13 articles — 🔴 1 critical , 🟡 7 important , 🟢 5 interesting
NVIDIA released Nemotron 3 Ultra, an open-weight Mixture-of-Experts model with 550 billion total parameters and 55B active per token. The model targets long agentic workflows with up to 30% lower cost compared to other leading open models. It is available in Ollama, and vLLM provided Day-0 serving support.
The scientific paper arXiv:2606.05555 argues that the main driver of scalable multitask reinforcement learning is representation learning, not model-based control. The model-free algorithm MR.Q combines predictive representations with a high-capacity value function and outperforms world-model methods at lower compute cost. The result holds even without planning, on continuous control tasks.
The scientific paper arXiv:2606.05661 introduces CL-Bench, the first hard, expert-validated benchmark that measures whether LLM systems actually improve through sequential experience. Evaluation of frontier models reveals problems with knowledge reuse and overfitting to immediate observations. Even purpose-built memory architectures underperform compared to simple in-context learning.
IBM and Google Cloud announced a strategic partnership described as a multi-billion-dollar opportunity, launching a new Google Cloud Practice that combines IBM's industry expertise with the Google Cloud Gemini Enterprise Agent Platform. IBM builds AI agents on the IBM Consulting Advantage platform optimized for Gemini Enterprise across eight sectors. The goal is to move clients from AI pilots to production deployment.
In a blog post by Neil Dahlke, LangChain argues that model neutrality — the ability to work across different AI models — matters more for flexibility and resilience than neutrality toward a cloud provider. The approach enables avoiding vendor lock-in at the model level and adapting as AI capabilities evolve.
OpenAI is introducing a new memory system for ChatGPT called 'Dreaming', designed to better remember user preferences and keep context fresh and relevant across conversations. The goal is to retain relevant information between sessions. Implementation details are limited for now, as the announcement is concise.
The scientific paper arXiv:2606.05610 introduces a framework that uses empirical scaling laws to predict optimal hyperparameters during continued pre-training of LLMs. A two-stage approach derives compute→settings functions via proxy models, then estimates the equivalent pre-training compute from validation loss. It achieves comparable or better performance while reducing the cost of hyperparameter search by up to 90%, in a model-agnostic way.
Anthropic released Claude Code v2.1.163 with Managed Settings for enterprise version-locking, a new /plugin list command, and improvements to the Stop and SubagentStop hooks. The release also brings several practical additions and a series of bugfixes, including a fix for hook conditions that incorrectly triggered on every Bash command.
NIST presented Safe Step, an AI system based on reinforcement learning that computes the safest fire evacuation routes in a building one step at a time. The model learns from floor plans and NIST's fire simulations and uses live sensor data to continuously update routes. It chooses routes with the lowest dose of toxic gases and predicts fire spread, but for now works only for single-story floor plans.
CHASE is a closed-loop framework in which an attacker and a defender model co-evolve through reinforcement learning. The attacker uses GRPO to rewrite prompts while preserving intent, while the defense is strengthened through two-stage training. The result is a 43.2% reduction in vulnerability score with a zero false-refusal rate on benign inputs.
A paper on arXiv shows that safety alignment paradoxically creates vulnerability in large language models. The 'Posterior Attack' is a single-query jailbreak that exploits the model's ability to recognize harmful content. It was tested on 30 open-source LLMs and on frontier models such as GPT-5 and Claude 4.6.
![Editorial illustration: FLUX.2 [klein] ships pre-installed on ASUS ProArt laptops](https://images.24-ai.news/black-forest-labs-flux-outpainting-600.webp)
