Test-Time Discovery (TTD) addresses a practical problem: after a model is deployed, it must handle a continuous stream of unlabeled test data, recognizing classes it was trained on while also flagging anything new. The challenge is to do this without retraining or access to labeled examples from the test distribution.

According to the abstract, existing TTD methods typically depend on visual clustering or prototype comparison. The new paper proposes "relational semantic grounding" as a different strategy, presumably using relationships between semantic concepts to guide discovery. However, the abstract is truncated, so the exact mechanism and experimental results are not described in the available text.

What is clear is the intended contribution: moving beyond purely visual cues toward a semantically grounded approach for distinguishing known from novel classes in an open-ended test stream.