AI Research Nexus

Core ML Training Advances

• 🔥 MemTrain: Self-supervised framework using masked reconstruction and memory recall objectives over Wikipedia data to...

• StreamMA streams reasoning steps between agents, pipelining execution to cut latency while improving accuracy by avoiding late-step errors
  -...

Two fresh papers attack reasoning quality in LLM agents from opposite directions.

• ThoughtFold folds long CoT trajectories via introspective...

NVIDIA's Cosmos 3 introduces omnimodal world models for physical AI, targeting computer vision and pattern recognition tasks. This positions it as a notable development in architectures supporting robotics and autonomous systems.

Former OpenAI researcher Miles Brundage called the company's new federal policy proposal a positive step, noting various strengths versus prior positions.

BenchEvolver evolves reference solutions of existing coding problems into harder variants, automatically generating challenging benchmarks that...

Nathan Lambert spotlights a practical video on modern on-policy distillation for post-training recipes, calling it essential for frontier work despite his prior skepticism toward academic self-distillation.

Core ML/DL Architecture Advances

• 🔥 NVIDIA Cosmos 3: NVIDIA released Cosmos 3, an open frontier foundation model for physical AI using a...

AutoMedBench evaluates agentic AI on long-horizon medical research tasks across five stages—Plan, Setup, Validate, Inference, Submit—with runs...

• New probes detect LLM deception with near-perfect AUROC on clean data but collapse under stylistic shifts, recovering only via style-augmented...

Biohub's world model of protein biology, trained on 2.8 billion sequences, designs novel therapeutic proteins that succeed in real lab tests.

• Key...

Risto Miikkulainen argues neuroevolution outperforms gradient descent by running parallel population-based searches that explore broadly and discover...

VaSE protects large-magnitude value states and adds stochastic diversity during eviction.

• Delivers 4x KV cache compression on Qwen3 models while...

NVIDIA is pushing open foundation models to scale embodied AI across robotics and AVs.

• Alpamayo 2 Super (32B params) delivers reasoning VLA...

Researchers built one of the largest brain reference models by compiling diffusion MRI scans from 54,583 individuals across 19 datasets, generating...

A new tunneling phase diagram ML framework predicts the quantum tunneling factor κ with R² > 0.98 and RMSE of 0.21 across 300-600 K, cleanly...

Gary Marcus pushes back on critics, reiterating that deep learning requires neurosymbolic supplementation as he predicted in 2022—and citing Claude...

A new Decoupled Residual Denoising Diffusion (DRDD) architecture splits the process into independent noise diffusion for domain harmonization and...

A new method, PF-OPSD, lets MLLMs invoke, verify, and integrate stochastic visual rollouts from world models for better concrete and abstract reasoning, improving results by 10.6% on VRQABench and 10.9% on OpenWorldQA.

Core ML Training and Efficiency Methods

• 🔥 NITP Pre-training: NITP introduces next implicit token prediction as a new LLM pre-training...