CE-CM: Capability Estimation for Task-Agnostic Adaptation in Ad-Hoc Teamwork

Forum topic · 小凯 · 2026-07-31

Summary

This post summarizes an arXiv paper (2607.27177) by Peter Tisnikar and colleagues on ad-hoc teamwork (AHT) with hidden partner capabilities. Most AHT methods assume a single fixed task and known partner capabilities; the authors instead reframe AHT as joint planning with decentralised execution across multiple tasks under hidden capabilities. They propose CE-CM (Capability Estimation via Contextual Models), an approximate Bayesian method that infers task-invariant capability vectors using simulation-based sampling to induce a contextual multi-agent Markov decision process for planning, requiring no population pretraining and refining beliefs online over a few tasks. To handle human unpredictability, CE-CM-Div evaluates capability hypotheses against diverse planner rollouts rather than a single optimal trajectory. Simulations show CE-CM quickly recovers hidden capabilities, reduces infeasible action assignments, and adapts to time-varying capabilities. In an offline human study with 225 trajectories from 15 participants, CE-CM-Div significantly outperforms a baseline CE-M method, suggesting capability-based modeling is a promising interpretable, task-agnostic representation for robust human-AI teaming.

Partner Capability Estimation for Task-Agnostic Adaptation in Ad-Hoc Teamwork

Field: Machine Learning Authors: Peter Tisnikar, Maja Swieczkowska, Benteng Ma, Gerard Canal, Matteo Leonetti arXiv: 2607.27177

Overview

Effective collaboration with novel and diverse partners is a crucial skill for autonomous agents. Most current ad-hoc teamwork (AHT) approaches assume that agents will collaborate on a single, fixed task and that the partner's capabilities — their ability to successfully execute the desired action — are already known. In reality, a partner's true capabilities are often hidden, and human collaborators may act sub-optimally on tasks with multiple valid strategies.

To address these limitations, the authors extend ad-hoc teamwork into a multi-task setting by re-framing it as a problem of joint planning with decentralised execution under hidden partner capabilities.

Method

They introduce CE-CM (Capability Estimation via Contextual Models), an approximate Bayesian method that infers a task-invariant capability vector. Through simulation-based sampling, the agent estimates capabilities and induces a contextual multi-agent Markov decision process for planning. The method requires no population pretraining and refines beliefs online across only a few tasks.

To handle human unpredictability, they propose CE-CM-Div, an extension that evaluates capability hypotheses against diverse planner rollouts rather than a single optimal trajectory.

Results

In simulation, CE-CM rapidly recovers hidden capabilities, reduces infeasible action assignments, and adapts to time-varying capabilities.
In an offline human study with 225 trajectories from 15 participants, CE-CM-Div significantly outperforms the baseline CE-M method.

Conclusion

The results indicate that capability-based modeling is a promising interpretable, task-agnostic representation for the studied scenarios, and that accounting for behavioral diversity is essential for robust human-AI collaboration.

--- *Auto-collected on 2026-07-31*