[KongchangAI]
BenchmarkAnyGoal基准

AnyGoal

A benchmark and open-source library for evaluating the compositional generalization ability of world models. By constructing goal-shifted task variants, it tests whether a model can recombine state factors and actions seen during training to complete new planning tasks.

Timeline (last 90 days)

Oct 20

AnyGoal is a benchmark that studies the generalization capability of world models from a compositional generalization perspective.

Unverified50%
Oct 20

AnyGoal constructs "goal-shift" task variants built on widely used world model environments.

Unverified50%
Oct 20

In AnyGoal tasks, individual state factors and actions remain within the training support, but their combinations rarely appear in expert trajectories.

Unverified50%
Oct 20

AnyGoal is a public benchmark and open-source library.

Unverified50%
Oct 20

The researchers analyzed dynamic failures and representation failures, involving different interaction mechanisms and evaluation metrics.

Unverified50%

All Facts (5)

Source Articles