Developer Tools B

Matrix-update non-Euclidean optimizers like Muon and Scion train transformers well but lack theory. The paper offers a heavy-tailed analysis justifying Muon's empirical advantages over Euclidean methods.

Modern image classifiers widely use global average pooling followed by a linear head. The paper shows this linearity makes GAP-based classifiers behave as multi-instance learners, prompting a rethink of global average pooling.

Autonomous robots may need decisions based on evidence no longer visible. For delayed-evidence tasks, where an early cue disappears before a later decision, TRACE introduces trajectory-routed causal memory for visuomotor imitation.

Safe RL usually relies on soft-constrained policy optimization without hard guarantees. This work proposes an approach that is provably safe while remaining scalable.

BayLing-Duplex enables native full-duplex speech interaction with a single autoregressive LLM, letting it listen and speak simultaneously. It handles natural phenomena such as overlap and hesitation.

Machine unlearning removes learned data from models, but auditing its behavior is itself costly. The paper shows that behavioral audits of unlearning incur a privacy cost.

Macrocyclic peptides are promising for intracellular targets but require joint control of non-natural chemistry, ring topology, and permeability. PepALD generates them using autoregressive latent diffusion.

Knowing how a model will perform in a new environment before deployment helps prevent harm. The paper evaluates predictions under two common sources of degradation: distribution shift and selection bias.

The paper analyzes nonlinear two-time-scale stochastic approximation, revealing a sharp phase transition under contractive assumptions and showing how to beat it.

A growing line of work equips LLM agents with graph neural networks as callable tools. The paper finds that agents defer blindly to these GNN tools, and that stronger backbones tend to defer even more.

The paper investigates the fragility of post-hoc explanations in audio deepfake detection. Introducing a psychoacoustic framework beyond image-style Lp metrics, it shows attributions can be manipulated while predictions remain unchanged.

As clinical language models enter high-stakes decision support, the paper audits how demographic associations are encoded in ClinicalBERT and shape its language predictions.