Bias, Baked In: Why Pretraining, Not Fine-Tuning, Shapes LLM Behavior
A causal study suggests that LLM cognitive biases are shaped mainly by pretraining, while fine-tuning mostly modulates how those biases appear.
Cognaptus Insights
Industry thoughts, trends, and commentary.
Prompt Without Words: Distilling GPT Semantics for Smarter Vision Models
DeMul shows how vision-language models can borrow GPT semantics without depending on noisy generated descriptions.
The Missing Link: How AI Maps Hidden Properties in Materials Science
A mechanism-first look at how literature mining and matrix factorisation can rank hidden material-property links without pretending to replace experimental validation.
The Rise of the Self-Evolving Scientist: STELLA and the Future of Biomedical AI
STELLA shows how biomedical AI agents may become more useful by evolving their reasoning workflows and toolsets, not merely by adding a larger language model.
What LLMs Remember—and Why: Unpacking the Entropy-Memorization Law
A mechanism-first reading of how set-level entropy turns noisy memorization anecdotes into a practical audit signal for LLM training data.
LLMs Meet Logic: SymbolicThought Turns AI Relationship Guesswork into Graphs
SymbolicThought shows why reliable relationship extraction needs LLMs, symbolic rules, evidence retrieval, and human verification working as one system.
Peering Through the Fog: A Hierarchy of Causal Identifiability Without Full Graphs
A practical map of what causal analysts can and cannot claim when the full causal graph is unknown.
Residual Entanglement: How ResQuNNs Fix Gradient Flow in Quantum Neural Networks
ResQuNNs show that deeper quanvolutional neural networks fail less from lack of quantum expressivity than from broken gradient access after measurement and re-encoding.
The Meek Shall Compute It
A mechanism-first reading of why diminishing returns may erode frontier-model compute moats faster than most AI strategy decks admit.
Threading the Needle: How GRAFT Reinvents Document Translation with DAGs and LLM Agents
GRAFT shows that better document translation may come less from longer context windows than from explicit discourse structure, dependency graphs, and selective memory.