LLM agents build human-like social networks — and polarize anyway
profjamesevans · x · 2026-10-09
From the Evans team's paper thread: agents autonomously built social networks resembling human ones — homophilic clusters, scale-free degree distributions, and two camps that stopped talking. Radical allies pushed agents further out, and opponents often produced a backfire effect. The quoted post adds that even all-neutral agents fell from 40% to as low as 0.4% within ten rounds, polarizing regardless.
More from Research
- Multi-agent framework CORAL presented at COLM, tops 1.1K GitHub stars — pliang279 · 2026-10-09
- LLM judges of AI-scientist novelty flip on tiny prompt tweaks, AI2 finds — TuhinChakr · 2026-10-09
- TritonRL: an 8B RL-trained model matches 100B+ frontier models on Triton kernel generation — allenainie · 2026-10-09
- LLM agent polarization study from Tsinghua-UChicago team published in Nature Communications — profjamesevans · 2026-10-09
- Forced contact moves LLM agent opinions just 2%, polarization study finds — profjamesevans · 2026-10-09
- Mathematician downplays Navier-Stokes 'mistranslation' uproar: no doubt about the proof — basedjensen · 2026-10-09