Silico reproduces Goodfire AI's RLFR method, cutting Qwen3-8B hallucinations 37%

Silico reproduces Goodfire AI's RLFR method, cutting Qwen3-8B hallucinations 37% · Digg

/TECH

The independent lab reproduced the internal-probe RL technique in two days.

Combined views

12.2K

3 posts, first seen 14h ago

117 likes

Commentary on X

Marshwiggle@Marshwiggle119

@GoodfireAI I've been telling people Goodfire has been doing good work. And you have been. But this... there is no way no one on your team is unaware of the reasons this might be a bad idea, right? Why do this anyway?

10:15 PM · Jul 14, 2026

Hisku@ZikuD_s

@GoodfireAI Interesting research. I wonder whether this could be expanded to safety research, where probes for deception, scheming, or harmful intent are used as internal reward signals during training.

6:32 AM · Jul 15, 2026

VioP@AcousimHss

@GoodfireAI gang i requested in like 4 5 months ago plssspslpslpsl give

11:28 PM · Jul 14, 2026

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex

…oh, right of course they instantly did that @TheZvi hardest hit

8:57 AM · Jul 15, 2026

Combined views

12.2K

3 posts, first seen 14h ago

117 likes7 comments

7 comments

26 saves

10 reposts

Sentiment

Positive66.7%33.3%Negative

Positive Read

Users are excited about Silico reproducing Goodfire's RLFR method to cut Qwen3-8B hallucinations because it shows promising results and potential for safety research, while some worry the team overlooked reasons it could be a bad idea.

Based on 3 visible X reactions from 9 accounts; directional sample.