Sebastian Cygert

Wojciech Zarzecki, Jan Dubiński, Sebastian Cygert

Benchmark contamination, where evaluation examples appear in a model's training data, threatens the validity of LLM assessment. Statistical tools for detecting training-data membership exist, but have been validated almost exclusively in controlled academic regimes: large, homogeneous pre-training corpora and transparent, single-stage training pipelines. Whether these methods remain reliable in realistic auditing scenarios remains unclear. We identify two under-studied failure modes: distribution shift, which arises when suspect and validation sets violate the IID assumption, and scale constraints, which arise because benchmarks are orders of magnitude smaller than pre-training corpora. We systematically evaluate three leading paradigms: LLM Dataset Inference, Post-Hoc Dataset Inference, and CoDeC across 27 models from multiple families (including Pythia, OLMo~2, and specialised cultural and medical LLMs) and scales (up to 27B). We then further extend our analysis to frontier industry models. Across 335 evaluations, only 199 yield correct outcomes. LLM Dataset Inference results in false positives under distribution shift, Post-Hoc Dataset Inference is underpowered at benchmark scale, and CoDeC provides only coarse provenance signals that are insufficient to verify individual benchmark splits. Our results reveal a systematic reliability gap between controlled validation and practical benchmark auditing, and show that statistical detection cannot yet replace transparent data provenance. We open-source our benchmark for further research.

Valeriya Khan, Sebastian Cygert, Kamil Deja et al.

In this work, we improve the generative replay in a continual learning setting to perform well on challenging scenarios. Current generative rehearsal methods are usually benchmarked on small and simple datasets as they are not powerful enough to generate more complex data with a greater number of classes. We notice that in VAE-based generative replay, this could be attributed to the fact that the generated features are far from the original ones when mapped to the latent space. Therefore, we propose three modifications that allow the model to learn and generate complex data. More specifically, we incorporate the distillation in latent space between the current and previous models to reduce feature drift. Additionally, a latent matching for the reconstruction and original data is proposed to improve generated features alignment. Further, based on the observation that the reconstructions are better for preserving knowledge, we add the cycling of generations through the previously trained model to make them closer to the original data. Our method outperforms other generative replay methods in various scenarios. Code available at https://github.com/valeriya-khan/looking-through-the-past.

Daniel Marczak, Bartłomiej Twardowski, Tomasz Trzciński et al.

This paper introduces a continual learning approach named MagMax, which utilizes model merging to enable large pre-trained models to continuously learn from new data without forgetting previously acquired knowledge. Distinct from traditional continual learning methods that aim to reduce forgetting during task training, MagMax combines sequential fine-tuning with a maximum magnitude weight selection for effective knowledge integration across tasks. Our initial contribution is an extensive examination of model merging techniques, revealing that simple approaches like weight averaging and random weight selection surprisingly hold up well in various continual learning contexts. More importantly, we present MagMax, a novel model-merging strategy that enables continual learning of large pre-trained models for successive tasks. Our thorough evaluation demonstrates the superiority of MagMax in various scenarios, including class- and domain-incremental learning settings. The code is available at this URL: https://github.com/danielm1405/magmax.

Filip Szatkowski, Mateusz Pyla, Marcin Przewięźlikowski et al.

In this work, we investigate exemplar-free class incremental learning (CIL) with knowledge distillation (KD) as a regularization strategy, aiming to prevent forgetting. KD-based methods are successfully used in CIL, but they often struggle to regularize the model without access to exemplars of the training data from previous tasks. Our analysis reveals that this issue originates from substantial representation shifts in the teacher network when dealing with out-of-distribution data. This causes large errors in the KD loss component, leading to performance degradation in CIL models. Inspired by recent test-time adaptation methods, we introduce Teacher Adaptation (TA), a method that concurrently updates the teacher and the main models during incremental training. Our method seamlessly integrates with KD-based CIL approaches and allows for consistent enhancement of their performance across multiple exemplar-free CIL benchmarks. The source code for our method is available at https://github.com/fszatkowski/cl-teacher-adaptation.

Damian Sójka, Sebastian Cygert, Bartłomiej Twardowski et al.

Test-time adaptation is a promising research direction that allows the source model to adapt itself to changes in data distribution without any supervision. Yet, current methods are usually evaluated on benchmarks that are only a simplification of real-world scenarios. Hence, we propose to validate test-time adaptation methods using the recently introduced datasets for autonomous driving, namely CLAD-C and SHIFT. We observe that current test-time adaptation methods struggle to effectively handle varying degrees of domain shift, often resulting in degraded performance that falls below that of the source model. We noticed that the root of the problem lies in the inability to preserve the knowledge of the source model and adapt to dynamically changing, temporally correlated data streams. Therefore, we enhance the well-established self-training framework by incorporating a small memory buffer to increase model stability and at the same time perform dynamic adaptation based on the intensity of domain shift. The proposed method, named AR-TTA, outperforms existing approaches on both synthetic and more real-world benchmarks and shows robustness across a variety of TTA scenarios. The code is available at https://github.com/dmn-sjk/AR-TTA.

Grzegorz Rypeść, Sebastian Cygert, Tomasz Trzciński et al.

Exemplar-Free Class Incremental Learning (EFCIL) tackles the problem of training a model on a sequence of tasks without access to past data. Existing state-of-the-art methods represent classes as Gaussian distributions in the feature extractor's latent space, enabling Bayes classification or training the classifier by replaying pseudo features. However, we identify two critical issues that compromise their efficacy when the feature extractor is updated on incremental tasks. First, they do not consider that classes' covariance matrices change and must be adapted after each task. Second, they are susceptible to a task-recency bias caused by dimensionality collapse occurring during training. In this work, we propose AdaGauss -- a novel method that adapts covariance matrices from task to task and mitigates the task-recency bias owing to the additional anti-collapse loss function. AdaGauss yields state-of-the-art results on popular EFCIL benchmarks and datasets when training from scratch or starting from a pre-trained backbone. The code is available at: https://github.com/grypesc/AdaGauss.

Grzegorz Rypeść, Daniel Marczak, Sebastian Cygert et al.

Generalized Continual Category Discovery (GCCD) tackles learning from sequentially arriving, partially labeled datasets while uncovering new categories. Traditional methods depend on feature distillation to prevent forgetting the old knowledge. However, this strategy restricts the model's ability to adapt and effectively distinguish new categories. To address this, we introduce a novel technique integrating a learnable projector with feature distillation, thus enhancing model adaptability without sacrificing past knowledge. The resulting distribution shift of the previously learned categories is mitigated with the auxiliary category adaptation network. We demonstrate that while each component offers modest benefits individually, their combination - dubbed CAMP (Category Adaptation Meets Projected distillation) - significantly improves the balance between learning new information and retaining old. CAMP exhibits superior performance across several GCCD and Class Incremental Learning scenarios. The code is available at https://github.com/grypesc/CAMP.

Daniel Marczak, Sebastian Cygert, Tomasz Trzciński et al.

In the field of continual learning, models are designed to learn tasks one after the other. While most research has centered on supervised continual learning, there is a growing interest in unsupervised continual learning, which makes use of the vast amounts of unlabeled data. Recent studies have highlighted the strengths of unsupervised methods, particularly self-supervised learning, in providing robust representations. The improved transferability of those representations built with self-supervised methods is often associated with the role played by the multi-layer perceptron projector. In this work, we depart from this observation and reexamine the role of supervision in continual representation learning. We reckon that additional information, such as human annotations, should not deteriorate the quality of representations. Our findings show that supervised models when enhanced with a multi-layer perceptron head, can outperform self-supervised models in continual representation learning. This highlights the importance of the multi-layer perceptron projector in shaping feature transferability across a sequence of tasks in continual learning. The code is available on github: https://github.com/danielm1405/sl-vs-ssl-cl.

Sebastian Cygert, Damian Sójka, Tomasz Trzciński et al.

Test-Time Adaptation (TTA) has recently emerged as a promising strategy for tackling the problem of machine learning model robustness under distribution shifts by adapting the model during inference without access to any labels. Because of task difficulty, hyperparameters strongly influence the effectiveness of adaptation. However, the literature has provided little exploration into optimal hyperparameter selection. In this work, we tackle this problem by evaluating existing TTA methods using surrogate-based hp-selection strategies (which do not assume access to the test labels) to obtain a more realistic evaluation of their performance. We show that some of the recent state-of-the-art methods exhibit inferior performance compared to the previous algorithms when using our more realistic evaluation setup. Further, we show that forgetting is still a problem in TTA as the only method that is robust to hp-selection resets the model to the initial state at every step. We analyze different types of unsupervised selection strategies, and while they work reasonably well in most scenarios, the only strategies that work consistently well use some kind of supervision (either by a limited number of annotated test samples or by using pretraining data). Our findings underscore the need for further research with more rigorous benchmarking by explicitly stating model selection strategies, to facilitate which we open-source our code.

Alicja Dobrzeniecka, Filip Szatkowski, Sebastian Cygert et al.

While retrieval is a core function of vision-language models, continually updating these models for retrieval tasks remains critically underexplored. Existing work often approaches continual retrieval through the lens of class-incremental learning (CIL), evaluating both standard CIL methods and retrieval-oriented adaptations in settings that may not fully capture the retrieval-specific dynamics. To address this, we introduce a new, principled evaluation framework for continual multimodal retrieval (CMR) spanning diverse visual domains, and systematically evaluate common approaches within this setting. Our empirical analysis shows that standard CIL methods fail to yield meaningful gains in our more challenging scenario. Therefore, we propose Dynamic Adapter Routing (DAR), a novel approach based on adapters selected through prototype-based routing and combined via model merging.DAR achieves superior performance over the previous baselines and demonstrates strong generalization under out-of-distribution evaluation. Our results highlights the unique challenges of CMR and encourages further research in this direction.

Guangyan Zhang, Thomas Merritt, Manuel Sam Ribeiro et al.

Neural text-to-speech systems are often optimized on L1/L2 losses, which make strong assumptions about the distributions of the target data space. Aiming to improve those assumptions, Normalizing Flows and Diffusion Probabilistic Models were recently proposed as alternatives. In this paper, we compare traditional L1/L2-based approaches to diffusion and flow-based approaches for the tasks of prosody and mel-spectrogram prediction for text-to-speech synthesis. We use a prosody model to generate log-f0 and duration features, which are used to condition an acoustic model that generates mel-spectrograms. Experimental results demonstrate that the flow-based model achieves the best performance for spectrogram prediction, improving over equivalent diffusion and L1 models. Meanwhile, both diffusion and flow-based prosody predictors result in significant improvements over a typical L2-trained prosody models.

Damian Sójka, Yuyang Liu, Dipam Goswami et al.

The goal of the challenge is to develop a test-time adaptation (TTA) method, which could adapt the model to gradually changing domains in video sequences for semantic segmentation task. It is based on a synthetic driving video dataset - SHIFT. The source model is trained on images taken during daytime in clear weather. Domain changes at test-time are mainly caused by varying weather conditions and times of day. The TTA methods are evaluated in each image sequence (video) separately, meaning the model is reset to the source model state before the next sequence. Images come one by one and a prediction has to be made at the arrival of each frame. Each sequence is composed of 401 images and starts with the source domain, then gradually drifts to a different one (changing weather or time of day) until the middle of the sequence. In the second half of the sequence, the domain gradually shifts back to the source one. Ground truth data is available only for the validation split of the SHIFT dataset, in which there are only six sequences that start and end with the source domain. We conduct an analysis specifically on those sequences. Ground truth data for test split, on which the developed TTA methods are evaluated for leader board ranking, are not publicly available. The proposed solution secured a 3rd place in a challenge and received an innovation award. Contrary to the solutions that scored better, we did not use any external pretrained models or specialized data augmentations, to keep the solutions as general as possible. We have focused on analyzing the distributional shift and developing a method that could adapt to changing data dynamics and generalize across different scenarios.

Daniel Marczak, Simone Magistri, Sebastian Cygert et al.

Model merging integrates the weights of multiple task-specific models into a single multi-task model. Despite recent interest in the problem, a significant performance gap between the combined and single-task models remains. In this paper, we investigate the key characteristics of task matrices -- weight update matrices applied to a pre-trained model -- that enable effective merging. We show that alignment between singular components of task-specific and merged matrices strongly correlates with performance improvement over the pre-trained model. Based on this, we propose an isotropic merging framework that flattens the singular value spectrum of task matrices, enhances alignment, and reduces the performance gap. Additionally, we incorporate both common and task-specific subspaces to further improve alignment and performance. Our proposed approach achieves state-of-the-art performance on vision and language tasks across various sets of tasks and model scales. This work advances the understanding of model merging dynamics, offering an effective methodology to merge models without requiring additional training. Code is available at https://github.com/danielm1405/iso-merging .

Maciej Chrabąszcz, Aleksander Szymczyk, Marcin Sendera et al.

Large Reasoning Models (LRMs) introduce new opportunities for safety monitoring through their Chain of Thought (CoT) reasoning. However, CoT is not always faithful to the model's final output, undermining its reliability as a monitoring tool. To address this, we investigate the hidden representations of LRMs to determine whether future behavior can be predicted from prompt and CoT representations. By evaluating a probe at each generated token, we construct a probe trajectory, the continuous evolution of a concept's probability across the reasoning process. We find that future model behavior is more distinguishable when examined over the full trajectory than from a single static prediction. To characterize these temporal dynamics, we extract signal-processing features that capture volatility, trend, and steady-state behavior, significantly improving the separation of future model states. We also present two methodological insights. First, template-based training data achieves near-parity with dynamically generated model responses, eliminating the need for a costly initial inference and labeling. Second, the choice of pooling operation is critical: average-pooling and last-token methods collapse to near-random performance, while max-pooling achieves up to 95% AUROC and yields stable probe trajectories. Using four datasets and four reasoning models across the domains of safety and mathematics, we demonstrate that trajectory features encode task-specific dynamics that improve outcome separability. These findings establish probe trajectories as a complementary framework for monitoring LRM behavior. Warning: This article contains potentially harmful content.

Dariusz Piotrowski, Renard Korzeniowski, Alessio Falai et al.

In this work, we introduce a framework for cross-lingual speech synthesis, which involves an upstream Voice Conversion (VC) model and a downstream Text-To-Speech (TTS) model. The proposed framework consists of 4 stages. In the first two stages, we use a VC model to convert utterances in the target locale to the voice of the target speaker. In the third stage, the converted data is combined with the linguistic features and durations from recordings in the target language, which are then used to train a single-speaker acoustic model. Finally, the last stage entails the training of a locale-independent vocoder. Our evaluations show that the proposed paradigm outperforms state-of-the-art approaches which are based on training a large multilingual TTS model. In addition, our experiments demonstrate the robustness of our approach with different model architectures, languages, speakers and amounts of data. Moreover, our solution is especially beneficial in low-resource settings.

Damian Sójka, Sebastian Cygert, Marc Masana

We introduce PACE, a backpropagation-free continual test-time adaptation system that directly optimizes the affine parameters of normalization layers. Existing derivative-free approaches struggle to balance runtime efficiency with learning capacity, as they either restrict updates to input prompts or require continuous, resource-intensive adaptation regardless of domain stability. To address these limitations, PACE leverages the Covariance Matrix Adaptation Evolution Strategy with the Fastfood projection to optimize high-dimensional affine parameters within a low-dimensional subspace, leading to superior adaptive performance. Furthermore, we enhance the runtime efficiency by incorporating an adaptation stopping criterion and a domain-specialized vector bank to eliminate redundant computation. Our framework achieves state-of-the-art accuracy across multiple benchmarks under continual distribution shifts, reducing runtime by over 50% compared to existing backpropagation-free methods.

Piotr Bilinski, Thomas Merritt, Abdelhamid Ezzerg et al.

Creating realistic and natural-sounding synthetic speech remains a big challenge for voice identities unseen during training. As there is growing interest in synthesizing voices of new speakers, here we investigate the ability of normalizing flows in text-to-speech (TTS) and voice conversion (VC) modes to extrapolate from speakers observed during training to create unseen speaker identities. Firstly, we create an approach for TTS and VC, and then we comprehensively evaluate our methods and baselines in terms of intelligibility, naturalness, speaker similarity, and ability to create new voices. We use both objective and subjective metrics to benchmark our techniques on 2 evaluation tasks: zero-shot and new voice speech synthesis. The goal of the former task is to measure the precision of the conversion to an unseen voice. The goal of the latter is to measure the ability to create new voices. Extensive evaluations demonstrate that the proposed approach systematically allows to obtain state-of-the-art performance in zero-shot speech synthesis and creates various new voices, unobserved in the training set. We consider this work to be the first attempt to synthesize new voices based on mel-spectrograms and normalizing flows, along with a comprehensive analysis and comparison of the TTS and VC modes.

Grzegorz Statkiewicz, Alicja Dobrzeniecka, Karolina Seweryn et al.

Most vision-language models (VLMs) are trained on English-centric data, limiting their performance in other languages and cultural contexts. This restricts their usability for non-English-speaking users and hinders the development of multimodal systems that reflect diverse linguistic and cultural realities. In this work, we reproduce and adapt the LLaVA-Next methodology to create a set of Polish VLMs. We rely on a fully automated pipeline for translating and filtering existing multimodal datasets, and complement this with synthetic Polish data for OCR and culturally specific tasks. Despite relying almost entirely on automatic translation and minimal manual intervention to the training data, our approach yields strong results: we observe a +9.5% improvement over LLaVA-1.6-Vicuna-13B on a Polish-adapted MMBench, along with higher-quality captions in generative evaluations, as measured by human annotators in terms of linguistic correctness. These findings highlight that large-scale automated translation, combined with lightweight filtering, can effectively bootstrap high-quality multimodal models for low-resource languages. Some challenges remain, particularly in cultural coverage and evaluation. To facilitate further research, we make our models and evaluation dataset publicly available.

Marcin Osial, Bartosz Wójcik, Bartosz Zieliński et al.

While transfer learning is an advantageous strategy, it overlooks the opportunity to leverage knowledge from numerous available models online. Addressing this multi-source transfer learning problem is a promising path to boost adaptability and cut re-training costs. However, existing approaches are inherently coarse-grained, lacking the necessary precision for granular knowledge extraction and the aggregation efficiency required to fuse knowledge from either a large number of source models or those with high parameter counts. We address these limitations by leveraging Singular Value Decomposition (SVD) to first decompose each source model into its elementary, rank-one components. A subsequent aggregation stage then selects only the most salient components from all sources, thereby overcoming the previous efficiency and precision limitations. To best preserve and leverage the synthesized knowledge base, our method adapts to the target task by fine-tuning only the principal singular values of the merged matrix. In essence, this process only recalibrates the importance of top SVD components. The proposed framework allows for efficient transfer learning, is robust to perturbations both at the input level and in the parameter space (e.g., noisy or pruned sources), and scales well computationally.

Maciej Chrabąszcz, Filip Szatkowski, Bartosz Wójcik et al.

With the rise of LLMs, ensuring model safety and alignment has become a critical concern. While modern instruction-finetuned LLMs incorporate alignment during training, they still frequently require moderation tools to prevent unsafe behavior. The most common approach to moderation are guard models that flag unsafe inputs. However, guards require costly training and are typically limited to fixed-size, pre-trained options, making them difficult to adapt to evolving risks and resource constraints. We hypothesize that instruction-finetuned LLMs already encode safety-relevant information internally and explore training-free safety assessment methods that work with off-the-shelf models. We show that simple prompting allows models to recognize harmful inputs they would otherwise mishandle. We also demonstrate that safe and unsafe prompts are distinctly separable in the models' latent space. Building on this, we introduce the Latent Prototype Moderator (LPM), a training-free moderation method that uses Mahalanobis distance in latent space to assess input safety. LPM is a lightweight, customizable add-on that generalizes across model families and sizes. Our method matches or exceeds state-of-the-art guard models across multiple safety benchmarks, offering a practical and flexible solution for scalable LLM moderation.

Grzegorz Rypeść, Sebastian Cygert, Valeriya Khan et al.

Class-incremental learning is becoming more popular as it helps models widen their applicability while not forgetting what they already know. A trend in this area is to use a mixture-of-expert technique, where different models work together to solve the task. However, the experts are usually trained all at once using whole task data, which makes them all prone to forgetting and increasing computational burden. To address this limitation, we introduce a novel approach named SEED. SEED selects only one, the most optimal expert for a considered task, and uses data from this task to fine-tune only this expert. For this purpose, each expert represents each class with a Gaussian distribution, and the optimal expert is selected based on the similarity of those distributions. Consequently, SEED increases diversity and heterogeneity within the experts while maintaining the high stability of this ensemble method. The extensive experiments demonstrate that SEED achieves state-of-the-art performance in exemplar-free settings across various scenarios, showing the potential of expert diversification through data in continual learning.

Sebastian Cygert, Andrzej Czyzewski

Recent years have seen impressive progress in visual recognition on many benchmarks, however, generalization to the real-world in out-of-distribution setting remains a significant challenge. A state-of-the-art method for robust visual recognition is model ensembling. however, recently it was shown that similarly competitive results could be achieved with a much smaller cost, by using multi-input multi-output architecture (MIMO). In this work, a generalization of the MIMO approach is applied to the task of object detection using the general-purpose Faster R-CNN model. It was shown that using the MIMO framework allows building strong feature representation and obtains very competitive accuracy when using just two input/output pairs. Furthermore, it adds just 0.5\% additional model parameters and increases the inference time by 15.9\% when compared to the standard Faster R-CNN. It also works comparably to, or outperforms the Deep Ensemble approach in terms of model accuracy, robustness to out-of-distribution setting, and uncertainty calibration when the same number of predictions is used. This work opens up avenues for applying the MIMO approach in other high-level tasks such as semantic segmentation and depth estimation.

Sebastian Cygert, Bartłomiej Wróblewski, Karol Woźniak et al.

While recent computer vision algorithms achieve impressive performance on many benchmarks, they lack robustness - presented with an image from a different distribution, (e.g. weather or lighting conditions not considered during training), they may produce an erroneous prediction. Therefore, it is desired that such a model will be able to reliably predict its confidence measure. In this work, uncertainty estimation for the task of semantic segmentation is evaluated under a varying level of domain shift: in a cross-dataset setting and when adapting a model trained on data from the simulation. It was shown that simple color transformations already provide a strong baseline, comparable to using more sophisticated style-transfer data augmentation. Further, by constructing an ensemble consisting of models using different backbones and/or augmentation methods, it was possible to improve significantly model performance in terms of overall accuracy and uncertainty estimation under the domain shift setting. The Expected Calibration Error (ECE) on challenging GTA to Cityscapes adaptation was reduced from 4.05 to the competitive value of 1.1. Further, an ensemble of models was utilized in the self-training setting to improve the pseudo-labels generation, which resulted in a significant gain in the final model accuracy, compared to the standard fine-tuning (without ensemble).

Sebastian Cygert, Andrzej Czyżewski

Model compression techniques allow to significantly reduce the computational cost associated with data processing by deep neural networks with only a minor decrease in average accuracy. Simultaneously, reducing the model size may have a large effect on noisy cases or objects belonging to less frequent classes. It is a crucial problem from the perspective of the models' safety, especially for object detection in the autonomous driving setting, which is considered in this work. It was shown in the paper that the sensitivity of compressed models to different distortion types is nuanced, and some of the corruptions are heavily impacted by the compression methods (i.e., additive noise), while others (blur effect) are only slightly affected. A common way to improve the robustness of models is to use data augmentation, which was confirmed to positively affect models' robustness, also for highly compressed models. It was further shown that while data imbalance methods brought only a slight increase in accuracy for the baseline model (without compression), the impact was more striking at higher compression rates for the structured pruning. Finally, methods for handling data imbalance brought a significant improvement of the pruned models' worst-detected class accuracy.