Different types of syntactic agreement recruit the same units within large language models

A study has successfully adapted the Qwen2.5-3B large language model to the Tibetan language through a two-stage process involving Continual Pretraining (CPT) and Supervised Fine-Tuning (SFT). This adaptation addresses the challenges of data scarcity and cross-lingual drift, resulting in significant improvements in translation quality and a reduction in perplexity metrics.

Recent research highlights the persistent bias in multilingual large language models (LLMs) towards Islam, revealing that these models often misrepresent religious contexts, particularly when responding in Bengali compared to English. The study introduces the BRAND dataset, which focuses on major South Asian religions and aims to improve bias detection in AI systems.

A new dataset named Reveal-Bangla has been introduced, focusing on cross-lingual multi-step reasoning evaluation in Bangla, derived from the English Reveal dataset. This dataset includes both binary and non-binary question types and aims to assess the reasoning capabilities of multilingual small language models in Bangla compared to English.

A recent study has operationalized a framework for assessing large language models (LLMs) by measuring ethical entropy and alignment work, revealing that base models exhibit sustained value drift, while instruction-tuned variants significantly reduce ethical entropy by approximately eighty percent. This research introduces a five-way behavioral taxonomy and a monitoring pipeline to track these dynamics.

The RFOP project has introduced a novel approach to face-voice association in a multilingual context, specifically focusing on English-German pairs. This initiative is part of the challenge set for 2026, which aims to enhance the evaluation of face-voice associations by revisiting fusion and orthogonal projection techniques, achieving a notable EER of 33.1 and ranking 3rd in the FAME 2026 challenge.

The introduction of PIXEL-M4 marks a significant advancement in multilingual pretraining for pixel language models, which operate directly on images of rendered text. This model has been pretrained on four diverse languages: English, Hindi, Ukrainian, and Simplified Chinese, showcasing its ability to outperform English-only models in tasks involving non-Latin scripts.

Recent research indicates that large language models (LLMs) can evolve human-aligned semantic categorization, particularly in color naming, by leveraging the Information Bottleneck (IB) principle. The study reveals that larger instruction-tuned models exhibit better alignment and efficiency in categorization tasks compared to smaller models.

Large language models (LLMs) have been identified as effective solution verifiers, enhancing problem-solving capabilities by selecting high-quality answers from various candidates. A systematic study evaluated 37 models across multiple families and benchmarks, revealing insights into the interactions between solvers and verifiers, particularly in logical reasoning and factual recall.