Greedy, but Not Blind: Teaching Optimization to Listen
A mechanism-first reading of LEG, a hybrid LLM-and-greedy optimization framework that lets qualitative advice influence facility planning without surrendering coverage guarantees.
Cognaptus Insights
Industry thoughts, trends, and commentary.
Houston, We Have a Benchmark: When Agentic AI Meets Orbital Reality
AstroReason-Bench shows why agentic AI needs physics-aware simulators, structured planning workflows, and specialized optimizers before it can handle real operational planning.
Probe, Then Commit: Why Solver Tuning Finally Grew Up
A practical reading of the Probe and Solve Algorithm, a two-phase method for tuning constraint programming solvers under real time budgets.
Punching Above Baselines: When Boxing Strategy Learns to Differentiate
BoxMind shows that applied AI becomes useful when perception, prediction, and intervention are joined into a closed operational loop.
Think-with-Me: When LLMs Learn to Stop Thinking
A mechanism-first reading of Think-with-Me, a test-time intervention framework that turns LLM reasoning from uncontrolled token generation into a feedback-guided control loop.
When LLMs Read the Room: Predictive Process Monitoring Without the Data Buffet
A mechanism-first reading of why LLMs can predict process outcomes from tiny event logs, and why the advantage depends on semantics rather than spreadsheet magic.
Fish in the Ocean, Not Needles in the Haystack
A mechanism-first reading of SIN-Bench, and why enterprise AI evaluation must move from answer accuracy to auditable evidence chains.
One-Shot Brains, Fewer Mouths: When Multi-Agent Systems Learn to Stop Talking
A mechanism-first reading of TOPODIM, a multi-agent framework that replaces chatty coordination with sparse, task-specific topology generation.
Redundancy Overload Is Optional: Finding the FDs That Actually Matter
Why redundancy-driven top-k functional dependency discovery is not just faster FD mining, but a cleaner way to decide which database constraints deserve attention.
Seeing Is Not Thinking: Teaching Multimodal Models Where to Look
LaViT shows why multimodal models can copy answers without inheriting visual grounding, and why enterprise AI teams should audit where models look, not only what they say.