AI Research Brief

Archives
Log in
Subscribe
September 4, 2026

Stable Agent Protocols and 18% Cheaper Loops

  • Separating Control Flow From Prompts Delivers 100% Protocol Validity. Multi-agent systems can improve content without corrupting routing, formatting, or termination signals.
  • Language Is Becoming The Control Layer For Video Worlds. H3-World controls characters, cameras, and time spans while training only 0.199% of its parameters.
  • A Drone Can Reach Its Target And Still Fail. DroneCATS shows that termination, protocol compliance, and multi-view decisions also determine autonomy.
  • Acting Like A Student Does Not Mean Simulating One Accurately. StudentSim separates behavioral fidelity from responsiveness to instruction.
  • Looped Layers Save 6.8%–18.0% Of Training Compute Under Matched Budgets. SMELT offers a fairer test of shared layers and sparse experts.

Also Notable

  • Bind Cross-Modal Evidence Into Event Memories Before Retrieval. EM²Mem organizes long-video memory around events instead of disconnected frames, subtitles, and summaries.
  • Unified Understanding And Generation Do Not Automatically Help Each Other. This controlled study separates cooperation from capacity competition across representations, tasks, and systems.
  • One Successful Safety Test Does Not Prove A Deployment Is Safer. The better question is how much harmful assistance an adaptive attacker can eventually obtain.
  • Multi-Day Coding Agents Depend On Improvement Loops That Accumulate Without Regressing. Harness-of-Harness studies sustained autonomous planning, implementation, and testing.
  • SAR-To-Optical Conversion Must Treat Codec Fidelity As A Design Variable. ReFlowSET improves SAR-to-EO generation through representation alignment.
  • Persistent Agents Need Identity, Memory, And Code Outside The Current Runtime. This boundary preserves continuity across model, framework, and server changes.
  • Cultural Fine-Tuning Can Raise Average Alignment Scores While Erasing Internal Differences. Evaluation should measure both cultural fit and diversity.
  • Short-Drama Evaluation Should Track Errors Across The Full Production Chain. DramaChain Bench covers scripts, storyboards, keyframes, and final shot synthesis.

Read the full edition →

Don't miss what's next. Subscribe to AI Research Brief:
← Newer Production Traffic, Video Pretraining, and Smarter Distillation Older → OPSA Lifts AIME24 by 35.41 Points
Powered by Buttondown, the easiest way to start and grow your newsletter.