The Big Picture
Estimating a partner’s stable capabilities from a few joint tasks lets an AI avoid assigning infeasible work and reuse that knowledge across new tasks — but you must handle diverse, suboptimal human behaviour for robust results.
ON THIS PAGE
Key Findings
Learning an explicit, task‑independent model of what a partner can and cannot do (their capabilities) helps an agent adapt across different tasks and reduces infeasible action assignments. The capability estimator (CE-CM) recovers discrete capability vectors from only a small number of interactions and improves coordination in well-specified domains. When partners behave suboptimally or multiple strategies are possible, evaluating capability guesses against many plausible behaviours (many plausible behaviours) is needed to avoid brittle inferences. Capabilities constrain feasible plans but do not fully predict which plan a partner will choose when preferences or conventions matter. capability vectors
Test your agentsValidate against real scenarios
By the Numbers
1TidyUP used a 20-dimensional binary capability vector to represent what a partner can do, enabling fine-grained inference of limits.
2Collected 225 human gameplay trajectories from 15 participants in Overcooked, showing substantial divergence from planner-generated single-rollout predictions.
3Evaluation used sequences of 5 observed tasks (and 9 distinct Overcooked tasks from 3 layouts × 3 recipes); CE-CM infers useful capability estimates from only a few tasks in these settings.
Why It Matters
Engineers building collaborative agents who must work with unknown humans or other agents will gain an interpretable interpretable capability models to learn partner limits and avoid giving impossible tasks. Technical leaders evaluating agent reliability and researchers studying generalisation in teamwork can use capability models to transfer partner knowledge across tasks and spot when behaviour diversity requires richer models.
Key Figures
Fig 1: Figure 1 : Overview of CE-CM. After completing a task with a partner, the agent uses the observed joint trajectory and the task’s known goal to infer what the partner is capable of doing. It samples candidate capability vectors c ^ 1 , … , c ^ N \hat{c}_{1},\dots,\hat{c}_{N} , simulates the joint trajectory that each would induce in the observed task, and compares the simulated trajectories with the observation. Capability hypotheses that explain the observation are used to update the agent’s belief c ∗ c^{*} , producing an explicit partner model ℳ c ∗ ( ⋅ ) \mathcal{M}_{c^{*}}(\cdot) . The model is then used by the agent to compute a joint plan ρ ∗ \rho^{*} for the new task. Whilst the agent computes a joint plan, it only executes its own actions. After the new task is completed, its joint trajectory becomes new evidence, and the inference-planning loop repeats.
Fig 2: Figure 2 : TidyUP domain. Left: one of the possible starting states, where the blanket and a dirty bowl are on the couch, the dirty dish is in the study, and the toothbrush is in the kitchen. Right: possible goal state where the blanket is on the bed, the bowl and the dish are clean and in the kitchen, and the toothbrush is in the bathroom.
Fig 3: (a) High-capability type
Fig 4: (a) High-capability type
Ready to evaluate your AI agents?
Learn how ReputAgent helps teams build trustworthy AI through systematic evaluation.
Learn MoreYes, But...
The method requires the agent to simulate joint plans under candidate capabilities, which increases computational cost and depends on having a planner or model of the task. Capability estimates explain feasibility but do not capture partner preferences or conventions, so improved capability knowledge may not reduce coordination errors when many valid strategies exist. planning Scaling joint planning to teams larger than two agents is hard; practical deployment will need approximations, role factoring, or preference inference added on top of capabilities.
Deep Dive
Representing a teammate by what they are capable of (a task‑invariant binary vector) gives an explicit, reusable model that an AI can update after each joint task. CE-CM (Capability Estimation via Contextual Models) uses a simulate‑and‑compare loop: sample candidate capability vectors, simulate the joint trajectory each would produce under the known task goal, and keep hypotheses whose simulated trajectories match the observed one. The retained samples are aggregated into a posterior-like estimate and converted to a point capability estimate that conditions future joint planning, letting the agent avoid assigning infeasible actions. safer coordination. Human and simulated experiments show where this helps and where it falls short. In a fully specified household domain (TidyUP) with 20 capability bits, CE-CM quickly recovers partner limits and reduces infeasible assignments. In a more ambiguous cooking domain (Overcooked) many joint behaviours achieve similar outcomes, and single-rollout comparisons are brittle: humans often follow diverse, suboptimal strategies. CE-CM-Div extends the method by scoring capability hypotheses against sets of diverse planner rollouts rather than a single prediction, substantially improving inference on human data. The upshot: capability models are an interpretable, task-agnostic tool for transfer and safer coordination, but robust human teaming also requires modelling behavioural diversity and, ideally, partner preferences.
Test your agentsValidate against real scenarios
Credibility Assessment:
Authors affiliated with King's College London (a recognized institution) but individual h-indices are low and it is an arXiv preprint, so reasonably credible but not top-tier.