minutes of planning, to taking action on the field. While the continuous communication format fostered more fluid behavioral patterns than the planning-only communication format of the defusal task, the strict 17-minute time limit imposed a constraint on behavioral expression that was absent in the defusal task, where unlimited planning sessions were allowed. This highlights how task structure shapes the manifestation of team behaviors and demonstrates TRIBE's ability to capture these dynamics.
Study 3 · 8 intervention conditions · 252 transitions per agent
Study 3 assigned teams to eight intervention conditions: six AI agents, a human advisor, and no advisor. Because TRIBE derives clusters from communication alone and never sees trial score, it measures what an intervention did to a team's behavior without knowing how that agent decided to intervene.
Splitting each trial into 10% segments and treating departure from the initial cluster as the event, median survival is 1.0 window. Most teams abandon their opening communication pattern between the 10% and 20% marks. That matches the task structure, where 2 minutes of planning precede 15 minutes of field execution, so the shift TRIBE detects is planning giving way to action rather than noise.
Each agent's influence is an 8×8 transition matrix over the eight clusters. The chart traces the probability of a team occupying one of the three best-performing clusters across ten segments. Toggle agents to compare trajectories.
Most movement happens by step 2 or 3. No cluster showed absorption above 90% self-transition, so teams stay behaviorally dynamic throughout, and every cluster can reach every other.
Each transition is classified as positive change, negative change, good stable, or bad stable. Net change is positive minus negative.
| Agent | Bad stable | Good stable | Negative | Positive | Net |
|---|---|---|---|---|---|
| USC | 22.5% | 25.0% | 20.0% | 32.5% | +12.5 |
| Human | 15.7% | 28.3% | 25.2% | 30.7% | +5.5 |
| CRA | 15.1% | 20.2% | 32.5% | 32.1% | −0.4 |
| UAZ | 13.8% | 28.9% | 28.0% | 29.4% | −1.4 |
| DOLL | 26.5% | 13.7% | 30.8% | 29.1% | −1.7 |
| CMURI | 15.2% | 37.8% | 24.4% | 22.6% | −1.8 |
| SIFT | 13.1% | 18.8% | 35.0% | 33.1% | −1.9 |
CMURI keeps good teams in place most often at 37.8%, but when its interventions do change behavior the net effect is negative, suggesting successful teams dismiss the advice. USC intervenes far less often, 80 times against CMURI's 217, yet produces more than twice the net positive impact of the human advisor and roughly three times the average rank improvement.
| Agent | χ² | p | Significant |
|---|---|---|---|
| DOLL | 121.64 | <0.0001 | Yes |
| USC | 120.68 | <0.0001 | Yes |
| CRA | 118.69 | <0.0001 | Yes |
| SIFT | 102.88 | 0.0001 | Yes |
| UAZ | 97.41 | 0.0009 | Yes |
| CMURI | 85.92 | 0.0126 | Yes |
| Human | 57.53 | 0.3113 | No |
Every AI agent diverges significantly from the no-advisor condition. The human advisor does not, tracking natural team dynamics instead, though a larger sample could still reveal a difference.
| Agent | Steps from worst to best cluster |
|---|---|
| No-Advisor | 8.1 |
| USC | 11.2 |
| Human | 12.7 |
| CMURI | 20.1 |
Unguided teams reaching the best cluster fastest looks like evidence against intervention. It is not. No-advisor received 16 top-tier trials and none from the worst cluster, while USC received 13 low-tier trials. The imbalance is team composition, not intervention effect. Detecting this by score alone would be circular, since labeling high-scoring teams as stronger and then citing those scores as evidence explains nothing. TRIBE resolves it because clusters come from communication and are independent of score.
Adjusted R² throughout
Adding clusters raises R² mechanically, since each additional cluster costs a regression degree of freedom. Every figure below is adjusted R² so configurations with different k stay comparable. Each stochastic method was refit across 100 random seeds with hyperparameters chosen by unsupervised criteria rather than against trial score.
| Method | k | Adjusted R² |
|---|---|---|
| GMM | 16 | 25.4% |
| K-means | 12 | 21.8% |
| BGMM | 10 | 20.1% |
| Hierarchical | 3 | 18.3% |
| DTW | 6–20 | 14.7–15.1% |
Internal metrics (BIC, silhouette, gap statistic) all pointed to k of 3 or 4. External validation against performance pointed the opposite way, with k of 9 to 16 predicting better. A permutation test at the selected configuration shuffled trial scores 1,000 times and never exceeded 0.152, mean 0.028, against the observed 0.254 (p < 0.001).
| Pipeline | Adjusted R² | Runtime |
|---|---|---|
| TF-IDF (SVD 12 dims, k=8) | 6.7% | — |
| Baseline TRIBE: LDA(12) + K-means(8) | 18.5% | 120 min |
| Optimal TRIBE: NMF(6) + GMM(16) | 25.4% | 5 min |
Replacing LDA with NMF and k-means with GMM improves prediction by 37.3% and runs 24 times faster. TF-IDF trailing the baseline confirms topic structure carries signal beyond raw lexical features. All temporal analysis above uses LDA and k-means so intervention comparisons stay on one footing.
Llama 3.3 70B was asked to extract behavioral patterns from the same communications under three prompting regimes. Free labeling produced generic labels, all positive regardless of performance. Domain-informed prompting still labeled the worst cluster "Good Communication". Only after being handed cluster score means and rankings did it produce partly relevant labels, and even then Cluster 4 came back as "Rank-4 Team".
| Cl. | Free labeling | Domain informed | Performance informed |
|---|---|---|---|
| 1 | Specialized Coordination | Strong Collaboration | Strategic Execution |
| 2 | Highly Coordinated | High Quality | Effective Coordination |
| 3 | Efficient Rescue | Outstanding Performance | Focused Approach |
| 4 | Dynamic Teamwork | Exceptional Coordination | Rank-4 Team |
| 5 | Strategic Prioritization | Above Average | Limited Progress |
| 6 | Role Based Comm. | Excellent Teamwork | Inefficient Response |
| 7 | Real time Collaboration | Very Effective | Disorganized Rescue |
| 8 | Task Oriented Dialogue | Good Communication | Chaotic Efforts |
The claim is deliberately narrow. This is one open-weight model under three prompting regimes, not a verdict on LLMs generally. For real-time, on-device deployment inside an agent, a lightweight statistical pipeline extracted performance-predictive structure that this configuration did not, despite vastly greater scale.
Four datasets · identical LDA + k-means pipeline
| Domain | ANOVA F | p | Adjusted R² | Best−worst gap | Cohen's d |
|---|---|---|---|---|---|
| Study 4 (search and defusal) | 11.95 | <0.001 | 44.3% | 610 pts | 1.82 |
| Study 3 (search and rescue) | 8.27 | <0.001 | 18.5% | 214 pts | 0.94 |
| DELI (Wason selection) | 4.10 | <0.001 | 14.8% | 28.3% | 0.51 |
| GAP (survival ranking) | 1.42 | 0.253 | 28.6% | 23 pts | 0.68 |
The spread is the finding. TRIBE's predictive strength tracks how much behavioral freedom a task allows. Study 4 lets teams return to the shop and re-plan without limit, so behavior surfaces and adjusted R² reaches 44.3%, with 97% of worst-cluster teams below the best cluster's average. Study 3's 17-minute cap on planning and execution together compresses that room. DELI has one logically deducible answer, bounding behavioral variation and explaining the lowest figure. GAP shows a promising 28.6% but fails significance at 28 discussions, which is small-sample inflation rather than a result.
TRIBE's domain-agnostic architecture means the same system that analyzes military planning sessions can assess corporate strategy meetings or emergency response coordination. The specific topics and patterns differ, but the underlying method remains constant: discover what teams discuss, measure how they discuss it, correlate patterns with outcomes, predict future performance.
The analyses demonstrate TRIBE's core capabilities and suggest broader applications.
Automatic Discovery: TRIBE found the critical patterns without being programmed to look for them. The system wasn't told that role discussion hurts performance or that completion focus helps. It discovered these relationships from data alone.
Dimensional Simplification: From 13 topics and countless communication patterns, TRIBE identified critical axes of variation. This reduction from high-dimensional data to actionable insight is what makes the system practical.
Quantitative Precision: The variance explanations aren't estimates—they're mathematical relationships. Given a team's behavioral scores, we can predict performance within calculable error bounds.
Domain Transfer: While different tasks show different patterns, the same architecture applies anywhere teams communicate and performance can be measured. Different domains will reveal different critical dimensions, but the method remains constant.
For more detailed technical information about TRIBE and its applications, please refer to the following publications:
1. Jalal-Kamali, A., Gurney, N.M. and Pynadath, D.V., 2025, May. Predicting Team Performance from Communications in Simulated Search-and-Rescue. In Proceedings of the 24th International Conference on Autonomous Agents and Multiagent Systems (pp. 2565-2567).
View Publication →
2. Jalal-Kamali, A., 2024. Team-based Raw Interactions to Behavioral Ensembles (TRIBE): Predicting Team Performance from Communication Patterns (Doctoral dissertation). University of Southern California.
View Dissertation →