Knowledge distillation compresses a larger neural model (teacher) into smaller, faster student models by training them to match teacher outputs. However, the internal computational transformations that occur during this process remain poorly understood. We apply techniques from mechanistic interpretability to analyze how internal circuits, representations, and activation patterns differ between teacher and student. Focusing on GPT2-small and its distilled counterpart DistilGPT2, we find that student models reorganize, compress, and discard teacher components, often resulting in stronger reliance on fewer individual components. To quantify functional alignment beyond output similarity, we introduce an alignment metric based on influence-weighted component similarity, validated across multiple tasks. Our findings reveal that while knowledge distillation preserves broad functional behaviors, it also causes significant shifts in internal computation, with important implications for the robustness and generalization capacity of distilled models.
Sparse and delayed reward functions pose a significant obstacle for real-world Reinforcement Learning (RL) applications. In this work, we propose Attention-based REward Shaping (ARES), a general and robust algorithm which uses a transformer's attention mechanism to generate shaped rewards and create a dense reward function for any environment. ARES requires a set of episodes and their final returns as input. It can be trained entirely offline and is able to generate meaningful shaped rewards even when using small datasets or episodes produced by agents taking random actions. ARES is compatible with any RL algorithm and can handle any level of reward sparsity. In our experiments, we focus on the most challenging case where rewards are fully delayed until the end of each episode. We evaluate ARES across a diverse range of environments, widely used RL algorithms, and baseline methods to assess the effectiveness of the shaped rewards it produces. Our results show that ARES can significantly improve learning in delayed reward settings, enabling RL agents to train in scenarios that would otherwise require impractical amounts of data or even be unlearnable. To our knowledge, ARES is the first approach that works fully offline, remains robust to extreme reward delays and low-quality data, and is not limited to goal-based tasks.
We present GNN-Suite, a robust modular framework for constructing and benchmarking Graph Neural Network (GNN) architectures in computational biology. GNN-Suite standardises experimentation and reproducibility using the Nextflow workflow to evaluate GNN performance. We demonstrate its utility in identifying cancer-driver genes by constructing molecular networks from protein-protein interaction (PPI) data from STRING and BioGRID and annotating nodes with features from the PCAWG, PID, and COSMIC-CGC repositories. Our design enables fair comparisons among diverse GNN architectures including GAT, GAT3H, GCN, GCN2, GIN, GTN, HGCN, PHGCN, and GraphSAGE and a baseline Logistic Regression (LR) model. All GNNs were configured as standardised two-layer models and trained with uniform hyperparameters (dropout = 0.2; Adam optimiser with learning rate = 0.01; and an adjusted binary cross-entropy loss to address class imbalance) over an 80/20 train-test split for 300 epochs. Each model was evaluated over 10 independent runs with different random seeds to yield statistically robust performance metrics, with balanced accuracy (BACC) as the primary measure. Notably, GCN2 achieved the highest BACC (0.807 +/- 0.035) on a STRING-based network, although all GNN types outperformed the LR baseline, highlighting the advantage of network-based learning over feature-only approaches. Our results show that a common framework for implementing and evaluating GNN architectures aids in identifying not only the best model but also the most effective means of incorporating complementary data. By making GNN-Suite publicly available, we aim to foster reproducible research and promote improved benchmarking standards in computational biology. Future work will explore additional omics datasets and further refine network architectures to enhance predictive accuracy and interpretability in biomedical applications.
Clustering tabular data remains a significant open challenge in data analysis and machine learning. Unlike for image data, similarity between tabular records often varies across datasets, making the definition of clusters highly dataset-dependent. Furthermore, the absence of supervised signals complicates hyperparameter tuning in deep learning clustering methods, frequently resulting in unstable performance. To address these issues and reduce the need for per-dataset tuning, we adopt an emerging approach in deep learning: zero-shot learning. We propose ZEUS, a self-contained model capable of clustering new datasets without any additional training or fine-tuning. It operates by decomposing complex datasets into meaningful components that can then be clustered effectively. Thanks to pre-training on synthetic datasets generated from a latent-variable prior, it generalizes across various datasets without requiring user intervention. To the best of our knowledge, ZEUS is the first zero-shot method capable of generating embeddings for tabular data in a fully unsupervised manner. Experimental results demonstrate that it performs on par with or better than traditional clustering algorithms and recent deep learning-based methods, while being significantly faster and more user-friendly.
As the number of rooftop photovoltaic (PV) installations increases, aggregators and system operators are required to monitor and analyze these systems, raising the challenge of integration and management of large, spatially distributed time-series data that are both high-dimensional and affected by missing values. In this work, a probabilistic entity embedding-based clustering framework is proposed to address these problems. This method encodes each PV system's characteristic power generation patterns and uncertainty as a probability distribution, then groups systems by their statistical distances and agglomerative clustering. Applied to a multi-year residential PV dataset, it produces concise, uncertainty-aware cluster profiles that outperform a physics-based baseline in representativeness and robustness, and support reliable missing-value imputation. A systematic hyperparameter study further offers practical guidance for balancing model performance and robustness.
The rapid decline of Arctic sea ice resulting from anthropogenic climate change poses significant risks to indigenous communities, ecosystems, and the global climate system. This situation emphasizes the immediate necessity for precise seasonal sea ice forecasts. While dynamical models perform well for short-term forecasts, they encounter limitations in long-term forecasts and are computationally intensive. Deep learning models, while more computationally efficient, often have difficulty managing seasonal variations and uncertainties when dealing with complex sea ice dynamics. In this research, we introduce IceMamba, a deep learning architecture that integrates sophisticated attention mechanisms within the state space model. Through comparative analysis of 25 renowned forecast models, including dynamical, statistical, and deep learning approaches, our experimental results indicate that IceMamba delivers excellent seasonal forecasting capabilities for Pan-Arctic sea ice concentration. Specifically, IceMamba outperforms all tested models regarding average RMSE and anomaly correlation coefficient (ACC) and ranks second in Integrated Ice Edge Error (IIEE). This innovative approach enhances our ability to foresee and alleviate the effects of sea ice variability, offering essential insights for strategies aimed at climate adaptation.
In this work, we propose a computationally efficient algorithm for visual policy learning that leverages differentiable simulation and first-order analytical policy gradients. Our approach decouple the rendering process from the computation graph, enabling seamless integration with existing differentiable simulation ecosystems without the need for specialized differentiable rendering software. This decoupling not only reduces computational and memory overhead but also effectively attenuates the policy gradient norm, leading to more stable and smoother optimization. We evaluate our method on standard visual control benchmarks using modern GPU-accelerated simulation. Experiments show that our approach significantly reduces wall-clock training time and consistently outperforms all baseline methods in terms of final returns. Notably, on complex tasks such as humanoid locomotion, our method achieves a $4\times$ improvement in final return, and successfully learns a humanoid running policy within 4 hours on a single GPU.
The growing use of smartphones and IoT devices necessitates efficient time-series analysis on resource-constrained hardware, which is critical for sensing applications such as human activity recognition and air quality prediction. Recent efforts in hardware-aware neural architecture search (NAS) automate architecture discovery for specific platforms; however, none focus on general time-series analysis with edge deployment. Leveraging the problem-solving and reasoning capabilities of large language models (LLM), we propose MONAQ, a novel framework that reformulates NAS into Multi-Objective Neural Architecture Querying tasks. MONAQ is equipped with multimodal query generation for processing multimodal time-series inputs and hardware constraints, alongside an LLM agent-based multi-objective search to achieve deployment-ready models via code generation. By integrating numerical data, time-series images, and textual descriptions, MONAQ improves an LLM's understanding of time-series data. Experiments on fifteen datasets demonstrate that MONAQ-discovered models outperform both handcrafted models and NAS baselines while being more efficient.
Editing images using natural language instructions has become a natural and expressive way to modify visual content; yet, evaluating the performance of such models remains challenging. Existing evaluation approaches often rely on image-text similarity metrics like CLIP, which lack precision. In this work, we introduce a new benchmark designed to evaluate text-guided image editing models in a more grounded manner, along two critical dimensions: (i) functional correctness, assessed via automatically generated multiple-choice questions that verify whether the intended change was successfully applied; and (ii) image content preservation, which ensures that non-targeted regions of the image remain visually consistent using an object-aware masking technique and preservation scoring. The benchmark includes over 1000 high-quality editing examples across 20 diverse content categories, each annotated with detailed editing instructions, evaluation questions, and spatial object masks. We conduct a large-scale study comparing GPT-Image-1, the latest flagship in the text-guided image editing space, against several state-of-the-art editing models, and validate our automatic metrics against human ratings. Results show that GPT-Image-1 leads in instruction-following accuracy, but often over-modifies irrelevant image regions, highlighting a key trade-off in the current model behavior. GIE-Bench provides a scalable, reproducible framework for advancing more accurate evaluation of text-guided image editing.
In recent years, Recurrent Spiking Neural Networks (RSNNs) have shown promising potential in long-term temporal modeling. Many studies focus on improving neuron models and also integrate recurrent structures, leveraging their synergistic effects to improve the long-term temporal modeling capabilities of Spiking Neural Networks (SNNs). However, these studies often place an excessive emphasis on the role of neurons, overlooking the importance of analyzing neurons and recurrent structures as an integrated framework. In this work, we consider neurons and recurrent structures as an integrated system and conduct a systematic analysis of gradient propagation along the temporal dimension, revealing a challenging gradient vanishing problem. To address this issue, we propose the Skip Recurrent Connection (SRC) as a replacement for the vanilla recurrent structure, effectively mitigating the gradient vanishing problem and enhancing long-term temporal modeling performance. Additionally, we propose the Adaptive Skip Recurrent Connection (ASRC), a method that can learn the skip span of skip recurrent connection in each layer of the network. Experiments show that replacing the vanilla recurrent structure in RSNN with SRC significantly improves the model's performance on temporal benchmark datasets. Moreover, ASRC-SNN outperforms SRC-SNN in terms of temporal modeling capabilities and robustness.
Large multimodal models (LMMs) now excel on many vision language benchmarks, however, they still struggle with human centered criteria such as fairness, ethics, empathy, and inclusivity, key to aligning with human values. We introduce HumaniBench, a holistic benchmark of 32K real-world image question pairs, annotated via a scalable GPT4o assisted pipeline and exhaustively verified by domain experts. HumaniBench evaluates seven Human Centered AI (HCAI) principles: fairness, ethics, understanding, reasoning, language inclusivity, empathy, and robustness, across seven diverse tasks, including open and closed ended visual question answering (VQA), multilingual QA, visual grounding, empathetic captioning, and robustness tests. Benchmarking 15 state of the art LMMs (open and closed source) reveals that proprietary models generally lead, though robustness and visual grounding remain weak points. Some open-source models also struggle to balance accuracy with adherence to human-aligned principles. HumaniBench is the first benchmark purpose built around HCAI principles. It provides a rigorous testbed for diagnosing alignment gaps and guiding LMMs toward behavior that is both accurate and socially responsible. Dataset, annotation prompts, and evaluation code are available at: https://vectorinstitute.github.io/HumaniBench
Recent advancements in Large Language Models (LLMs) and their multimodal extensions (MLLMs) have substantially enhanced machine reasoning across diverse tasks. However, these models predominantly rely on pure text as the medium for both expressing and structuring reasoning, even when visual information is present. In this work, we argue that language may not always be the most natural or effective modality for reasoning, particularly in tasks involving spatial and geometrical information. Motivated by this, we propose a new paradigm, Visual Planning, which enables planning through purely visual representations, independent of text. In this paradigm, planning is executed via sequences of images that encode step-by-step inference in the visual domain, akin to how humans sketch or visualize future actions. We introduce a novel reinforcement learning framework, Visual Planning via Reinforcement Learning (VPRL), empowered by GRPO for post-training large vision models, leading to substantial improvements in planning in a selection of representative visual navigation tasks, FrozenLake, Maze, and MiniBehavior. Our visual planning paradigm outperforms all other planning variants that conduct reasoning in the text-only space. Our results establish Visual Planning as a viable and promising alternative to language-based reasoning, opening new avenues for tasks that benefit from intuitive, image-based inference.
Emotion understanding is a critical yet challenging task. Recent advances in Multimodal Large Language Models (MLLMs) have significantly enhanced their capabilities in this area. However, MLLMs often suffer from hallucinations, generating irrelevant or nonsensical content. To the best of our knowledge, despite the importance of this issue, there has been no dedicated effort to evaluate emotion-related hallucinations in MLLMs. In this work, we introduce EmotionHallucer, the first benchmark for detecting and analyzing emotion hallucinations in MLLMs. Unlike humans, whose emotion understanding stems from the interplay of biology and social learning, MLLMs rely solely on data-driven learning and lack innate emotional instincts. Fortunately, emotion psychology provides a solid foundation of knowledge about human emotions. Building on this, we assess emotion hallucinations from two dimensions: emotion psychology knowledge and real-world multimodal perception. To support robust evaluation, we utilize an adversarial binary question-answer (QA) framework, which employs carefully crafted basic and hallucinated pairs to assess the emotion hallucination tendencies of MLLMs. By evaluating 38 LLMs and MLLMs on EmotionHallucer, we reveal that: i) most current models exhibit substantial issues with emotion hallucinations; ii) closed-source models outperform open-source ones in detecting emotion hallucinations, and reasoning capability provides additional advantages; iii) existing models perform better in emotion psychology knowledge than in multimodal emotion perception. As a byproduct, these findings inspire us to propose the PEP-MEK framework, which yields an average improvement of 9.90% in emotion hallucination detection across selected models. Resources will be available at https://github.com/xxtars/EmotionHallucer.
Recent advances in vision language models (VLMs) have enabled broad progress in the general medical field. However, pathology still remains a more challenging subdomain, with current pathology specific VLMs exhibiting limitations in both diagnostic accuracy and reasoning plausibility. Such shortcomings are largely attributable to the nature of current pathology datasets, which are primarily composed of image description pairs that lack the depth and structured diagnostic paradigms employed by real world pathologists. In this study, we leverage pathology textbooks and real world pathology experts to construct high-quality, reasoning-oriented datasets. Building on this, we introduce Patho-R1, a multimodal RL-based pathology Reasoner, trained through a three-stage pipeline: (1) continued pretraining on 3.5 million image-text pairs for knowledge infusion; (2) supervised fine-tuning on 500k high-quality Chain-of-Thought samples for reasoning incentivizing; (3) reinforcement learning using Group Relative Policy Optimization and Decoupled Clip and Dynamic sAmpling Policy Optimization strategies for multimodal reasoning quality refinement. To further assess the alignment quality of our dataset, we propose PathoCLIP, trained on the same figure-caption corpus used for continued pretraining. Comprehensive experimental results demonstrate that both PathoCLIP and Patho-R1 achieve robust performance across a wide range of pathology-related tasks, including zero-shot classification, cross-modal retrieval, Visual Question Answering, and Multiple Choice Question. Our project is available at the Patho-R1 repository: https://github.com/Wenchuan-Zhang/Patho-R1.
Comprehensive assessment of the various aspects of the brain's microstructure requires the use of complementary imaging techniques. This includes measuring the spatial distribution of cell bodies (cytoarchitecture) and nerve fibers (myeloarchitecture). The gold standard for cytoarchitectonic analysis is light microscopic imaging of cell-body stained tissue sections. To reveal the 3D orientations of nerve fibers, 3D Polarized Light Imaging (3D-PLI) has been introduced as a reliable technique providing a resolution in the micrometer range while allowing processing of series of complete brain sections. 3D-PLI acquisition is label-free and allows subsequent staining of sections after measurement. By post-staining for cell bodies, a direct link between fiber- and cytoarchitecture can potentially be established within the same section. However, inevitable distortions introduced during the staining process make a nonlinear and cross-modal registration necessary in order to study the detailed relationships between cells and fibers in the images. In addition, the complexity of processing histological sections for post-staining only allows for a limited number of samples. In this work, we take advantage of deep learning methods for image-to-image translation to generate a virtual staining of 3D-PLI that is spatially aligned at the cellular level. In a supervised setting, we build on a unique dataset of brain sections, to which Cresyl violet staining has been applied after 3D-PLI measurement. To ensure high correspondence between both modalities, we address the misalignment of training data using Fourier-based registration methods. In this way, registration can be efficiently calculated during training for local image patches of target and predicted staining. We demonstrate that the proposed method enables prediction of a Cresyl violet staining from 3D-PLI, matching individual cell instances.
Industry-scale recommender systems face a core challenge: representing entities with high cardinality, such as users or items, using dense embeddings that must be accessible during both training and inference. However, as embedding sizes grow, memory constraints make storage and access increasingly difficult. We describe a lightweight, learnable embedding compression technique that projects dense embeddings into a high-dimensional, sparsely activated space. Designed for retrieval tasks, our method reduces memory requirements while preserving retrieval performance, enabling scalable deployment under strict resource constraints. Our results demonstrate that leveraging sparsity is a promising approach for improving the efficiency of large-scale recommenders. We release our code at https://github.com/recombee/CompresSAE.
Vision-and-Language Navigation (VLN) is a core task where embodied agents leverage their spatial mobility to navigate in 3D environments toward designated destinations based on natural language instructions. Recently, video-language large models (Video-VLMs) with strong generalization capabilities and rich commonsense knowledge have shown remarkable performance when applied to VLN tasks. However, these models still encounter the following challenges when applied to real-world 3D navigation: 1) Insufficient understanding of 3D geometry and spatial semantics; 2) Limited capacity for large-scale exploration and long-term environmental memory; 3) Poor adaptability to dynamic and changing environments.To address these limitations, we propose Dynam3D, a dynamic layered 3D representation model that leverages language-aligned, generalizable, and hierarchical 3D representations as visual input to train 3D-VLM in navigation action prediction. Given posed RGB-D images, our Dynam3D projects 2D CLIP features into 3D space and constructs multi-level 3D patch-instance-zone representations for 3D geometric and semantic understanding with a dynamic and layer-wise update strategy. Our Dynam3D is capable of online encoding and localization of 3D instances, and dynamically updates them in changing environments to provide large-scale exploration and long-term memory capabilities for navigation. By leveraging large-scale 3D-language pretraining and task-specific adaptation, our Dynam3D sets new state-of-the-art performance on VLN benchmarks including R2R-CE, REVERIE-CE and NavRAG-CE under monocular settings. Furthermore, experiments for pre-exploration, lifelong memory, and real-world robot validate the effectiveness of practical deployment.
Vision-language models (VLMs) have shown remarkable progress in offline tasks such as image captioning and video question answering. However, real-time interactive environments impose new demands on VLMs, requiring them to generate utterances that are not only semantically accurate but also precisely timed. We identify two core capabilities necessary for such settings -- $\textit{perceptual updating}$ and $\textit{contingency awareness}$ -- and propose a new benchmark task, $\textbf{Temporally-Grounded Language Generation (TGLG)}$, to evaluate them. TGLG requires models to generate utterances in response to streaming video such that both content and timing align with dynamic visual input. To support this benchmark, we curate evaluation datasets from sports broadcasting and egocentric human interaction domains, and introduce a new metric, $\textbf{TRACE}$, to evaluate TGLG by jointly measuring semantic similarity and temporal alignment. Finally, we present $\textbf{Vision-Language Model with Time-Synchronized Interleaving (VLM-TSI)}$, a model that interleaves visual and linguistic tokens in a time-synchronized manner, enabling real-time language generation without relying on turn-based assumptions. Experimental results show that VLM-TSI significantly outperforms a strong baseline, yet overall performance remains modest -- highlighting the difficulty of TGLG and motivating further research in real-time VLMs. Code and data available $\href{https://github.com/yukw777/tglg}{here}$.
The assessment of evaluation metrics (meta-evaluation) is crucial for determining the suitability of existing metrics in text-to-image (T2I) generation tasks. Human-based meta-evaluation is costly and time-intensive, and automated alternatives are scarce. We address this gap and propose CROC: a scalable framework for automated Contrastive Robustness Checks that systematically probes and quantifies metric robustness by synthesizing contrastive test cases across a comprehensive taxonomy of image properties. With CROC, we generate a pseudo-labeled dataset (CROC$^{syn}$) of over one million contrastive prompt-image pairs to enable a fine-grained comparison of evaluation metrics. We also use the dataset to train CROCScore, a new metric that achieves state-of-the-art performance among open-source methods, demonstrating an additional key application of our framework. To complement this dataset, we introduce a human-supervised benchmark (CROC$^{hum}$) targeting especially challenging categories. Our results highlight robustness issues in existing metrics: for example, many fail on prompts involving negation, and all tested open-source metrics fail on at least 25% of cases involving correct identification of body parts.
Contrastive learning (CL) is a prevalent technique for training embedding models, which pulls semantically similar examples (positives) closer in the representation space while pushing dissimilar ones (negatives) further apart. A key source of negatives are 'in-batch' examples, i.e., positives from other examples in the batch. Effectiveness of such models is hence strongly influenced by the size and quality of training batches. In this work, we propose 'Breaking the Batch Barrier' (B3), a novel batch construction strategy designed to curate high-quality batches for CL. Our approach begins by using a pretrained teacher embedding model to rank all examples in the dataset, from which a sparse similarity graph is constructed. A community detection algorithm is then applied to this graph to identify clusters of examples that serve as strong negatives for one another. The clusters are then used to construct batches that are rich in in-batch negatives. Empirical results on the MMEB multimodal embedding benchmark (36 tasks) demonstrate that our method sets a new state of the art, outperforming previous best methods by +1.3 and +2.9 points at the 7B and 2B model scales, respectively. Notably, models trained with B3 surpass existing state-of-the-art results even with a batch size as small as 64, which is 4-16x smaller than that required by other methods.