AI Briefing#llm
The latest AI and developer news about #llm, with the original source and a short summary.
The latest AI and developer news about #llm, with the original source and a short summary.
Anthropic has discovered 'J-space,' where LLMs internally form concepts before generating language.
Microsoft research found that AI agents showed better performance using conventional CLI formats than when command-line arguments were replaced with JSON payloads.
NVIDIA has released Nemotron-3 Puzzle-75B, an MoE-based model that dramatically improves inference efficiency.
The Chinese government is reviewing measures to restrict overseas users' access to major AI models developed domestically.
The paper proposes a new methodology that separates memorization and generalization in language models to measure model capacity.
Musinsa introduced Prompt Caching and cut LLM API costs by 64% while achieving a 98% cache hit rate.
Through the Jacobian Lens technique, the region where an LLM's conscious reasoning occurs, called J-Space, was identified.
It has been revealed that LLM safety alignment is not distributed across the entire model but is concentrated in specific neurons, meaning that manipulating just a single neuron can disable safety guardrails.
We propose FlowEval, a framework that measures whether LLM-generated UIs support natural interaction flows like real websites.
Weblica provides a reproducible and scalable environment for training visual web agents through HTTP caching and LLM-based synthesis.
DynaMiCS is a dynamic mixture optimization technique that helps boost performance in a specific domain while preserving core capabilities such as existing knowledge and safety.
PyTorch Monarch now supports AMD GPU's ROCm environment, enabling uninterrupted, resilient distributed training even in the event of node failures.
The declining open source share in the enterprise AI market isn't a defeat for open source, but a transition to optimized models as AI use cases enter a maturity stage.
This introduces a technique that uses a small LLM to remove 68% of RAG context while maintaining 96% recall.
Countries are focusing on building sovereign AI ecosystems based on their own data and infrastructure to drive economic growth and national security.
Sberbank has released the GigaChat 3.5 model with 432B parameters along with a GGUF quantized version.
We propose the K-Merge framework, which continually expands an LLM's capabilities without data even in storage-constrained on-device environments.
The performance of AI agents is determined not by the model itself, but by how the 'harness'—the loop that drives the model—is designed.
It explains the hardware setup and Docker configuration needed to run SOTA LLMs and STT models in a local environment.
This work proposes a method that improves the efficiency and accuracy of ASR error correction by leveraging a seq2seq model lighter than an LLM.
The LongCat 2.0 model, with 1.6T parameters, has been open-sourced under the MIT license.
The PACE framework accurately predicts performance on costly LLM agent benchmarks using only a small amount of data.
We propose a new training objective to address the embedding condensation phenomenon that occurs in small language models.
A deterministic folding engine that improves context management efficiency for AI agents has been released.