AGI Advance: Weekly AI & AGI Insights (June 16, 2026)

This week, we're highlighting how Turing delivered an ad storyboard dataset, where evaluators produced original shot product advertisement concepts grounded in real product images and descriptions. This required creative reasoning structured to reflect how effective advertising actually works, from product introduction through feature demonstration to brand reinforcement.

Here's what we delivered:

• 7,500+ original shot descriptions, each covering scene setting, character actions, product placement, and camera motion drawn from a defined taxonomy with every concept grounded strictly in provided product inputs and no invented features permitted
• 6,000+ voice-over lines matched to specific product features or benefits visible in each shot, validated against SOTA video generation models to confirm storyboards translated into coherent video sequences without hallucination
• 90%+ quality score across delivered tasks, enforced through a four-dimension rubric covering conceptual accuracy, visual clarity and creativity, ad structure and purpose, and technical adherence

💡 Training multimodal models to generate video concepts from product images requires creative reasoning grounded in real inputs, structured to an advertising arc, and validated against actual video generation outputs. 

→ Read the full case study

• Agentic Discovery of Neural Architectures: AIRA-Compose and AIRA-Design
  Researchers at Meta introduce AIRA-Compose and AIRA-Design, two agent-based frameworks that explore whether AI agents can autonomously design better foundation model architectures and training systems. Instead of relying on human-crafted Transformer designs, agents search over architectural primitives, generate new model structures, implement novel attention mechanisms, and optimize training loops.
  Using these frameworks, agents discovered 14 new architectures (AIRAformers and AIRAhybrids) that outperform strong baselines such as Llama 3.2 and Composer-found models on validation loss, downstream tasks, and scaling efficiency. In mechanistic design tasks, agents achieved near-human state-of-the-art performance on Long Range Arena benchmarks and optimized language model training scripts beyond published Autoresearch baselines.
• HyperAgents
  Researchers from Meta, UBC, and collaborators introduce HyperAgents, a self-referential agent framework that can improve not only how it performs tasks, but also how it generates future improvements. Unlike prior self-improving systems that rely on fixed, handcrafted meta-level mechanisms, HyperAgents unify a task agent and a meta agent into a single editable program, enabling metacognitive self-modification.
  The resulting system, DGM-H (Darwin Gödel Machine with HyperAgents), demonstrates sustained self-improvement across coding, paper review, robotics reward design, and Olympiad-level math grading. It improves paper-review accuracy from 0.0 to 0.71, robotics reward-design performance from 0.06 to 0.372, and learns transferable self-improvement strategies such as persistent memory, performance tracking, bias detection, and compute-aware planning that generalize across domains.
• SkillFactory: Self-Distillation For Learning Cognitive Behaviors
  Researchers from NYU and Toyota Research Institute introduce SkillFactory, a self-distillation framework that teaches language models cognitive reasoning skills such as reflection, verification, and retrying without relying on stronger teacher models. The method restructures a model’s own sampled solutions and self-critiques into "silver" training traces, then uses SFT followed by RL to reinforce effective skill usage.
  Experiments show that SkillFactory improves generalization to harder reasoning tasks, increases explicit use of verification and retry behaviors, and helps models retain stronger out-of-domain performance after RL. On Countdown reasoning tasks, SkillFactory + RL outperformed RL-only and matched or exceeded strong distillation baselines, while also producing longer, more structured reasoning traces.

🔹 ICML 2026 — International Conference on Machine Learning
📍 Seoul, South Korea | 🗓️ July 6-11

ICML is one of the world’s leading machine learning conferences, highlighting frontier research across AI, data science, and applied domains from vision to robotics.

Turing is leading the charge in bridging AI research with real-world applications. Subscribe to AGI Advance for weekly insights into breakthroughs, research, and industry shifts that matter.

[Subscribe & Read More]