A deployed classifier that keeps learning from its own predictions has a hidden weak point: the confidence score that decides which predictions to trust is produced by the same model that is being updated. When the incoming stream contains objects from classes the model has never seen, self-training pushes the model to become more confident on everything. That includes those unknown objects, so the gate that was supposed to keep them out slowly opens. We propose to break this loop by splitting the model's output into two separate quantities, what it believes and how much evidence it has for that belief, and to forbid self-training from changing the second. The evidence level then stays an honest gate, because the pseudo-label loss can never optimize it.
Concretely, we use evidential deep learning, where a network outputs non-negative "votes" per class that parameterize a Dirichlet distribution (a distribution over class-probability vectors) instead of a single softmax vector. The total vote count measures how much evidence the model has, and the normalized votes measure which class it favors. Test-time adaptation (updating a model on unlabeled test data as it arrives) is allowed to sharpen the normalized votes of trusted samples, but it is constrained to leave their total evidence unchanged. For samples with little evidence, adaptation removes what evidence they have, pushing them toward the "I don't know" state rather than toward an arbitrary known class. The claim is that this evidence-conservation constraint, not the evidential parameterization alone, is what keeps open-set pseudo-label noise from poisoning a long adaptation stream.
Why the obvious combination is not enough
The research file points to an apparently open gap. Evidential networks separate "no evidence" (vacuity, $u = K/S$ for $K$ classes and total Dirichlet strength $S$) from "conflicting evidence," yet no test-time adaptation method gates its updates on vacuity. Two prior works motivate this. ANEDL uses an evidential head to find outliers in offline open-set semi-supervised learning and reports outlier AUROC of 91.1% against OpenMatch's 86.8% on CIFAR-100. DEAR uses evidential uncertainty to reject unknown video actions, but the model is frozen at deployment. Neither updates online, so neither faces a gate that drifts as the model changes.
We expect the obvious move, plugging vacuity into an entropy-minimizing adapter like TENT, to fail, and the reason is the core of this proposal. Entropy minimization sharpens predictions, and with softplus or exponential evidence the cheapest way to sharpen is to inflate the top class's votes. That raises $S$ on every sample the loss touches, lowers vacuity on unknowns, and admits them into the next update. The file lists evidence-scale drift under online shift as an open problem of evidential learning. Here the drift is caused by the adaptation itself, which makes the gate self-defeating. A second failure runs the other way: covariate corruption lowers evidence for all inputs, including known-class ones, so a fixed vacuity threshold stalls adaptation exactly when it is needed. This is the "transient noise versus persistent shift" problem from the gated-updating literature.
Core Claim
If the self-training loss can change only the shape of the Dirichlet and never its strength, then vacuity remains a gate the adaptation cannot game. Under that constraint, open-set samples can be driven to zero evidence without distorting the class ranking of known samples. This is something softmax cannot do, because in softmax "uncertain" and "flat logits" are the same degree of freedom. We therefore predict that an evidence-conserving adapter keeps both in-distribution accuracy and open-set AUROC stable over long streams. We also predict that a vacuity-gated adapter without conservation shows rising evidence on unknown samples and decaying AUROC over rounds.
Research Program
The source model is trained once with a single evidential head. This avoids the softmax-plus-evidential dual head that ANEDL needs and flags as unresolved. It uses the standard evidential likelihood plus DEAR's evidential uncertainty calibration term, which penalizes being confident and wrong. At test time we track a slow exponential moving average $\bar{\ell}_t$ of the batch-median $\log S$. The average is slow so that uniform corruption shifts the reference rather than closing the gate, while a subset of unknowns within a batch still stands out. Each sample receives a soft gate
$$w(x) = \sigma\big((\log S(x) - \bar{\ell}_t - \delta)/\tau\big),$$
where $\sigma$ is the logistic function, $\delta$ is an offset and $\tau$ a temperature, both fixed on held-out source data. The adaptation objective on a batch $B$ is
$$\mathcal{L} = \sum_{x \in B} w(x)\, H\!\left(\tfrac{\alpha(x)}{S(x)}\right) + \lambda \sum_{x \in B} (1-w(x))\, \mathrm{KL}\big(\mathrm{Dir}(\tilde\alpha(x)) \,\|\, \mathrm{Dir}(\mathbf{1})\big) + \mu \Big(\tfrac{1}{|B|}\sum_{x \in B} w(x)\,[\log S_\theta(x) - \log S_0(x)]\Big)^2 .$$
In this objective, $\alpha$ is the vector of Dirichlet parameters and $H$ is the entropy of the expected class probabilities. $\tilde\alpha$ removes the top-class evidence, as in standard evidential regularization, so the KL term suppresses only non-supporting evidence. $\mathrm{Dir}(\mathbf{1})$ is the flat, all-ignorance Dirichlet. $\log S_0(x)$ is the strength recorded for $x$ at its first, pre-update forward pass, cached from that same pass so that no second network is needed. The last term is the conservation constraint. We will compare this penalty against a hard version that projects each gradient orthogonal to $\nabla_\theta \log S$ on gated-in samples. Updates are restricted to normalization affine parameters, as in TENT, so the parameter budget matches existing adapters.
The analysis side is as important as the method. For every run we log the gate open-rate on true unknowns, the mean $\log S$ of unknowns and of knowns over time, and in-distribution accuracy and AUROC per adaptation round. These logs let us observe the self-opening gate directly, rather than infer it from end metrics.
Baselines
The do-nothing reference is the frozen evidential source model with vacuity rejection, together with its softmax twin. It tells us how much any adaptation helps. The naive adapter is TENT, whose failure is documented. Wisdom of Crowds TTA reports that TENT's error on CIFAR-10-C rises to 85.22% over 50 rounds, while their confidence-difference filter holds it at 15.77%. UniEnt reports that TENT's CIFAR-100-C AUROC is 65.00%.
The strongest competitors differ by metric. On AUROC it is UniEnt and UniEnt+, which gate with a Gaussian mixture fit on prototype cosine similarity and maximize entropy on detected unknowns; they report CIFAR-100-C AUROC of 89.28%. On in-distribution accuracy it is LU-OSTTA, which reports 89.34% accuracy on CIFAR-10-C and 42.78% on Tiny-ImageNet-C. On long-stream stability it is the Wisdom of Crowds filter, with 15.77% error at 50 rounds on CIFAR-10-C. We also run TENT with independent sigmoid heads from A Closer Look at InD vs. OOD Accuracy, for which no reported result was found on Lacuna, and TENT optimized with CAdam, for which no reported result was found on Lacuna either.
For fairness, all methods share the same source backbone per dataset, the same updated parameter set where the method allows it, the same batch size, the same stream order and the same tuning budget of hyperparameter trials chosen on a held-out corruption type. We rerun every baseline rather than copy numbers, and we report cited figures only as sanity checks.
Two ablations isolate the key ingredient. The first is the obvious idea: vacuity-gated entropy minimization with no conservation term. The second is conservation on a softmax model, where logsumexp energy plays the role of $\log S$. Beating the baselines means matching or exceeding UniEnt's AUROC and LU-OSTTA's accuracy in OSCR, the open-set classification rate that combines both, while showing no upward drift in error across the 50-round protocol. Beating the first ablation by a clear margin in late-round AUROC is the evidence that conservation is the operative idea.
Why not just…?
Why not just threshold softmax confidence or entropy? The Wisdom of Crowds paper shows that static MSP and entropy thresholds admit confident open-set objects and let TENT collapse. A softmax cannot express "no evidence" apart from "ambiguous," as the $[-100,-100,-100]$ versus $[10,10,0]$ example in the file illustrates. We include this gate as a baseline, and we expect it to fail on ImageNet-O-C, which was built to trigger high softmax confidence.
Why not just use sigmoid heads? This is a real alternative, because sigmoids also let every class score be low at once. The TMLR study reports that sigmoid heads cost in-distribution accuracy. Our argument is that the Dirichlet keeps a normalized ranking separate from its strength, so suppressing unknowns should not cost that accuracy. If sigmoid TENT with our conservation penalty applied to summed activations matches us, we will report that the Dirichlet adds nothing beyond parameterization.
Why not just gate on energy scores with the energy detached from the loss? This is the most serious threat. Exponential evidence makes $S$ nearly the exponentiated energy, so conservation might work equally well on a softmax model. The second ablation tests exactly this. If it ties, the finding is still useful: conservation is the fix, and evidential training matters only if its calibration loss gives a better-separated source $S$. We would test that second possibility by comparing source-model AUROC between the two parameterizations before any adaptation.
Why not just use the Wisdom of Crowds dual-pass filter? It already stabilizes long streams, but it compares softmax confidence against a frozen copy that may itself be overconfident on unknowns. It also doubles memory and forward passes. Our method caches $\log S_0$ from a single pass. We compare stability at equal wall-clock time and report memory use.
Why not just use a confidence-gated optimizer like CAdam? CAdam withholds updates only when gradient and momentum disagree. Open-set objects that recur across consecutive batches produce persistent, consistent gradients, which its momentum will eventually follow. The OOD-only batch intervals in our protocol test this directly.
Proposed Experiments
Streams follow the TMLR protocol. CIFAR-10-C and CIFAR-100-C are the known-class data, with SVHN-C as far unknowns. Tiny-ImageNet-C is the known-class data following UniEnt and LU-OSTTA. ImageNet-C is the known-class data with ImageNet-O-C and Textures-C as unknowns. Unknown proportions range from 25% to 75%, and OOD-only intervals are included. We add a near-unknown setting: train on 80 CIFAR-100 classes, as in ANEDL's split, and inject the corrupted held-out 20 into the stream. This targets the semantic-proximity case in which prototype mixtures should struggle. Every configuration also runs under the 50-round long-term protocol of the Wisdom of Crowds paper. Metrics are in-distribution accuracy, AUROC, OSCR, per-round curves, and latency per batch.
The decisive check is the time series of mean $\log S$ on true unknowns. Without conservation, we predict it will rise across rounds as late-round AUROC falls. With conservation, it should stay flat or fall. If the unconserved evidential gate shows no such rise, the premise of the proposal is wrong, and the simpler vacuity-gated adapter is the right method to publish. The second check is the near-unknown split. Here we predict the largest margin over UniEnt, because a bimodal similarity mixture has little room to separate near classes, while per-sample evidence relative to a stream reference does not rely on that bimodality.
References
- ANEDL: Adaptive Negative Evidential Deep Learning for Open-Set Semi-supervised Learning
- Evidential Deep Learning for Open Set Action Recognition
- Towards Open-Set Test-Time Adaptation Utilizing the Wisdom of Crowds in Entropy Minimization
- Unified Entropy Optimization for Open-Set Test-Time Adaptation
- Learning from Unknown for Open-Set Test-Time Adaptation
- A Closer Look at In-Distribution vs. Out-of-Distribution Accuracy for Open-Set Test-time Adaptation
- CAdam: Confidence-Based Optimization for Online Learning