Research map / Large language models / Language, safety and interpretability
Measuring LLM Novelty As The Frontier Of Original And High-Quality Output
Vishakh Padmakumar, Chen Yueh-Han, Jane Pan, Valerie Chen, He He
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| scale OLMo from 1B to 7B | OLMo · TinyStories and MacGyver | novelty: +19% quality TinyStories, +39% quality MacGyver | Improved |
| scale OLMo-2 from 7B to 32B | OLMo-2 · TinyStories, CoPoet, MacGyver (n=4) | novelty: +1.8%, -9%, +21% | Mixed |
| add post-training (SFT+DPO) to OLMo-7B | OLMo-7B · TinyStories | novelty delta (n=5): -0.026 -> +0.044 | Improved |
| add post-training (SFT+DPO) to OLMo-7B | OLMo-7B · CoPoet | novelty delta (n=5): -0.105 -> +0.231 | Improved |
| add post-training (SFT+DPO) to OLMo-7B | OLMo-7B · MacGyver | novelty delta (n=5): -0.294 -> -0.168 | Improved |
| add post-training (SFT+DPO+RLVR) to OLMo-2-7B | OLMo-2-7B · TinyStories | novelty delta (n=5): +0.225 -> +0.378 | Improved |
| improve base model from OLMo-7B to OLMo-2-7B | OLMo · CoPoet | n-gram originality (n=5): 0.413 -> 0.772 | Improved |
| improve base model from OLMo-7B to OLMo-2-7B | OLMo · MacGyver | n-gram originality (n=5): 0.520 -> 0.850 | Improved |
| increase sampling temperature from 0.5 to 2.0 (ablation) | OLMo-7B and Pythia-12B · TinyStories, CoPoet, MacGyver | novelty: originality rises, quality falls | Mixed |
| add novel in-context examples (ablation) | OLMo-7B · CoPoet | novelty delta (n=5): +0.050 | Improved |
| add novel in-context examples (ablation) | OLMo-7B · TinyStories | novelty delta (n=5): -0.003 | No clear effect |
| add asking for novelty to prompt (ablation) | OLMo-7B-Instruct · MacGyver | novelty delta (n=5): -0.074 | Hurt |
Closest papers by meaning
- Within-Model vs Between-Prompt Variability in Large Language Models for Creative Tasks (NeurIPS 2026 · similarity 0.90)
- How new data permeates LLM knowledge and how to dilute it (ICLR 2025 · similarity 0.88)
- Benchmarking LLMs' Judgments with No Gold Standard (ICLR 2025 · similarity 0.88)
- Turning Up the Heat: Min-p Sampling for Creative and Coherent LLM Outputs (ICLR 2025 · similarity 0.88)
- Accelerating Unbiased LLM Evaluation via Synthetic Feedback (ICML 2025 · similarity 0.88)
- Quality-Diversity through AI Feedback (ICLR 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
