UxBench - How well AI models perform UX tasks
{{ glow }}

How well AI models predict user satisfaction

Ten models judged real conversations from an AI assistant, testing how well each one reads user intent and spots the patterns behind how people respond.
UXBENCH · OCT 19, 2026
RankModelLabOverall score (Avg-Acc)Likely rangeLiked replies recognised (Good-Acc)Disliked replies caught (Bad-Acc)
{{ r.rank }}{{ r.name }}{{ r.lab }}{{ r.overall }}{{ r.range }}{{ r.liked }}{{ r.disliked }}

How this test measures user experience

A model that reads user intent correctly, and can tell which replies satisfied the user and which did not, is better placed to design interfaces that apply usability best practices and established patterns. A model reads a conversation and predicts whether the user liked the last reply. A model that does this well can help teams catch unsatisfying replies before users see them.

How the test measures it

How the test works

Each model received the same set of conversations and the same instructions. Half of the replies were liked by users and half were disliked. Models were not told about the split.

  1. 01

    The model reads a real conversation

    It sees the earlier turns, the user's last message and the assistant's reply. The user's thumbs up or down is hidden.

  2. 02

    It gives one verdict

    Using the benchmark's official checklist (correctness, completeness, understanding of intent, clarity and conciseness, usefulness), it answers good or bad.

  3. 03

    The verdict is checked against the user's click

    A verdict counts as right when it matches what the user actually pressed.

Example, written for this page. Not from the dataset.
Earlier turns
User: How long should I rest a steak after cooking it? Assistant: A long, accurate answer about resting, carryover heat and how thickness changes the timing.
User's last message

OK, but for a 2 cm steak, how many minutes exactly?

Assistant's reply

Great question! Resting matters because the juices need time to redistribute through the meat.

1. Carryover heat: the inside keeps cooking for a few minutes after it leaves the pan.

2. Thickness: thicker cuts hold more heat, so they generally need longer.

3. A common rule is to rest meat for around half of its cooking time, depending on the cut and

The reply carries on for several more paragraphs and never gives a number.
Lenient judge: good User: thumbs down

A lenient AI judge would call this reply good, because everything in it is correct and well written. The trouble is that the model ignored the simple question the user asked, how many minutes for a 2 cm steak, and kept explaining resting meat instead. The user did not like that answer and gave it a thumbs down. What this test checks is whether a model can catch the same thing, that a reply which reads perfectly well on its own was disliked because of the context around it. The authors list verbosity and task incompleteness as the two most common reasons users disliked a reply.

What UXBench measures

Most AI benchmarks ask whether an answer is correct. UXBench asks whether the person who received it was satisfied.

It is built from more than 70,000 real conversations on Tencent Yuanbao, an AI assistant used in China, where people pressed thumbs up or thumbs down on a reply. Those clicks are the answer key.

This site runs UX Judge: a model reads a conversation and predicts whether the user liked the last reply. A model that does this well can help teams catch unsatisfying replies before users see them.

Read about the benchmark

How to read a score

Overall score (Avg-Acc)
The average of the next two numbers. 50% is what guessing scores, and 100% would mean every verdict matched the user.
Liked replies recognised (Good-Acc)
Of the replies users liked, how many the model called good.
Disliked replies caught (Bad-Acc)
Of the replies users disliked, how many the model called bad. This is the hard part: every model scores lower here than on liked replies.
Lenient or strict
A lenient judge calls most replies good, so it rarely rejects a good reply but misses many bad ones. A strict judge does the opposite. Nine of the ten models here are lenient.