Why AI models flop easy tests: how LLM benchmarks work

10 steps · about 4 h 35 min · Beginner

  1. [1hr Talk] Intro to Large Language Models (video, Andrej Karpathy (YouTube), 1 h)

    You need a working picture of what an LLM is and does before you can judge how it is tested.

  2. SimpleBench (article, SimpleBench authors, 15 min)

    This is the benchmark in the headline: trick questions and spatial, temporal and social reasoning that most people get right.

  3. SimpleBench (reference, Epoch AI, 10 min)

    It shows the exact prompt and setup, so you can see that a score is the result of a specific procedure.

  4. Chain-of-Thought Prompting Elicits Reasoning in Large Language Models (article, Jason Wei et al., arXiv, 30 min)

    SimpleBench's prompt asks models to think step by step, and this paper explains why that helps reasoning.

  5. A Not So Simple Way to Beat Simple Bench (article, Sane & McLean, arXiv, 25 min)

    It is a real case study of how models fail SimpleBench questions and how changing the approach changes the result.

  6. LLM Evaluation Benchmarks: What They Measure and What They Miss (article, LangChain, 15 min)

    A top benchmark score doesn't guarantee the model will do your task well, and this article shows where the gaps are.

  7. Don't Make Your LLM an Evaluation Benchmark Cheater (article, arXiv, 30 min)

    It introduces Goodhart's law and data leakage, two reasons a benchmark score can overstate ability.

  8. ReasonBENCH: Benchmarking the (In)Stability of LLM Reasoning (article, arXiv, 25 min)

    It shows that a single score hides run-to-run variance, which is why one flop or one win proves little.

  9. Inspect AI (doc, UK AI Security Institute, 20 min)

    The official docs break an evaluation into datasets, solvers and scorers, the parts of any benchmark.

  10. lm-evaluation-harness: A framework for few-shot evaluation of language models (repo, EleutherAI (GitHub), 45 min)

    Last, you can read and run a widely used open evaluation harness yourself on a small model.