Key Takeaways 

• TeamFusion is a multi-agent framework that supports team decision-making by creating AI proxy agents grounded in each member’s preferences, so that humans can make better-informed decisions together. 

• Across evaluations in civic decision-making and visual design, TeamFusion improves representativeness and informativeness compared to direct aggregation approaches. 

• In a study with professional designers, over 75% of proxy agent commentary was rated as representative of participants’ reasoning, and TeamFusion-generated designs became the team’s top-ranked option in 42–50% of cases. 

• In a small pilot study, teams using TeamFusion reached decisions faster while reporting higher satisfaction and perceived representativeness (descriptive results from a small-scale pilot). 

Research by Jiale Liu (Penn State), Victor S. Bursztyn (Adobe Research), Lin Ai (Columbia University), Haoliang Wang (Adobe Research), Sunav Choudhary (Adobe Research), Saayan Mitra (Adobe Research), and Qingyun Wu (Penn State / AG2ai) | Research to be presented at ACL 2026 

📄 Read the full paper on arXiv. 

This work was conducted in part during an internship collaboration with Adobe Research. The system described here is experimental and does not represent a current Adobe product feature. 

When teams make decisions together, disagreement carries information. It reflects different interpretations of the same problem, different priorities, and different definitions of success. 

Most AI tools do not handle this well. They tend to compress multiple viewpoints into a single summary, smoothing over differences in ways that can make the result easier to read but less useful for decision-making. Important tensions disappear, along with the reasoning behind them. 

In our study, this challenge appeared clearly. When professional designers were given the same creative brief, their rankings were indistinguishable from random chance in 70% of cases under standard agreement metrics. This pattern points to something fundamental: divergence is common in open-ended work, not an exception to be corrected. 

TeamFusion, a new experimental technology to be published at ACL 2026, is built to work with this reality. Instead of collapsing perspectives, it models the process through which teams surface disagreements, examine trade-offs, and move toward outcomes they can support. 

How TeamFusion works 

TeamFusion creates a proxy agent for each team member, grounded in that person’s expressed preferences, role, and communication style. These agents act as stand-ins, articulating and defending individual viewpoints during a structured discussion. 

For example, in a design task, one agent might prioritize brand consistency, another visual novelty, and another accessibility. Rather than blending these perspectives prematurely, TeamFusion lets them interact. 

Agents take turns in a structured dialogue, ensuring each perspective is heard and reducing dominant-voice effects common in both human meetings and AI-generated debates. After discussion, a remixing agent synthesizes the exchange into an editable deliverable that captures key trade-offs and reasoning. This output can then be refined further. 

Evaluating representation in civic decision-making 

We evaluated TeamFusion on the DeliberationBank benchmark, which contains public opinion comments on technology and policy topics. 

Across 500 simulated teams and multiple language models, TeamFusion improved representativeness while also outperforming baseline approaches on informativeness and policy usefulness. These gains held across different team sizes and improved further with iterative refinement. 

Visual design with professional teams 

We also evaluated TeamFusion in a more subjective and creative domain: visual design. 

Across 50 advertising scenarios, nine professional designers provided rankings and justifications. Over 75% of agent-generated comments were rated as representative of the designers’ own reasoning, with a mean score of 4.06 out of 5. 

This sense of representation translated into measurable alignment. A standard agreement metric (Kendall’s W) increased from 0.37 to 0.43 after using TeamFusion. TeamFusion-generated designs were selected as the team’s top choice in 42–50% of cases because they reflected perspectives the team itself preferred. 

Why this matters 

Many real-world decisions require reconciling fundamentally different viewpoints. 

TeamFusion points toward collaborative systems that make discussion more structured, visible, and inclusive. By preserving disagreement rather than flattening it, such systems can help teams reach decisions that better reflect their members. 

Design principles and limitations 

TeamFusion is designed to support human teams. It keeps human judgment central by making intermediate reasoning transparent and allowing outputs to be reviewed, edited, and refined. 

The system has limitations. It assumes a flat team structure and may struggle in zero-sum conflicts where preferences cannot be reconciled. Proxy agents approximate participant perspectives based on available inputs and may not capture all nuances of human judgment. 

Ethics and responsible use 

TeamFusion is designed to support creative teams in their decision-making, keeping human judgment at the center of how the system is used. Human participants remain the decision-makers, determining whether and how outputs are applied. 

To support transparency and accountability, TeamFusion preserves the provenance of its outputs, including the prompts and multi-agent discussions that produced them. This allows teams to review, question, and adapt results while maintaining human-in-the-loop control. This work reflects Adobe’s broader commitment to developing AI systems that augment human creativity and decision-making while remaining transparent and controllable. 

The full paper is available on arXiv and will be presented at ACL 2026.