Research map / Large language models / Instruction and fine-tuning
The Unlocking Spell on Base LLMs: Rethinking Alignment via In-Context Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Analyzes token distribution shift between base and aligned LLMs, then proposes URIAL, a tuning-free alignment method using a system prompt and three restyled constant in-context examples on base LLMs, evaluated on a new 1,000-instruction multi-aspect benchmark (Just-Eval-Instruct).
Problem
Alignment tuning (SFT and RLHF) is costly, and it is unclear how it transforms base LLMs; the paper asks whether base LLMs can be aligned without any weight tuning.
Approach
Token-level analysis shows alignment mainly shifts stylistic tokens, so a constant system prompt plus a few stylistically curated in-context examples can align base LLMs purely through ICL, matching or beating SFT/RLHF models built on strong bases.
Key results
- URIAL with Mistral-7b (K=3) scores 4.63 average, above Mistral-7b-Instruct (SFT) at 4.44.
- URIAL with Llama2-70b (K=3) scores 4.74 average, above Llama2-70b-chat (RLHF) at 4.67, close to gpt-3.5-turbo (4.75).
- URIAL with Llama2-7b (K=3) scores 4.33 average, versus 2.90 zero-shot and 3.72 retrieval ICL, approaching SFT/RLHF 7B models (4.46-4.47).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace zero-shot templated prompting with URIAL using three restyled examples and a system prompt | Llama-2-7B · URIAL | average score: 2.90 -> 4.33 | Helped |
| Replace vanilla ICL with URIAL using three restyled examples and a system prompt | Llama-2-7B · URIAL | average score: 3.18 -> 4.33 | Helped |
| Replace retrieval ICL with URIAL using three restyled constant examples and a system prompt | Llama-2-7B · URIAL | average score: 3.72 -> 4.33 | Helped |
| Use URIAL instead of SFT | Mistral-7B · URIAL | average score: 4.44 -> 4.63 | Helped |
| Use URIAL instead of RLHF | Llama-2-70B · URIAL | average score: 4.67 vs 4.74 | Helped |
| Increase URIAL examples from three to eight (ablation) | Llama-2-7B · URIAL | safety: 4.29 to 4.97 | Helped |
| Increase URIAL examples from three to eight (ablation) | Llama-2-7B · URIAL | other aspects | Hurt |
| Increase URIAL examples from three to eight (ablation) | Mistral-7B · URIAL | safety | Helped |
| Increase URIAL examples from three to eight (ablation) | Mistral-7B · URIAL | engagement | Helped |
| Increase URIAL examples from three to eight (ablation) | Mistral-7B · URIAL | average score: 4.63 vs. 4.62 | No clear effect |
| Use one URIAL example instead of three (ablation) | Llama-2-70B · URIAL | average score: 4.62 | No clear effect |
| Use a different set of three URIAL examples (ablation) | Mistral-7B · URIAL | average score: 4.63 vs. 4.64 | No clear effect |
Surprising findings
- Base and aligned models share top-ranked tokens at most positions; 77.7% of tokens are unshifted, and aligned tokens are usually within the base model's top 5.
- Stylistic, well-written static examples outperform dynamically retrieved semantically similar examples.
Limitations
- Open-source models still lag GPTs on coding, math, and STEM topics.
- Evaluation relies on GPT-4/ChatGPT judges with 1-5 scores and a modest human study of 100 examples.
- Tuning may still be necessary for tasks such as coding, mathematics, and interactive agents.
Experimental setup
Datasets: AlpacaEval MT-Bench LIMA HH-RLHF-redteam MaliciousInstruct open-instruct UltraChat
Benchmarks: Just-Eval-Instruct
Metrics: helpfulness score (1-5) clarity score factuality score depth score engagement score safety score average score response length
Baselines: Zero-shot templated prompting Vanilla ICL Retrieval ICL Vicuna-7b Llama2-7b-chat Mistral-7b-Instruct Llama2-70b-chat gpt-3.5-turbo gpt-4
Builds on: LIMA Retrieval ICL Llama-2-chat system prompt AlpacaEval MT-Bench
Closest papers by meaning
- Align-Pro: A Principled Approach to Prompt Optimization for LLM Alignment (AAAI 2025 · similarity 0.91)
- Alignment Dynamics in LLM Fine-Tuning (NeurIPS 2026 · similarity 0.91)
- LLM Alignment as Retriever Optimization: An Information Retrieval Perspective (ICML 2025 · similarity 0.91)
- Your Weak LLM is Secretly a Strong Teacher for Alignment (ICLR 2025 · similarity 0.91)
- Beyond Imitation: Leveraging Fine-grained Quality Signals for Alignment (ICLR 2024 · similarity 0.90)
- LLM Alignment--Utility Asymmetry under Semantic-Preserving Transformations (NeurIPS 2026 · similarity 0.90)
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.