←
Blog

Improving Qwen3.8-27B using GRPO - Frist JetBlue vs WebInfinity

Jeung ParkSung Ho ChoiJoannier PinalesJJonathan Beltran

Overview

Frist is creating data infrastructure to enable next generation computer use agents and coding models. To advance the frontier, we build reinforcement learning environments at internet scale.

Researchers believe reducing the sim2real gap in environments is important for training models through reinforcement learning. This article explores how much fidelity and completeness contribute to performance.

For training, we have two phases. First, we generate grounding datasets (SFT). Second, we train the models using GRPO, full parameter fine tuning. This follows best practices established by UI-Tars, Web-Arena, and Microsoft Fara.

This is a training recipe for CUA, and the same protocol can be applied to browser use, since the Frist applications have super high DOM fidelity.

Model performance gains on three tiers of environments

We trained the model on 3 environment variations, each with different levels of completeness and fidelity. Then we evaluated model accuracy on the live site for flight booking.

  • WebInfinity - open source low fidelity RL environment creator
  • JB1 – a high fidelity replica with a few missing pages and components
  • JB2 – a high fidelity replica that most closely matches the live website

For live site task completion compared to the base model:

  • WebInfinity model is 2x more accurate
  • JB1 model is 2.75x more accurate
  • JB2 model is 4x more accurate

Full results generated from 32 flight booking attempts:

ModelHomeOutboundReturnCartTravelerSeatsReview & PayFinal success
JetBlue v290.6%84.4%75.0%75.0%43.8%34.4%28.1%25.00%
JetBlue v178.1%78.1%75.0%71.9%56.3%34.4%21.9%17.19%
WebInfinity78.1%68.8%50.0%46.9%21.9%15.6%15.6%12.50%
Vanilla65.6%53.1%34.4%34.4%21.9%9.4%6.3%6.25%
Share of the 32 attempts per model that reached each stage of the booking. Stage counts exclude errors from earlier stages cumulatively. Final success averages both judges’ full-task verdicts, including final-page requirements.

Fidelity comparison: WebInfinity vs Frist JetBlue

Fidelity difference between WebInfinity clone and Frist JetBlue clone on a fare selection page:

Frist JetBlue
The Frist JetBlue clone’s fare selection page: a nonstop DCA to SJU flight with Core and EvenMore fares and four fare cards, Blue Basic through Blue Extra, each with its inclusions and price.
WebInfinity
The WebInfinity clone’s fare selection page: three generic fare cards, Main Base, Main and Main Flex, above a flight list with Main and EvenMore prices.

We build the highest quality RL environments for consumer and enterprise applications. Research teams interested in collaborating, please reach out: jeung@frist.ai

We make the observable world verifiable.