Frontier Lab Notes

Evaluation Benchmarks and Scientific Method

Evaluation is how LLM research knows whether it made progress. Bad evals create fake progress. Good evals reveal capability, failure modes, and scaling behavior.

Primary sources:

Why Evaluation Is Hard

LLMs are broad systems. A single score can hide:

Types Of Evaluation

Loss Evaluation

Measures predictive performance on held-out text.

Pros:

Cons:

Multiple-Choice Benchmarks

Examples: MMLU-style tasks.

Pros:

Cons:

Generative Reasoning Benchmarks

Examples: GSM8K, MATH, code tasks.

Pros:

Cons:

Human Or Model Preference Evals

Pros:

Cons:

Systems Evals

Measure:

These are essential for Frontier Model Systems and Inference.

Contamination

Benchmark contamination happens when eval examples appear in training data.

Research defense:

Prompt Sensitivity

A model's score can change with:

Always log prompts.

Capability Versus Elicitation

If a model fails, there are multiple explanations:

This connects to Post Training Alignment and Reasoning.

Scientific Method For LLM Research

For any claim:

  1. Define the bottleneck.
  2. Choose a baseline.
  3. Isolate one variable.
  4. Run enough seeds or sizes to avoid noise.
  5. Evaluate across domains.
  6. Include systems cost.
  7. Check contamination.
  8. Inspect failures.
  9. State what would falsify the claim.

A Good Eval Suite For Architecture Work

Use a mix:

Research Questions

Related