Multimodal Unlearning Evaluations Miss Key Failure Modes
A new paper argues that current tests for multimodal unlearning only check whether models stop outputting target information, overlooking deeper issues.
Research Digest
Preprints and papers, rewritten in plain English.
Categorical Framework Recasts Parameter-Efficient Fine-Tuning as Maps
A new arXiv preprint argues that PEFT methods should be compared as maps in weight space, not just by parameter count and benchmark scores.
Sizing Inputs for Neural Surrogates of Sparse Physical Fields
A new preprint examines how practitioners should choose the spatial window and temporal history for neural surrogates that forecast physical fields from sparse sensors.
Gaussian-width bounds for offline-to-online linear bandits
A new arXiv paper analyses how offline data can improve online stochastic linear bandits when only a small number of feature directions are poorly covered.
Transformers as Learned Approximations to Hybrid Probabilistic Inference
A new arXiv paper proposes that Transformers can be understood as learned approximations to a hybrid of loopy belief propagation and expectation-maximization.
Structure-Disjoint Splits Drift as Molecular Libraries Grow
A new preprint argues that structure-disjoint splits in molecular machine learning are not stable benchmarks because their construction depends on the size of the library they are built from.
Primal-Dual Flow Matching for Sample-Wise Constrained Generation
The paper introduces a primal-dual flow matching framework for generating data that meets sample-wise constraints with a user-specified probability.
Target-Constrained Pooling Aims to Resolve Trajectory Extrapolation
A new method uses related completed trajectories to constrain the underdetermined problem of extrapolating a partial trajectory.
Evidential fusion improves emotion recognition under contradictory sensor data
New evidential transformers handle conflicting facial and vocal cues without overconfident predictions.
AlgoPerf V1 Update Aims to Make Optimizer Comparisons Fairer
New benchmark results from AlgoPerf V1 tackle the confounding effects of hyperparameter tuning and model co-design in evaluating deep learning optimizers.
Double-Take Memory Method Boosts Long-Context Recall
A new memory mechanism aims to keep distant information usable in long sequences by improving capacity, binding, and selectivity.
Kolmogorov-Arnold Operator Learns Physical Laws in Closed Form
A new neural operator aims to generalize beyond training data while exposing learned physical laws in closed form.
Federated Learning Gets Privacy and Robustness for Fault Diagnosis
A new framework trains multiclass fault-diagnosis models across distributed sites while handling noisy, sensitive data without centralising it.
Random Mask Beats Static Gradient Attribution for LoRA Continual Learning
A new arXiv preprint reports that a density-matched random mask outperforms static gradient attribution for selecting LoRA B-matrix parameters in continual learning.
Attributing Agent Failures Without Step-Level Labels
A new method identifies the decisive step in failed LLM agent trajectories without requiring expensive step-level error labels.
Iterative Refinement Proposed to Scale RNA Foundation Models
A new approach aims to decouple model size from inference depth so RNA models can capture both local and long-range features.
How Calendar Features Are Encoded in Time Series Foundation Models
A new arXiv paper examines the role of calendar information as a categorical covariate in time series foundation models.
PANDA: Dual-Path Hypergraph Network Blends Priors and Attention
A proposed architecture for hypergraph neural networks combines prior-guided and attentional paths to overcome the pitfalls of fixed structural weights.