model evaluation llm
**LLM Capability Elicitation and Evaluation** is the **systematic process of measuring what a language model can and cannot do** — including prompt engineering for evaluation, avoiding contamination, and interpreting benchmark results correctly.
**The Evaluation Challenge**
- LLMs are sensitive to prompt formatting — same capability, different prompt → different score.
- Benchmark contamination: Training data may include test examples.
- Prompt sensitivity: "Answer:" vs. "The answer is:" can change accuracy by 10%.
- True vs. elicited capability: Model may know but fail to express correctly.
**Evaluation Methodologies**
**Few-Shot Prompting for Evaluation**:
- Include K examples in prompt before the test question.
- K=0 (zero-shot): Tests true generalization.
- K=5 (5-shot): Helps model understand format — reveals more capability.
- GPT-3 paper: 5-shot outperforms 0-shot by 20+ points on many benchmarks.
**Chain-of-Thought Evaluation**:
- Complex reasoning: CoT prompting ("think step by step") reveals reasoning.
- Direct answer vs. CoT: 65% → 92% on GSM8K for GPT-4.
**Contamination Detection**
- n-gram overlap: Check if test questions appear in training data.
- Membership inference: Does model complete test examples unusually well?
- Dynamic benchmarks: New questions generated after model's training cutoff.
- LiveBench: Continuously updated benchmark with recent data.
**Evaluation Dimensions**
| Dimension | Key Benchmarks |
|-----------|---------------|
| Knowledge | MMLU, ARC |
| Reasoning | GSM8K, MATH, BBH |
| Code | HumanEval, SWE-bench |
| Instruction following | IFEval, MT-Bench |
| Safety | TruthfulQA, AdvGLUE |
**Human Evaluation**
- Automated benchmarks miss: Fluency, creativity, factual grounding, tone.
- Chatbot Arena (LMSYS): Blind pairwise comparison — Elo rating from human preferences.
- Most reliable ranking but expensive and slow.
Robust LLM evaluation is **a critical and unsolved problem in AI** — with models increasingly exceeding benchmark saturation, understanding the gap between benchmark performance and real-world capability requires ever more sophisticated evaluation methodologies that resist gaming and contamination.