Shared Workspace Human-AI Collaboration: Synergy vs Overhead

The paper "Searching for Synergy in Shared Workspace Human-AI Collaboration," by Nachiket Kotalwar, Rohini Das, and Carolyn Rose, submitted 16 June 2026, finds that adding simulated human collaborators can lower team performance unless teams have explicit coordination scaffolds. The authors ran 1,482 sessions in the Collaborative Gym environment on DiscoveryBench tasks and report that a scaffolding design combining shared group memory with simulated human-in-the-loop approval gates raised mean performance, most clearly in three-person teams.

How did the authors test shared-workspace human-AI teams?

The authors evaluated teams in the Collaborative Gym environment using DiscoveryBench tasks across 1,482 sessions, varying team size and the presence of simulated human collaborators. They used simulated participants to represent additional human collaborators and measured how team performance changed as collaborators were added and as coordination scaffolds were introduced. The paper is 13 pages long and contains five figures and three tables documenting experimental conditions and outcomes.

The experimental setup isolates two failure modes. First, adding relevant collaborators sometimes lowered performance because teams lacked structure to route expertise and assign responsibility. Second, the authors tested scaffolding interventions intended to reduce this process loss. The primary manipulation paired shared group memory with human-in-the-loop gates that required approval from a designated simulated participant before selected actions were executed.

What scaffolding improved outcomes?

Scaffolding that combined shared group memory with simulated human-in-the-loop gates increased mean performance, with the clearest gains in three-person teams. The authors report that this combination produced clearer responsibility signals and stronger routing of expertise to team actions, which reduced coordination overhead that otherwise offset the value of additional collaborators.

In practice the scaffold required a designated simulated participant to approve certain actions, and the shared memory tracked group state to make contributions and approvals visible. The paper frames this as a solution to two problems: determining who should act and ensuring the team’s expertise reaches the right decision points. The measured improvements appear most pronounced when the team size increased to three participants, where unstructured addition of collaborators otherwise created the largest process losses.

Why it matters

The study shows that capability alone does not guarantee better human-AI team outcomes. Even when collaborators are relevant, teams can lose performance to coordination friction. The result shifts attention from focusing solely on agent capability to designing interaction structure, such as shared memory and selective approval, that routes expertise and signals responsibility.

This matters for researchers building multiagent or human-AI systems, and for practitioners deploying collaborative tools: without lightweight scaffolds, adding more agents or people can become a liability rather than an asset.

What to watch

The paper was accepted at the ICML 2026 Workshop on Human-AI Co-Creativity, where more detailed results and discussion of the five figures and three tables should appear. Future confirmations to watch for include replication of these scaffolding gains in real human-in-the-loop studies and evaluations across other task suites beyond DiscoveryBench.

Bibliographic note: the arXiv submission id is arXiv:2606.18413, version v1, and the submission date is 16 June 2026.