Research map / Large language models / Alignment and preferences
Maximizing Mutual Information Between Prompt and Response Improves LLM Performance With No Additional Data
HyunJi Nam, Haoran Li, Natasha Jaques
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 |
|---|---|---|---|
| train with MIPO contrastive DPO pairs on top of personalized prompting | Qwen2.5-1.5B-Instruct · Multifaceted Bench | win rate: 39.29 -> 74.60 | Improved |
| train with MIPO on top of personalized prompting | Llama-3.2-1B-Instruct · Community Alignment | win rate: 78.00 -> 93.67 | Improved |
| train with MIPO on top of personalized prompting for the 7B model | Qwen2.5-7B-Instruct · PRISM | win rate: 58.25 -> 58.76 | No clear effect |
| add MIPO contrastive pairs instead of SFT on the model's own chosen responses | Llama-3.2-1B-Instruct · ARC-Easy | accuracy: 34.13 -> 51.87 | Improved |
| train with MIPO without any verifier | Llama-3.2-1B-Instruct · Reasoning/MCQ average (GSM, SVAMP, ARC-E, ARC-C) | average accuracy: 37.72 -> 45.25 | Improved |
| train with MIPO on strongly math-tuned 7B model | Qwen2.5-7B-Instruct · Reasoning/MCQ average (GSM, SVAMP, ARC-E, ARC-C) | average accuracy: 91.71 -> 92.22 | No clear effect |
| MIPO versus RLVR with ground-truth rewards | Qwen2.5-3B-Instruct · Reasoning/MCQ average (GSM, SVAMP, ARC-E, ARC-C) | average accuracy: 81.45 -> 87.86 | Improved |
| SFT on the model's own outputs without contrast (ablation) | Llama-3.2-1B-Instruct · Reasoning/MCQ average (GSM, SVAMP, ARC-E, ARC-C) | average accuracy: 37.72 -> 35.48 | Hurt |
| replace reference-policy mutual information (DPO) with online MI reward via PPO (ablation) | Qwen2.5-1.5B-Instruct · Community Alignment | win rate: 78.83 -> 39.67 | Hurt |
| add RLAIF reward at 1B scale | Llama-3.2-1B-Instruct · PRISM | win rate: 72.17 -> 56.87 | Hurt |
| apply MIPO to improve output diversity | Qwen 1.5B · Personalization tasks | Self-BLEU-4: 0.310 -> 0.256 | Improved |
| use random or missing-context negatives instead of reshuffling chosen responses (ablation) | MIPO · Personalized instruction-following | win rate: random +1.3-13%, missing context +3-30% over reshuffling | Improved |
Closest papers by meaning
- Self-Boosting Large Language Models with Synthetic Preference Data (ICLR 2025 · similarity 0.89)
- Reward-Augmented Data Enhances Direct Preference Alignment of LLMs (ICML 2025 · similarity 0.88)
- 3D-Properties: Identifying Challenges in DPO and Charting a Path Forward (ICLR 2025 · similarity 0.88)
- Scaling Prompt Synthesis for Large Language Model Reasoning (ICML 2026 · similarity 0.88)
- Test-Time Preference Optimization: On-the-Fly Alignment via Iterative Textual Feedback (ICML 2025 · similarity 0.88)
- Less is More: Improving LLM Alignment via Preference Data Selection (NeurIPS 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.
