This video provides a comprehensive overview of methods for evaluating Large Language Models (LLMs) on subjective language generation tasks, moving beyond simple benchmark scores. It categorizes evaluation techniques into three main types: reference-based evaluation (using metrics like BLEU, ROUGE, BERTScore, and COMET), human evaluation (including pairwise comparisons like LMSys Chatbot Arena and addressing biases with style control), and using LLMs themselves as judges (assessing outputs based on predefined criteria or pairwise preference, while also considering and mitigating potential biases). The video highlights the pros and cons of each approach, discussing aspects like cost, nuance, and the evolving role of AI in its own assessment, as exemplified by the MT-Bench findings where GPT-4 showed high agreement with human judgments.
A Guide to Evaluating Large Language Models on Subjective Tasks
00:00:02 Introduction to LLM Benchmarks & Evaluation Challenges: The video begins by noting the increasing benchmark scores of new LLM releases (e.g., MMLU, GPQA for multiple choice; HumanEval for code). It highlights the difficulty in evaluating subjective tasks, such as explaining a concept clearly to a specific audience, and sets out to provide an overview of evaluation methods for such tasks.
00:00:52 Sponsor Segment: VoiceWriter.io: Introduces VoiceWriter, an AI-powered tool designed to help users write faster using voice. It features real-time grammar correction, punctuation addition, advanced speech recognition, and a Chrome extension. A free trial is offered.
This method is useful when a reference (ideal output) is available, common in tasks like summarization and machine translation. It's best for well-defined answers with limited variability.
00:01:52 BLEU and ROUGE: These are traditional metrics that calculate the overlap of n-grams between the model's output and the reference. They perform surface-level matching and are not very nuanced, treating tokens as simply correct or incorrect without judging the degree of error.
00:02:39 Advanced Model-Based Reference Scoring: More sophisticated approaches use language models.
00:02:45 BERTScore: Employs an encoder LM to get embeddings for words in both candidate and reference sentences, then scores the similarity of these embedding sequences.
00:03:09 COMET and BLURT: These metrics involve training a model to predict how a human would rate the similarity between two sentences. COMET, for example, encodes both sentences and uses a regression model for this prediction.
Involves humans directly assessing LLM outputs. Methods include rating output quality or, increasingly popular, pairwise comparison where users select the preferred of two model responses.
00:03:54 LMSys Chatbot Arena: An open-source platform where users query two random models and pick the better response. Aggregated ratings generate an ELO score, creating a model leaderboard.
00:04:42 Pros and Cons: While providing valuable insight, human evaluation is expensive and annotators aren't perfect, often influenced by superficial aspects like output length, formatting (e.g., use of bold text, emojis).
00:05:31 Solution - Style Control: To address biases, a regression model can be fitted to account for style attributes (length, formatting) that might influence user preference, thereby disentangling style from true content quality. This showed some models (e.g., Claude 3.5) perform better after style control, while others (e.g., GPT-4o mini) perform worse, indicating their wins were sometimes due to formatting.
This approach uses another LLM to evaluate the output(s) of one or more LLMs.
00:06:57 Setup: The judge LLM can pick the better of two outputs or rate outputs on predefined criteria (e.g., helpfulness, relevance, accuracy, depth, creativity). Multi-factor rating provides deeper insights. (Example given: VoiceWriter uses LLMs to evaluate style adherence and content preservation for its grammar correction tool).
00:08:07 Validation: The MT-Bench paper found that GPT-4's agreement with human judgments (85%) was even higher than human-human agreement (81%), suggesting LLMs can be reliable judges.
00:08:39 Potential Biases: LLM judges can suffer from biases similar to humans, such as length bias (favoring longer responses), position bias (favoring the first response shown), or self-preference bias (a model rating its own outputs higher). These need to be controlled.
00:09:09 Cost and Utility: While using LLMs for evaluation has a cost, it's typically cheaper than human evaluation, with a range of models offering different cost-quality trade-offs. LLMs are also useful for evaluating structured short answers where rule-based evaluation is hard (e.g., OmniMath benchmark for mathematical expressions).
00:09:53 Key Takeaway: Evaluating LLMs on subjective tasks requires moving beyond simple metrics to a combination of reference-based, human, and LLM-as-judge evaluations. Each method has its strengths, weaknesses, and associated costs, and understanding biases (and how to control for them) is crucial for accurate assessment.