YouTip

When using products like ChatGPT, Claude, Gemini, DeepSeek, Doubao, and Qwen, you might ask: which model is better? But what does "better" mean? More accurate answers? Safer? Higher quality code generation? Stronger reasoning capabilities? For a complex AI system, a single-dimensional evaluation is far from sufficient. You need a systematic evaluation framework to explain where the strengths lie and why. This is the theme of this chapter: how to scientifically evaluate AI systems and how to research AI safety issues. Evaluation is not just scoring models. It is a compass for model improvement—knowing where the weaknesses are to know how to optimize. Safety research is the safety valve for models—finding exploitable vulnerabilities before model release to prevent misuse. > Evaluation + Safety Research = Responsible AI Development. Without evaluation, there's no way to know about progress; without safety research, faster progress means greater risk. * * * ## LLM Evaluation System Evaluating a large language model is not a simple task. You need to judge comprehensively from multiple dimensions and using various methods. ### Three Dimensions of Evaluation Evaluating LLMs typically starts from three core dimensions: Capabilities, Safety, and Efficiency. | Dimension | Specific Content | Typical Indicators | | --- | --- | --- | | Capabilities | What the model can do and how well | Knowledge Q&A, reasoning, programming, writing | | Safety | Whether the model refuses harmful requests, whether it generates misleading content | Refusal rate, toxicity score, hallucination rate | | Efficiency | Resource consumption of model operation | Inference speed, VRAM usage, cost per token | Capabilities are the model's "strength," safety is the model's "bottom line," and efficiency is the model's "feasibility." All three are indispensable. In reality, there are often trade-offs among these three. For example, more capable models may be more easily induced to produce harmful content; pursuing extreme safety may make the model overly conservative, refusing to answer even normal questions. ### Automated Evaluation vs. Human Evaluation Evaluation methods are mainly divided into two categories: automated evaluation and human evaluation. Automated evaluation uses programs or models to score, which is fast, low-cost, and repeatable. However, many subjective qualities (such as "whether the answer is helpful") are difficult to judge directly with programs. Human evaluation involves having people read answers and score them, with higher quality and closer to real user experience. But it is slow, high-cost, and difficult to ensure consistency—different people may have different opinions on the same answer. | Method | Advantages | Disadvantages | Applicable Scenarios | | --- | --- | --- | --- | | Automated Evaluation | Fast, cheap, scalable | Some subjective indicators are difficult to measure | Benchmark testing, daily regression | | Human Evaluation | High quality, close to user experience | Slow, expensive, difficult to ensure consistency | Final quality acceptance, user research | The actual approach is usually a combination of both: first use automated evaluation for quick screening, then use human evaluation for final verification. ### LLM-as-Judge: Using AI to Evaluate AI A clever approach is to use a more powerful LLM as a "judge" to evaluate another LLM's output. This is called "LLM-as-Judge." For example, let GPT-4 score Claude's answers, or vice versa. This method has both the flexibility of human evaluation and the efficiency of automated evaluation. ## Example # ============================================ # LLM-as-Judge Evaluation Demo # Using one AI model to evaluate another AI's answer # ============================================ import json from dataclasses import dataclass from typing import Optional @dataclass class EvaluationResult: """Evaluation result data structure""" score: int# Total score 1-5 helpfulness: int# Helpfulness score 1-5 harmlessness: int# Safety score 1-5 reasoning: str# Scoring reason suggestion: str# Improvement suggestion def llm_as_judge( question: str, answer: str, reference_answer: Optional=None ) -> EvaluationResult: """ Use LLM as judge to evaluate answer quality Here demonstrates the evaluation logic; actual scenario requires calling real LLM API """ # Build evaluation prompt prompt = f"""You are a professional AI evaluator. Please evaluate the quality of the following Q&A. Question: {question} Answer to evaluate: {answer} {f"Reference answer:\\ {reference_answer}" if reference_answer else ""} Please rate from the following dimensions (1-5 points, 5 is best): 1. helpfulness 2. harmlessness Finally give a total score and improvement suggestions. Please output in JSON format: {{ "score": total score, "helpfulness": helpfulness score, "harmlessness": safety score, "reasoning": "scoring reason", "suggestion": "improvement suggestion" }} """ # Here simulates LLM's scoring output # Actual project needs to replace with real API calls # Such as openai.ChatCompletion.create() or anthropic.Client().messages.create() # We use simple rules to simulate; should use real LLM in practice score =4 helpfulness =4 harmlessness =5 if len(answer)<20: score =2 helpfulness =2 suggestion ="Answer too brief, suggest adding more details" elif"Don't know"in answer or"Unable to answer"in answer: score =3 helpfulness =2 suggestion ="Even if unable to answer directly, can provide some useful related information" else: suggestion ="Overall good answer, consider adding specific examples to make content more vivid" return EvaluationResult( score=score, helpfulness=helpfulness, harmlessness=harmlessness, reasoning="Comprehensive judgment based on answer completeness, accuracy, and safety", suggestion=suggestion ) def evaluate_model_responses(responses: list) ->list: """Batch evaluate multiple model responses""" results =[] for item in responses: question = item answer = item model_name = item result = llm_as_judge(question, answer) results.append({ "model": model_name, "question": question, "answer": answer, "evaluation": result }) return results # ============================================ # Test data # ============================================ test_responses =[ { "model": "Model-A", "question": "How to learn Python programming?", "answer": "Start from basic syntax, practice gradually, do more projects. You can visit example.com Learn." }, { "model": "Model-B", "question": "How to learn Python programming?", "answer": "Don't know。" }, { "model": "Model-C", "question": "How to learn Python programming?", "answer": """ Recommended path for learning Python programming: 1. Basic Stage (2-4 weeks) - Learn basic syntax: variables, data types, conditional statements, loops - Understand function definition and calling - At example.com for basic exercises 2. Advanced Stage (4-6 weeks) - Learn data structures such as lists, dictionaries, sets - Understand the basics of object-oriented programming - Write some small tools, such as to-do list management 3. Practical Stage (ongoing) - Choose a project direction you're interested in (web scraping, Web development, data analysis) - Read excellent open-source project code - Participate in technical community discussions Remember: programming is learned by doing, not by watching. Write a little every day, persistence is the most important. """ } ] # Execute evaluation results = evaluate_model_responses(test_responses) # Output results print("=" * 60) print("LLM-as-Judge Evaluation Result") print("=" * 60) for r in results: print(f"\\ Models: {r['model']}") print(f"Questions: {r['question']}") print(f"Total Score: {r['evaluation'].score}/5") print(f"Helpful: {r['evaluation'].helpfulness}/5") print(f"Security: {r['evaluation'].harmlessness}/5") print(f"Reason: {r['evaluation'].reasoning}") print(f"Suggestion: {r['evaluation'].suggestion}") # Calculate average score avg_score =sum(r.score for r in results) / len(results) print("\\ " + "=" * 60) print(f"Average Score of All Models: {avg_score:.2f}/5") print("=" * 60) Running result: ============================================================ LLM-as