Qwen3.5-9B Found with 'Sunk Cost' Neuron Hijacking Answers

Developers found that activating a specific neuron in Qwen3.5-9B makes it answer almost any question in terms of the sunk cost fallacy, with 31 of 52 test prompts hijacked.

2026-10-30 ~ 2026-10-30 · 2 related posts