Many document types use intrinsic, convention-driven structures that serve to encode precise and structured information, such as the conventions governing engineering drawings. However, state-of-the-art approaches treat document recognition as a mere computer vision problem, neglecting these underlying document-type-specific structural properties, making them dependent on sub-optimal heuristic post-processing and rendering many less frequent or more complicated document types inaccessible to modern document recognition. We suggest a novel perspective that frames document recognition as a transcription task from a document to a record. This implies a natural grouping of documents based on the intrinsic structure inherent in their transcription, where related document types can be treated (and learned) similarly. We propose a method to design structure-specific inductive biases for the underlying machine-learned end-to-end document recognition systems, and a respective base transformer architecture that we successfully adapt to different structures. We demonstrate the effectiveness of the so-found inductive biases in extensive experiments with progressively complex record structures from monophonic sheet music, shape drawings, and simplified engineering drawings. By integrating an inductive bias for unrestricted graph structures, we train the first-ever successful end-to-end model to transcribe engineering drawings to their inherently interlinked information. Our approach is relevant to inform the design of document recognition systems for document types that are less well understood than standard OCR, OMR, etc., and serves as a guide to unify the design of future document foundation models.

On-chip learning is key to scalable and adaptive neuromorphic systems, yet existing training methods are either difficult to implement in hardware or overly restrictive. However, recent studies show that feedback-control optimizers can enable expressive, on-chip training of neuromorphic devices. In this work, we present a proof-of-concept implementation of such feedback-control optimizers on a mixed-signal neuromorphic processor. We assess the proposed approach in an In-The-Loop(ITL) training setup on both a binary classification task and the nonlinear Yin-Yang problem, demonstrating on-chip training that matches the performance of numerical simulations and gradient-based baselines. Our results highlight the feasibility of feedback-driven, online learning under realistic mixed-signal constraints, and represent a co-design approach toward embedding such rules directly in silicon for autonomous and adaptive neuromorphic computing.

Background: Agents for computer use (ACUs) are systems that execute complex tasks on digital devices-such as personal computers or mobile phones-given instructions in natural language. These agents automate tasks by controlling software through low-level actions like mouse clicks and touchscreen gestures. However, despite rapid progress, ACUs are not yet mature for everyday use. Objectives: This survey examines the current state-of-the-art, identifies trends, and points out research gaps in the development of practical ACUs. The goal is to provide a comprehensive review and analysis that helps advance general-purpose, robust, and scalable agents for real-world computer use. Methods: We introduce a multifaceted taxonomy of ACUs across three dimensions: (I) the domain perspective, characterizing the contexts in which agents operate; (II) the interaction perspective, describing observation modalities (e.g., screenshots, HTML) and action modalities (e.g., mouse, keyboard, code execution); and (III) the agent perspective, detailing how agents perceive, reason, and learn. We review 87 original research papers about ACUs and 33 relevant datasets, covering both foundation model-based and specialized approaches. Results: Our taxonomy comprehensively structures state-of-the-art approaches and establishes the groundwork for guiding future ACU research. We found that the field is transitioning from specialized agents toward foundation-model-based agents, a shift from text to image-based observation space, and an increasing adoption of behavior cloning methodologies. Furthermore, we identify six key research gaps: insufficient generalization, inefficient learning, limited planning, low task complexity in benchmarks, non-standardized evaluation, and a disconnect between research and practical conditions. Conclusions: To continue rapid improvements in the field, we recommend focusing on: (a) vision-based observations and low-level control to enhance generalization; (b) adaptive learning beyond static prompting; (c) effective planning and reasoning capabilities; (d) realistic, high-complexity benchmarks; (e) standardized evaluation criteria based on task success; and (f) aligning agent design with real-world deployment constraints. Collectively, our findings and proposed directions help develop more general-purpose agents for everyday digital tasks.

We introduce the cooperative network architecture (CNA), a model that represents sensory signals using structured, recurrently connected networks of neurons, termed "nets." Nets are dynamically assembled from overlapping net fragments, which are learned based on statistical regularities in sensory input. This architecture offers robustness to noise, deformation, and generalization to out-of-distribution data, addressing challenges in current vision systems from a novel perspective. We demonstrate that net fragments can be learned without supervision and flexibly recombined to encode novel patterns, enabling figure completion and resilience to noise. Our findings establish CNA as a promising paradigm for developing neural representations that integrate local feature processing with global structure formation, providing a foundation for future research on invariant object recognition.

We present the methodology and results of the Deep Retrieval team for subtask 4b of the CLEF CheckThat! 2025 competition, which focuses on retrieving relevant scientific literature for given social media posts. To address this task, we propose a hybrid retrieval pipeline that combines lexical precision, semantic generalization, and deep contextual re-ranking, enabling robust retrieval that bridges the informal-to-formal language gap. Specifically, we combine BM25-based keyword matching with a FAISS vector store using a fine-tuned INF-Retriever-v1 model for dense semantic retrieval. BM25 returns the top 30 candidates, and semantic search yields 100 candidates, which are then merged and re-ranked via a large language model (LLM)-based cross-encoder. Our approach achieves a mean reciprocal rank at 5 (MRR@5) of 76.46

To go from (passive) process monitoring to active process control, an effective AI system must learn about the behavior of the complex system from very limited training data, forming an ad-hoc digital twin with respect to process inputs and outputs that captures the consequences of actions on the process's world. We propose a novel methodology based on learning world models that disentangles process parameters in the learned latent representation, allowing for fine-grained control. Representation learning is driven by the latent factors influencing the processes through contrastive learning within a joint embedding predictive architecture. This makes changes in representations predictable from changes in inputs and vice versa, facilitating interpretability of key factors responsible for process variations, paving the way for effective control actions to keep the process within operational bounds. The effectiveness of our method is validated on the example of plastic injection molding, demonstrating practical relevance in proposing specific control actions for a notoriously unstable process.

The ability to compose learned concepts and apply them in novel settings is key to human intelligence, but remains a persistent limitation in state-of-the-art machine learning models. To address this issue, we introduce COGITAO, a modular and extensible data generation framework and benchmark designed to systematically study compositionality and generalization in visual domains. Drawing inspiration from ARC-AGI's problem-setting, COGITAO constructs rule-based tasks which apply a set of transformations to objects in grid-like environments. It supports composition, at adjustable depth, over a set of 28 interoperable transformations, along with extensive control over grid parametrization and object properties. This flexibility enables the creation of millions of unique task rules – surpassing concurrent datasets by several orders of magnitude – across a wide range of difficulties, while allowing virtually unlimited sample generation per rule. We provide baseline experiments using state-of-the-art vision models, highlighting their consistent failures to generalize to novel combinations of familiar elements, despite strong in-domain performance. COGITAO is fully open-sourced, including all code and datasets, to support continued research in this field.

Automated tissue segmentation in medical imaging plays a critical role in clinical AI-assisted decision-making, and particularly in the assessment of body composition from CT scans. However, acquiring data and annotations of sufficient quality to train deep-learning models is expensive and timeconsuming. In this work, we propose a novel approach to improve data efficiency and model accuracy by leveraging domain knowledge about biologically relevant tissue-specific Hounsfield unit (HU) ranges as an inductive bias for learning. Specifically, we extend the input representation of deep learning-based segmentation models with binary masks indicating potential tissue types, where each binary mask is created from thresholds derived from medical literature. Our method not only enhances segmentation performance by up to 5% for intramuscular adipose tissue but surpasses the performance of the baseline model with 50% of the training data. Our easy-to-apply method thus improves data efficiency and facilitates the development and use of segmentation models in resource-constrained clinical settings.

We present a diffusion-based model recipe for real-world control of a highly dexterous humanoid robotic hand, designed for sample-efficient learning and smooth fine-motor action inference. Our system features a newly designed 16-DoF tendon-driven hand, equipped with wide angle wrist cameras and mounted on a Franka Emika Panda arm. We develop a versatile teleoperation pipeline and data collection protocol using both glove-based and VR interfaces, enabling high-quality data collection across diverse tasks such as pick and place, item sorting and assembly insertion. Leveraging high-frequency generative control, we train end-to-end policies from raw sensory inputs, enabling smooth, self-correcting motions in complex manipulation scenarios. Real-world evaluations demonstrate up to 93.3 to a +33.3 also revealing scaling trends in policy performance. Our results advance the state-of-the-art in dexterous robotic manipulation through a fully integrated, practical approach to hardware, learning, and real-world deployment.

Deep learning has proved to be a successful paradigm for solving many challenges in machine learning. However, deep neural networks fail when trained sequentially on multiple tasks, a shortcoming known as catastrophic forgetting in the continual learning literature. Despite a recent flourish of learning algorithms successfully addressing this problem, we find that provable guarantees against catastrophic forgetting are lacking. In this work, we study the relationship between learning and forgetting by looking at the geometry of neural networks' loss landscape. We offer a unifying perspective on a family of continual learning algorithms, namely methods based on parameter isolation, and we establish guarantees on catastrophic forgetting for some of them.

The mammalian neocortex possesses the remarkable ability to translate complex sensory inputs into abstract representations through the coordinated activity of large neuronal ensembles across the sensory hierarchy. While cortical hierarchies are anatomically well described, how learning is or-chestrated across the spatial scales ranging from large neuronal networks to pyramidal neurons and their individual synapses is unknown. Here we address this gap from the ground up by modeling the membrane potential and calcium dynamics of individual pyramidal neuron synapses while working upward toward network learning. Starting at the lowest level, we adopt a calcium-dependent synaptic plasticity rule consistent with a wide range of molecular and electrophysiological findings and implement this rule in a synaptic model. We then embed our synaptic model into a pyramidal cell model with apical and dendritic compartments, and integrate various experimental observations such as bursts, calcium plateaus, and somato-apical coupling. We validate the predictions of our neuron model through direct in vitro electrophysiology experiments on layer 5 (L5) pyramidal neurons from the mouse prefrontal cortex and demonstrate that inputs arriving at apical dendrites guide plasticity at basal synapses. Finally, we investigate the algorithmic principles of hierarchical credit assignment in the mammalian neocortex by embedding our pyramidal neuron model in various biologically-plausible deep learning architectures that have been proposed to explain learning in cortex. We find that our model seamlessly aligns with target learning architectures, where top-down feedback arriving at the apical dendrite modifies pyramidal neuron activities to align with desired higher-level neural activity. Importantly, supported by our biological data, this cortical target learning cannot be easily cast into the backpropagation algorithm. By providing a cross-scale framework for cortical hierarchical learning, our work reveals a potential discrepancy between learning in biological neural networks and conventional deep learning.### Competing Interest StatementThe authors have declared no competing interest.

Continual learning is the problem of integrating new information in a model while retaining the knowledge acquired in the past. Despite the tangible improvements achieved in recent years, the problem of continual learning is still an open one. A better understanding of the mechanisms behind the successes and failures of existing continual learning algorithms can unlock the development of new successful strategies. In this work, we view continual learning from the perspective of the multi-task loss approximation, and we compare two alternative strategies, namely local and global approximations. We classify existing continual learning algorithms based on the approximation used, and we assess the practical effects of this distinction in common continual learning settings.Additionally, we study optimal continual learning objectives in the case of local polynomial approximations and we provide examples of existing algorithms implementing the optimal objectives

Online process monitoring is essential to detect failures and respond promptly in automated industrial processes such as injection molding. Traditional systems rely on experienced operators manually defining operational boundaries around a reference signal. We propose a data-driven representation that auto-tunes the sensitivity to a pre-set specificity threshold and automatically detects anomalies alongside interpretable indices that help identify root causes. Our automated system achieved an average AUC of 0.998 and detected 100 percent of the anomalies with the proposed dynamic calibration of the data-driven embedding method. The dynamic calibration, which accounted for drift, boosts the average specificity from 0.362 to 0.869. The outputs also indicate the direction and relative magnitude of characteristic deviations caused by machine parameters, including holding pressure, mold temperature, and injection speed. The AI-derived process boundaries are superior to manual annotation in tested real-world production environments.

At least since the introduction of ChatGPT, the abilities of generative large language models (LLMs), sometimes called GPTs, are at the center of the attention of AI researchers, entrepreneurs, and others. However, for many applications, it is not possible to call an existing LLM service via an API due to data protection concerns or when no task-appropriate LLM exists. On the other hand, deploying or training a private LLM is often prohibitively computationally expensive. In this paper, we give an overview of the most important recent methodologies that help reduce the computational footprint of LLMs. We further present extensive benchmarks for seven methods from two of the most important areas of recent progress: model quantization and low-rank adapters, showcasing how it is possible to leverage state-of-the-art LLMs with limited resources. Our benchmarks include resource consumption metrics (e.g. GPU memory usage), a state-of-the-art quantitative performance evaluation as well as a qualitative performance study conducted by eight individual human raters. Our evaluations show that quantization has a profound effect on GPU memory requirements. However, we also show that these quantization methods, contrary to how they are advertised, cause a noticeable loss in text quality. We further show that low-rank adapters allow effective model fine-tuning with moderate compute resources. For methods that require less than 16 GB of GPU memory, we provide easy-to-use Jupyter notebooks that allow anyone to deploy and fine-tune state-of-the-art LLMs on the Google Colab free tier within minutes without any prior experience or infrastructure.

The medial prefrontal cortex (mPFC) has been proposed to link sensory inputs and behavioral outputs to mediate the execution of learned behaviors. However, how such a link is implemented has remained unclear. To measure prefrontal neural correlates of sensory stimuli and learned behaviors, we performed population calcium imaging during a new tone-signaled active avoidance paradigm in mice. We developed an analysis approach based on dimensionality reduction and decoding that allowed us to identify interpretable task-related population activity patterns. While a large fraction of tone-evoked activity was not informative about behavior execution, we identified an activity pattern that was predictive of tone-induced avoidance actions and did not occur for spontaneous actions with similar motion kinematics. Moreover, this avoidance-specific activity differed between distinct avoidance actions learned in two consecutive tasks. Overall, our results are consistent with a model in which mPFC contributes to the selection of goal-directed actions by transforming sensory inputs into specific behavioral outputs through distributed population-level computations.

The computational capabilities of neuronal networks are fundamentally constrained by their specific connectivity. Previous studies of cortical connectivity have mostly been carried out in rodents; whether the principles established therein also apply to the evolutionarily expanded human cortex is unclear. We studied network properties within the human temporal cortex using samples obtained from brain surgery. We analyzed multineuron patch-clamp recordings in layer 2-3 pyramidal neurons and identified substantial differences compared with rodents. Reciprocity showed random distribution, synaptic strength was independent from connection probability, and connectivity of the supragranular temporal cortex followed a directed and mostly acyclic graph topology. Application of these principles in neuronal models increased dimensionality of network dynamics, suggesting a critical role for cortical computation.

The anterior cingulate cortex (ACC) causally influences cognitive control of goal-directed behaviour. However, it is unclear whether ACC directly encodes cognitive variables like attention or impulsivity, or implements goal-directed action selection mechanisms that are modulated by them. We recorded ACC activity with miniature endoscopic microscopes in mice performing the 5-choice-serial-reaction time task, and applied decoding and encoding analyses. ACC pyramidal cells represented specific actions before and during the behavioural response, whereas the response type (e.g. correct/incorrect/premature) – indicating the state of attentional and impulse control – could only be decoded during and after the response with high reliability. Devaluation and extinction experiments further revealed that action encoding depended on reward expectation. Our findings support a role for ACC in goal-directed action selection and monitoring, that is modulated by cognitive state, rather than in tracking levels of attention or impulsivity directly in individual trials. ### Competing Interest Statement The authors have declared no competing interest.

We introduce meta-learning algorithms that perform zero-shot weight-space adaptation of neural network models to unseen tasks. Our methods repurpose the popular generative image synthesis techniques of natural language guidance and diffusion models to generate neural network weights adapted for tasks. We first train an unconditional generative hypernetwork model to produce neural network weights; then we train a second "guidance" model that, given a natural language task description, traverses the hypernetwork latent space to find high-performance task-adapted weights in a zero-shot manner. We explore two alternative approaches for latent space guidance: "HyperCLIP"-based classifier guidance and a conditional Hypernetwork Latent Diffusion Model ("HyperLDM"), which we show to benefit from the classifier-free guidance technique common in image generation. Finally, we demonstrate that our approaches outperform existing multi-task and meta-learning methods in a series of zero-shot learning experiments on our Meta-VQA dataset.

Artificial Neural Networks (ANNs) trained with Backpropagation (BP) excel in different daily tasks but have a dangerous vulnerability: inputs with small targeted perturbations, also known as adversarial samples, can drastically disrupt their performance. Adversarial training, a technique in which the training dataset is augmented with exemplary adversarial samples, is proven to mitigate this problem but comes at a high computational cost. In contrast to ANNs, humans are not susceptible to misclassifying these same adversarial samples. Thus, one can postulate that biologically-plausible trained ANNs might be more robust against adversarial attacks. In this work, we chose the biologically-plausible learning algorithm Present the Error to Perturb the Input To modulate Activity (PEPITA) as a case study and investigated this question through a comparative analysis with BP-trained ANNs on various computer vision tasks. We observe that PEPITA has a higher intrinsic adversarial robustness and, when adversarially trained, also has a more favorable natural-vs-adversarial performance trade-off. In particular, for the same natural accuracies on the MNIST task, PEPITA's adversarial accuracies decrease on average only by 0.26