![Spectra demonstrates improved convergence loss and downstream performance across a range of learning rates-[latex] \eta \in \{8 \times 10^{-4}, 1 \times 10^{-3}, 5 \times 10^{-3}, 1 \times 10^{-2}\} [/latex]-establishing its advantage in optimization regimes.](https://arxiv.org/html/2602.11185v1/figures/lr_ablation.png)
Researchers have discovered a key pattern in the gradients of large language models and developed a new optimizer designed to exploit this structure for faster, more efficient training.
A new technique fortifies the privacy of network communities by subtly altering both the connections and characteristics of nodes, making them harder for machine learning models to identify.
Researchers have developed a new framework, AssetFormer, that uses the power of autoregressive transformers to create customizable 3D models from text prompts.
![The expert model classifies textual tokens as either factual content or hallucination, assigning a binary label of [latex]0[/latex] to indicate factual accuracy and [latex]1[/latex] to denote fabricated content.](https://arxiv.org/html/2602.11167v1/figures/example_hallucination_marking.png)
New research sheds light on why large language models confidently generate incorrect statements, and how internal analysis can reveal patterns in these ‘hallucinations’.
Researchers have developed a novel framework that combines the benefits of additive models with the power of expert systems to achieve both high accuracy and clear interpretability.
New research pinpoints how users behave when they become overly dependent on conversational AI, offering critical insights for building more responsible interfaces.
A new study reveals that common data imbalance issues significantly hinder the performance of deep learning models used to identify software vulnerabilities.
![A system trains itself to reason by internally rewarding lines of thought that bolster its confidence in a correct answer, circumventing the limitations of external verification-a process demonstrated by a model that eschews reliance on externally defined rewards or pre-authored reasoning traces, instead generating reasoning [latex]z[/latex] based solely on a question [latex]x[/latex] and a reference answer [latex]y^{\star}[/latex].](https://arxiv.org/html/2602.11549v1/x1.png)
New research explores a method for training large language models to develop robust reasoning skills by rewarding internally consistent thought processes.
A new framework uses the power of large language models to create adaptable anomaly detection systems that generalize across diverse and unseen graph datasets.
New research shows that influencing the learning of investment agents can overcome common hurdles in climate-focused financial modeling and lead to more effective sustainability outcomes.