Professor of Psychology and Computer Science at USC
Morteza Dehghani
Loading...
By injecting "hostility vectors" into the model's activations, we could causally shift agents from neutral to aggressive behavior.
This proves these internal representations aren't just noise—they functionalize behavior.
A new video via Vahid Online shows Islamic Republic security forces shooting at protestors in Tehran during a protest on January 8.
Using mechanistic interpretability, we found the LLM encodes "realistic threat," "symbolic threat," and "hostility" as distinct internal activation states.
We then "steered" these states...
This framework allows us to bridge the gap between individual psychology and macro-level social outcomes in ways previously impossible.
In our 2x2 simulation, Realistic Threat (safety/resources) was the heavy hitter. It was the most reliable predictor of actual behavioral escalation.
Symbolic Threat drove ingroup bias and "hateful" speech but rarely escalated conflict on its own.
Key finding: "Hateful language" was a transient symptom, not the cause. The true engine of long-term escalation was persistent ingroup bias.
Structural factors like segregation didn't change the psychology but concentrated hostility within majority groups.
My earliest memory: pedaling my bike at age 5 to warn my parents about the Basijis. At 6, I heard my mother’s screams as she was arrested for nail polish.
I tell my kids the story of Zahhak—the tyrant who fed on youth until he was overthrown. The Zahhak of our time is falling 🕊️
#IranRevolution2026
The Islamic regime has killed 12k people in 5 days. Iranians are hurting; we’ve been hurting for 150 years. If you have an Iranian colleague, check on them. No unsolicited advice—just kindness and space. They may kill our brothers and sisters, but never the dream of freedom.
Led by @sabdurah.bsky.social w/ Farzan Karimi-Malekabadi, Chenxiao Yu, @nourkteily.bsky.social
Here is the link to the paper: arxiv.org/abs/2512.17066
Morteza Dehghani
Morteza Dehghani
Yashar Ali 🐘
Morteza Dehghani
How do resource fears (realistic threat) vs. value clashes (symbolic threat) drive war & peace?
We built a virtual society of 25 autonomous agents using the Park et al. (2023) framework to find out, using a "minimal groups" paradigm (Group A vs. B).
But first, we looked under the hood. 🧠
