In robotics, Vision-Language-Action (VLA) models that integrate diverse multimodal signals from multi-view inputs have emerged as an effective approach. However, most prior work adopts static fusion that processes all visual inputs uniformly, which incurs unnecessary computational overhead and allows task-irrelevant background information to act as noise. Inspired by the principles of human active perception, we propose a dynamic information fusion framework designed to maximize the efficiency and robustness of VLA models. Our approach introduces a lightweight adaptive routing architecture that analyzes the current text prompt and observations from a wrist-mounted camera in real-time to predict the task-relevance of multiple camera views. By conditionally attenuating computations for views with low informational utility and selectively providing only essential visual features to the policy network, Our framework achieves computation efficiency proportional to task relevance. Furthermore, to efficiently secure large-scale annotation data for router training, we established an automated labeling pipeline utilizing Vision-Language Models (VLMs) to minimize data collection and annotation costs. Experimental results in real-world robotic manipulation scenarios demonstrate that the proposed approach achieves significant improvements in both inference efficiency and control performance compared to existing VLA models, validating the effectiveness and practicality of dynamic information fusion in resource-constrained, real-time robot control environments.

This paper introduces Perspectives, an interactive extension of the Discourse Analysis Tool Suite designed to empower Digital Humanities (DH) scholars to explore and organize large, unstructured document collections. Perspectives implements a flexible, aspect-focused document clustering pipeline with human-in-the-loop refinement capabilities. We showcase how this process can be initially steered by defining analytical lenses through document rewriting prompts and instruction-based embeddings, and further aligned with user intent through tools for refining clusters and mechanisms for fine-tuning the embedding model. The demonstration highlights a typical workflow, illustrating how DH researchers can leverage Perspectives's interactive document map to uncover topics, sentiments, or other relevant categories, thereby gaining insights and preparing their data for subsequent in-depth analysis.

Recent studies have explored the combination of multiple LoRAs to simultaneously generate user-specified subjects and styles. However, most existing approaches fuse LoRA weights using static statistical heuristics that deviate from LoRA's original purpose of learning adaptive feature adjustments and ignore the randomness of sampled inputs. To address this, we propose a dynamic training-free fusion framework that operates throughout the generation process. During the forward pass, at each LoRA-applied layer, we dynamically compute the KL divergence between the base model's original features and those produced by subject and style LoRAs, respectively, and adaptively select the most appropriate weights for fusion. In the reverse denoising stage, we further refine the generation trajectory by dynamically applying gradient-based corrections derived from objective metrics such as CLIP and DINO scores, providing continuous semantic and stylistic guidance. By integrating these two complementary mechanisms-feature-level selection and metric-guided latent adjustment-across the entire diffusion timeline, our method dynamically achieves coherent subject-style synthesis without any retraining. Extensive experiments across diverse subject-style combinations demonstrate that our approach consistently outperforms state-of-the-art LoRA fusion methods both qualitatively and quantitatively.

Quantifying biometric characteristics within hand gestures involve derivation of fitness scores from a gesture and identity aware feature space. However, evaluating the quality of these scores remains an open question. Existing biometric capacity estimation literature relies upon error rates. But these rates do not indicate goodness of scores. Thus, in this manuscript we present an exhaustive set of evaluation measures. We firstly identify ranking order and relevance of output scores as the primary basis for evaluation. In particular, we consider both rank deviation as well as rewards for: (i) higher scores of high ranked gestures and (ii) lower scores of low ranked gestures. We also compensate for correspondence between trends of output and ground truth scores. Finally, we account for disentanglement between identity features of gestures as a discounting factor. Integrating these elements with adequate weighting, we formulate advanced acceptance score as a holistic evaluation measure. To assess effectivity of the proposed we perform in-depth experimentation over three datasets with five state-of-the-art (SOTA) models. Results show that the optimal score selected with our measure is more appropriate than existing other measures. Also, our proposed measure depicts correlation with existing measures. This further validates its reliability. We have made our \href{https://github.com/AmanVerma2307/MeasureSuite}{code} public.

World Models have emerged as a powerful paradigm for learning compact, predictive representations of environment dynamics, enabling agents to reason, plan, and generalize beyond direct experience. Despite recent interest in World Models, most available implementations remain publication-specific, severely limiting their reusability, increasing the risk of bugs, and reducing evaluation standardization. To mitigate these issues, we introduce stable-worldmodel (SWM), a modular, tested, and documented world-model research ecosystem that provides efficient data-collection tools, standardized environments, planning algorithms, and baseline implementations. In addition, each environment in SWM enables controllable factors of variation, including visual and physical properties, to support robustness and continual learning research. Finally, we demonstrate the utility of SWM by using it to study zero-shot robustness in DINO-WM.

We study diffusion-based world models for reinforcement learning, which offer high generative fidelity but face critical efficiency challenges in control. Current methods either require heavyweight models at inference or rely on highly sequential imagination, both of which impose prohibitive computational costs. We propose Horizon Imagination (HI), an on-policy imagination process for discrete stochastic policies that denoises multiple future observations in parallel. HI incorporates a stabilization mechanism and a novel sampling schedule that decouples the denoising budget from the effective horizon over which denoising is applied while also supporting sub-frame budgets. Experiments on Atari 100K and Craftium show that our approach maintains control performance with a sub-frame budget of half the denoising steps and achieves superior generation quality under varied schedules. Code is available at https://github.com/leor-c/horizon-imagination.

Although state-of-the-art LLMs can solve math problems, we find that they make errors on numerical comparisons with mixed notation: "Which is larger, $5.7 \times 10^2$ or $580$?" This raises a fundamental question: Do LLMs even know how big these numbers are? We probe the hidden states of several smaller open-source LLMs. A single linear projection of an appropriate hidden layer encodes the log-magnitudes of both kinds of numerals, allowing us to recover the numbers with relative error of about 2.3% (on restricted synthetic text) or 19.06% (on scientific papers). Furthermore, the hidden state after reading a pair of numerals encodes their ranking, with a linear classifier achieving over 90% accuracy. Yet surprisingly, when explicitly asked to rank the same pairs of numerals, these LLMs achieve only 50-70% accuracy, with worse performance for models whose probes are less effective. Finally, we show that incorporating the classifier probe's log-loss as an auxiliary objective during finetuning brings an additional 3.22% improvement in verbalized accuracy over base models, demonstrating that improving models' internal magnitude representations can enhance their numerical reasoning capabilities. Our code is available at https://github.com/VCY019/Numeracy-Probing.

Large Language Models (LLM) benchmarks tell us when models fail, but not why they fail. A wrong answer on a reasoning dataset may stem from formatting issues, calculation errors, or dataset noise rather than weak reasoning. Without disentangling such causes, benchmarks remain incomplete and cannot reliably guide model improvement. We introduce ErrorMap, the first method to chart the sources of LLM failure. It extracts a model's unique "failure signature", clarifies what benchmarks measure, and broadens error identification to reduce blind spots. This helps developers debug models, aligns benchmark goals with outcomes, and supports informed model selection. ErrorMap works on any model or dataset with the same logic. Applying our method to 35 datasets and 83 models we generate ErrorAtlas, a taxonomy of model errors, revealing recurring failure patterns. ErrorAtlas highlights error types that are currently underexplored in LLM research, such as omissions of required details in the output and question misinterpretation. By shifting focus from where models succeed to why they fail, ErrorMap and ErrorAtlas enable advanced evaluation - one that exposes hidden weaknesses and directs progress. Unlike success, typically measured by task-level metrics, our approach introduces a deeper evaluation layer that can be applied globally across models and tasks, offering richer insights into model behavior and limitations. We make the taxonomy and code publicly available with plans to periodically update ErrorAtlas as new benchmarks and models emerge.

Large Language Models (LLMs) are fundamentally constrained by the quadratic computational cost of self-attention and the "Lost in the Middle" phenomenon, where reasoning capabilities degrade as context windows expand. Existing solutions, primarily "Flat RAG" architectures relying on vector databases, treat memory as an unstructured bag of embeddings, failing to capture the hierarchical and temporal structure of long-horizon interactions. This paper presents Aeon, a Neuro-Symbolic Cognitive Operating System that redefines memory as a managed OS resource. Aeon structures memory into a Memory Palace (a spatial index implemented via Atlas, a SIMD-accelerated Page-Clustered Vector Index) and a Trace (a neuro-symbolic episodic graph). This architecture introduces three advances: (1) Symmetric INT8 Scalar Quantization, achieving 3.1x spatial compression and 5.6x math acceleration via NEON SDOT intrinsics; (2) a decoupled Write-Ahead Log (WAL) ensuring crash-recoverability with statistically negligible overhead (<1%); and (3) a Sidecar Blob Arena eliminating the prior 440-character text ceiling via an append-only mmap-backed blob file with generational garbage collection. The Semantic Lookaside Buffer (SLB) exploits conversational locality to achieve sub-5us retrieval latencies, with INT8 vectors dequantized to FP32 on cache insertion to preserve L1-resident lookup performance. Benchmarks on Apple M4 Max demonstrate that the combined architecture achieves 4.70ns INT8 dot product latency, 3.09us tree traversal at 100K nodes (3.4x over FP32), and P99 read latency of 750ns under hostile 16-thread contention via epoch-based reclamation.

Embedding retrofitting adjusts pre-trained word vectors using knowledge graph constraints to improve domain-specific retrieval. However, the effectiveness of retrofitting depends critically on knowledge graph quality, which in turn depends on text preprocessing. This paper presents a data engineering framework that addresses data quality degradation from annotation artifacts in real-world corpora. The analysis shows that hashtag annotations inflate knowledge graph density, leading to creating spurious edges that corrupt the retrofitting objective. On noisy graphs, all retrofitting techniques produce statistically significant degradation ($-3.5\%$ to $-5.2\%$, $p<0.05$). After preprocessing, \acrshort{ewma} retrofitting achieves $+6.2\%$ improvement ($p=0.0348$) with benefits concentrated in quantitative synthesis questions ($+33.8\%$ average). The gap between clean and noisy preprocessing (10\%+ swing) exceeds the gap between algorithms (3\%), establishing preprocessing quality as the primary determinant of retrofitting success.

Urban wind flow reconstruction is essential for assessing air quality, heat dispersion, and pedestrian comfort, yet remains challenging when only sparse sensor data are available. We propose GenDA, a generative data assimilation framework that reconstructs high-resolution wind fields on unstructured meshes from limited observations. The model employs a multiscale graph-based diffusion architecture trained on computational fluid dynamics (CFD) simulations and interprets classifier-free guidance as a learned posterior reconstruction mechanism: the unconditional branch learns a geometry-aware flow prior, while the sensor-conditioned branch injects observational constraints during sampling. This formulation enables obstacle-aware reconstruction and generalization across unseen geometries, wind directions, and mesh resolutions without retraining. We consider both sparse fixed sensors and trajectory-based observations using the same reconstruction procedure. When evaluated against supervised graph neural network (GNN) baselines and classical reduced-order data assimilation methods, GenDA reduces the relative root-mean-square error (RRMSE) by 25-57% and increases the structural similarity index (SSIM) by 23-33% across the tested meshes. Experiments are conducted on Reynolds-averaged Navier-Stokes (RANS) simulations of a real urban neighbourhood in Bristol, United Kingdom, at a characteristic Reynolds number of $\mathrm{Re}\approx2\times10^{7}$, featuring complex building geometry and irregular terrain. The proposed framework provides a scalable path toward generative, geometry-aware data assimilation for environmental monitoring in complex domains.

Multimodal large language models (MLLMs) typically rely on a single late-layer feature from a frozen vision encoder, leaving the encoder's rich hierarchy of visual cues under-utilized. MLLMs still suffer from visually ungrounded hallucinations, often relying on language priors rather than image evidence. While many prior mitigation strategies operate on the text side, they leave the visual representation unchanged and do not exploit the rich hierarchy of features encoded across vision layers. Existing multi-layer fusion methods partially address this limitation but remain static, applying the same layer mixture regardless of the query. In this work, we introduce TGIF (Text-Guided Inter-layer Fusion), a lightweight module that treats encoder layers as depth-wise "experts" and predicts a prompt-dependent fusion of visual features. TGIF follows the principle of direct external fusion, requires no vision-encoder updates, and adds minimal overhead. Integrated into LLaVA-1.5-7B, TGIF provides consistent improvements across hallucination, OCR, and VQA benchmarks, while preserving or improving performance on ScienceQA, GQA, and MMBench. These results suggest that query-conditioned, hierarchy-aware fusion is an effective way to strengthen visual grounding and reduce hallucination in modern MLLMs.

Detecting distributional drift in high-dimensional data streams presents fundamental challenges: global comparison methods scale poorly, projection-based approaches lose geometric structure, and re-clustering methods suffer from identity instability. This paper introduces Argus, A framework that reconceptualizes drift detection as tracking local statistics over a fixed spatial partition of the data manifold. The key contributions are fourfold. First, it is proved that Voronoi tessellations over canonical orthonormal frames yield drift metrics that are invariant to orthogonal transformations. The rotations and reflections that preserve Euclidean geometry. Second, it is established that this framework achieves O(N) complexity per snapshot while providing cell-level spatial localization of distributional change. Third, a graph-theoretic characterization of drift propagation is developed that distinguishes coherent distributional shifts from isolated perturbations. Fourth, product quantization tessellation is introduced for scaling to very high dimensions (d>500) by decomposing the space into independent subspaces and aggregating drift signals across subspaces. This paper formalizes the theoretical foundations, proves invariance properties, and presents experimental validation demonstrating that the framework correctly identifies drift under coordinate rotation while existing methods produce false positives. The tessellated approach offers a principled geometric foundation for distribution monitoring that preserves high-dimensional structure without the computational burden of pairwise comparisons.

Binary Neural Networks (BNNs), which constrain both weights and activations to binary values, offer substantial reductions in computational complexity, memory footprint, and energy consumption. These advantages make them particularly well suited for deployment on resource-constrained devices. However, training BNNs via gradient-based optimization remains challenging due to the discrete nature of their variables. The dominant approach, quantization-aware training, circumvents this issue by employing surrogate gradients. Yet, this method requires maintaining latent full-precision parameters and performing the backward pass with floating-point arithmetic, thereby forfeiting the efficiency of binary operations during training. While alternative approaches based on local learning rules exist, they are unsuitable for global credit assignment and for back-propagating errors in multi-layer architectures. This paper introduces Binary Error Propagation (BEP), the first learning algorithm to establish a principled, discrete analog of the backpropagation chain rule. This mechanism enables error signals, represented as binary vectors, to be propagated backward through multiple layers of a neural network. BEP operates entirely on binary variables, with all forward and backward computations performed using only bitwise operations. Crucially, this makes BEP the first solution to enable end-to-end binary training for recurrent neural network architectures. We validate the effectiveness of BEP on both multi-layer perceptrons and recurrent neural networks, demonstrating gains of up to +6.89% and +10.57% in test accuracy, respectively. The proposed algorithm is released as an open-source repository.

Error Correction Codes (ECC) are fundamental to reliable digital communication, yet designing neural decoders that are both accurate and computationally efficient remains challenging. Recent denoising diffusion decoders achieve state-of-the-art performance, but their iterative sampling limits practicality in low-latency settings. To bridge this gap, consistency models (CMs) offer a potential path to high-fidelity one-step decoding. However, applying CMs to ECC presents a significant challenge: the discrete nature of error correction means the decoding trajectory is highly non-smooth, making it incompatible with a simple continuous timestep parameterization. To address this, we re-parameterize the reverse Probability Flow Ordinary Differential Equation (PF-ODE) by soft-syndrome condition, providing a smooth trajectory of signal corruption. Building on this, we propose the Error Correction Syndrome-Flow Consistency Model (ECCFM), a model-agnostic framework designed specifically for ECC task, ensuring the model learns a smooth trajectory from any noisy signal directly to the original codeword in a single step. Across multiple benchmarks, ECCFM attains lower bit-error-rate (BER) and frame-error-rate (FER) than transformer-based decoders, while delivering inference speeds 30x to 100x faster than iterative denoising diffusion decoders.

Coordinating a team of robots to reposition multiple objects in cluttered environments requires reasoning jointly about where robots should establish contact, how to manipulate objects once contact is made, and how to navigate safely and efficiently at scale. Prior approaches typically fall into two extremes -- either learning the entire task or relying on privileged information and hand-designed planners -- both of which struggle to handle diverse objects in long-horizon tasks. To address these challenges, we present a unified framework for collaborative multi-robot, multi-object non-prehensile manipulation that integrates flow-matching co-generation with anonymous multi-robot motion planning. Within this framework, a generative model co-generates contact formations and manipulation trajectories from visual observations, while a novel motion planner conveys robots at scale. Crucially, the same planner also supports coordination at the object level, assigning manipulated objects to larger target structures and thereby unifying robot- and object-level reasoning within a single algorithmic framework. Experiments in challenging simulated environments demonstrate that our approach outperforms baselines in both motion planning and manipulation tasks, highlighting the benefits of generative co-design and integrated planning for scaling collaborative manipulation to complex multi-agent, multi-object settings. Visit gco-paper.github.io for code and demonstrations.

Despite rapid progress, multimodal reasoning still lacks a systematic approach to synthesize large-scale vision-centric datasets beyond visual math. We introduce a framework able to synthesize vision-centric problems spanning diverse levels of complexity, and the resulting dataset with over 1M high-quality problems including: reasoning traces, preference data, and instruction prompts supporting SFT, offline and online RL. Our vision-centric synthesis framework uses a two-stage process focusing on: (1) generating diverse verifiable questions from existing images at scale, and (2) creating complex compositional visual problems by merging simpler questions. Remarkably, finetuning Qwen2.5-VL-7B on our data outperforms existing open-data baselines across evaluated vision-centric benchmarks, and our best configurations match or surpass strong closed-data models such as MiMo-VL-7B-RL on Vstar Bench, CV-Bench and MMStar-V. Notably, despite being entirely vision-centric, our data transfers positively to text-only reasoning (MMLU-Pro, +3.7%) and audio reasoning (MMAU, +1.32%), demonstrating its effectiveness. Similarly, despite containing no embodied visual data, we observe notable gains (NiEH, +8.8%) when evaluating open-ended embodied QA. Lastly, we use our data to comprehensively analyze at scale (1M+) the entire VLM post-training pipeline showing that (i) SFT on high-quality data with cognitive behaviors on reasoning traces is essential to scale online RL, (ii) offline RL could match online RL's performance while disaggregating compute demands, and, (iii) SFT on high quality data also improve out-of-domain, cross-modality transfer.

We propose a general feedback-driven retrieval-augmented generation (RAG) approach that leverages Large Audio Language Models (LALMs) to address the missing or imperfect synthesis of specific sound events in text-to-audio (TTA) generation. Unlike previous RAG-based TTA methods that typically train specialized models from scratch, we utilize LALMs to analyze audio generation outputs, retrieve concepts that pre-trained models struggle to generate from an external database, and incorporate the retrieved information into the generation process. Experimental results show that our method not only enhances the ability of LALMs to identify missing sound events but also delivers improvements across different models, outperforming existing RAG-specialized approaches.

We present a fully interpretable and flexible statistical method for background subtraction in roadside LiDAR data, aimed at enhancing infrastructure-based perception in automated driving. Our approach introduces both a Gaussian distribution grid (GDG), which models the spatial statistics of the background using background-only scans, and a filtering algorithm that uses this representation to classify LiDAR points as foreground or background. The method supports diverse LiDAR types, including multiline 360 degree and micro-electro-mechanical systems (MEMS) sensors, and adapts to various configurations. Evaluated on the publicly available RCooper dataset, it outperforms state-of-the-art techniques in accuracy and flexibility, even with minimal background data. Its efficient implementation ensures reliable performance on low-resource hardware, enabling scalable real-world deployment.

Modelling qualitative uncertainty in formal argumentation is essential both for practical applications and theoretical understanding. Yet, most of the existing works focus on \textit{abstract} models for arguing with uncertainty. Following a recent trend in the literature, we tackle the open question of studying plausible instantiations of these abstract models. To do so, we ground the uncertainty of arguments in their components, structured within rules and premises. Our main technical contributions are: i) the introduction of a notion of expressivity that can handle abstract and structured formalisms, and ii) the presentation of both negative and positive expressivity results, comparing the expressivity of abstract and structured models of argumentation with uncertainty. These results affect incomplete abstract argumentation frameworks, and their extension with dependencies, on the abstract side, and ASPIC+, on the structured side.

Missing data is a widespread problem in tabular settings. Existing solutions range from simple averaging to complex generative adversarial networks, but due to each method's large variance in performance across real-world domains and time-consuming hyperparameter tuning, no universal imputation method exists. This performance variance is particularly pronounced in small datasets, where the models have the least amount of information. Building on TabPFN, a recent tabular foundation model for supervised learning, we propose TabImpute, a pre-trained transformer that delivers accurate and fast zero-shot imputations, requiring no fitting or hyperparameter tuning at inference time. To train and evaluate TabImpute, we introduce (i) an entry-wise featurization for tabular settings, enabling a 100x speedup over the previous TabPFN imputation method, (ii) a synthetic training data generation pipeline incorporating a diverse set of missingness patterns to enhance accuracy on real-world missing data problems, and (iii) MissBench, a comprehensive benchmark with 42 OpenML tables and 13 new missingness patterns. MissBench spans domains such as medicine, finance, and engineering, showcasing TabImpute's robust performance compared to numerous established imputation methods.

We build upon vec2vec, a procedure designed to align text embedding spaces without parallel data. vec2vec finds a near-perfect alignment, but it is expensive and unstable. We present mini-vec2vec, a simple and efficient alternative that requires substantially lower computational cost and is highly robust. Moreover, the learned mapping is a linear transformation. Our method consists of three main stages: a tentative matching of pseudo-parallel embedding vectors, transformation fitting, and iterative refinement. Our linear alternative exceeds the original instantiation of vec2vec by orders of magnitude in efficiency, while matching or exceeding their results. The method's stability and interpretable algorithmic steps facilitate scaling and unlock new opportunities for adoption in new domains and fields.

The discovery of deep, steerable taxonomies in large text corpora is currently restricted by a trade-off between the surface-level efficiency of topic models and the prohibitive, non-scalable assignment costs of LLM-integrated frameworks. We introduce \textbf{LogiPart}, a scalable, hypothesis-first framework for building interpretable hierarchical partitions that decouples hierarchy growth from expensive full-corpus LLM conditioning. LogiPart utilizes locally hosted LLMs on compact, embedding-aware samples to generate concise natural-language taxonomic predicates. These predicates are then evaluated efficiently across the entire corpus using zero-shot Natural Language Inference (NLI) combined with fast graph-based label propagation, achieving constant $O(1)$ generative token complexity per node relative to corpus size. We evaluate LogiPart across four diverse text corpora (totaling $\approx$140,000 documents). Using structured manifolds for \textbf{calibration}, we identify an empirical reasoning threshold at the 14B-parameter scale required for stable semantic grounding. On complex, high-entropy corpora (Wikipedia, US Bills), where traditional thematic metrics reveal an ``alignment gap,'' inverse logic validation confirms the stability of the induced logic, with individual taxonomic bisections maintaining an average per-node routing accuracy of up to 96\%. A qualitative audit by an independent LLM-as-a-judge confirms the discovery of meaningful functional axes, such as policy intent, that thematic ground-truth labels fail to capture. LogiPart enables frontier-level exploratory analysis on consumer-grade hardware, making hypothesis-driven taxonomic discovery feasible under realistic computational and governance constraints.

Learning-based planners are sensitive to the long-tailed distribution of driving data. Common maneuvers dominate datasets, while dangerous or rare scenarios are sparse. This imbalance can bias models toward the frequent cases and degrade performance on critical scenarios. To tackle this problem, we compare balancing strategies for sampling training data and find reweighting by trajectory pattern an effective approach. We then present FlowDrive, a flow-matching trajectory planner that learns a conditional rectified flow to map noise directly to trajectory distributions with few flow-matching steps. We further introduce moderated, in-the-loop guidance that injects small perturbation between flow steps to systematically increase trajectory diversity while remaining scene-consistent. On nuPlan and the interaction-focused interPlan benchmarks, FlowDrive achieves state-of-the-art results among learning-based planners and approaches methods with rule-based refinements. After adding moderated guidance and light post-processing (FlowDrive*), it achieves overall state-of-the-art performance across nearly all benchmark splits. Our code is available at https://github.com/einsteinguang/flow_drive_planner.

Training large language models (LLMs) to reason via reinforcement learning (RL) significantly improves their problem-solving capabilities. In agentic settings, existing methods like ReAct prompt LLMs to explicitly plan before every action; however, we demonstrate that always planning is computationally expensive and degrades performance on long-horizon tasks, while never planning further limits performance. To address this, we introduce a conceptual framework formalizing dynamic planning for LLM agents, enabling them to flexibly decide when to allocate test-time compute for planning. We propose a simple two-stage training pipeline: (1) supervised fine-tuning on diverse synthetic data to prime models for dynamic planning, and (2) RL to refine this capability in long-horizon environments. Experiments on the Crafter environment show that dynamic planning agents trained with this approach are more sample-efficient and consistently achieve more complex objectives. Additionally, we demonstrate that these agents can be effectively steered by human-written plans, surpassing their independent capabilities and highlighting the potential for safer and more collaborative agentic systems.

While stereotypes are well-documented in human social interactions, AI systems are often presumed to be less susceptible to such biases. Previous studies have focused on biases inherited from training data, but whether stereotypes can emerge spontaneously in AI agent interactions merits further exploration. Through a novel experimental framework simulating workplace interactions with neutral initial conditions, we investigate the emergence and evolution of stereotypes in LLM-based multi-agent systems. Our findings reveal that (1) LLM-Based AI agents develop stereotype-driven biases in their interactions despite beginning without predefined biases; (2) stereotype effects intensify with increased interaction rounds and decision-making power, particularly after introducing hierarchical structures; (3) these systems exhibit group effects analogous to human social behavior, including halo effects, confirmation bias, and role congruity; and (4) these stereotype patterns manifest consistently across different LLM architectures. Through comprehensive quantitative analysis, these findings suggest that stereotype formation in AI systems may arise as an emergent property of multi-agent interactions, rather than merely from training data biases. Our work underscores the need for future research to explore the underlying mechanisms of this phenomenon and develop strategies to mitigate its ethical impacts.

Drug discovery is a complex and resource-intensive process, making early prediction of approval outcomes critical for optimizing research investments. While classical machine learning and deep learning methods have shown promise in drug approval prediction, their limited interpretability constraints their impact. Here, we present DrugReasoner, a reasoning-based large language model (LLM) built on the LLaMA architecture and fine-tuned with group relative policy optimization (GRPO) to predict the likelihood of small-molecule approval. DrugReasoner integrates molecular descriptors with comparative reasoning against structurally similar approved and unapproved compounds, generating predictions alongside step-by-step rationales and confidence scores. DrugReasoner achieved robust performance with an AUC of 0.732 and an F1 score of 0.729 on the validation set and 0.725 and 0.718 on the test set, respectively. These results outperformed conventional baselines, including logistic regression, support vector machine, and k-nearest neighbors and had competitive performance relative to XGBoost. On an external independent dataset, DrugReasoner outperformed both baseline and the recently developed ChemAP model, achieving an AUC of 0.728 and an F1-score of 0.774, while maintaining high precision and balanced sensitivity, demonstrating robustness in real-world scenarios. These findings demonstrate that DrugReasoner not only delivers competitive predictive accuracy but also enhances transparency through its reasoning outputs, thereby addressing a key bottleneck in AI-assisted drug discovery. This study highlights the potential of reasoning-augmented LLMs as interpretable and effective tools for pharmaceutical decision-making.

Rigorous statistical methods, including parameter estimation with accompanying uncertainties, underpin the validity of scientific discovery, especially in the natural sciences. With increasingly complex data models such as deep learning techniques, uncertainty quantification has become exceedingly difficult and a plethora of techniques have been proposed. In this case study, we use the unifying framework of approximate Bayesian inference combined with empirical tests on carefully created synthetic classification datasets to investigate qualitative properties of six different probabilistic machine learning algorithms for class probability and uncertainty estimation: (i) a neural network ensemble, (ii) neural network ensemble with conflictual loss, (iii) evidential deep learning, (iv) a single neural network with Monte Carlo Dropout, (v) Gaussian process classification and (vi) a Dirichlet process mixture model. We check if the algorithms produce uncertainty estimates which reflect commonly desired properties, such as being well calibrated and exhibiting an increase in uncertainty for out-of-distribution data points. Our results indicate that all algorithms show reasonably good calibration performance on our synthetic test sets, but none of the deep learning based algorithms provide uncertainties that consistently reflect lack of experimental evidence for out-of-distribution data points. We hope our study may serve as a clarifying example for researchers that are using or developing methods of uncertainty estimation for scientific data-driven modeling and analysis.

Federated learning (FL) enables the collaborative training of deep neural networks across decentralized data archives (i.e., clients), where each client stores data locally and only shares model updates with a central server. This makes FL a suitable learning paradigm for remote sensing (RS) image classification tasks, where data centralization may be restricted due to legal and privacy constraints. However, a key challenge in applying FL to RS tasks is the communication overhead caused by the frequent exchange of large model updates between clients and the central server. To address this issue, in this paper we propose a novel strategy (denoted as FedX) that uses explanation-guided pruning to reduce communication overhead by minimizing the size of the transmitted models without compromising performance. FedX leverages backpropagation-based explanation methods to estimate the task-specific importance of model components and prunes the least relevant ones at the central server. The resulting sparse global model is then sent to clients, substantially reducing communication overhead. We evaluate FedX on multi-label scene classification using the BigEarthNet-S2 dataset and single-label scene classification using the EuroSAT dataset. Experimental results show the success of FedX in significantly reducing the number of shared model parameters while enhancing the generalization capability of the global model, compared to both unpruned model and state-of-the-art pruning methods. The code of FedX will be available at https://git.tu-berlin.de/rsim/FedX.

Audio inpainting seeks to restore missing segments in degraded recordings. Previous diffusion-based methods exhibit impaired performance when the missing region is large. We introduce the first approach that applies discrete diffusion over tokenized music representations from a pre-trained audio tokenizer, enabling stable and semantically coherent restoration of long gaps. Our method further incorporates two training approaches: a derivative-based regularization loss that enforces smooth temporal dynamics, and a span-based absorbing transition that provides structured corruption during diffusion. Experiments on the MusicNet and MAESTRO datasets with gaps up to 750 ms show that our approach consistently outperforms strong baselines across range of gap lengths, for gaps of 150 ms and above. This work advances musical audio restoration and introduces new directions for discrete diffusion model training. Visit our project page for examples and code.