Training Transformers on algorithmic tasks frequently demonstrates an intriguing abrupt learning phenomenon: an extended performance plateau followed by a sudden, sharp improvement. This work investigates the underlying mechanisms for such dynamics, primarily in shallow Transformers. We reveal that during the plateau, the model often develops an interpretable partial solution while simultaneously exhibiting a strong repetition bias in their outputs. This output degeneracy is accompanied by internal representation collapse, where hidden states across different tokens become nearly parallel. We further identify the slow learning of optimal attention maps as a key bottleneck. Hidden progress in attention configuration during the plateau precedes the eventual rapid convergence, and directly intervening on attention significantly alters plateau duration and the severity of repetition bias and representational collapse. We validate that these identified phenomena-repetition bias and representation collapse-are not artifacts of toy setups but also manifest in the early pre-training stage of large language models like Pythia and OLMo.
Probabilistic forecasts are typically obtained using state-of-the-art statistical and machine learning models, with model parameters estimated by optimizing a proper scoring rule over a set of training data. If the model class is not correctly specified, then the learned model will not necessarily issue forecasts that are calibrated. Calibrated forecasts allow users to appropriately balance risks in decision making, and it is particularly important that forecast models issue calibrated predictions for extreme events, since such outcomes often generate large socio-economic impacts. In this work, we study how the loss function used to train probabilistic forecast models can be adapted to improve the reliability of forecasts made for extreme events. We investigate loss functions based on weighted scoring rules, and additionally propose regularizing loss functions using a measure of tail miscalibration. We apply these approaches to a hierarchy of increasingly flexible forecast models for UK wind speeds, including simple parametric models, distributional regression networks, and conditional generative models. We demonstrate that state-of-the-art models do not issue calibrated forecasts for extreme wind speeds, and that the calibration of forecasts for extreme events can be improved by suitable adaptations to the loss function during model training. This, however, introduces a trade-off between calibrated forecasts for extreme events and calibrated forecasts for more common outcomes.
Estimating conditional average treatment effects (CATE) from observational data involves modeling decisions that differ from supervised learning, particularly concerning how to regularize model complexity. Previous approaches can be grouped into two primary "meta-learner" paradigms that impose distinct inductive biases. Indirect meta-learners first fit and regularize separate potential outcome (PO) models and then estimate CATE by taking their difference, whereas direct meta-learners construct and directly regularize estimators for the CATE function itself. Neither approach consistently outperforms the other across all scenarios: indirect learners perform well when the PO functions are simple, while direct learners outperform when the CATE is simpler than individual PO functions. In this paper, we introduce the Hybrid Learner (H-learner), a novel regularization strategy that interpolates between the direct and indirect regularizations depending on the dataset at hand. The H-learner achieves this by learning intermediate functions whose difference closely approximates the CATE without necessarily requiring accurate individual approximations of the POs themselves. We demonstrate empirically that intentionally allowing suboptimal fits to the POs improves the bias-variance tradeoff in estimating CATE. Experiments conducted on semi-synthetic and real-world benchmark datasets illustrate that the H-learner consistently operates at the Pareto frontier, effectively combining the strengths of both direct and indirect meta-learners.
The EUNIS habitat classification is crucial for categorising European habitats, supporting European policy on nature conservation and implementing the Nature Restoration Law. To meet the growing demand for detailed and accurate habitat information, we provide spatial predictions for 260 EUNIS habitat types at hierarchical level 3, together with independent validation and uncertainty analyses. Using ensemble machine learning models, together with high-resolution satellite imagery and ecologically meaningful climatic, topographic and edaphic variables, we produced a European habitat map indicating the most probable EUNIS habitat at 100-m resolution across Europe. Additionally, we provide information on prediction uncertainty and the most probable habitats at level 3 within each EUNIS level 1 formation. This product is particularly useful for both conservation and restoration purposes. Predictions were cross-validated at European scale using a spatial block cross-validation and evaluated against independent data from France (forests only), the Netherlands and Austria. The habitat maps obtained strong predictive performances on the validation datasets with distinct trade-offs in terms of recall and precision across habitat formations.
A compelling portrayal of characters is essential to the success of narrative writing. For readers, appreciating a character's traits requires the ability to infer their evolving beliefs, desires, and intentions over the course of a complex storyline, a cognitive skill known as Theory-of-Mind (ToM). Performing ToM reasoning in prolonged narratives requires readers to integrate historical context with current narrative information, a task at which humans excel but Large Language Models (LLMs) often struggle. To systematically evaluate LLMs' ToM reasoning capability in long narratives, we construct LitCharToM, a benchmark of character-centric questions across four ToM dimensions from classic literature. Further, we introduce EvolvTrip, a perspective-aware temporal knowledge graph that tracks psychological development throughout narratives. Our experiments demonstrate that EvolvTrip consistently enhances performance of LLMs across varying scales, even in challenging extended-context scenarios. EvolvTrip proves to be particularly valuable for smaller models, partially bridging the performance gap with larger LLMs and showing great compatibility with lengthy narratives. Our findings highlight the importance of explicit representation of temporal character mental states in narrative comprehension and offer a foundation for more sophisticated character understanding. Our data and code are publicly available at https://github.com/Bernard-Yang/EvolvTrip.
Traditional Retrieval-Augmented Generation (RAG) systems employ brute-force inner product search to retrieve the top-k most similar documents, then combined with the user query and passed to a language model. This allows the model to access external knowledge and reduce hallucinations. However, selecting an appropriate k value remains a significant challenge in practical applications: a small k may fail to retrieve sufficient information, while a large k can introduce excessive and irrelevant content. To address this, we propose a hierarchical clustering-based retrieval method that eliminates the need to predefine k. Our approach maintains the accuracy and relevance of system responses while adaptively selecting semantically relevant content. In the experiment stage, we applied our method to a Taiwanese legal dataset with expert-graded queries. The results show that our approach achieves superior performance in expert evaluations and maintains high precision while eliminating the need to predefine k, demonstrating improved accuracy and interpretability in legal text retrieval tasks. Our framework is simple to implement and easily integrates with existing RAG pipelines, making it a practical solution for real-world applications under limited resources.
Measuring how semantics of words change over time improves our understanding of how cultures and perspectives change. Diachronic word embeddings help us quantify this shift, although previous studies leveraged substantial temporally annotated corpora. In this work, we use a corpus of 9.5 million Croatian news articles spanning the past 25 years and quantify semantic change using skip-gram word embeddings trained on five-year periods. Our analysis finds that word embeddings capture linguistic shifts of terms pertaining to major topics in this timespan (COVID-19, Croatia joining the European Union, technological advancements). We also find evidence that embeddings from post-2020 encode increased positivity in sentiment analysis tasks, contrasting studies reporting a decline in mental health over the same period.
$E(3)$-equivariant neural networks have demonstrated success across a wide range of 3D modelling tasks. A fundamental operation in these networks is the tensor product, which interacts two geometric features in an equivariant manner to create new features. Due to the high computational complexity of the tensor product, significant effort has been invested to optimize the runtime of this operation. For example, Luo et al. (2024) recently proposed the Gaunt tensor product (GTP) which promises a significant speedup. In this work, we provide a careful, systematic analysis of a number of tensor product operations. In particular, we emphasize that different tensor products are not performing the same operation. The reported speedups typically come at the cost of expressivity. We introduce measures of expressivity and interactability to characterize these differences. In addition, we realized the original implementation of GTP can be greatly simplified by directly using a spherical grid at no cost in asymptotic runtime. This spherical grid approach is faster on our benchmarks and in actual training of the MACE interatomic potential by 30\%. Finally, we provide the first systematic microbenchmarks of the various tensor product operations. We find that the theoretical runtime guarantees can differ wildly from empirical performance, demonstrating the need for careful application-specific benchmarking. Code is available at \href{https://github.com/atomicarchitects/PriceofFreedom}{https://github.com/atomicarchitects/PriceofFreedom}
Reconstructing 3D scenes from unconstrained image collections poses significant challenges due to variations in appearance. In this paper, we propose Scalable Micro-macro Wavelet-based Gaussian Splatting (SMW-GS), a novel method that enhances 3D reconstruction across diverse scales by decomposing scene representations into global, refined, and intrinsic components. SMW-GS incorporates the following innovations: Micro-macro Projection, which enables Gaussian points to sample multi-scale details with improved diversity; and Wavelet-based Sampling, which refines feature representations using frequency-domain information to better capture complex scene appearances. To achieve scalability, we further propose a large-scale scene promotion strategy, which optimally assigns camera views to scene partitions by maximizing their contributions to Gaussian points, achieving consistent and high-quality reconstructions even in expansive environments. Extensive experiments demonstrate that SMW-GS significantly outperforms existing methods in both reconstruction quality and scalability, particularly excelling in large-scale urban environments with challenging illumination variations. Project is available at https://github.com/Kidleyh/SMW-GS.
Deep neural networks have been shown to achieve exceptional performance for computer vision tasks like image recognition, segmentation, and reconstruction or denoising. Here, we evaluate the ultimate performance limits of deep convolutional neural network models for image reconstruction, by comparing them against the standard quantum limit set by shot-noise and the Heisenberg limit on precision. We train U-Net models on images of natural objects illuminated with coherent states of light, and find that the average mean-squared error of the reconstructions can surpass the standard quantum limit, and in some cases reaches the Heisenberg limit. Further, we train models on well-parameterized images for which we can calculate the quantum Cram\'er-Rao bound to determine the minimum possible measurable variance of an estimated parameter for a given probe state. We find the mean-squared error of the model predictions reaches these bounds calculated for the parameters, across a variety of parameterized images. These results suggest that deep convolutional neural networks can learn to become the optimal estimators allowed by the laws of physics, performing parameter estimation and image reconstruction at the ultimate possible limits of precision for the case of classical illumination of the object.
The growing volume of omics and clinical data generated for neurodegenerative diseases (NDs) requires new approaches for their curation so they can be ready-to-use in bioinformatics. NeuroEmbed is an approach for the engineering of semantically accurate embedding spaces to represent cohorts and samples. The NeuroEmbed method comprises four stages: (1) extraction of ND cohorts from public repositories; (2) semi-automated normalization and augmentation of metadata of cohorts and samples using biomedical ontologies and clustering on the embedding space; (3) automated generation of a natural language question-answering (QA) dataset for cohorts and samples based on randomized combinations of standardized metadata dimensions and (4) fine-tuning of a domain-specific embedder to optimize queries. We illustrate the approach using the GEO repository and the PubMedBERT pretrained embedder. Applying NeuroEmbed, we semantically indexed 2,801 repositories and 150,924 samples. Amongst many biology-relevant categories, we normalized more than 1,700 heterogeneous tissue labels from GEO into 326 unique ontology-aligned concepts and enriched annotations with new ontology-aligned terms, leading to a fold increase in size for the metadata terms between 2.7 and 20 fold. After fine-tuning PubMedBERT with the QA training data augmented with the enlarged metadata, the model increased its mean Retrieval Precision from 0.277 to 0.866 and its mean Percentile Rank from 0.355 to 0.896. The NeuroEmbed methodology for the creation of electronic catalogues of omics cohorts and samples will foster automated bioinformatic pipelines construction. The NeuroEmbed catalogue of cohorts and samples is available at https://github.com/JoseAdrian3/NeuroEmbed.
While modern segmentation models often prioritize performance over practicality, we advocate a design philosophy prioritizing simplicity and efficiency, and attempted high performance segmentation model design. This paper presents SimpleUNet, a scalable ultra-lightweight medical image segmentation model with three key innovations: (1) A partial feature selection mechanism in skip connections for redundancy reduction while enhancing segmentation performance; (2) A fixed-width architecture that prevents exponential parameter growth across network stages; (3) An adaptive feature fusion module achieving enhanced representation with minimal computational overhead. With a record-breaking 16 KB parameter configuration, SimpleUNet outperforms LBUNet and other lightweight benchmarks across multiple public datasets. The 0.67 MB variant achieves superior efficiency (8.60 GFLOPs) and accuracy, attaining a mean DSC/IoU of 85.76%/75.60% on multi-center breast lesion datasets, surpassing both U-Net and TransUNet. Evaluations on skin lesion datasets (ISIC 2017/2018: mDice 84.86%/88.77%) and endoscopic polyp segmentation (KVASIR-SEG: 86.46%/76.48% mDice/mIoU) confirm consistent dominance over state-of-the-art models. This work demonstrates that extreme model compression need not compromise performance, providing new insights for efficient and accurate medical image segmentation. Codes can be found at https://github.com/Frankyu5666666/SimpleUNet.
This work presents a generalizable framework to transfer relative depth to metric depth. Current monocular depth estimation methods are mainly divided into metric depth estimation (MMDE) and relative depth estimation (MRDE). MMDEs estimate depth in metric scale but are often limited to a specific domain. MRDEs generalize well across different domains, but with uncertain scales which hinders downstream applications. To this end, we aim to build up a framework to solve scale uncertainty and transfer relative depth to metric depth. Previous methods used language as input and estimated two factors for conducting rescaling. Our approach, TR2M, utilizes both text description and image as inputs and estimates two rescale maps to transfer relative depth to metric depth at pixel level. Features from two modalities are fused with a cross-modality attention module to better capture scale information. A strategy is designed to construct and filter confident pseudo metric depth for more comprehensive supervision. We also develop scale-oriented contrastive learning to utilize depth distribution as guidance to enforce the model learning about intrinsic knowledge aligning with the scale distribution. TR2M only exploits a small number of trainable parameters to train on datasets in various domains and experiments not only demonstrate TR2M's great performance in seen datasets but also reveal superior zero-shot capabilities on five unseen datasets. We show the huge potential in pixel-wise transferring relative depth to metric depth with language assistance. (Code is available at: https://github.com/BeileiCui/TR2M)
Large Language Models (LLMs) have shown strong potential for recommendation by framing item prediction as a token-by-token language generation task. However, existing methods treat all item tokens equally, simply pursuing likelihood maximization during both optimization and decoding. This overlooks crucial token-level differences in decisiveness-many tokens contribute little to item discrimination yet can dominate optimization or decoding. To quantify token decisiveness, we propose a novel perspective that models item generation as a decision process, measuring token decisiveness by the Information Gain (IG) each token provides in reducing uncertainty about the generated item. Our empirical analysis reveals that most tokens have low IG but often correspond to high logits, disproportionately influencing training loss and decoding, which may impair model performance. Building on these insights, we introduce an Information Gain-based Decisiveness-aware Token handling (IGD) strategy that integrates token decisiveness into both tuning and decoding. Specifically, IGD downweights low-IG tokens during tuning and rebalances decoding to emphasize tokens with high IG. In this way, IGD moves beyond pure likelihood maximization, effectively prioritizing high-decisiveness tokens. Extensive experiments on four benchmark datasets with two LLM backbones demonstrate that IGD consistently improves recommendation accuracy, achieving significant gains on widely used ranking metrics compared to strong baselines.
As large language models (LLMs) are trained on massive datasets, they have raised significant privacy and ethical concerns due to their potential to inadvertently retain sensitive information. Unlearning seeks to selectively remove specific data from trained models, such as personal information or copyrighted content. Current approaches targeting specific output sequences at the token level often fail to achieve complete forgetting and remain susceptible to prompt rephrasing. We propose Align-then-Unlearn, a novel framework that performs unlearning in the semantic embedding space rather than directly on output tokens. Align-then-Unlearn first augments the LLM with an embedding prediction module trained to anticipate future context representations. Unlearning is then achieved by fine-tuning the model to minimize the similarity between these predicted embeddings and a target embedding that represents the concept to be removed. Initial results show that Align-then-Unlearn effectively removes targeted knowledge with minimal degradation in overall model utility. These findings suggest that embedding-based unlearning offers a promising and robust approach to removing conceptual knowledge. Our code is available at https://github.com/ExplainableML/align-then-unlearn.
ADMM is a popular method for federated deep learning which originated in the 1970s and, even though many new variants of it have been proposed since then, its core algorithmic structure has remained unchanged. Here, we take a major departure from the old structure and present a fundamentally new way to derive and extend federated ADMM. We propose to use a structure called Bayesian Duality which exploits a duality of the posterior distributions obtained by solving a variational-Bayesian reformulation of the original problem. We show that this naturally recovers the original ADMM when isotropic Gaussian posteriors are used, and yields non-trivial extensions for other posterior forms. For instance, full-covariance Gaussians lead to Newton-like variants of ADMM, while diagonal covariances result in a cheap Adam-like variant. This is especially useful to handle heterogeneity in federated deep learning, giving up to 7% accuracy improvements over recent baselines. Our work opens a new Bayesian path to improve primal-dual methods.
Sequential recommender systems aim to model users' evolving preferences by capturing patterns in their historical interactions. Recent advances in this area have leveraged deep neural networks and attention mechanisms to effectively represent sequential behaviors and time-sensitive interests. In this work, we propose C-TLSAN (Content-Enhanced Time-Aware Long- and Short-Term Attention Network), an extension of the TLSAN architecture that jointly models long- and short-term user preferences while incorporating semantic content associated with items, such as product descriptions. C-TLSAN enriches the recommendation pipeline by embedding textual content linked to users' historical interactions directly into both long-term and short-term attention layers. This allows the model to learn from both behavioral patterns and rich item content, enhancing user and item representations across temporal dimensions. By fusing sequential signals with textual semantics, our approach improves the expressiveness and personalization capacity of recommendation systems. We conduct extensive experiments on large-scale Amazon datasets, benchmarking C-TLSAN against state-of-the-art baselines, including recent sequential recommenders based on Large Language Models (LLMs), which represent interaction history and predictions in text form. Empirical results demonstrate that C-TLSAN consistently outperforms strong baselines in next-item prediction tasks. Notably, it improves AUC by 1.66%, Recall@10 by 93.99%, and Precision@10 by 94.80% on average over the best-performing baseline (TLSAN) across 10 Amazon product categories. These results highlight the value of integrating content-aware enhancements into temporal modeling frameworks for sequential recommendation. Our code is available at https://github.com/booml247/cTLSAN.
Language models can retain dangerous knowledge and skills even after extensive safety fine-tuning, posing both misuse and misalignment risks. Recent studies show that even specialized unlearning methods can be easily reversed. To address this, we systematically evaluate many existing and novel components of unlearning methods and identify ones crucial for irreversible unlearning. We introduce Disruption Masking, a technique in which we only allow updating weights, where the signs of the unlearning gradient and the retaining gradient are the same. This ensures all updates are non-disruptive. Additionally, we identify the need for normalizing the unlearning gradients, and also confirm the usefulness of meta-learning. We combine these insights into MUDMAN (Meta-Unlearning with Disruption Masking and Normalization) and validate its effectiveness at preventing the recovery of dangerous capabilities. MUDMAN outperforms the prior TAR method by 40\%, setting a new state-of-the-art for robust unlearning.
A widely recognized limitation of molecular prediction models is their reliance on structures observed in the training data, resulting in poor generalization to out-of-distribution compounds. Yet in drug discovery, the compounds most critical for advancing research often lie beyond the training set, making the bias toward the training data particularly problematic. This mismatch introduces substantial covariate shift, under which standard deep learning models produce unstable and inaccurate predictions. Furthermore, the scarcity of labeled data, stemming from the onerous and costly nature of experimental validation, further exacerbates the difficulty of achieving reliable generalization. To address these limitations, we propose a novel meta-learning-based approach that leverages unlabeled data to interpolate between in-distribution (ID) and out-of-distribution (OOD) data, enabling the model to meta-learn how to generalize beyond the training distribution. We demonstrate significant performance gains over state-of-the-art methods on challenging real-world datasets that exhibit substantial covariate shift.
Parameter Efficient FineTuning (PEFT), such as Low-Rank Adaptation (LoRA), aligns pre-trained Large Language Models (LLMs) to particular downstream tasks in a resource-efficient manner. Because efficiency has been the main metric of progress, very little attention has been put in understanding possible catastrophic failures. We uncover one such failure: PEFT encourages a model to search for shortcut solutions to solve its fine-tuning tasks. When very small amount of tokens, e.g., one token per prompt, are correlated with downstream task classes, PEFT makes any pretrained model rely predominantly on that token for decision making. While such spurious tokens may emerge accidentally from incorrect data cleaning, it also opens opportunities for malevolent parties to control a model's behavior from Seamless Spurious Token Injection (SSTI). In SSTI, a small amount of tokens correlated with downstream classes are injected by the dataset creators. At test time, the finetuned LLM's behavior can be controlled solely by injecting those few tokens. We apply SSTI across models from three families (Snowflake Arctic, Apple OpenELM, and Meta LLaMA-3) and four diverse datasets (IMDB, Financial Classification, CommonSense QA, and Bias in Bios). Our findings reveal three astonishing behaviors. First, as few as a single token of SSTI is sufficient to steer a model's decision making. Second, for light SSTI, the reliance on spurious tokens is proportional to the LoRA rank. Lastly, with aggressive SSTI, larger LoRA rank values become preferable to small rank values as it makes the model attend to non-spurious tokens, hence improving robustness.