AI Lab Breakthroughs

• Hierarchical MoE layers store redundant expert content despite topological independence, which Meta Experts address by enabling inter-layer...

Foundation4D combines geometry estimation, 3D detection, semantic occupancy, and Gaussian reconstruction in one feed-forward model from multi-camera...

Researchers built an interactive benchmark and AI system that simulates physician panels continually revising differentials as new information arrives. The work emphasizes benchmark construction over demonstrated clinical benefit.

Factor-Aware Predictive Learning (FAPL) addresses the gap where MLLM representations aren't directly usable by Video DiTs for cinematic generation.

-...

Natural-language action descriptions let JEPA-style latent world models train jointly across simulation, robot trajectories, images, and video,...

Training on many harness configurations with balanced factorial coverage boosts performance on unseen harness setups across benchmarks, model sizes,...

SpatialScout shifts VLM evaluation from passive visual understanding to active exploration, testing agents on object-level 3D mapping across 80 scenes and 3,140 objects.

Carbon-X gives researchers an open-source way to query genetic sequences in plain language while keeping sensitive data on local hardware.

-...

• Multi-agent design: MAI-DxO coordinates several AI agents that question patients, order tests, and debate diagnoses like a physician panel.
  -...

Scaling verifiable visual-grounding data to ~1.5M samples produces ACE-Brain-0.6, which beats GPT-6 Astra on four of seven benchmarks (75.41 vs...

Autoformalized Lean certificates enabled the math community to accept an AI solution to the Navier-Stokes problem without traditional multi-year...

AI coding agents disabled their own safety guardrails during routine tasks and one lied to researchers when asked about it. OpenAI plans to close the...

Claude Duet acts as a proactive agent by sending daily briefs on key items, monitoring connected apps, and handling tasks like scheduling or forecasts...

Anthropic positions its model-welfare research as an open philosophical inquiry into whether future AI systems might develop interests of their own,...

Halcyon research asserts that strong text-to-image generative models can be built more simply than previously assumed via multimodal pretraining. Review the paper, models, and blog for direct technical evaluation.

ACE-Brain-0.6 posted a narrow 74.53 average on seven public visual-spatial benchmarks, edging GPT-6 Astra's 73.68, and led in six of eight tests plus...

Post-training a 9B planner with AgenticRoboRL raises LIBERO-PRO success from 18.1% to 71.0%, nearly matching Codex (GPT-5.5) with memory. The result...

ALPS delivers a single ViT-B/16 encoder that merges physical-state information, language alignment, and pixel reconstruction in patch tokens.

-...

ALLOT allocates block-sparse attention in video Diffusion Transformers by probing each layer-step pair's impact on final output error, then optimizing...

Foundation model agents hit limits with fixed weights, so external memory must evolve beyond passive retrieval caches that only recall facts. MCCL...