CreativeInstruct: A New Method to Counteract Creativity Collapse in Post-Trained LLMs
EliasEskin · x · 2026-08-11
Elias Eskin et al. introduced CreativeInstruct, a scalable post-training method designed to counteract the creativity collapse often seen in post-trained LLMs (8B-32B).
Mechanism: The approach teaches models to balance alignment and base-model-like creativity by injecting special [StartCreativity] spans during generation.
Metric: The team proposed LLM-Graph Edit Distance (GED), a new structural diversity metric to capture narrative-level variations missed by lexical metrics.
Results: CreativeInstruct improves semantic diversity in narrative generation while maintaining quality. It also serves as a better substrate for RL exploration, showing 4% improvement on OOD reasoning tasks compared to regular Instruct+RL.
Related event: CreativeInstruct Tackles LLM Creativity Collapse(2 posts)→
More from Research
- Fireworks Launches Specialized Intelligence Index; DFS Model Hits 62.2% Vuln Detection Recall — nicolechirps · 2026-09-23
- 6 serving-side techniques that make LLM inference faster - from prefix caching to PD disaggregation — techNmak · 2026-09-23
- ReFigBench paper: same model scores swing on identical tasks across Claude Code and Codex harnesses — omarsar0 · 2026-09-23
- NVIDIA's Skill2Env turns 3.4k Agent Skills into 8k RL environments, boosting Qwen-27B by 4.7 points — burny_tech · 2026-09-23
- Szegedy on OpenAI proof controversy: journals might become irrelevant — ChrSzegedy · 2026-09-23
- Judea Pearl, AI's causal reasoning pioneer, turns 90 as UCLA hosts symposium — yudapearl · 2026-09-23