arXiv subjects

André Jonasson

Publications and source records attributed to André Jonasson.

3 recordsLinked to original sources

arXiv2026

Shieldstral

We introduce Shieldstral, a 3B-parameter policy-adaptive multimodal safety classifier that matches or outperforms models nearly 7$\times$ its size on text safety benchmarks and sets a new state of the art on multimodal safety classification. Shieldstral formulates content moderation as a binary question-answering task. This simple formulation unifies diverse moderation tasks into a single yes/no problem, enabling heterogeneous safety datasets with divergent taxonomies to be consolidated under one training framework. We present the data construction recipe, covering curation and generation of approximately 54.1M samples and a fine-grained evaluation set to evaluate policy adaptability. Together, these enable a small adaptive model to match or outperform much larger models.

cs.CL↗

arXiv2025

Rotary Offset Features in Large Language Models

Transformer-based Large Language Models (LLMs) rely on positional encodings to provide sequence position information to their attention mechanism. Rotary Positional Encodings (RoPE), which encode relative position by rotating queries and keys, have become widely used in modern LLMs. We study the features and patterns that emerge in queries and keys when using rotary embeddings and introduce the concept of rotary offset features. Our analysis reveals that these features, which frequently exhibit large activations and are often interpreted as outliers, arise consistently across layers, attention heads, and model architectures. We derive bounds predicting which rotary frequencies give rise to rotary offset features and the minimum angle between the query-key pairs for these features. We verify our predictions empirically across models of different sizes and architectures.

cs.CL↗

arXiv2021

AM-RRT*: Informed Sampling-based Planning with Assisting Metric

In this paper, we present a new algorithm that extends RRT* and RT-RRT* for online path planning in complex, dynamic environments. Sampling-based approaches often perform poorly in environments with narrow passages, a feature common to many indoor applications of mobile robots as well as computer games. Our method extends RRT-based sampling methods to enable the use of an assisting distance metric to improve performance in environments with obstacles. This assisting metric, which can be any metric that has better properties than the Euclidean metric when line of sight is blocked, is used in combination with the standard Euclidean metric in such a way that the algorithm can reap benefits from the assisting metric while maintaining the desirable properties of previous RRT variants - namely probabilistic completeness in tree coverage and asymptotic optimality in path length. We also introduce a new method of targeted rewiring, aimed at shortening search times and path lengths in tasks where the goal shifts repeatedly. We demonstrate that our method offers considerable improvements over existing multi-query planners such as RT-RRT* when using diffusion distance as an assisting metric; finding near-optimal paths with a decrease in search time of several orders of magnitude. Experimental results show planning times reduced by 99.5% and path lengths by 9.8% over existing real-time RRT planners in a variety of environments.

cs.RO↗