Tool Up or Tap Out: How Multi-TAG Elevates Math Reasoning with Smarter LLM Workflows
Multi-TAG shows that better LLM reasoning comes from structured tool arbitration, not merely from adding another calculator to the prompt.
Cognaptus Insights
Industry thoughts, trends, and commentary.
All Eggs, One Basket: When Diversification Backfires in Risk Modeling
A mechanism-first reading of when extremely heavy-tailed risks can make diversification more dangerous than concentrated exposure.
Boxed In, Cashed Out: Deep Gradient Flows for Fast American Option Pricing
A mechanism-first look at how Time Deep Gradient Flow adapts neural PDE solvers to American-option pricing, and where the speed claim is genuinely useful.
Divide, Route, and Conquer: DriftMoE's Smart Take on Concept Drift
DriftMoE reframes concept drift as a routing problem, showing how lightweight expert allocation can compete with heavier adaptive ensembles while exposing a sharp weakness on imbalanced streams.
Factor Factory: How LLMs Are Reinventing Sparse Portfolio Optimization
A mechanism-first reading of EFS, an LLM-guided evolutionary framework that turns sparse portfolio optimisation into adaptive alpha-factor search.
From Sobol to Sinkhorn: A Transport Revolution in Sensitivity Analysis
A practical reading of gsaot, the R package that turns Optimal Transport-based sensitivity analysis into a usable workflow for correlated inputs, multivariate outputs, and existing simulation datasets.
One Model to Train Them All: How OmniTrain Rethinks Open-Vocabulary Detection
A clear operator’s reading of OmniTrain’s unified training recipe for open-vocabulary detection, where the value is less model sprawl and more diagnosable training.
Speed Bumps and Swells: Rethinking Optimal Trading with Stochastic Volatility
A mechanism-first reading of how multiscale stochastic volatility changes optimal trading speed, target inventory, and the practical value of tractable execution rules.
Stacking Alpha: How HARLF's Three-Tier Reinforcement Learner Beats the Market
A mechanism-first reading of HARLF, a hierarchical reinforcement-learning portfolio system that combines market metrics and FinBERT sentiment while leaving important deployment questions unresolved.
The Sentiment Edge: How FinDPO Trains LLMs to Think Like Traders
FinDPO shows how preference alignment can turn financial sentiment classification into more resilient portfolio signals, while leaving the usual live-trading dragons very much alive.