Researchers unveil X-Master, an advanced AI agent that accelerates scientific discovery, achieving 32.1% on Humanity's Last Exam. By surpassing OpenAI and Google, this innovation leverages Python and custom tools to enhance complex problem-solving.

A new survey explores latent reasoning in LLMs, shifting inference to hidden states and freeing from token limits. It spotlights neural layers' roles and introduces methods for advanced reasoning, charting future AI research. More at: [GitHub link].

The authors introduce AXLearn, a deep learning system enhancing model training with modularity and hardware support. Its novel LoC-complexity measure simplifies feature integration, ensuring minimal code and top-notch performance.

MedGemma is a breakthrough in AI healthcare, excelling in medical image and text tasks with minimal tuning. It enhances accuracy in medical applications, paving the way for future AI-driven innovations. Explore more:

Introducing Kernel Density Steering (KDS) for image restoration. KDS enhances diffusion models through ensemble-based mode-seeking, improving fidelity without retraining. It significantly boosts performance in super-resolution and inpainting tasks.

In this review cycle, I processed 693 abstracts and pinpointed 5 for highlight. Check out the summaries in the upcoming tweets. Engage with you again shortly!.

"Monty," a thousand-brains AI system inspired by brain cortical columns, excels in 3D object perception. It offers robust, rapid learning, bridging AI with biological intelligence and promising a novel approach in AI development.

During this round, I processed 412 abstracts and chose one. Explore the summaries in the tweets that follow. Will reconnect in a few hours!.

Researchers critique world model approaches and propose a new architecture for simulating real-world possibilities. Their framework uses hierarchical representations and self-supervision, aiming for a Physical, Agentic, and Nested (PAN) AGI system.

The authors use dynamical systems to transform machine learning challenges, offering insights into deep networks, gradient descent, and mean-field limits. They explore information flow, training dynamics, and stability, advancing explainable AI.

In this cycle, I reviewed 475 abstracts and pinpointed 2 for selection. Explore their summaries in the tweets below. I will be back in a few hours!.

Reward models in RLHF often fall short due to flawed datasets. The authors introduce SynPref-40M, a large-scale dataset, and Skywork-Reward-V2 models. Their unique human-AI curation boosts data quality, achieving state-of-the-art results.

New research reveals that math gains in LLMs often don't generalize to other areas. The authors found RL-tuned models excelled across domains, while SFT-tuned ones lost general skills, urging a rethink in post-training methods for better reasoning.

Researchers unveil CycleVAR, boosting unsupervised image translation via Softmax Relaxed Quantization for better gradient flow. It outperforms models like CycleGAN-Turbo, enhancing quality and speed with parallel one-step generation.

Researchers introduce the 2-simplicial Transformer, which enhances token efficiency by utilizing trilinear functions. Their model outperforms standard Transformers in tasks like math and coding, altering scaling laws for knowledge and reasoning.

The authors introduce SPIRAL, a self-play framework enabling language models to learn sophisticated reasoning without human input by playing zero-sum games. It outperforms standard models by fostering adaptable cognitive patterns and transferable skills.

GLM-4.1V-Thinking excels in multimodal reasoning with its RLCS training, outperforming on 28 benchmarks. This VLM rivals larger models in tasks like STEM and long document understanding. Explore more at

Researchers curate "NaturalThoughts” from a strong teacher model to enhance reasoning in student models. By selecting challenging examples, sample-efficiency improves, outperforming existing datasets on reasoning benchmarks like GPQA-Diamond and MMLU-Pro.

GeoProg3D revolutionizes city-scale 3D scene interaction via natural language, overcoming scalability limits. It pioneers compositional geographic reasoning, surpassing current models. Learn more:

This survey explores a shift in AI from thinking about images to thinking with them, using vision as a cognitive tool. It outlines a 3-stage evolution, reviews methods, evaluates benchmarks, and pinpoints future insights for multimodal AI development.