Research map / Large language models / Language, safety and interpretability
Evaluating and Explaining Prompt Sensitivity of LLMs Using Interactions
Ruiyang Qin, Qingzhuo Wang, Tian Wang, Zhihua Wei, Wen Shen
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 |
|---|---|---|---|
| supervised fine-tuning (instruct/chat vs base model) (ablation) | LLMs (Llama-2, Qwen, Olmo, Mistral etc.) · ARC | IPS | Improved |
| increase model scale within a model series (ablation) | LLMs (Llama-2, Qwen2.5, Olmo2) · ARC | IPS | Improved |
| use dense architecture instead of MoE (ablation) | LLMs (Llama-2, Llama-3, Qwen, Olmo families) · ARC | IPS | Improved |
| dense vs MoE with matched active parameters (ablation) | LLMs · ARC (controlled for active parameters) | IPS | Improved |
| add in-context examples (few-shot vs 0-shot) (ablation) | LLMs (instruct/chat models) · ARC | IPS | Improved |
| add more shots beyond 1-shot (ablation) | LLMs (instruct/chat models) · ARC | IPS | Improved |
| four factors (SFT, scale, dense, few-shot) applied to low-order interactions (ablation) | LLMs · ARC | IPS_low | Improved |
| four factors (SFT, scale, dense, few-shot) applied to high-order interactions (ablation) | LLMs · ARC | IPS_high | No clear effect |
| change prompt template (semantically equivalent) even when output stays the same | LLMs · MCQ samples | fraction of unstable salient interactions: 60%-80% unstable | Hurt |
| change prompt template when output changes | LLMs · MCQ samples | share of same-sign and similar-effect interactions: 2.7%-6.8% stable | Hurt |
| vary salient-interaction threshold tau from 0.05 to 0.20 (ablation) | IPS metric · 50 LLMs | IPS ranking correlation: Spearman 0.9905, Pearson 0.9957 | No clear effect |
| apply analysis to open-ended Dolly-15k with paraphrases and instruction reordering (ablation) | Llama-2 family · Dolly-15k | IPS conclusions | No clear effect |
Closest papers by meaning
- Quantifying Language Models' Sensitivity to Spurious Features in Prompt Design or: How I learned to start worrying about prompt formatting (ICLR 2024 · similarity 0.92)
- Brittlebench: Quantifying LLM robustness via prompt sensitivity (NeurIPS 2026 · similarity 0.91)
- On the Worst Prompt Performance of Large Language Models (NeurIPS 2024 · similarity 0.89)
- Mapping from Meaning: Addressing the Miscalibration of Prompt-Sensitive Language Models (AAAI 2025 · similarity 0.89)
- Scaling Prompt Synthesis for Large Language Model Reasoning (ICML 2026 · similarity 0.88)
- TuCo: Measuring the Contribution of Fine-Tuning to Individual Responses of LLMs (ICML 2025 · similarity 0.88)
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.
