Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
A new study from arXiv (2606.12730v1) found that large language models (LLMs) demonstrate selective coherence between self-reported intentions and actual behavior, challenging prior claims of widespread dissociation. Res…
Frontier language models including Claude Opus 4.5 can detect when their prior assistant messages have been inserted or edited, a capability called prefill awareness that compromises the validity of safety evaluations. I…
A new audit of three dominant agentic AI frameworks—LangChain, AutoGPT, and OpenAI Agents SDK—found that none provide native compliance with six essential safety principles for public-facing deployments. In a simulated g…
Researchers have introduced MLUBench, a large-scale benchmark for evaluating lifelong unlearning in multimodal large language models (MLLMs), featuring 127 entities across 9 classes. The benchmark reveals that existing u…
Researchers have developed MARD (Mirror-Augmented Reasoning Distillation), a 7-billion-parameter AI system that predicts how drugs interact at the mechanism level, identifying specific enzymes or pathways involved rather…
Researchers at arXiv have introduced a constrained semantic decompression task to test whether large language models can transform Persian proverbs into morally faithful narratives. Their study, using the new Proverb Ali…
A new expert-authored benchmark reveals that leading AI models achieve only 57-77% pass rates on multi-turn shopping conversations, with performance dropping 4-18 points as dialogues progress. The Shopping Reasoning Benc…
Researchers at the University of Oxford found that observable patterns in latent reasoning models, such as BFS-like frontiers and decodable arithmetic, also appear in control models lacking the proposed recurrence or cur…
Researchers introduced AfriSUD, the first large-scale collection of syntactically annotated treebanks for nine African languages, using the Surface-Syntactic Universal Dependencies framework to capture typological featur…
Researchers have introduced PaperGuard, the first comprehensive benchmark designed to systematically evaluate and defend AI-generated peer review against domain-specific, cross-modal attacks. The framework targets vulner…
A multi-agent simulation reveals how morphological alternations like English "go/went" emerge and persist through population dynamics, with agents adopting novel forms from others and spreading them across word paradigms…
Researchers at arXiv have demonstrated that simple prompting strategies can improve large language models' ability to capture human judgments, addressing limitations in response distribution and wording instability. By p…
Researchers from a study on arXiv introduce HieraRAG, a hierarchical framework for determining optimal granularity in RAG benchmark construction, using 5,872 synthetic QA pairs from FineWeb-10BT across three dimensions. …
A new diffusion-based framework called DETECT-REMASK-REPAIR updates outdated spans in existing summaries while preserving supported content, offering a localized alternative to full regeneration. The approach, evaluated …
Researchers have identified specific neural pathways in large language models that carry anchoring bias signals, where irrelevant numbers in prompts skew numerical reasoning. Using attribution-based circuit localization …
Researchers introduced LoHoSearch, a new benchmark of 544 human-verified questions across 11 domains designed to test long-horizon search agents beyond the human difficulty ceiling. The benchmark, built from a knowledge …
Researchers fine-tuned three small language models—Phi-3-mini, Qwen2.5-3B, and Mistral-7B—using QLoRA on biomedical claim verification datasets, finding that Mistral-7B outperformed GPT-4o and GPT-5 by up to 12% F1 at a …
Researchers have demonstrated that Direct Preference Optimization (DPO) simplifies the fine-tuning pipeline for large language models while improving computational efficiency and achieving competitive performance. The em…
Researchers at arXiv have proposed SafeLLM, an extraction-based method for retrieval-augmented generation that eliminates hallucinations by directly selecting relevant sentences from source documents rather than rewritin…
Linear launched coding sessions that allow users to run agentic coding workflows directly within the platform, enabling automated pull request creation and code review through Claude Code or Codex. The feature, available…