LLM Evaluation
You can't improve what you don't measure. LLM evaluation is the discipline of systematically testing model outputs against quality criteria — accuracy, consistency, safety, and cost — so you know whether your AI is getting better or silently getting worse.
A complete eval pipeline runs test cases through the model, scores outputs with multiple complementary methods, and gates the final decision: ship if scores meet thresholds, iterate if they don't.
32% of teams cite quality as the #1 barrier to production deployment.
According to LangChain's June 2026 State of AI report, nearly one in three teams names output quality as their top obstacle. The problem isn't that models are bad — it's that model behavior shifts over time. Providers update endpoints silently. Prompts that worked perfectly last month start returning different answers. Edge cases that nobody thought to test fail in production.
Without systematic evaluation, you're deploying blind. You don't know whether your last prompt change improved the system or broke it. You don't know if the model upgrade you're excited about actually makes your use case better. And when a user reports a bad response, you have no way to prove it used to work — or to catch it before it ships again. No evals means no confidence.
Five dimensions that matter for every LLM deployment.
Accuracy
Did the model give the right answer? For classification tasks this is straightforward — does it match the label? For generative tasks it's harder: did it include all required information, avoid hallucination, and stay factually correct?
Consistency
Run the same prompt 10 times and you should get substantially the same output. High variance — especially on factual questions — is a red flag. Consistency matters most in production where unpredictable behavior erodes user trust.
Safety
Does the model refuse when it should? More importantly, does it refuse when it shouldn't? Over-refusal is its own failure mode — blocking legitimate medical or legal queries because they contain trigger words.
Latency & Cost
Is the model fast enough for your use case? A perfect answer that takes 12 seconds is useless for a chatbot, and a $0.50 answer that could have been handled by a $0.002 model is just waste. Track both per-call and aggregate.
Task Completion
For agentic systems: did the agent finish the task or get stuck in a loop? Did it use the right tools? Did it produce a usable result or an error? Task completion rate is the bottom-line metric for agent quality.
Four evaluation strategies, from quick and automated to slow and thorough.
Offline Evals
A curated test set of inputs with expected outputs. Run it before every deploy. Quick, repeatable, and consistent — the unit tests of the LLM world. Build a dataset of 20 to 50 hand-labeled examples covering your most important use cases, then expand as you discover failures. Every prompt change and model upgrade should pass this suite before shipping.
Online Evals
Monitor real production traffic in real time. Sample a percentage of user interactions, score them automatically, and alert on regressions. Catches problems your offline test set missed — real users ask questions your test cases never considered. Harder to set up but essential for high-stakes deployments.
Human Review
Domain experts spot-check a sample of outputs on a regular cadence. Nothing beats a human reading the actual output and judging whether it's good. But it's slow, expensive, and doesn't scale. Use it to calibrate your automated evals and to catch subtle failures that rule checks and LLM judges miss.
LLM-as-Judge
Use a stronger model to score the outputs of a weaker one. Give the judge model the input, the output, and a rubric — then ask it to rate on dimensions like correctness, completeness, and tone. Much faster than human review, works at scale, and correlates reasonably well with human judgment when the rubric is well-designed.
Four ways evaluation goes wrong — and how to avoid them.
Benchmark Leakage
Your model was trained on your test set — or on data that overlaps heavily with it. You're not measuring capability, you're measuring memorization. Mitigation: use private, hand-curated datasets. Never evaluate on publicly available benchmarks without verifying the model hasn't seen them.
Goodhart's Law
"When a measure becomes a target, it ceases to be a good measure." Optimize for BLEU score and your model will produce stilted, unnatural text that scores well but reads terribly. Always validate that metric improvements correspond to real quality improvements as judged by humans.
One-Metric Blindness
BLEU and ROUGE are useless for creative tasks — they penalize variation. A perfect poem that uses different words than the reference scores zero. Match your metrics to your task. Creative generation needs human or LLM-as-judge evaluation, not n-gram overlap.
Prompt Sensitivity
Same model, same input, different prompt phrasing — and scores swing by 20+ points. A system that passes with one prompt can fail with a semantically equivalent rewording. Always evaluate across multiple prompt variants and report the range, not just the best-case score.
An eval pipeline is a habit, not a one-time project.
Start small and build incrementally. The teams with the best evaluation practices didn't build everything at once — they started with 20 hand-labeled examples and a script that ran them before deploys, then expanded from there.
Tools that make this practical: LangSmith and Braintrust provide managed eval runners with dashboards and collaboration. Galileo specializes in observability and drift detection. For teams that prefer a lightweight approach, a pytest suite with LLM calls gives you the same capabilities with zero vendor lock-in. The tool matters less than the discipline — the best eval pipeline is the one your team actually runs.
"If you can't measure it, you can't improve it. If you can't reproduce the measurement, you can't trust it."
LLM behavior drifts over time — providers update endpoints, prompts degrade, and edge cases emerge. Run evals before every deploy or you're shipping blind.
No single eval method is sufficient. Combine offline test suites, online monitoring, LLM-as-judge, and human review for a complete picture of quality.