Skip to content

Preference and RLHF tasks: teaching a model what good looks like

Pathwize EditorialExpert data and AI evaluation2 min read
InsightPathwize AI
Guides

Pairwise comparison is the workhorse of RLHF. How preference tasks work, why the rationale matters as much as the choice, and how to design them well.

The preference task is the workhorse of RLHF. An expert is shown two model outputs for the same prompt and asked a deceptively simple question: which is better, and why? Collect enough of these pairwise judgments and you can train a reward model that captures what people actually prefer, then use it to steer the model itself.

Why pairwise, not a score

You could ask an expert to score each answer out of ten, but people are inconsistent at absolute scores and drift over a shift. Comparison is far more stable: humans are good at saying which of two things is better even when they struggle to put a number on either. Pairwise data is the signal reward models are built to learn from.

The rationale is not optional

A bare choice (A over B) tells you what an expert preferred but not why. The rationale is where the value is. It surfaces the tie-breaker the expert applied, catches the case where they picked correctly for the wrong reason, and gives your team something to audit when two experts disagree. A preference task without a rationale is throwing away half the signal.

Designing a good preference task

Keep the two outputs genuinely comparable: same prompt, similar length, no giveaway formatting that biases the choice. Tell experts what dimension to weigh (helpfulness, safety, factual accuracy) so different raters optimise the same thing. And allow a considered tie, because forcing a choice between two equal answers manufactures noise.

Use more than one rater per item where the judgment is subtle, and treat disagreement as information: a pair that splits your experts is often a pair worth a second look.

Where preference data goes wrong

The classic failure is contaminated feedback: a contributor quietly routes the task through an LLM and submits the model's preference as their own. You then train your model to imitate another model while believing you captured human judgment. The defence is provenance and real expertise, knowing who judged each pair, under what instructions, and that they were qualified to.

On Pathwize

A preference task on Pathwize routes each pair to credential-verified experts and records the choice, the rationale, and who made it, as a signed trail. You get comparison data you can defend, not just a spreadsheet of A-or-B. Create one from the AI tasks console and set how many experts should judge each item.

Frequently asked questions

What is a preference task in RLHF?+

A task where an expert compares two model outputs for the same prompt and picks the better one, with a rationale. Enough of these pairwise judgments train a reward model that captures human preference.

Why use pairwise comparison instead of scoring each answer?+

People are inconsistent at absolute scores but stable at comparisons. Pairwise judgments are more reliable and are the exact signal reward models are designed to learn from.

Why does the rationale matter?+

The rationale reveals the tie-breaker the expert applied, catches right-answer-wrong-reason cases, and gives you something to audit when experts disagree. Without it you keep the choice but lose why.

Related reading

See Pathwize on your own data

Source verifiable expert data with provenance built in, EU-native and audit-ready.

Book a demo
← All posts

Related stories

GuidesPathwize AI

How to evaluate an AI data vendor: a due-diligence checklist

The questions that separate a data vendor you can defend to an auditor from one that will cost you a re-labelling project six months in.

Pathwize AIGuides

How to run a human-data pilot before you commit

A pilot is the cheapest way to learn whether a data partner can actually do your hardest work. Here is how to design one that tells you something real.

GuidesPathwize AI

The four types of AI data tasks, explained

Preference, evaluation, red teaming and domain judgment. What each task type does, when to use it, and how they fit together in a data programme.