According to arxiv.org, researchers have introduced Safety-aware Contrastive Decoding (SafeCoDe), a framework designed to improve context-aware safety decisions in Multimodal Large Language Models (MLLMs). The paper, accepted to EMNLP 2026, describes SafeCoDe as “a lightweight and model-agnostic decoding framework that dynamically adjusts token generation based on multimodal context.” The system operates through contrastive decoding that contrasts real and Gaussian-noised images, combined with global-aware token modulation to adapt refusals based on safety predictions.
In separate research also published on arxiv.org and accepted to Findings of EMNLP 2026, researchers examined cross-lingual knowledge transfer in large reasoning models. The study analyzed performance on ECLeKTic and MultiLoKo datasets, finding that “script match - not language or family - is the primary predictor of knowledge transfer failure.” The researchers developed synthetic training samples to help models better reason about transliteration ambiguities during inference.
Additionally, according to arxiv.org, research accepted at the Third Conference on Language Modeling (COLM 2026) introduced methods for improving Retrieval-Augmented Generation (RAG). The work demonstrates that language models “internally encode whether retrieved evidence is sufficient to support answering,” using hidden activations and attention-derived features to classify information as sufficient, insufficient, or conflicting. The feature-based approach “consistently outperforms prompting-based baselines” across 16 language models.