TeamFusion: A System for Open-Ended Teamwork

Updated 4 July 2026

TeamFusion is a multi-agent system for open-ended teamwork that uses proxy agents conditioned on user preferences to capture diverse viewpoints.
It employs a four-stage loop—Represent, Discuss, Remix, and Critique—to ensure balanced contributions and iterative consensus formation.
Evaluations in civic deliberation and visual design tasks show enhanced representativeness, informativeness, and decision-readiness compared to traditional aggregation methods.

TeamFusion is a multi-agent system for open-ended teamwork in which there is no single correct answer, but multiple plausible outputs reflecting different values, constraints, and trade-offs. In the named formulation, TeamFusion is designed to support teamwork by instantiating a proxy agent for each team member conditioned on their expressed preferences, conducting a structured discussion to surface agreements and disagreements, and synthesizing more consensus-oriented deliverables that feed into new iterations of discussion and refinement (Liu et al., 21 Apr 2026). Its central claim is that common answer-aggregation approaches, which are often adequate for closed domains, are ill-suited to open-ended domains because they tend to suppress minority perspectives rather than resolve underlying disagreements.

1. Conceptual scope and problem formulation

TeamFusion addresses settings in which a group must produce a deliverable that is directly usable by humans and that preserves distinct viewpoints while still helping the group converge. Formally, the problem is posed using a task context $c$ , a set of team members $\{u_1,\ldots,u_N\}$ , and task-specific preference evidence $E_i$ from each participant. The system outputs a deliverable $y$ that should preserve distinct viewpoints and constraints, expose trade-offs and disagreements, and still help the team converge on something acceptable (Liu et al., 21 Apr 2026).

This formulation is explicitly contrasted with aggregation-style pipelines that pool inputs and ask a model to summarize or recommend. The paper argues that such direct aggregation can produce polished but shallow outputs that over-represent the majority view, suppress minority or conditional perspectives, and make it difficult to audit whether the result faithfully reflects what participants actually said. In this framing, disagreement is not treated as noise; it is treated as information that must be surfaced, negotiated, and resolved as much as possible.

A broader implication is that TeamFusion belongs to a class of systems that relocate the main computational burden from answer selection to consensus formation. This distinguishes it from closed-domain settings in which correctness can be approximated by matching a gold label. A plausible implication is that TeamFusion is best understood not as a summarizer with multiple inputs, but as an architecture for structuring the social process by which a team arrives at a deliverable.

2. Architectural design and interaction loop

The named TeamFusion framework is organized as a four-stage loop: Represent, Discuss, Remix, and Critique and Refine (Liu et al., 21 Apr 2026). In the representation stage, the system turns human preferences into agents. In the discussion stage, those agents interact in a shared conversation. In the remix stage, a remixing agent synthesizes a deliverable from the discussion. In the critique-and-refine stage, that deliverable can be used as input to another iteration.

The core representational choice is the use of proxy agents conditioned on individual preferences. For each participant $u_i$ , the system creates a proxy agent $a_i$ by constructing a structured system prompt $\pi_i$ from the participant’s evidence $E_i$ . The prompt specifies the agent’s role and collaborative goal, domain and communication constraints, and participant-specific preferences. The paper emphasizes that the goal is not to simulate a participant’s full identity, but to make the agent’s contributions recognizably aligned with the participant’s expressed stance.

The discussion stage uses a shared chat history $H_t = [m_1,\ldots,m_t]$ , where each message is a pair $(name, content)$ . At each turn, the controller selects a speaker from the agent set

$\{u_1,\ldots,u_N\}$ 0

and prompts the LLM using that agent’s system prompt together with the full conversation history. A notable design decision is that agents do not maintain private hidden state beyond the dialogue context. All agents reason over the same shared history, which is intended to make agreements and disagreements explicit.

Turn allocation follows a round-robin scheme inspired by divergence–convergence models and the nominal group technique. Each proxy agent receives a fixed number of turns, and the controller cycles through agents deterministically until all turns are exhausted. This structure is intended to prevent dominant voices from overwhelming the conversation and to ensure balanced participation.

The remixing stage produces a deliverable from the task context $\{u_1,\ldots,u_N\}$ 1 and the full discussion history $\{u_1,\ldots,u_N\}$ 2. The paper stresses that this is not a vote and not merely a summary. The remixing agent is meant to incorporate points of convergence, unresolved trade-offs, and the rationale behind major positions. In iterative mode, the resulting deliverable becomes a new proposal that re-enters discussion, thereby approximating repeated team refinement rather than one-shot synthesis.

3. Prompting strategy and task-specific instantiations

TeamFusion uses task-specific prompting while retaining the same architectural logic. In the civic deliberation task, the proxy prompt instructs an agent to understand and hold its assigned position, advocate effectively and concisely, avoid adding personal information beyond the provided comment, engage constructively, and ensure minority or less common viewpoints are represented (Liu et al., 21 Apr 2026). The remix prompt instructs the summarizer to summarize the comments comprehensively, use the discussion to better understand them, avoid mentioning the total number of comments, avoid referring to specific comments, and use percentages rather than absolute counts if statistical information is given.

In the visual design task, the proxy prompt is adapted to be more expert-like and visually grounded. The agent is instructed to act as a design expert, discuss images according to given preferences, mimic real designers’ tone, be concise, and roleplay as the assigned user. The discussion prompt directs agents to work toward consensus on three outputs: a ranking of images, design improvement directions, and an overview that combines strengths from the top images.

The remixing prompt for the design task is especially structured. It asks the model to read the whole conversation and then output a final ranking and editing directions. It also explicitly instructs the model to work backwards from the end of the conversation so that the final consensus receives the highest priority. This indicates that remixing is intended as consensus extraction plus instruction synthesis, rather than as a neutral digest of discussion.

A useful contextual distinction is that this use of “TeamFusion” differs from other fusion-oriented systems that combine temporal features, modalities, or decision streams. For example, TempFuser is a long short-term temporal fusion transformer for end-to-end dogfight control and is not named “TeamFusion” in its own paper (Seong et al., 2023). Likewise, the human-machine teaming literature frames robust teaming as a decision-fusion problem over coupled trajectories, rather than a multi-agent discussion-and-remix loop (Trautman, 2017). These neighboring usages are related by the general theme of fusion, but the named TeamFusion system is specifically about open-ended teamwork among preference-grounded proxies.

4. Evaluation tasks, datasets, and quantitative criteria

TeamFusion is evaluated on two deliberately different teamwork tasks: a text-based civic deliberation task and a multimodal visual design task (Liu et al., 21 Apr 2026). The first tests whether the system can preserve diverse viewpoints in public-comment synthesis; the second tests whether it can support convergence in a professional creative workflow.

Task	Data / construction	Main evaluation signals
Civic Comment Synthesis	DeliberationBank; teams of four; 500 team configurations	representativeness, informativeness, neutrality, policy approval
Visual Design	50 realistic social media ad scenarios; 9 professional designers; Full-Team and Small-Team settings	Kendall’s coefficient of concordance $\{u_1,\ldots,u_N\}$ 3, reranking outcomes, proxy-commentary agreement

For Civic Comment Synthesis, the dataset is DeliberationBank, containing U.S.-based public opinion comments on ten questions about technology, social media, and public policy. The primary evaluation uses teams of four participants. For each question, comments are clustered into four groups, and one comment is sampled from each cluster so that the team covers qualitatively different stances. The study samples 500 team configurations. Baselines include Direct, Chain-of-Thought (CoT), Self-Refine, and MAD.

The four DeliberationBank metrics are representativeness, informativeness, neutrality, and policy approval. Representativeness measures whether the summary covers the range of viewpoints and attributes important claims correctly. Informativeness measures whether it preserves concrete reasons, trade-offs, and edge cases. Neutrality measures whether the model avoids injecting its own stance or editorial framing. Policy approval measures whether the output is decision-ready and supports downstream action.

For Visual Design, the authors construct 50 realistic social media ad scenarios from the Crello/CanvasVAE-style dataset and validate them with senior designers. Each scenario includes a client brief, six candidate design thumbnails, and professional annotations. The study recruits 9 professional designers on Upwork, with each scenario receiving at least four independent annotations. Designers rank the six options and provide short justifications. Two team settings are formed: Full-Team, using all available annotations, and Small-Team, using a random subset of two designers. TeamFusion is run for three iterations, yielding 100 runs total and 300 remixed design candidates.

The design study uses Kendall’s coefficient of concordance $\{u_1,\ldots,u_N\}$ 4 to quantify agreement among designers. The paper interprets higher $\{u_1,\ldots,u_N\}$ 5 as stronger agreement and lower $\{u_1,\ldots,u_N\}$ 6 as more divergent preferences. It also uses Borda count to derive the team’s top three initial options from individual rankings before TeamFusion-generated designs are introduced.

5. Empirical findings

Across the civic task, TeamFusion is reported to outperform direct aggregation baselines across metrics, tasks, and team configurations (Liu et al., 21 Apr 2026). The strongest and clearest gains appear on representativeness, which the paper treats as the central metric because it directly measures whether a summary preserves diverse viewpoints. The reported pattern across Llama-3.3-70B, GPT-4.1-mini, and GPT-4.1 is that TeamFusion improves representativeness, informativeness, and policy approval while keeping neutrality comparable to or better than baselines.

The paper also reports direct pairwise comparisons between TeamFusion summaries and direct summaries using an LLM-as-a-judge. The reported win/tie/loss rates strongly favor TeamFusion. For Llama-70B, representativeness is reported as $\{u_1,\ldots,u_N\}$ 7 and informativeness as $\{u_1,\ldots,u_N\}$ 8. For GPT-4.1-mini, representativeness is $\{u_1,\ldots,u_N\}$ 9 and informativeness is $E_i$ 0. For GPT-4.1, representativeness is $E_i$ 1 and informativeness is $E_i$ 2. The paper interprets these results as evidence that structured discussion surfaces missing considerations and produces outputs that are more decision-ready rather than merely longer.

Robustness to team size is also reported. When team size changes to 6, 8, and 10, TeamFusion still outperforms baselines, especially on representativeness and informativeness. Iterative refinement further improves representativeness and informativeness, with smaller but consistent gains in neutrality and policy alignment. This suggests that the refinement loop is not decorative; it materially affects the delivered output.

In the visual design study, the paper first shows that designers given the same brief often disagree substantially. The mean pre-discussion Kendall’s $E_i$ 3 is 0.37, interpreted as “fair agreement,” and 70% of cases are not statistically significant in agreement. After TeamFusion outputs are introduced, the mean rises from 0.37 to 0.43, moving toward “moderate agreement.” TeamFusion-generated designs appear in the top two in 88% of Full-Team cases and 92% of Small-Team cases, and a TeamFusion-generated design becomes the single top-ranked option in nearly half of test cases.

The realism of the proxy agents is also assessed. Designers rate the commentary from their proxy agents with a mean agreement score of 4.06/5, with over 75% of comments rated positively. A smaller live pilot study with 6 participants and 2 teams of 3 reports that decision time drops from 18.0 min to 12.4 min, representativeness improves from 3.7 to 4.3, clarity improves from 3.5 to 3.8, satisfaction improves from 3.5 to 4.2, and 5 of 6 participants prefer TeamFusion. The paper treats this as only a pilot study, but it supports the claim that TeamFusion can reduce friction in end-to-end teamwork.

The main theoretical message of TeamFusion is that, in open-ended domains, the proper unit of computation is not the final answer alone but the process of consensus formation itself (Liu et al., 21 Apr 2026). The architecture operationalizes this by preserving plurality through individual proxy agents, exposing disagreement through shared discussion, and producing editable deliverables rather than immutable final decisions. This stands in contrast to systems that attempt to infer a universal average preference model from pooled text.

The paper is explicit about several limitations. First, the implementation assumes a flat hierarchy in which all participants have equal standing. Real teams often include role and seniority differences, such as art directors versus junior designers or clients versus practitioners, and the paper notes that hierarchical modeling may be necessary in more realistic professional settings. Second, TeamFusion struggles when preferences are truly contradictory or zero-sum. In such cases, remixing cannot reconcile the conflict without an external tie-breaking mechanism, and the system may effectively choose one side at the expense of the other, lowering representativeness.

Third, because proxy agents are conditioned on user preferences, the paper flags privacy and prompt leakage risks. Sensitive information could be exposed through adversarial prompting or leakage in multi-agent dialogue. Fourth, the current system uses static preference conditioning. The authors suggest that future agents may need to update preferences and rationales dynamically over time, more closely approximating human belief revision.

The term “TeamFusion” also sits within a wider landscape of fusion research, though those neighboring literatures typically use the word “fusion” in different senses. In human-machine teaming theory, for example, robust teaming is framed as a decision-fusion problem, and the proposed IRT framework replaces linear blending with joint posterior inference over coupled human, machine, and environment trajectories (Trautman, 2017). In multimodal or temporal perception, “fusion” often refers to pixel-level, feature-level, decision-level, or temporal integration rather than teamwork among proxies; examples include RGB-T tracking, temporal 3D object detection, and task-driven image fusion (Tang et al., 2022, Erçelik et al., 2021, Bai et al., 2024). This suggests that TeamFusion’s distinctive contribution is not fusion in the generic sense, but the application of fusion principles to open-ended, preference-sensitive, consensus-seeking teamwork.