Structural Debugging for Chain-of-Thought Graphs
Chain-of-thought traces can be inspected as dependency graphs instead of prose, making circular support, abandoned threads, unsupported conclusions, and contradictions easier to find.
LLM
Articles connected to this term.
Executable Metaphors: Compiling Analogy Into Prototype Code
Executable Metaphors treats a natural-language analogy as a compact architecture brief, then generates scaffolded code, build files, documentation, and a repair loop for rapid prototyping.
The Unaskable Question
A probing tool asks language models structurally impossible questions and classifies how they fail.
Evolving Better Prompts
A genetic prompt run improved average fitness from 0.887 to 0.926 in four generations using LLM-based mutation and crossover.