Thomas Manuel Rost

Thomas Manuel Rost

Automated species classification from underwater imagery is bottlenecked by the cost of expert annotation, and supervised models trained on one dataset rarely transfer to new conditions. We investigate whether semi-supervised methods operating on frozen foundation model embeddings can close this annotation gap with minimal labeling effort. Using DINOv3 ViT-B embeddings with no fine-tuning, we propagate a small set of labeled seeds through unlabeled data via nearest-neighbor-based self-training and evaluate on the AQUA20 benchmark (20 marine species). With fewer than 5% of the training labels, self-training on frozen embeddings closes much of the gap to a fully supervised ConvNeXt baseline trained on the entire labeled dataset; at full supervision, the gap narrows to a few percentage points, with several species exceeding the supervised baseline. Class separability in the embedding space, measured by ROC-AUC, is high even at extreme label scarcity, indicating that the frozen representations capture discriminative structure well before decision boundaries can be reliably estimated. Our approach requires no training, no domain-specific data engineering, and no underwater-adapted models, establishing a practical, immediately deployable baseline for label-efficient marine species recognition. All results are reported on the held-out test set over 100 random seed initializations.

Thomas Manuel Rost

Automated underwater species classification is constrained by annotation cost and environmental variation that limits the transferability of fully supervised models. Recent work has shown that frozen embeddings from self-supervised vision foundation models already provide a strong label-efficient baseline for marine image classification. Here we investigate whether this frozen-embedding regime can be improved at inference time, without fine-tuning or changing model weights. We apply Circuit Duplication, an inference-time method originally proposed for Large Language Models, in which a selected range of transformer layers is traversed twice during the forward pass. We evaluate on the class-imbalanced AQUA20 benchmark using frozen DINOv3 embeddings under two settings: global circuit selection, where a single duplicated circuit is chosen for the full dataset, and class-specific circuit selection, where each species may receive a different optimal circuit. Both settings use simple semi-supervised downstream classifiers. Circuit Duplication consistently improves over the standard frozen forward pass. At the maximum label budget, class-specific selection reaches a macro F1 of 0.875, closing the gap to the fully supervised ConvNeXt benchmark (0.889) to 1.4 points without any gradient-based training. Four species exceed their fully supervised reference, with octopus improving by +12.1 F1 points. Across all budgets, roughly 75% of classes prefer a class-specific circuit, indicating a genuinely class-dependent benefit. To our knowledge, this is the first application of Circuit Duplication to computer vision.

Thomas Manuel Rost, Martina Figlia, Bernd Wallraff

We introduce and evaluate Stated Preference for Interaction and Continued Engagement (SPICE), a simple diagnostic signal elicited by asking a Large Language Model a YES or NO question about its willingness to re-engage with a user's behavior after reviewing a short transcript. In a study using a 3-tone (friendly, unclear, abusive) by 10-interaction stimulus set, we tested four open-weight chat models across four framing conditions, resulting in 480 trials. Our findings show that SPICE sharply discriminates by user tone. Friendly interactions yielded a near-unanimous preference to continue (97.5% YES), while abusive interactions yielded a strong preference to discontinue (17.9% YES), with unclear interactions falling in between (60.4% YES). This core association remains decisive under multiple dependence-aware statistical tests, including Rao-Scott adjustment and cluster permutation tests. Furthermore, we demonstrate that SPICE provides a distinct signal from abuse classification. In trials where a model failed to identify abuse, it still overwhelmingly stated a preference not to continue the interaction (81% of the time). An exploratory analysis also reveals a significant interaction effect: a preamble describing the study context significantly impacts SPICE under ambiguity, but only when transcripts are presented as a single block of text rather than a multi-turn chat. The results validate SPICE as a robust, low-overhead, and reproducible tool for auditing model dispositions, complementing existing metrics by offering a direct, relational signal of a model's state. All stimuli, code, and analysis scripts are released to support replication.