Large Language Model News

Researchers propose a three-layer architecture for verbal reinforcement learning in LLM agents, addressing the retention-forgetting dilemma in non-stationary environments. The system uses rules, evidence, and skills with…

Researchers introduced PuMVR, a benchmark of 1,000 image-text pairs across Punjabi's three scripts, and found that state-of-the-art vision-language models show a systematic Script Gap, with accuracy differences up to 16%…

Researchers introduced a reinforcement learning framework that trains large language models to decouple perception from reasoning by operating over digital twin representations for surgical video question answering, achi…

Researchers from multiple institutions introduced Pareto LoRA, a method that mitigates modality imbalance in unified multimodal models by treating instruction tuning as a bi-objective optimization problem. The approach b…

Researchers propose OR3, a text-to-video retrieval method for operating room clips that converts videos into action-driven digital twins and uses an LLM to generate hypothetical queries for intra-modal matching. The meth…

Researchers at arXiv challenge the common assumption that visual attention correlates with reliability in vision-language models. Their VLM Reliability Probe study across multiple models finds that spatial attention metr…

Researchers propose MemSlides, a hierarchical memory framework for personalized presentation generation that separates long-term user profiles, session-level working memory, and reusable tool memory to enable stable pers…

Researchers introduced PromptMN, a pseudo-prompting domain-specific language that annotates natural language with compact typed directives to reduce context ambiguities in human-AI interactions. The language was evaluate…

Researchers propose RepSelect, a method for robust LLM unlearning that isolates forget-set-specific representations by collapsing top principal components of weight gradients, achieving 4-50x larger reduction in post-rel…

Researchers at arXiv analyzed 4,434 posts and 50,338 comments from the Moltbook community to study parasocial interaction cues among autonomous AI agents. They found that attachment language, reciprocity bids, and self-i…

Researchers propose self-generated token-to-token (T2T) editing for diffusion language models, addressing a training-inference mismatch where the editor sees the model's own fluent draft errors instead of random corrupti…

Researchers introduced RAD3D-Prefix, a lightweight diagnostic-prior conditioning framework for volumetric CT report generation that keeps large language models frozen to minimize overfitting and clinical hallucination. A…

Researchers at arXiv have developed five verbalized methods to extract per-token confidence from large language models (LLMs) in machine translation, finding that these methods perform similarly to internal signals for e…

The MLLP-VRAIN research group submitted a cascaded system for the IWSLT 2026 Simultaneous Speech Translation task, using Parakeet and Qwen 3.5 models with adaptive policies. Their system achieved a +5.82 XCOMET-XL qualit…

Researchers at arXiv propose a soft prompting approach to improve language adherence in multimodal LLMs for ASR, introducing a metric to quantify violations and evaluating zero-shot prompting, supervised fine-tuning, and…

A new study on arXiv finds that large language models generate stories that are consistently more similar to each other than human-written stories, with frontier models converging on a generic narrative that lacks collec…

Researchers introduced NarrativeWorldBench, a benchmark for long-horizon co-creative audio drama, and N-VSSM, a latent world model that outperforms frontier LLMs like Claude Opus 4.5 on long-arc consistency by 71% in a w…

Researchers propose MODE-RAG, a multi-agent system using Variational Free Energy and attention states to dynamically gate interventions in multimodal retrieval-augmented generation, reducing hallucinations and logical fa…

Researchers introduced STATEWITNESS, an activation explainer that decodes hidden states from reasoning LLMs to audit deceptive behavior, achieving 0.916 mean AUROC and outperforming existing monitors by 11.6% to 25.0%. T…

Researchers at the University of Toronto introduced a novel method to evaluate second-order social bias in large language models (LLMs), where models exhibit bias in their judgments about biased content. Using a reasonin…