Skip to content
Trending storyDeveloping

Peking University and Samsung build RealtimeBench and TORSI for real-time game agents

1 article1 sourcesince Oct 16Last article 3h ago ·

Overview

AISummary of 1 article

Peking University and Samsung researchers built RealtimeBench, 180 tasks across six games including Mario, Snake and Tetris, that keep running during inference.

In paused evaluations, the slow reasoning model GPT-6.1-sol beat the faster model Jev in five of six games, but in live play it scored 1.51 out of 100 against Jev's 13.51, according to a report by Rohan Paul on X.

The researchers' TORSI method has the slow model analyze Jev's failures between games and write instructions that Jev follows during play. After two rounds, Jev's average score rose to 24.19, the report says.

Written by AI from the articles below · updated Oct 16, 3:45 PM ET

Check the sources:

Latest developmentOct 16, 3:13 PM ET
Peking University and Samsung paper: slow reasoning model coaches fast model in real-time games

Article timeline

The articles in this story. Times are ET.

Oct 16
  1. Rohan PaulX
    Peking University and Samsung paper: slow reasoning model coaches fast model in real-time games

    AIPeking University and Samsung researchers built RealtimeBench, 180 tasks across six games that keep running during inference. In paused evaluations GPT-6.1-sol beat the faster Jev in five of six games, but scored 1.51 versus Jev's 13.51 out of 100 in live play. Their TORSI method has the slow model analyze Jev's failures between games and write instructions for play, raising Jev's average score to 24.19 after two rounds.

    Image from @rohanpaul_ai's post

Heat trend

Current comparable heat
9
Peak comparable heat
10Oct 16, 4:00 PM ET
Change in past 24 hours
–

The trend compares only continuously observed participants, so its scope may be smaller than the current heat total. Hover or tap for hourly heat; use arrow keys with a keyboard.