Facthink Facthink
Example · a real run

One hypothesis, fully traced.

Oriented from research interests that touch multiple domains.

HOW IT STARTED
RESEARCHER'S EXPERTISE
LLM Compression
Research interest (CROSS-DOMAIN)
Representation Geometry

Two phrases steered hypothesis expansion toward where these two fields meet.

↓
A

Generated hypothesis

Generated hypothesis #0481

Rotor-style compression for LLM attention layers outperforms low-rank approximation when the layers' singular values are spread fairly evenly.

Novelty0.58
Testability0.70
Fit to the topic0.82
Overlap0.67
from a single run that involves
68
sources
36
nodes
B

The causal graph

The mechanism the engine assembled — every concept node coloured by its provenance, every edge by its committed status, converging on the final hypothesis.

Causal graph: concept nodes coloured by provenance, edges by committed status, leading to the final hypothesis
C

Involved Concepts from Multiple Domains 

What each term means in this context, and where it's commonly used.

Concept in hypothesis
What it means here
Widely used in
Low-rank approximation
Replace a large matrix with a smaller rank-k version.
Numerical linear algebra, recommender systems, ML compression, signal processing, statistics
Rotor
Geometric-algebra representation of rotation.
Geometric algebra, Clifford algebra, robotics, graphics, physics, some 3D / computer vision work
Representation geometry
Study the geometric structure of hidden states and transformations.
Mechanistic interpretability, deep-learning theory, VLM / LLM analysis