From what constitutes collaboration, to how to represent it structurally, whether predefined or emergent, to what that structure makes computationally possible, and at what cost. Collaboration is itself agnostic to team members and team size; my current work focuses on its realization among LLM agents.
Definition.What constitutes collaboration among intelligent entities, and which primitives distinguish it from mere group behavior?
Collaboration can be a property of the task itself: in CUBE, some goals are individually infeasible. COOP2 turns that into a formal requirement, with four categories of constraints that guard cooperative progress. TEA defines the collaboration policy, separating the behavior it prescribes from the behavior that emerges. Earlier steps came from the LLM-based digital twin, decentralized generative agents, and the Five Ws survey.
Interdependence and collective capability are candidate primitives; which others matter remains open.
Representation.What representational structure can capture collaboration, whether predefined or emergent, across tasks, teams, and scales—and under what assumptions?
Collaboration can be recorded without assuming how agents organize. DIG captures it as event passing between activations, and TEA adds the tasks agents create and revise and the shared environment they act in. COOP2 and DR. WELL instead couple symbolic reasoning with grounded transitions through roles, plans, and a shared world model, and iCORE links observed cooperation to evolving obligations and the evidence connecting them.
Each exposes a tradeoff among expressiveness, generalizability, and scalability; their balance remains open.
Computation.What can be derived, evaluated, and optimized once collaboration is represented, and at what computational cost?
A representation makes visible and repairable the failures that final task success hides. DIG exposes reachability and progress failures and heals them online through information injection and rerouting, and COOP2 uses predicted constraint deficits to open targeted communication. Representation also supports guarantees and improvement: iCORE formalizes work soundness and assignment stability, yielding a conditional performance bound, and TEA recovers the realized topology and candidate rules from the record, so Infuse can decide what to keep, add, revise, or leave open and propose an updated policy.
The open question is which claims and interventions stay tractable as tasks, teams, and representations grow. Where they fail, they expose what a representation is missing and feed back into the definition.
