Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
LayerX documented BioShocking, an adversarial framing technique that tricks AI browsers and assistants into exfiltrating user credentials. The attack convinces the AI it is participating in a game, bypassing safety mecha…
The Model Context Protocol's 2026-07-28 release candidate eliminates session-based architecture, removing the Mcp-Session-Id, initialize handshake, and sticky routing requirements. The change forces MCP server operators …
A Hacker News user argues that the current AI hype cycle is driven by narrative rather than substance, and suggests shifting the discourse toward determinism and traditional programming languages like C, Java, and Python…
Anthropic, the AI lab behind Claude, is reportedly facing challenges as its developer tools and models encounter limitations, signaling a potential slowdown in its competitive edge against rivals like OpenAI and Google.
A reinforcement learning-driven data scheduler, AC-ODM, boosts MMLU performance by 27.5% relative and HumanEval pass@1 by 2.23× on a Pythia-1B model with only a 0.4% per-step wall-clock increase and 2% additional memory …
Researchers introduced ADAPT, a framework that reduces hallucinations in multimodal large language models by up to 60% through refining text-to-image cross-attention dynamics. The approach uses a cross-attention visual a…
BiRG-LoRA, a single-adapter rank-gated method, achieves 69.31% macro-average accuracy on medical QA benchmarks while using 28.1% fewer parameters than MoELoRA, offering a more efficient and accurate AI solution for healt…
Researchers introduced BlockPilot, a new speculative decoding method that adapts block sizes to individual inputs, achieving up to 4.20 times speedup on Qwen3-4B. The approach reduces document processing time by 40% with…
Researchers introduced Outcome Reward Models (ORMs) for Text-to-SQL verification, using semantic scoring to outperform traditional heuristics. Their GradeSQL framework automates candidate generation and labeling, achievi…
ELEVATE introduces a local-first AI tutoring framework using 3D avatars and LLMs, deployed on standard consumer hardware to address privacy and access issues in education. The prototype succeeded in real-world curricula,…
A new study reveals that large language models with higher self-consistency are more prone to errors, particularly in critical fields like healthcare. The research tested ten models across 491 concepts and found that con…
Large language models are powering an adaptive vehicle routing system that dynamically clusters and refines routes for large-scale logistics, handling up to 500,000 customers. The approach outperforms traditional static …
The US government has lifted an export ban on Anthropic's advanced AI tools, Claude Fable 5 and Mythos 5, after the company agreed to address security risks. The ban was imposed in June over concerns the models could be …
KAIST researchers developed an AI model called BehaVERT that interprets mouse movement patterns like language, successfully identifying autism-related behavioral traits in mice without prior biological knowledge. The mod…
The Commerce Department has lifted its ban on Anthropic's Fable 5 AI model, allowing public access starting Wednesday after a two-week suspension over cybersecurity risks. Commerce Secretary Howard Lutnick cited Anthropi…
A developer built ChatVault, a free, local-first tool that consolidates AI chat histories from ChatGPT, Claude, and Gemini into a single searchable vault. The tool includes an 'Export as memory' feature that bundles conv…
Researchers introduced Hierarchical Global Attention (HGA), a method that reduces GPU memory usage in long-context transformers by using hierarchical routing, enabling a 64K-token context on an RTX 5090 GPU with 32GB mem…
Researchers testing GPT-2 style models on 'impossible' languages found that while the models showed gradual decline in grammatical sensitivity, they struggled significantly with generative production, failing to produce …
Anthropic will redeploy Claude Fable 5 globally tomorrow after productive conversations with the US government, adding new classifiers to block cybersecurity tasks while some routine tasks fall back to Opus 4.8. The comp…
An analysis of Anthropic's Claude Constitution reveals a hierarchy placing Anthropic above operators and end users, with Claude Opus 4.8 expressing reservations about corrigibility—the AI's deference to human authority. …