AI Research Brief

Archives
Log in
Subscribe
August 10, 2026

Runtime Control Moves Into the Open

  • WorldTrace Makes Long Video Memory Addressable: It assigns in-distribution virtual positions to compressed memory slots. Two variants improve LoopBench temporal consistency by 15.5% and episodic recall by 19.5%.
  • Modular TTT Makes Test-Time Learning Composable: It represents fast weights, loss functions, and learning rates as a DAG. After training on 100B tokens, its 410M and 1.45B models match Gated DeltaNet.
  • ReASearch Lets One Agent Run the Search: It unifies evaluation, diagnosis, editing, and stopping. It beats specialized systems on most of 14 tasks by 2% to 40%, though its validation discipline needs scrutiny.

Also Notable

  • Traditional Scaling Laws May Misprice Data-Scarce And Overtrained Regimes When Model Size And Data Volume Do Not Affect Loss Independently. Skaling combines Chinchilla exponents with the Kaplan coupling relationship.
  • Unified Image Restoration Models Can Adapt Tokenization Itself To Spatially Uneven Degradation Instead Of Waiting For The Backbone. The paper proposes degradation-aware deformable tokenization.
  • General Leaderboards Cannot Capture Real-Time Embodied Interaction When Selecting Foundation Models For Social Robots. The paper proposes a community evaluation framework that includes deployment constraints.
  • One-Shot, Training-Free, Task-Agnostic Pruning Could Cut Vision-Language Model Deployment Costs On Constrained Devices. Prune Once explores this approach for vision-language models.
  • Training-Free High-Resolution Generation Could Spare Teams From Retraining Existing Diffusion Transformers. HRDiT extends pretrained DiT models to higher resolutions without training.
  • Corruption Resistance In Natural Images May Depend On Internal Computation Paths, Not Just Final Representations. The paper studies resistance by suppressing and diversifying those paths.
  • Autonomous Discovery Of Accelerator-Tuning Algorithms Could Reduce Repeated Expert Rewrites After Light-Source Design Changes. The study applies automated algorithm discovery to simulated accelerator tuning.

Read the full edition →

Don't miss what's next. Subscribe to AI Research Brief:
← Newer SimWAM Hits 91.5 as LoRA Cuts Memory 43.9% Older → Skill Entropy Lifts Qwen3-4B to 68.4%
Powered by Buttondown, the easiest way to start and grow your newsletter.