LongCat-Image Technical Report

A recent study explores sound symbolism, revealing how Multimodal Large Language Models (MLLMs) interpret auditory information in human languages. The research introduces LEX-ICON, a dataset comprising 8,052 words and 2,930 pseudo-words across four languages, examining MLLMs' phonetic iconicity through phoneme-level attention scores.

SwissGov-RSD has been introduced as the first naturalistic, document-level, cross-lingual dataset designed for recognizing semantic differences across documents in multiple languages, including English, German, French, and Italian. This dataset includes 224 multi-parallel documents annotated at the token level by human annotators, addressing a previously underexplored area in text generation evaluation and multilingual content alignment.

GUMBridge has been introduced as a new resource for bridging anaphora, encompassing 16 diverse genres of English. This corpus aims to provide comprehensive coverage of the phenomenon, which involves understanding references in discourse that depend on previous entities, such as identifying 'the door' as belonging to 'a house.'

A new benchmark corpus for Telugu-English speech translation, named TeluguST-46, has been developed, comprising 46 hours of manually verified data. This initiative addresses the underexplored area of speech translation for Telugu, a language spoken by over 80 million people, and includes a systematic evaluation of various translation architectures, highlighting the performance of IndicWhisper + IndicMT and finetuned SeamlessM4T models.

A systematic assessment has been conducted on Chinese language models (LMs) using the ZhoBLiMP benchmark, which includes over 100 linguistic minimal pairs. The study reveals that LMs struggle with certain linguistic constructs in Chinese, such as anaphors and quantifiers, even with models up to 32 billion parameters. A new metric, sub-linear length normalized log-probabilities (SLLN-LP), was introduced to address biases in sentence length.

Structure-inducing Language Models (SiLM) have been trained from scratch using three different architectures: Structformer, UDGN, and GPST, focusing on their syntactic generalization capabilities and performance across various NLP tasks. The study evaluates the models on their induced syntactic representations, grammaticality judgment tasks, and training dynamics, revealing no single architecture excels across all metrics.

The TRepLiNa method, which combines Centered Kernel Alignment (CKA) and REPINA, has been introduced to enhance low-resource machine translation, particularly for Indian languages like Mundari, Santali, and Bhili, using the Aya-23 8B model. This approach aims to improve translation quality from low-resource languages to high-resource languages such as Hindi and English.