Planners, Meet Your Smart Sidekick
SMARTAPS shows how LLMs can make advanced planning systems more usable by routing planner questions to expert-built optimisation tools rather than pretending to optimise from vibes.
Cognaptus Insights
Industry thoughts, trends, and commentary.
Steering by the Token: How GRAINS Turns Attribution into Alignment
GRAINS shows how token-level attribution can become a practical inference-time steering mechanism for safer LLM and VLM behaviour without retraining.
Structure Matters: Externalities and the Hidden Logic of GNN Decisions
GraphEXT reframes GNN explanation around structural externalities, showing why node importance depends on changing graph coalitions rather than isolated feature scores.
The LoRA Mirage: Why Lightweight Finetuning Isn't Lightweight on Privacy
LoRA makes model adaptation cheaper, but LoRA-Leak shows that released adapters can still expose whether sensitive records were used in fine-tuning.
The Most Dangerous Query Is the One You Don't Question
VeriMinder shows why enterprise NL2SQL systems need a pre-query reasoning layer, not just better SQL generation.
The Two Minds of Finance: Testing LLMs for Divergence and Discipline
A finance benchmark shows why strategic AI evaluation must test both imaginative scenario generation and disciplined answer selection.
Trained on Tickers, Tuned for Trust: The New Frontier of FinTech AI
A practical map of financial foundation models across language, time-series, and visual-language systems, and what they can realistically do for financial engineering.
Forecasting a Smarter Planet: How EarthLink Reimagines Climate Science with Self-Evolving AI Agents
EarthLink shows how climate AI becomes useful when it orchestrates auditable workflows, not when it merely chats confidently about weather.
From Cora to Cosmos: How PyG 2.0 Scales GNNs for the Real World
PyG 2.0 matters because it turns graph neural networks from benchmark experiments into modular infrastructure for heterogeneous, temporal, explainable, and production-scale graph learning.
GraphRAG Without the Drag: Scaling Knowledge-Augmented LLMs to Web-Scale
A mechanism-first analysis of how GeAR was adapted for million-passage GraphRAG by replacing offline triple extraction with online Wikidata alignment—and why that shortcut is useful but fragile.