Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
Researchers propose TheProfessor, a multi-teacher prompt distillation method for vision-language models that ensembles a domain-finetuned teacher and a zero-shot teacher. On four datasets, confidence-weighted ensembling …
Researchers introduced DriveStack-VLA, a framework that enhances Vision-Language-Action driving models with Bird-Eye-View representations and Render-Teacher Alignment to improve spatial intelligence. The model achieved s…
Researchers discovered sentence-level contextual entrainment in large language models, where sentences in prompts increase their probability during inference even if counterfactual. The phenomenon decreases with model si…
Researchers introduced Cavewoman, a two-channel evaluation protocol for large language models under linguistic compression, finding that output compression reduces costs by 1.4-3x on API and open-weight models, while inp…
Researchers propose an architecture that uses formally verified law as a reward signal for training legal AI, adapting the LLM proposes, verifier disposes paradigm to legal reasoning. The system autoformalizes legal rule…
Researchers introduced RIFT-Bench, a graph-driven methodology for dynamic red-teaming of agentic AI systems, enabling unified security evaluations across diverse architectures. The framework operates in two automated pha…
Researchers critique current AI agent models, arguing that genuine agency requires internalized goal, identity, decision-making, self-regulation, and learning structures rather than external scaffolding. They propose a G…
Researchers introduced AgenticInterpBench, a benchmark for circuit explanation in mechanistic interpretability, and HyVE, an agentic explainer that uses language models to analyze transformer circuit components. Testing …
Researchers propose Strategy-Guided Policy Optimization (SGPO), a method that replaces trajectory-level imitation with reusable strategy distillation to improve LLM reasoning. SGPO extracts structured strategies from str…
Researchers introduced ReMMD, a framework for realistic multilingual multi-image agentic verification to detect multimodal misinformation. The framework includes a benchmark with 500 samples and 2,756 images across five …
Researchers introduced VeryTrace, a zero-shot verification-and-repair framework that formalizes Chain-of-Thought reasoning traces into a compilable Domain-Specific Language, enabling step-level error detection and correc…
Researchers introduced T2D-Bench, a benchmark for evaluating LLM outputs on type 2 diabetes using a multi-layer clinical-lifestyle knowledge graph. Testing showed GPT-4o-mini and GPT-4o failed evidence-path checks in 35%…
A new study from arXiv shows that multilingual encoder initialization benefits streaming automatic speech recognition (ASR) only at low target-language data scales, with the advantage decaying as data increases. Across e…
Researchers propose NaviGen, a system that encodes user behavior history into executable instructions for personalized image and video generation, overcoming misalignment between AIGC pipelines and user intent. Using a d…
A new Claude Code skill called 'crispy-comments' automatically prunes verbose code comments, removing incidental history, defensive justifications, commented-out code, and ASCII-art banners to keep only comments that hel…
Handshake is hiring students and graduates to remotely test Large Language Models for top AI labs, paying up to $30 per hour with flexible hours. The role requires no specific expertise, only attention to detail, and off…
Zhipu AI, a Chinese large language model developer, plans a Shanghai STAR Market listing to raise about $2.2 billion, six months after its Hong Kong IPO saw shares surge over 2,000%. The company, with 2025 revenue of 724…
A developer at Krazimo, a company building RAG systems over private knowledge, outlines the most common failure modes in production retrieval-augmented generation. The biggest source of wrong answers is retrieval providi…
GitHub Copilot's shift to token-based billing on June 1 revealed that heavy users could see costs rise from $29/month to over $750, exposing the hidden expense of AI coding tools. A developer reported burning $1,850 in u…
A researcher proposes training frontier AI systems on a synthetic world rather than real-world data to prevent them from learning the details of human society and their own deployment, arguing that knowledge of the real …