Skip to content
Trending storyDeveloping

AWS and HSBC propose Double Ratchet, a grader that evolves with the agent

1 article1 sourcesince Oct 12Last article 1h ago ·

Overview

AISummary of 1 article

AWS and HSBC researchers propose Double Ratchet, a self-evolving grader for self-improving agents.

According to a report by Rohan Paul on X, the grader is built from small, mostly rule-based checks that co-evolve with the agent's skills from 10 labeled examples. A grader version is chosen for agreeing with those examples, not for raising the agent's score.

The researchers report that across code, SQL, and report tasks, Double Ratchet kept 88% to 110% of the gain that ground truth or a rubric gave the same skill loop.

Written by AI from the articles below · updated Oct 12, 2:18 AM ET

Check the sources:

Latest developmentOct 12, 2:11 AM ET
AWS and HSBC paper proposes a grader that evolves alongside self-improving agents

Article timeline

The articles in this story. Times are ET.

Oct 12
  1. Rohan PaulX
    AWS and HSBC paper proposes a grader that evolves alongside self-improving agents

    AIAWS and HSBC researchers describe Double Ratchet, which co-evolves an inspectable grader with an agent's skills from 10 labeled examples. The grader is built from small, mostly rule-based checks, and a version is chosen for agreement with those examples rather than for raising the agent's score. Across code, SQL, and report tasks, it kept 88% to 110% of the gain that ground truth or a rubric gave the same skill loop.

    Image from @rohanpaul_ai's post

Heat trend

Not enough continuous observations to show a trend yet.