NVIDIA introduces an open evaluation standard for the Nemotron 3 Nano model using the NeMo Evaluator, ensuring transparent, reproducible benchmarks through open configurations, logs, and structured workflows. This approach addresses challenges in verifying model improvements by enabling independent verification of results.

The article introduces on-policy supervised fine-tuning (SFT) as a simplified training method for large reasoning models, reducing computational costs and improving efficiency by focusing on truncation-based length penalties instead of complex RL-based multi-reward objectives. This approach maintains accuracy while cutting chain-of-thought lengths by up to 80% across five benchmarks.

The article introduces a framework to distinguish between missing knowledge and inaccessible knowledge in LLMs, revealing that recall, not encoding, is the main bottleneck for factuality. Using WikiProfile, it shows that while top models encode most facts, many errors stem from poor recall, especially for long-tail facts, and inference-time computation can improve recall.

OpenAI plans to test advertising in the U.S. for ChatGPT’s free and Go tiers to increase global access to AI while maintaining privacy, trust, and answer quality. This approach aims to balance affordability with service reliability.

AllMem introduces a hybrid architecture combining Sliding Window Attention (SWA) and non-linear Test-Time Training (TTT) memory networks to address computational and memory challenges in long-sequence tasks for Large Language Models (LLMs), enabling efficient scaling and reduced overhead. The framework also includes a memory-efficient fine-tuning strategy for adapting pre-trained models.

A study examines whether neural networks can generalize irregular morphological patterns like humans, finding that position-invariant transformers better capture L-shaped verb paradigms in Spanish even with limited training data. Sequential positional encoding models show partial success, highlighting the importance of encoding design in morphological learning.

ChatGPT Go is now globally accessible, offering enhanced features like GPT-5.2 Instant, increased usage limits, and extended memory, making advanced AI more affordable. These updates aim to expand AI adoption by reducing barriers to entry for users worldwide.

BotzoneBench introduces a scalable evaluation framework for Large Language Models (LLMs) using fixed hierarchies of skill-calibrated AI anchors, enabling stable, interpretable assessments of strategic reasoning across diverse games. It addresses limitations of existing benchmarks by providing linear-time measurement and cross-temporal consistency, revealing performance disparities and strategic behaviors in LLMs.

The article introduces a multi-agent framework for medical AI that combines fine-tuned LLMs (GPT, LLaMA, DeepSeek R1) with evidence retrieval and bias checks to enhance clinical query processing reliability. It uses a two-phase approach, including model fine-tuning on medical data and a modular pipeline with agents for reasoning, evidence grounding, and refinement, supported by safety mechanisms like uncertainty scoring and bias detection.

The article highlights undisclosed details in Elon Musk’s recent court filing that could significantly affect ongoing legal battles and public perception of his companies.