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 study of a production enterprise LLM assistant with 110 agents and 584 tools found that routing accuracy degrades by 16-23 percentage points as the catalog scales, driven by retrieval and confusion gaps. Embedding-base…
Researchers have theoretically analyzed how deep transformers represent hierarchical structures in language, constructing models whose depth grows linearly with grammar depth and whose neuron count scales with derivation…
Researchers have discovered that the key/value cache in large language models can be edited and composed like notebook notes, enabling direct modification of model outputs without full recomputation. The approach, valida…
A comparative study of classification models within the InferBERT framework for pharmacovigilance found that domain-specific pre-trained BioBERT outperformed larger models like Med-LLaMA in detecting causal adverse drug …
Researchers show that grokking in deep neural networks is caused by noise-driven escape from metastable phases during first-order L2 regularization phase transitions. Using linear DNNs, they demonstrate that SGD noise he…
Researchers formalized and verified four concurrency anomalies in multi-agent LLM systems, establishing a machine-checked consistency hierarchy from L0 to L4. They implemented verified detectors and runtimes in Rust, pre…
Researchers at arXiv have identified a conflict signature in language models that distinguishes deceptive outputs from honest errors, achieving 100% accuracy in detecting lies across multiple models including GPT-2, Qwen…
Researchers propose negative token filtering, a method enabling stable single-rollout training for critic-free reinforcement learning in large language models, outperforming group-based techniques on agentic tasks.
Ray Data LLM, a library for large-scale batch inference, achieves 2x throughput over vLLM's synchronous LLM engine in production-scale workloads by optimizing hardware utilization and providing fault tolerance. The syste…
A developer calculated the actual VRAM requirements for running Llama 3 8B and Gemma 2 9B locally, revealing that the KV cache can consume far more memory than the model weights, especially at longer context lengths. For…
Microsoft is evaluating DeepSeek's AI models for integration into its Copilot Cowork platform, according to sources familiar with the discussions. The move could provide a cost-effective alternative to OpenAI's models fo…
OpenAI researchers tested whether public chat data from WildChat can predict real-world AI misalignments, finding that deployment simulations using public conversations can estimate rates of undesirable model behavior, o…
A developer reports that while LLMs lower coding costs, they may increase debugging complexity after spending two hours debugging an iCloud synchronization issue in an iOS app. The bug stemmed from AI-generated Swift cod…
A developer demonstrated that gzip, a standard file compressor, can perform rudimentary language modeling by exploiting the compression-prediction equivalence. Using beam search over byte sequences, gzip generates text t…
A developer uses Qwen Code, an open-source agentic coding CLI, to build Achu, a desktop screenshot beautification app built with Electron, React, and TypeScript. The workflow relies on slash commands like /init, /plan, /…
Alibaba's Qwen team released a robotics suite with three foundation models for embodied AI: Qwen-RobotNav, Qwen-RobotManip, and Qwen-RobotWorld. The models align language with physical actions for navigation, manipulatio…
A developer from Gate of AI published a tutorial on fine-tuning large language models for specialized communication tasks, using homeless shelters as a case study. The tutorial demonstrates how to structure data from the…
A developer identified that document review, not code writing, is the new bottleneck in AI-assisted software development. To solve the imprecision of natural-language feedback on generated docs, they built MDXG Redline, …
KPMG, a global consulting firm with 250,000 employees, published an AI-generated report in October 2025 titled 'Total Experience: Redefining Excellence in the Age of Agentic AI.' An audit by GPTZero found that 40 of 45 c…
Oxlo.ai demonstrates how few-shot learning with large language models enables domain-specific classification without weight updates, using in-context learning to infer patterns from exemplars. The company highlights its …