A new arXiv paper argues that drug-target binding affinity prediction improves when models preserve both chemical and geometric details of the interacting molecules. The authors focus on representation fidelity as a way to capture the subtle structural features that determine molecular recognition. The work falls under machine learning research rather than a released product or model.
A new arXiv paper analyzes how representations inside transformer models change across layers, treating each learned update as a combination of a component that keeps the existing direction and one that shifts it elsewhere. The authors frame this as a functional geometry, aiming to explain what the model preserves versus reorients as information flows through the network.
A new arXiv paper examines a structural risk for agents that act on compressed summaries of their history: when the summary conflates histories that call for different optimal actions, no decision rule defined over that summary can avoid a persistent per-round loss. The authors propose sequential self-certification of representation adequacy, framed around achieving minimum task loss. The work is cross-listed in cs.AI and cs.LG.
A new arXiv paper reframes the AddUNet model through the lens of perfect reconstruction and describes how it can operate at full rate. The authors introduce a residual full-rate perfect-reconstruction design aimed at task-directed representation learning, building on the architecture's survivor-skip structure. The work is theoretical and architectural rather than a released product or benchmark result.
Researchers propose a representation learning framework built on SL(n) that operates in an intrinsic mixed-curvature space rather than relying on manually composed product manifolds. The approach aims to provide higher curvature capacity and deeper order-aware composition for capturing complex geometric structure. It is presented as an alternative to existing product-manifold methods that require hand-specified curvature combinations.
A new arXiv paper introduces MANAS-2, an EEG foundation model that reframes masked reconstruction as a constrained problem. The authors argue that optimizing reconstruction on low-SNR brain waveforms does not reliably yield the most useful latent representations. The model is presented as a step toward EEG pretraining objectives that better match downstream usefulness.
A new arXiv paper analyzes spectral methods for representation learning, where an encoder is trained once, frozen, and then reused by lightweight probes on downstream tasks. The authors derive tight convergence rates and propose a minimal, one-line correction that determines when off-the-shelf features suffice and when they need adjustment.
A new arXiv paper proposes using language guidance to learn material representations that transfer across different vision-based tactile sensors. Because each sensor produces distinct readings of the same material, the approach aims to close that gap by grounding learned features in shared language descriptions. The work targets robotic touch perception, where properties such as softness and texture are difficult to recover from vision alone.
A new arXiv preprint examines how procedural audio should be scaled when used as a data source for learning transferable audio representations. The authors also ask whether training choices originally developed on natural audio still hold when the source is procedurally generated. The work aims to clarify design principles for procedural audio pre-training, an area the authors say lacks settled guidance.
A new arXiv paper argues that standard urban region representation learning, which merges data such as mobility flows, points of interest and land-use, can be misled by confounding factors that create spurious correlations. The authors introduce a confounder-aware multi-view approach intended to improve downstream tasks like mobility analysis, public safety forecasting and service demand estimation. The work appears in the cs.AI and cs.LG listings.
A new arXiv preprint argues that strong benchmark performance does not by itself show that a model has learned compositional, environment-invariant features. The authors aim to disentangle measured accuracy from the underlying representations that support out-of-distribution generalisation, a capability often treated as a marker of biological intelligence. Their analysis is framed around how systems can transfer invariant properties from training mappings to novel compositions.
A preprint on arXiv introduces an approach called the Metric Slingshot, which frames navigational reuse as a way to achieve width-optimal structural decoupling in continual learning. The work draws on neuroscience findings about grid cells, place cells, and hippocampal indexing, which the brain uses for both spatial and non-spatial tasks. It argues that reusing navigation-related circuitry can help neural networks avoid interference across sequential tasks. Only the abstract excerpt is available, so full results and benchmarks are not yet assessed.
A new arXiv paper introduces a similarity measure built on divergence for contrastive learning with multiple augmented views. The authors note that earlier approaches combine views either in the loss or in the feature space, but largely restrict themselves to pairwise comparisons. Their method aims to capture relationships spanning more than two views at once. The work is labeled a cross-listing replacement on arXiv's machine learning section.
A preprint on arXiv cs.LG describes a method that learns representations of human cortical folding patterns, which form before birth and stay largely stable afterwards. The authors suggest these folding signatures could serve as early markers of neurodevelopment. The abstract is truncated, so reported results could not be fully assessed.
An arXiv paper presents CardioState-JEPA, a delay-aware cross-modal model that learns one shared representation of cardiac activity from ECG, photoplethysmography, and phonocardiography signals. The authors argue that current cardiac foundation models are tied to a single sensing modality, and their approach instead aligns the complementary views these three signals provide of the same heartbeat.