and “Jailbreak Distillation: Renewable Safety Benchmarking” by Jingyu Zhang, Benjamin Van Durme, @danielkhashabi.bsky.social, and more proposed a novel benchmark construction framework that “distills” jailbreak attacks into high-quality and easily-updatable safety 🦺 benchmarks: (7/7)

In “HiMATE: A Hierarchical Multi-Agent Framework for Machine Translation Evaluation,” @phikoehn.bsky.social, @shijiezhang.bsky.social, & more developed a hierarchical multi-agent system grounded in the MQM error typology, enabling granular evaluation of subtype errors: (6/7)

Weiting Tan, @phikoehn.bsky.social, and Meta AI researchers presented an audio-visual language model for expressive speech 🗣️ generation in ”Seeing is Believing: Emotion-Aware Audio-Visual Language Modeling for Expressive Speech Generation”: (5/7)

“Speech Vecalign: An Embedding-Based Method for Aligning Parallel Speech Documents” by alum Chutong Meng and @phikoehn.bsky.social presented a parallel speech 🗣️ document alignment method that monotonically aligns speech segment embeddings and doesn’t depend on text transcriptions 📄: (4/7)

In “Learn and Unlearn: Addressing Misinformation in Multilingual LLMs,” TaiMing Lu and @phikoehn.bsky.social investigated the propagation of information in multilingual large language models and evaluated the efficacy of various unlearning methods: (3/7)

Undergrad (!) Alessa Carbo and @enalisnick.bsky.social introduced a novel graph neural network that separates temporal dynamics from static #ASL 🤟 handshape configurations in “Improving Handshape Representations for Sign Language Processing: A Graph Neural Network Approach”: (2/7)