Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
An engineer discovered that multi-turn AI agents in tools like Claude Code and OpenCode lose context between turns because each subagent invocation creates a new session with no history. The infrastructure to preserve co…
A developer built five Claude automation agents that save hours weekly. The agents handle tasks like clipboard-based knowledge storage, email action item extraction, error diagnosis, meeting note summarization, and pre-P…
A coalition of state attorneys general has launched an investigation into OpenAI, the maker of ChatGPT, over potential consumer protection violations, according to a Wall Street Journal report. The probe focuses on wheth…
Researchers at arXiv propose PauseRec, a lightweight implicit reasoning paradigm for large language model-based generative recommendation, which outperforms explicit chain-of-thought methods by up to 6.22% while reducing…
Researchers found that repetition loops in Gemma 4 instruction-tuned LLMs can be fixed by editing as few as one neuron, but the fix does not resolve deeper 'doom loops' caused by missing knowledge. The study shows that w…
Researchers propose HiLo-Token, an input-adaptive token compression framework that reduces latency in Diffusion Transformers for image editing by allocating more tokens to high-frequency regions and fewer to low-frequenc…
Researchers propose a self-evolving framework for vision-language models (VLMs) that enables them to improve their question-asking ability without external supervision. The framework uses the VLM as both a proposer and f…
Researchers propose RT-VLA, a lightweight vision-language-action model for autonomous driving that uses knowledge distillation from the larger SimLingo model to achieve real-time performance. RT-VLA reduces inference lat…
Researchers introduced Poker Arena, a no-limit Texas Hold'em platform that evaluates LLMs across nine cognitive axes and three memory layers. Claude Opus 4.6 won the most chips but ranked fifth on mean axis score, reveal…
A study of LLM-as-a-Judge evaluations found that pairwise preferences flip 13.6% of the time on average, with some questions reaching a 56% flip rate, and GPT-4o-mini showed a significant first-position bias. The researc…
Researchers developed a Transformer-based scheduling policy for the open shop scheduling problem (OSSP) using deep reinforcement learning. The model, trained on small Taillard benchmark instances, generalizes to larger p…
Researchers propose UP-NRPA, an online framework using large language models for dynamic dialogue policy adaptation based on user portraits, achieving a 100% success rate in multiple dialogue tasks and a 56.41% increase …
Researchers propose Orchestra-o1, an omnimodal agent orchestration framework that enables efficient collaboration across text, image, audio, and video modalities. The framework introduces modality-aware task decompositio…
A new arXiv paper revisits the WorkBench benchmark for workplace agents, finding that the best agent in June 2026, Claude Opus 4.8, completes 89% of tasks with only 2.5% unintended harmful actions, up from GPT-4's 43% co…
Researchers introduced YeasierAgent, a new application-building paradigm that uses symbiotic agents, narrative worlds, and scene-aware interaction to create platform-agnostic agent-native applications. The system enables…
Researchers introduced MA-ProofBench, the first formal theorem-proving benchmark for mathematical analysis, containing 200 problems across two difficulty levels. Evaluations of leading LLMs showed poor performance, with …
Researchers propose using hyperdimensional computing (HDC) for structured querying on tabular data embeddings, addressing the lack of interpretable similarity scores in current methods. The HDC approach, based on Hologra…
Researchers introduced Risk-Aware Causal Gating (RACG), a framework that gates decisions from LLM agents based on causal risk rather than predictive confidence, reducing high-cost errors while preserving utility. The met…
Researchers at arXiv have developed a method using LLM representational geometry to predict compositional errors, finding that near-orthogonal concept encodings lead to successful composition while close encodings cause …
Researchers propose MINIM, a trusted local broker that performs privacy-aware minimization on client-side UI observations before transmission to remote LLM inference servers. The system learns dual scores for each UI ele…