The authors present a method that uses an Invertible Neural Network to map system states to a latent space where a hyper‑rectangular forward‑invariant set can be trained and verified. When verification succeeds, the set and its inverse image are proven forward‑invariant in the original state space. The approach was tested on 45 AI‑controlled systems across three control testbeds.
Key points
Certified forward‑invariant sets were found for all 45 systems, while a state‑of‑the‑art baseline found none.
The method was faster on 40 of the 45 systems, and the resulting set centers roughly matched domain‑expert preferences.
Sources
Russell Dupuis $7.5M Endowed Directorship at HMNTL
Russell Dupuis established a $7.5 million endowment for a new directorship at the Holonyak Micro & Nanotechnology Laboratory. Minjoo Larry Lee will serve as the inaugural Russell Dean Dupuis Director. The endowment provides flexible, unrestricted annual funding for the lab’s core research.
Key points
Endowment amount: $7.5 million
Inaugural director: Professor Minjoo Larry Lee
Sources
Distributionally Robust Survival Models – new framework announced
A novel distributionally robust framework for survival analysis is introduced. It addresses latent subpopulation shift and outlier contamination with an outer minimization and an inner maximization. The method works with non‑decomposable survival losses and the Cox partial log‑likelihood. Experiments show improved worst‑group performance and stable training on contaminated data.
Key points
Framework combines outer minimization to reduce contaminated sample influence and inner maximization to target the hardest subpopulation.
Alternating gradient algorithm uses KKT‑derived outer updates; experiments on simulated data and two benchmarks show better worst‑group results while keeping overall performance competitive.
Sources
McKinsey predicts 24% annual data center power demand growth to 2030
McKinsey’s Global Energy Perspective 2026 forecasts data center electricity demand to increase at a 24% compound annual growth rate through 2030. Power infrastructure availability will limit how much compute can be added. Operators are turning to on‑site generation and hybrid grid strategies to access power faster.
Key points
Growth: Data center power demand projected to grow 24% per year to 2030.
Supply mix: Over 60% of operators plan to combine on‑site generation with grid connections.
Sources
CRAFT method for compositional generalization in vision‑language‑action models
CRAFT transfers supervision from demonstrated skill executions to counterfactual instruction‑observation pairs using reusable skill representations. It addresses vision shortcuts where policies rely on visual cues instead of instructions. Experiments show higher success on undemonstrated skill combinations while keeping performance on demonstrated ones, across three VLA models, two simulation benchmarks, and a real‑robot test.
Key points
CRAFT creates counterfactual pairs by fixing observation and changing instruction to an undemonstrated combination.
CRAFT improves success on undemonstrated combinations in three VLA models and two simulation benchmarks, and also improves compositional generalization on a real robot.
Sources
- Original post
- Same Scene, Different Task: Skill Alignment for Compositional Generalization in - Linked page
- Linked resource - Linked in the post
Claude Haiku 5.5 Intelligence Index jump to 43
Claude Haiku 5.5 reports an Intelligence Index of 43, up from 17 in Haiku 4.5, a +26 increase. The model now ranks ahead of GLM-5.3-Flash, Gemini 3.8 Flash, and DeepSeek V4.1 Flash. It is priced in the same class as GPT-6 Luna.
Key points
Intelligence Index: 17 → 43 (+26)
Ahead of: GLM-5.3-Flash, Gemini 3.8 Flash, DeepSeek V4.1 Flash
