Public leaderboards like LM Arena are becoming unreliable proxies for model performance. Teams implicitly or explicitly "benchmark" by optimizing for specific test sets. The superior strategy is to focus on internal, proprietary evaluation metrics and use public benchmarks only as a final, confirmatory check, not as a primary development target.

Standard automated metrics like perplexity and loss measure a model's statistical confidence, not its ability to follow instructions. To properly evaluate a fine-tuned model, establish a curated "golden set" of evaluation samples to manually or programmatically check if the model is actually performing the desired task correctly.

When an LLM sees the same document thousands of times, it prioritizes memorizing that specific content to lower its training loss. This gives a false impression of progress, while the model fails to generalize to new, unseen data. A study of Google's T5 corpus found one sentence repeated over 60,000 times.

The gap between benchmark scores and real-world performance suggests labs achieve high scores by distilling superior models or training for specific evals. This makes benchmarks a poor proxy for genuine capability, a skepticism that should be applied to all new model releases.

Seemingly simple benchmarks yield wildly different results if not run under identical conditions. Third-party evaluators must run tests themselves because labs often use optimized prompts to inflate scores. Even then, challenges like parsing inconsistent answer formats make truly fair comparison a significant technical hurdle.

Metrics like BLEU and ROUGE compare word overlap, not meaning. An LLM's output like "the cat is on the bed" could be semantically opposite to the ground truth "the cat is on the mat" but still score high. This highlights the need for more sophisticated, meaning-aware evaluations like LLM-as-a-judge.

Do not blindly trust an LLM's evaluation scores. The biggest mistake is showing stakeholders metrics that don't match their perception of product quality. To build trust, first hand-label a sample of data with binary outcomes (good/bad), then compare the LLM judge's scores against these human labels to ensure agreement before deploying the eval.

Contrary to the "more data is better" mantra, scaling with bad data actively degrades model performance. Undeduplicated data makes models "forgetful" and less intelligent over time. You cannot overcome poor data quality simply by adding more compute; better, cleaner data is more effective.

A flawed or unsolvable benchmark task can function as a 'canary' or 'honeypot'. If a model successfully completes it, it's a strong signal that the model has memorized the answer from contaminated training data, rather than reasoning its way to a solution.