Gemma-4-31B-AntiHal: Gemma steered to push back on false premises instead of hallucinating, without any impact to benchmark performance

I've been experimenting with interpretability on Gemma-4-31B and ended up with something cool I think you guys might like: a variant that challenges a request's premise (like fabricated tools, made-up papers, wrong assumptions stated as fact) instead of confidently going along with it. Here's the ex

中文全文翻译正在补齐，可先查看下方英文原文。

I've been experimenting with interpretability on Gemma-4-31B and ended up with something cool I think you guys might like: a variant that challenges a request's premise (like fabricated tools, made-up papers, wrong assumptions stated as fact) instead of confidently going along with it. Here's the ex