Test‑Time Discovery (TTD) aims to recognize known classes and detect novel ones while a deployed model processes a continuous unlabeled test stream. Existing TTD approaches are overly conservative, causing known and unknown features to collapse together during test‑time adaptation, which makes separation difficult. By revisiting TTD from an energy perspective we observe that the natural energy gap $ΔE$ between known and unknown classes is unstable and tends to vanish. To address this, we introduce Energy Ring, which anchors the two groups on separate energy rings to preserve the gap. The method consists of:
- a lightweight adapter that enlarges $ΔE$ via norm‑based scaling while keeping the direction of known‑class features;
- a perturbation consistency module that applies small perturbations to unknown samples, forming compact clusters in feature space;
- energy‑guided confidence weighting during inference that adjusts prediction confidence based on energy values;
- a stability‑based rectification mechanism that detects and corrects unreliable updates.
Extensive experiments on several public benchmarks show that Energy Ring significantly outperforms existing TTD baselines in accuracy, recall, and F1 score.
Review