New scoring method helps AI generate more creative and original text

Large language models have become increasingly capable at generating creative content, yet they frequently fall into predictable patterns that limit their originality. Researchers have introduced a context-based scoring system designed to push these models toward both accurate and inventive outputs without sacrificing quality.

The challenge of creativity in AI#

When generative systems are asked to produce creative work—from poetry to novel problem solutions—they typically generate variations that cluster around their training distribution. Common approaches to boost diversity, such as raising the temperature parameter during sampling, often come at the cost of lower-quality outputs. This tension between maintaining value and encouraging novelty remains largely unresolved in AI system design.

A new scoring approach#

Drawing on principles from information theory, the proposed score evaluates generated text across two dimensions: value (accuracy and fidelity to the user's request) and originality (departure from learned patterns). The metric works by rewarding outputs that satisfy both criteria simultaneously—pushing models to explore beyond their default tendencies while staying true to the task at hand.

Practical application#

The researchers demonstrate how this scoring method can function as a reward signal in reinforcement learning frameworks, enabling fine-tuning of large language models to maximize the balance between quality and novelty. Testing across diverse creative domains—including poetry composition and mathematical problem-solving—the approach consistently produces solutions that are both more valuable and more original than baseline methods.

The work suggests a viable path for designing AI systems that generate genuinely creative outputs without compromising on correctness or utility.