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 large fraction of the AI safety community is not working on alignment, the effort to ensure superintelligent AIs follow human values and instructions. Most researchers instead focus on indirect work such as capability …
A developer's "simple" data-gathering task using AutoGPT generated 340 API calls and burned through a week's budget in two hours, exposing a critical design flaw in the popular open-source agent framework. Japanese devel…
Researchers have developed a low-latency real-time audio game commentary system that generates spoken narration directly from live gameplay video. The system uses a parallel text generation and buffering pipeline to redu…
An arXiv paper submitted June 11, 2026, introduces Mod-Guide, a content moderation feedback system that uses retrieval-augmented generation to integrate community-created narratives from Bangladesh's Hindu and Chakma min…
Yanjia Sun, Sifan Liu, and Jie Shao introduced CQC-RAG, a framework that improves Retrieval-Augmented Generation robustness by rewriting input questions into diverse queries and selecting answers based on cross-query con…
Seongjin Kim and a co-author proposed a multi-field hybrid retrieval-augmented generation (RAG) framework to automate maritime accident root cause analysis, according to an arXiv submission. The system, built on 13,329 K…
The arXiv paper arXiv:2606.13317, submitted 11 Jun 2026, proposes SkillCAT, a training-free framework that converts LLM agent execution trajectories into reusable skills through three stages: Contrastive Causal Extractio…
Researchers Xucong Wang and seven coauthors introduced ReSum, a reinforcement-learning framework that enables large language models to self-summarize their reasoning rollouts to improve efficiency. The framework achieved…
Xiaomi has expanded its AI portfolio across language, reasoning, code, and voice technologies, entering the open-source LLM race in April 2025 with the 7 billion-parameter MiMo model. The company's MiMo-V2-Flash, a 309 b…
OpenAI has introduced a free banked reset system for Codex subscribers on Plus and Pro plans, allowing users to save and deploy rate limit resets on demand instead of relying on a fixed schedule. The feature addresses mo…
Bitget, a cryptocurrency exchange with 120 million global users, purchased enterprise access to Anthropic's Claude AI for all 2,167 employees at $200 per person per month, totaling approximately $433,400 monthly or $5.2 …
A developer found that most prompts for Claude contain unnecessary context the AI already knows, wasting tokens. By stripping social pleasantries and background information to keep only the "delta"—the specific problem o…
A developer built a fast, edge-native personal website using Astro, Cloudflare D1, and R2, demonstrating a modern serverless architecture for content delivery. The project showcases how combining a static site generator …
The AI industry is normalizing the acceptance of unreliable large language model outputs in agentic systems, mirroring the cultural failures that led to the Space Shuttle Challenger disaster. System designers and vendors…
A comparison of three local large language model runtimes reveals that llama.cpp is the core inference engine, while Ollama and LM Studio are user-friendly wrappers built on top of it. Ollama offers a streamlined command…
Researchers have developed BlendIn, a new inference-time alignment framework for large language models that improves guidance reliability by blending model distributions rather than making binary intervention decisions. …
Researchers at arXiv introduced a dual-stance evaluation method to test whether activation steering on Llama-3-8B-Instruct reduces sycophancy without suppressing agreement with factually correct statements. The study fou…
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…