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 researchers testing adapter composition in large language models found that geometry-aware merging strategies, which enforce orthogonality or directional independence in parameter updates, provide no con…
A new study quantifies the rate at which undesirable behaviors transfer from teacher to student language models during distillation, a phenomenon known as subliminal learning. Researchers steered Llama-2-7B-Chat and Qwen…
Researchers have introduced Rotary Value Embeddings (RoVE), a parameter-free modification to Rotary Position Embeddings (RoPE) that makes value pathways position-sensitive by rotating values simultaneously with keys. Thi…
Researchers have developed FlowBank, a three-stage framework that optimizes LLM-based multi-agent systems by precomputing a compact bank of reusable workflows and adaptively selecting among them per query. The approach a…
Researchers have identified three critical limitations in existing preference optimization methods for medical Large Vision-Language Models (LVLMs), including their inability to distinguish clinically critical tokens fro…
Researchers have developed SalArt-VQA, a diagnostic benchmark to evaluate whether vision-language models (VLMs) truly understand salient artifacts in AI-generated images, rather than just correctly flagging their presenc…
Researchers introduced GRIP, a feedback-guided retrieval framework for multimodal in-context learning that selects examples based on their actual impact on model predictions rather than visual similarity. The system impr…
Researchers introduced PERIA, a tool-augmented visual agent designed to improve spatial reasoning in vision-language models by actively acquiring evidence and performing multi-step visual interactions. The agent uses two…
Researchers have developed L-VARC, a novel framework that enhances visual reasoning on the Abstraction and Reasoning Corpus (ARC) by integrating a language-guided Learning Using Privileged Information (LUPI) branch. The …
Researchers have identified that vision-language models (VLMs) often locate relevant text in images but fail to utilize it for answering questions, a phenomenon called "localization-without-utilization." To address this,…
Researchers have developed ToolSense, an open-source diagnostic framework that automatically generates three benchmarks to audit whether large language models truly understand their parametric tool knowledge. Testing on …
Arbor, a multi-agent framework developed by researchers, introduces structured tree search as a cognition layer for autonomous agents operating in large, stateful action spaces. The system maintains an explicit search tr…
Researchers introduced Pythagoras-Prover, a family of compute-efficient Lean theorem provers that achieve state-of-the-art performance with significantly fewer parameters. The 4B-parameter model surpassed DeepSeek-Prover…
Researchers have developed new testbeds to evaluate lie detectors for language models, finding that existing detection methods often fail when models are trained to hold false beliefs. The study tested four detectors acr…
TrajGenAgent, a hierarchical LLM-agent framework developed by researchers, generates realistic human mobility trajectories without model fine-tuning by using a two-stage orchestrator-worker design that synthesizes activi…
Researchers introduced Evoflux, an inference-time evolutionary search method that repairs executable tool workflows for compact language model agents. On MCP-Bench tasks involving 250 live tools, Evoflux improved executi…
Researchers at an academic medical center developed a pre-response classifier that predicts the risk of a user rejecting an LLM-generated clinical response, achieving an AUROC of 0.719 over 4.5 months of user feedback. T…
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…