CreativeInstruct: Scalable Post-Training to Cure LLM Creativity Collapse
EliasEskin · x · 2026-08-11
While post-training enhances LLM capabilities, it often causes output homogenization and creativity collapse. This paper introduces CreativeInstruct, a method to balance high quality with greater diversity.
Core Mechanism:
- Generates training data via routing between base and aligned models, tagging base model spans with special tokens.
- Post-trains the model to autonomously inject creativity at appropriate positions, eliminating the need for multi-model inference.
Evaluation & Results:
- Introduces LLM-Graph Edit Distance (GED), a novel metric capturing narrative-level structural diversity beyond lexical variation.
- Tests on 8B-32B models show improved narrative diversity, with human evaluators preferring its creativity in 70.3% of cases.
- Serves as a superior substrate for RL, boosting performance on AMC (4%) and MATH (5%) compared to standard post-trained checkpoints.
Related event: CreativeInstruct Tackles LLM Creativity Collapse(2 posts)→
More from Research
- iSDFT: open-source self-distillation method enables continual learning for LLMs — hbouammar · 2026-09-23
- The Polynomial Freiman-Ruzsa Theorem Leaves an Open Algorithmic Question — gautamcgoel · 2026-09-23
- Researcher Argues Parallel Agent Swarms Are a Weak Path to RSI — gleech · 2026-09-23
- New paper proves the long-standing Courtade–Kumar conjecture with multibit extensions — abeirami · 2026-09-23
- TimePre Paper Lands in TMLR: Reversible Normalization Fixes MCL Instability in Forecasting — _vztu · 2026-09-23
- Reproducible agent evals: harbor makes configs, trajectories and logs shareable — seanwbren · 2026-09-23