AI Research
We study Creative Quality Alignment (CQA) — using information theory to formalize what makes creative work "good" and to align AI systems accordingly. Our research program comprises a 3+1 paper series: one theoretical foundation, one experimental validation, one evaluation benchmark, and one expert-knowledge extraction paradigm.
Calibrated Surprise: An Information-Theoretic Account of Creative Quality
Preprint
Explains why strong stories often feel “unexpected, but right in hindsight.” This paper gives creative quality a theoretical foundation that can be discussed and measured.
[arXiv]
Creative Quality Alignment: Expert Tacit Knowledge Transfer via Chain-of-Thought Fine-Tuning
Preprint
Turns expert writers' judgment processes into chain-of-thought training data, then tests whether fine-tuning helps models better recognize story quality.
[arXiv]
QUIET: A Multi-Blank Cascaded Story Cloze Benchmark for LLM Creative Generation Capability
Preprint
Builds a story-cloze style benchmark where models must understand setup, character relationships, narrative direction, and later constraints.
[arXiv]
BC Protocol: Structured Dual-Expert Dialogue for Eliciting High-Quality Chain-of-Thought Post-Training Data
Preprint
Proposes a dual-expert dialogue method for surfacing a writer's hidden tradeoffs, follow-up questions, and judgments as learnable training data.
[arXiv]
AI Alignment
Creative Quality
Information Theory
Chain-of-Thought
Expert Knowledge Extraction
