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 arXiv preprint (2606.18164) by Niklas Forner and coauthors analyzes how transformers learn chain-of-thought state tracking in a solvable setting, training a simplified one-block transformer on permutation compositi…
SpaceX has acquired Anysphere, the maker of the AI coding agent Cursor, to build a proprietary AI coding agent model. The deal signals a shift toward vertical integration in developer tools, with plans to deepen model tr…
A Hacker News user asked the community to share their best Claude hacks, highlighting a workflow of uploading research papers to Claude to create a temporary domain expert.
AI memory systems for LLM agents fail at scale due to structural precision problems in similarity search, not incidental issues. Testing across three embedding models showed mean retrieval precision of 0.09, with no impr…
Local AI servers fail to answer company-specific questions because they lack access to business data. RAG (Retrieval Augmented Generation) and MCP (Model Context Protocol) solve this by giving AI permanent, private acces…
OpenAI's Codex model completed a single 25-hour coding run, consuming approximately 13 million tokens, after the company acquired the startup behind the technology to enhance its AI coding capabilities.
Google Cloud Platform (GCP) provides a multi-layered approach to DNS abuse mitigation using Cloud DNS, Cloud Armor, and the Gemma 4 AI model. Cloud DNS offers managed DNSSEC to prevent cache poisoning and spoofing, while…
Amazon Bedrock AgentCore, generally available since late 2025, enables developers to deploy AI agents in four commands using any framework and model, handling operational infrastructure like session management, authentic…
Researchers propose a self-evolving framework for legal case retrieval that uses an LLM-based agent to iteratively create and refine rewriting rules for BM25, outperforming non-evolutionary baselines on the LeCaRD-v2 ben…
Researchers tested whether language models can independently discover the concept of zero through arithmetic generalization. GPT-2-sized models failed at test time regardless of language pretraining, but improved after t…
Researchers propose structural uncertainty, a framework that measures consistency in LLM logical reasoning by analyzing self-preference-induced rankings over sampled solutions. Across five LLMs and eight benchmarks, comb…
Researchers introduced MemTrace, a benchmark that evaluates LLM agents' long-term memory by probing individual knowledge points across dimensions like memory age, question type, and evidence condition. Testing 13 memory-…
Researchers propose a layered architecture for distributed general-purpose agent networks, enabling heterogeneous AI agents to discover, trust, and cooperate on open-ended tasks. The framework addresses key challenges in…
A new study reveals that large language models exhibit a strong brand bias, recommending well-known brands 100% of the time when product specifications are identical, but this dominance can be broken by minor rating adva…
Researchers introduced MapSatisfyBench, a benchmark to evaluate how well large language model agents in map services recover implicit user needs from available information before responding. The benchmark, built from rea…
Researchers introduced the Simple Strands Agent (SSA) to minimize the 'intent-execution gap' between AI models and their harnesses, improving agent performance. Analyzing 138k trajectories, they found that fine-grained m…
Researchers introduced CEO-Bench, a multi-agent benchmark evaluating large language models on strategic resource reallocation across business units. Testing five frontier models revealed high structural validity but shar…
Researchers have developed a curriculum-grounded LLM-as-Judge pipeline for automated marking of exam questions, co-developed with an industrial partner. The system uses authorized curriculum artefacts and marking guideli…
Researchers introduced SEAGym, an evaluation environment for self-evolving LLM agents that measures agent harness updates across training, validation, test, replay, and cost records. Instantiating SEAGym on Terminal-Benc…
Researchers introduced the first foundation model-orchestrated workflow for crash safety design, enabling surrogate-assisted pedestrian protection exploration that reduces evaluation time from hours to seconds. The workf…