Your bookmarks

Loading your bookmarks

Latest updates

New courses, skills, book chapters, and site improvements will appear here.

Create free accountSign upDashboardDashboard

Capability elicitation

  • Failures
  • Research term
2 steps

Capability elicitation is trying different prompts, examples, tools, and setups on the same model to find the best performance it can reach on a task.

A bare request leaves the model to guess what you want and how to get there. Examples show the format and standard, splitting a task keeps each part small enough for one pass, and tools supply missing facts. Keeping the best of several answers makes use of how answers vary from try to try.

Where you’ll meet it

In your team's chat, a coworker posts that the AI can't read your scanned supplier invoices. They uploaded all two hundred in one go and got back a jumble of half-filled rows.

What to do

Before the team drops it, take twenty invoices, a few hard to read. Send them once with a plain request, once with two worked examples, and once five at a time. Compare the rows with the scans, and time each way, since the team runs it every month.

Easy to mix up

Fine-tuning trains the model further, so only people allowed to retrain it can do it. Capability elicitation leaves the model unchanged, so anyone who can send it requests can try it.

Check yourself

Two models tie on a test of pulling data from documents when each gets one plain prompt. Given the same two worked examples, one of them pulls well ahead. Which is better at the task?

Go deeperSafety tests aim high, and a model holding back could fool them

A safety test asks whether a model could help with something dangerous. Testers give the model the best setups they can build. Some go a step past elicitation and fine-tune it on a little task data, because anyone who obtains a model's weights, the numbers that make up the trained model, could do the same.

Sandbagging, a model holding back on purpose during a test, would hide that ceiling. Researchers study it by training models to hide a skill unless a password appears in the request, then measuring which elicitation methods bring the skill out anyway. The Elicitation Game is one such study.

Practice Modern AI

Or take the learning quiz on a mix of every set