Research map / Large language models / Alignment and preferences
LIMA: Less Is More for Alignment
Chunting Zhou, Pengfei Liu, Puxin Xu, Srini Iyer, Jiao Sun, Yuning Mao, Xuezhe Ma, Avia Efrat, Ping Yu, LILI YU, Susan Zhang, Gargi Ghosh and 3 more
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 |
|---|---|---|---|
| Use diverse prompts rather than homogeneous how-to prompts (ablation) | LLaMa 7B fine-tuned on 2,000 examples · Stack Exchange versus wikiHow | ChatGPT helpfulness score | Improved |
| Use quality-filtered responses rather than unfiltered responses (ablation) | LLaMa 7B fine-tuned on 2,000 Stack Exchange examples · Stack Exchange | ChatGPT helpfulness score: 0.5 point difference | Improved |
| Increase training-set quantity without increasing prompt diversity (ablation) | LLaMa 7B trained on quality-filtered Stack Exchange · Stack Exchange | ChatGPT helpfulness score: up to a 16-fold increase | No clear effect |
| Add 30 multi-turn dialogue chains (ablation) | LIMA trained from pretrained LLaMa · 10 live conversations | excellent responses: 45.2% to 76.1% | Improved |
| Add 30 multi-turn dialogue chains (ablation) | LIMA trained from pretrained LLaMa · 10 live conversations | failed turns: 15 fails per 42 turns to 1 fail per 46 | Improved |
| Add six examples with varied formatting constraints (ablation) | LIMA · Prompts requiring complex structure | response quality | Improved |
| Fine-tune on 1,000 curated examples rather than using only the pretrained model | LLaMa 65B · 300 test prompts | human preference versus GPT-4: 43% | Mixed |
| Fine-tune on 1,000 curated examples rather than using only the pretrained model | LLaMa 65B · 300 test prompts | human preference versus DaVinci003: 65% | Improved |
| Train on 1,000 curated examples rather than 52,000 Alpaca examples | LLaMa 65B · 300 test prompts | human preference | Improved |
Later papers that compared with this method
4 later accepted papers report LIMA as a baseline in their results tables; 4 of them report a better number on most of the results they share with it.
- Neuron-Aware Data Selection in Instruction Tuning for Large Language Models (ICLR 2026 · better on 1 of 1 shared result)
- Improving Data Efficiency via Curating LLM-Driven Rating Systems (ICLR 2025 · better on 2 of 2 shared results)
- Long Is More for Alignment: A Simple but Tough-to-Beat Baseline for Instruction Fine-Tuning (ICML 2024 · better on 4 of 4 shared results)
- Self-Alignment with Instruction Backtranslation (ICLR 2024 · better on 1 of 1 shared result)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- The Unlocking Spell on Base LLMs: Rethinking Alignment via In-Context Learning (ICLR 2024 · similarity 0.88)
- Sheared LLaMA: Accelerating Language Model Pre-training via Structured Pruning (ICLR 2024 · similarity 0.88)
- Reinforcement Mid-Training (ICLR 2026 · similarity 0.88)
- An Emulator for Fine-tuning Large Language Models using Small Language Models (ICLR 2024 · similarity 0.88)
- WizardLM: Empowering Large Pre-Trained Language Models to Follow Complex Instructions (ICLR 2024 · similarity 0.88)
- A Closer Look at the Limitations of Instruction Tuning (ICML 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.
