How Sparse is Your Thought? Cracking the Inner Logic of Chain-of-Thought Prompts
A mechanism-first reading of how chain-of-thought prompting reshapes internal model features, when it helps, and why smaller models may not benefit.
Cognaptus Insights
Industry thoughts, trends, and commentary.
Layers of Thought: How Hierarchical Memory Supercharges LLM Agent Reasoning
A practical reading of H-MEM, a hierarchical memory architecture that makes long-term LLM agents faster and more coherent by changing the shape of memory, not merely its size.
Noise-Canceling Finance: How the Information Bottleneck Tames Overfitting in Asset Pricing
A mechanism-first look at how information bottleneck regularisation may help high-capacity asset pricing models forget noise without throwing away return-relevant structure.
Numbers Don’t Speak for Themselves: How LLMs Interpret the Soul of Financial Reports
A controlled pilot study shows why selecting an LLM for financial-report analysis requires multiple evaluation lenses, not a single leaderboard score.
SIMURA Says: Don’t Guess, Simulate
A mechanism-first look at SiRA, the agent architecture that uses explicit world-model simulation to plan before acting.
Echo Chambers or Stubborn Minds? Simulating Social Influence with LLM Agents
A comparison-based reading of how different LLM agent types conform, polarize, or preserve dissent in synthetic forum discussions.
Echoes in the Algorithm: How GPT-4o's Stories Flatten Global Culture
A mechanism-first reading of how gpt-4o-mini produces culturally decorated but structurally standardised stories, and what that means for organisations using LLMs in global communication.
From Chaos to Care: Structuring LLMs with Clinical Guidelines
CliCARE shows why clinical AI needs patient timelines, guideline alignment, and expert-calibrated evaluation—not just larger context windows.
Judo, Not Armor: Strategic Deflection as a New Defense Against LLM Jailbreaks
A mechanism-first look at SDeflection, a jailbreak defence that redirects harmful prompts into safe adjacent answers instead of betting everything on refusals.
Mind the Gap: How AI Papers Misuse Psychology
A practical reading of how LLM research borrows psychology—and where citation turns into conceptual debt.