Structure Before Collapse: Transient semantic geometry in next-token prediction

Neural Collapse predicts that balanced one-hot classification pushes model representations to be equally far from each other; a symmetric configuration that depends only on the output label and ignores any semantic similarity in the inputs. This creates a puzzle: next-token prediction language models are trained predominantly (as context length increases) with one-hot labels: the same context is very unlikely to appear twice in training with different labels. However, they clearly learn latent structural features. That is, despite the one-hot training regime, a language model's contextual embeddings represent the fact that the next word in ''Mary broke the ___'' is likely to be filled by tokens in the latent classes of a) medium-sized, b) rigid, c) inanimate nouns. How does gradient descent find such categorical semantic structure when co-occurrence statistics collapse to one-hot sparsity, eliminating any shared next-tokens among different contexts? To investigate this tension we identify three synthetic controlled settings where inputs have latent semantic factors but are mapped to distinct one-hot labels. We find that semantic geometry emerges early in training, and that representations cluster by shared attributes despite receiving no explicit supervision to do so. This structure is transient: with sufficient capacity and time, the model eventually reaches the predicted symmetric state where all representations are equally separated. We study this phase transition through Gram matrix analysis and propose a preliminary modification to the commonly used unconstrained features model to capture the emergent semantic geometry.

Disentangling Latent Risk Pathways via Bayesian Hypergraph Inference

Electronic health records (EHR) pose large-scale multi-disease modeling problems in which many outcomes are rare and strongly influenced by shared risk factors. While modern approaches achieve strong predictive performance, they often treat diseases independently or rely on black-box architectures, offering limited insight into how risk factors organize disease risk and little principled uncertainty quantification. We introduce a Bayesian hypergraph inference framework that reframes multi-disease modeling around latent, risk-factor-modulated disease pathways. Risk factors act on hyperedges, latent disease subsets with shared risk patterns, allowing diseases to participate in multiple distinct pathways and enabling interpretable, higher-order structure beyond pairwise associations. A repulsion prior encourages parsimonious and identifiable structure, while posterior inference provides calibrated uncertainty over both disease groupings and risk-factor influence. To enable scalable inference on large EHR datasets, we develop a structured variational inference algorithm that preserves logical dependencies among hyperedge existence, disease membership, and pathway-level effects. Experiments on simulated data and UK Biobank demonstrate stable and interpretable disease pathway structure, well-calibrated uncertainty, improved estimation for rare diseases, and competitive predictive performance.

Workflow Closure Is Not Scientific Closure in Auto-Research Systems

This paper argues that workflow closure is not scientific closure in auto-research systems. Current systems can increasingly complete research-like loops internally, moving from idea generation to experiment execution, writing, and self-evaluation. That achievement is real, but it does not by itself give the resulting outputs scientific standing. We argue that trustworthy auto-research should not aim for autonomous self-sufficiency, but should aim for autonomous execution under non-autonomous epistemic control. Based on a survey of more than 100 recent papers and repositories in this rapidly emerging area, together with a structured audit of 21 representative systems, we diagnose a recurring and structurally connected failure pattern: objective collapse, in which single-proxy targets replace multi-objective scientific aims; validation collapse, in which internal self-evaluation replaces independent validation; and acceptance collapse, in which benchmark scores or publication-shaped artifacts replace mechanisms for domain-level critique, reuse, and integration. These collapses are not inherent limits of autonomy but correctable design choices. Accordingly, we outline potential remedies across objective signal, validation, and output pathway to spark community discussion.

Why Loss Re-weighting Works If You Stop Early: Training Dynamics of Unconstrained Features

The application of loss reweighting in modern deep learning presents a nuanced picture. While it fails to alter the terminal learning phase in overparameterized deep neural networks (DNNs) trained on high-dimensional datasets, empirical evidence consistently shows it offers significant benefits early in training. To transparently demonstrate and analyze this phenomenon, we introduce a small-scale model (SSM). This model is specifically designed to abstract the inherent complexities of both the DNN architecture and the input data, while maintaining key information about the structure of imbalance within its spectral components. On the one hand, the SSM reveals how vanilla empirical risk minimization preferentially learns to distinguish majority classes over minorities early in training, consequently delaying minority learning. In stark contrast, reweighting restores balanced learning dynamics, enabling the simultaneous learning of features associated with both majorities and minorities.

How Muon's Spectral Design Benefits Generalization: A Study on Imbalanced Data

The growing adoption of spectrum-aware matrix-valued optimizers such as Muon and Shampoo in deep learning motivates a systematic study of their generalization properties and, in particular, when they might outperform competitive algorithms. We approach this question by introducing appropriate simplifying abstractions as follows: First, we use imbalanced data as a testbed. Second, we study the canonical form of such optimizers, which is Spectral Gradient Descent (SpecGD) -- each update step is $UV^T$ where $U\Sigma V^T$ is the truncated SVD of the gradient. Third, within this framework we identify a canonical setting for which we precisely quantify when SpecGD outperforms vanilla Euclidean GD. For a Gaussian mixture data model and both linear and bilinear models, we show that unlike GD, which prioritizes learning dominant principal components of the data first, SpecGD learns all principal components of the data at equal rates. We demonstrate how this translates to a growing gap in balanced accuracy favoring SpecGD early in training and further show that the gap remains consistent even when the GD counterpart uses adaptive step-sizes via normalization. By extending the analysis to deep linear models, we show that depth amplifies these effects. We empirically verify our theoretical findings on a variety of imbalanced datasets. Our experiments compare practical variants of spectral methods, like Muon and Shampoo, against their Euclidean counterparts and Adam. The results validate our findings that these spectral optimizers achieve superior generalization by promoting a more balanced learning of the data's underlying components.

On the Geometry of Semantics in Next-token Prediction

Modern language models demonstrate a remarkable ability to capture linguistic meaning despite being trained solely through next-token prediction (NTP). We investigate how this conceptually simple training objective leads models to extract and encode latent semantic and grammatical concepts. Our analysis reveals that NTP optimization implicitly guides models to encode concepts via singular value decomposition (SVD) factors of a centered data-sparsity matrix that captures next-word co-occurrence patterns. While the model never explicitly constructs this matrix, learned word and context embeddings effectively factor it to capture linguistic structure. We find that the most important SVD factors are learned first during training, motivating the use of spectral clustering of embeddings to identify human-interpretable semantics, including both classical k-means and a new orthant-based method directly motivated by our interpretation of concepts. Overall, our work bridges distributional semantics, neural collapse geometry, and neural network training dynamics, providing insights into how NTP's implicit biases shape the emergence of meaning representations in language models.

DARE the Extreme: Revisiting Delta-Parameter Pruning For Fine-Tuned Models

Storing open-source fine-tuned models separately introduces redundancy and increases response times in applications utilizing multiple models. Delta-parameter pruning (DPP), particularly the random drop and rescale (DARE) method proposed by Yu et al., addresses this by pruning the majority of delta parameters--the differences between fine-tuned and pre-trained model weights--while typically maintaining minimal performance loss. However, DARE fails when either the pruning rate or the magnitude of the delta parameters is large. We highlight two key reasons for this failure: (1) an excessively large rescaling factor as pruning rates increase, and (2) high mean and variance in the delta parameters. To push DARE's limits, we introduce DAREx (DARE the eXtreme), which features two algorithmic improvements: (1) DAREx-q, a rescaling factor modification that significantly boosts performance at high pruning rates (e.g., >30 % on COLA and SST2 for encoder models, with even greater gains in decoder models), and (2) DAREx-L2, which combines DARE with AdamR, an in-training method that applies appropriate delta regularization before DPP. We also demonstrate that DAREx-q can be seamlessly combined with vanilla parameter-efficient fine-tuning techniques like LoRA and can facilitate structural DPP. Additionally, we revisit the application of importance-based pruning techniques within DPP, demonstrating that they outperform random-based methods when delta parameters are large. Through this comprehensive study, we develop a pipeline for selecting the most appropriate DPP method under various practical scenarios.

Implicit Geometry of Next-token Prediction: From Language Sparsity Patterns to Model Representations

Next-token prediction (NTP) over large text corpora has become the go-to paradigm to train large language models. Yet, it remains unclear how NTP influences the mapping of linguistic patterns to geometric properties of the resulting model representations. We frame training of large language models as soft-label classification over sparse probabilistic label vectors, coupled with an analytical approximation that allows unrestricted generation of context embeddings. This approach links NTP training to rank-constrained, nuclear-norm regularized optimization in the logit domain, offering a framework for analyzing the geometry of word and context embeddings. In large embedding spaces, we find that NTP implicitly favors learning logits with a sparse plus low-rank structure. While the sparse component captures the co-occurrence frequency of context-word pairs, the orthogonal low-rank component, which becomes dominant as training progresses, depends solely on the sparsity pattern of the co-occurrence matrix. Consequently, when projected onto an appropriate subspace, representations of contexts that are followed by the same set of next-tokens collapse, a phenomenon we term subspace-collapse. We validate our findings on synthetic and small-scale real language datasets. Finally, we outline potential research directions aimed at deepening the understanding of NTP's influence on the learning of linguistic patterns and regularities.

Learnable Community-Aware Transformer for Brain Connectome Analysis with Token Clustering

Neuroscientific research has revealed that the complex brain network can be organized into distinct functional communities, each characterized by a cohesive group of regions of interest (ROIs) with strong interconnections. These communities play a crucial role in comprehending the functional organization of the brain and its implications for neurological conditions, including Autism Spectrum Disorder (ASD) and biological differences, such as in gender. Traditional models have been constrained by the necessity of predefined community clusters, limiting their flexibility and adaptability in deciphering the brain's functional organization. Furthermore, these models were restricted by a fixed number of communities, hindering their ability to accurately represent the brain's dynamic nature. In this study, we present a token clustering brain transformer-based model ($\texttt{TC-BrainTF}$) for joint community clustering and classification. Our approach proposes a novel token clustering (TC) module based on the transformer architecture, which utilizes learnable prompt tokens with orthogonal loss where each ROI embedding is projected onto the prompt embedding space, effectively clustering ROIs into communities and reducing the dimensions of the node representation via merging with communities. Our results demonstrate that our learnable community-aware model $\texttt{TC-BrainTF}$ offers improved accuracy in identifying ASD and classifying genders through rigorous testing on ABIDE and HCP datasets. Additionally, the qualitative analysis on $\texttt{TC-BrainTF}$ has demonstrated the effectiveness of the designed TC module and its relevance to neuroscience interpretations.

Learning High-Order Relationships of Brain Regions

Discovering reliable and informative interactions among brain regions from functional magnetic resonance imaging (fMRI) signals is essential in neuroscientific predictions of cognition. Most of the current methods fail to accurately characterize those interactions because they only focus on pairwise connections and overlook the high-order relationships of brain regions. We delve into this problem and argue that these high-order relationships should be maximally informative and minimally redundant (MIMR). However, identifying such high-order relationships is challenging and highly under-explored. Methods that can be tailored to our context are also non-existent. In response to this gap, we propose a novel method named HyBRiD that aims to extract MIMR high-order relationships from fMRI data. HyBRiD employs a Constructor to identify hyperedge structures, and a Weighter to compute a weight for each hyperedge. HyBRiD achieves the MIMR objective through an innovative information bottleneck framework named multi-head drop-bottleneck with theoretical guarantees. Our comprehensive experiments demonstrate the effectiveness of our model. Our model outperforms the state-of-the-art predictive model by an average of 12.1%, regarding the quality of hyperedges measured by CPM, a standard protocol for studying brain connections.