Yağız Çalık on X: “Then we picked what to train on. With binary rewards, a task only teac…”

A fast, private reader for public X posts. No account, no tracking, no pop-ups.

PeekX is an independent tool and is not affiliated with, endorsed by or sponsored by X Corp. “X” and “Twitter” are trademarks of their respective owners. We only show posts that their authors made public and never access private or protected accounts. Downloaded media belongs to its creators — please respect their rights.

Username or post link

Home / @Weyaxi / Post

Yağız Çalık

@Weyaxi

Oct 8

Most generated tasks don't work. DeepSeek-V4-Pro wrote 66k candidates and only 1.5% made it in: a task counts only if its own reference solution passes its own tests, offline, inside the built image.

Tests average ~16 named assertions, so a failure names exactly what broke.

0:10

1 0 13

Yağız Çalık

@Weyaxi

Replying to @Weyaxi

Then we picked what to train on. With binary rewards, a task only teaches the model if it sometimes passes and sometimes fails: the gradient weight is p(1−p), zero at p=0 and p=1, highest at 0.50.

So we graded gemma-4-31B-it separately, in the bash scaffold we train it in,

Oct 8, 2026, 5:20 PM UTC

1 10

Download media

Every quality X serves, straight from X's media CDN. No watermark, no wait.

OriginalBEST

Photo · 1199×858

Save & share

Turn this post into a clean image, or copy it as text or Markdown.

Open on X

Yağız Çalık on X: “Then we picked what to train on. With binary rewards, a task only teac…” | PeekX