Skip to content
TheChatArchiveThe Chat Archive
o3-2025-04-16· OpenAIHugging Face LMArena Human Preference 140K

how are you evaluating llms?

From a Chatbot Arena battleThe same person sent this prompt to two anonymous models at once. Compare it with gpt-4.1-2025-04-14’s answer.
  1. User

    how are you evaluating llms?

  2. Assistant

    Evaluating a large-language model (LLM) is a bit like evaluating a new hire who’s supposed to be a polymath—there’s no single “exam.” Instead, we use a toolbox of complementary methods that check different facets of its behavior:

    🏃‍♂️ 1. Automated Benchmarks (speedy & repeatable)
    • Perplexity: “How surprised is the model by the next word?” (lower = better for language modeling).
    • BLEU / ROUGE / METEOR: n-gram overlap for translation & summarization.
    • BERTScore, COMET, BLEURT: embed­ding-based metrics that look at semantic similarity instead of raw overlap.
    • Task suites:
    – MMLU (massive multitask knowledge)
    – Super-/GLUE (reasoning & NLU)
    – BIG-Bench & BIG-Bench-Hard (long-tail tasks, reasoning)
    – TruthfulQA (hallucination detection)
    – GSM8K & MathQA (math reasoning)
    – HumanEval, MBPP, CodeBERTScore (coding)
    – ToxiGen, RealToxicityPrompts (toxicity)
    – HELM or LLM Harness from EleutherAI (wrapper around many tasks)

    🤲 2. Human Judgments (slow but gold-standard)
    • Pairwise preference ranking (“Which answer is better?”) — used for RLHF reward models and arenas such as LMSYS Chatbot Arena.
    • Scalar scoring grids (1-5 or 1-7) on helpfulness, correctness, coherence, style.
    • Expert audits for specific domains (medical, legal).
    • Blind A/B tests with production users (a la search-engine evaluations).

    🌈 3. Holistic or Multi-Axis Frameworks
    • HHH (Helpfulness, Honesty, Harmlessness) from Anthropic.
    • SAFE Completion & Constitutional AI (policy-based safety checks).
    • Responsible AI scorecards: bias, toxicity, stereotyping, privacy leakage, jailbreak resistance.

    🔬 4. Stress & Adversarial Testing
    • Prompt mutational fuzzing (typos, paraphrases, adversarial triggers).
    • Jailbreak red-teaming (trying to elicit disallowed content).
    • Data extraction / privacy leakage probes.

    📈 5. Live Metrics in the Wild
    • User thumbs-up / ‑down plus textual feedback.
    • Retention, task success rate, latency, cost per 1k tokens.
    • Online learning A/B tests (control vs. new model).

    🧮 6. Diagnostic & Attribution Tools
    • Attention/gradient-based attribution to see why the model produced a token.
    • Chain-of-thought faithfulness checks (is the rationale consistent with the final answer?).
    • Neuron activation studies & probing classifiers.

    How we combine them

    1. Quick automated sweep → narrow candidate models.
    2. Human eval (crowd + experts) on a sample → calibrate metrics & detect blind spots.
    3. Safety & red-team passes.
    4. Online A/B rollout with guardrails & logging.
    5. Continuous monitoring/retraining as usage drifts.

    If you’re planning an evaluation yourself, the key questions are:
    • Which capability do you care about (knowledge, reasoning, code, safety, cost)?
    • How much room do you have for human-in-the-loop vs. fully automated?
    • What risk tolerance/safety bar is required for your domain (e.g., healthcare vs. creative writing)?

    Let me know what slice of this is most relevant to you and I can zoom in or share code snippets/frameworks!

Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.