Recent research papers published on arxiv.org address critical challenges in large language model performance across safety, reasoning, and information retrieval.
According to a paper accepted at EMNLP 2026, researchers introduced Safety-aware Contrastive Decoding (SafeCoDe), a framework designed to improve context-aware safety decisions in Multimodal Large Language Models (MLLMs). The method addresses two key issues: oversensitivity, where models unjustifiably refuse benign queries, and undersensitivity, where they miss visually grounded risks. SafeCoDe operates through contrastive decoding that highlights tokens sensitive to visual context by contrasting real and Gaussian-noised images, combined with a global-aware token modulation strategy that adapts refusals based on predicted safety verdicts.
Separately, research accepted at EMNLP 2026 Findings examined cross-lingual knowledge transfer in large reasoning models. According to the study, script match—not language or family—is the primary predictor of knowledge transfer failure. The researchers found that providing key entities in their source language “disproportionately improves cross-script questions,” and developed a supervised fine-tuning approach to help models better reason about transliteration ambiguities.
Meanwhile, research accepted at COLM 2026 focused on Retrieval-Augmented Generation (RAG) reliability. According to the paper, researchers created a controlled benchmark to determine whether retrieved information is sufficient, insufficient, or conflicting. Testing across 16 language models, they found that hidden activation states from middle layers can reliably decode whether evidence supports answering a question.