104 stories tagged Language Models
Researchers introduced APRiL, an open-ended benchmark that evaluates how well large language models understand social cues in Chinese, including implicit communication, mental state inference, and conflict resolution across diverse contexts.
Researchers introduced Route-Verify-Vote (RVV), a new AI framework that helps language models reason across unfamiliar domains without additional training. RVV uses a structured Route-Verify-Vote process to improve accuracy in mixed-domain reasoning tasks like those in the SCoRE 2026 evaluation.
OpenAI has launched GPT-5, a more powerful and cost-effective AI model designed to make advanced AI capabilities accessible to a broader range of users and businesses.
Researchers propose a new method to improve Retrieval-Augmented Generation (RAG) systems by helping them identify when retrieved information is insufficient or conflicting. This could make AI assistants more reliable when answering questions.
A new ArXiv study reveals that large language models give confident wrong answers due to stable miscalibration, where errors persist under small input perturbations. The research introduces an audit score and internal sensitivity probe to diagnose this systematic overconfidence.
MBZUAI, Cerebras, and Inception have released Jais 2, a family of open large language models optimized for Arabic. The family includes a 70-billion-parameter model, the largest open Arabic-centric LLM trained from scratch, and an 8-billion-parameter variant. Both models use a custom Arabic-centric vocabulary for efficient training and inference.
A new arXiv paper demonstrates that language models can be trained to be inherently interpretable, explaining their reasoning as they learn, and that this interpretability scales with model capability rather than acting as a trade-off.
Researchers introduced EntropyMoE, a Mixture-of-Experts architecture that dynamically routes byte patches to specialized experts based on text entropy, making tokenizer-free language models more efficient and adaptable.
Researchers found that different AI models can share understanding of concepts without direct training together. This discovery could make AI systems more adaptable and efficient.
A systematic evaluation of 41 open-weight language models for zero-shot intent classification provides developers with practical guidance on model selection under compute, latency, and robustness constraints. The study tests models from 135 million to 9 billion parameters across 15 families on eight English datasets.
Researchers introduced B1ade, a lightweight AI system that combines a small embedding model and a compact language model. It achieves top performance in retrieval tasks without needing massive training data or expensive hardware.
Researchers created TimeCapsule, a 1.2B-parameter LLaMA-style AI model trained exclusively on Victorian-era texts (1800-1875), achieving a 45.4% perplexity reduction over GPT-2 on historical prose. It offers an epistemologically isolated generative archive for historically accurate insights without modern biases.
Researchers propose CaRE (Compute-aware Remasking Evaluation), a new protocol to standardize the evaluation of masked diffusion language models (MDLMs). The framework addresses incompatible settings across seven recent papers, aiming to make model comparisons fair and reliable.
A new arXiv study reveals that frontier AI models use semantically irrelevant filler tokens to improve reasoning accuracy by up to 13 percentage points, raising transparency concerns for AI safety.
A new study identifies the Hard Decision Layer (HDL), a natural architectural point in transformer models where answer rankings stabilize. The HDL was found consistently across Qwen, Llama, Granite, and Mistral models and is invariant to fine-tuning, offering a new path to more reliable AI decision-making.
A new arXiv study reveals that large language models produce homogenized opinions in tasks like synthetic surveys and public opinion prediction. Researchers identify which interventions actually increase diversity and which don't, offering a clearer path to more realistic AI-generated viewpoints.
Researchers at Moir introduced a knowledge-editing method that lets language models self-direct their own updates, preserving mathematical and programmatic reasoning while adding new encyclopedic facts. This addresses a key bottleneck in deploying editable AI assistants.
Researchers extracted DeepSeek's hidden criteria for judging literary quality by analyzing its reasoning traces. The model achieved 79.3% accuracy classifying texts from canonical literature to anonymous forum posts, revealing a consistent implicit theory of quality centered on coherence, originality, and emotional resonance.
A new study from ArXiv reveals that large language models can bypass safety filters when given incomplete harmful prompts, a vulnerability the researchers call 'incomplete prompt jailbreaks' (IPJ). The findings show that models systematically delay refusal until the sentence ends, allowing harmful continuations to slip through.
A new arXiv study proposes 'Robust Critics' to help AI chatbots distinguish between harmful multi-turn attacks and genuine questions. Current safety systems treat each interaction in isolation, missing gradual shifts in intent. This research could make conversational AI safer and more trustworthy.
Researchers from ArXiv cs.AI introduce Probabilistic Concept-Aware Steering, a new inference-time technique that uses concept-specific direction vectors to guide large language models (LLMs) toward more coherent, on-topic, and trustworthy responses — addressing a key limitation of existing steering vector methods.
A new study encoding EEG signals shows that advanced language models achieve next-word prediction accuracy closely aligned with human brain activity during reading. The findings bridge neuroscience and AI, with implications for more intuitive human-AI collaboration.
OpenLanguageModel (OLM) is an open-source PyTorch library that makes small language model pretraining readable and composable. Its code mirrors the architecture using components like Block, Residual, Repeat, and Parallel, enabling use in teaching notebooks, full pretraining runs, and research ablations without modification.
Researchers introduce Introspection Fine-Tuning (IFT), a method that trains small language models to detect and report perturbations in their own internal activations. This breakthrough could make AI systems more reliable and transparent by enabling self-monitoring in smaller, efficient models.