7 LLMs tested on FDT in twin prisoners' dilemmas show dramatic disagreement on action-FDT
jessi_cata · x · 2026-11-10
jessicata ran simulated almost-twin prisoners' dilemmas where seven LLMs were instructed to apply four decision theories. Models largely agreed on CDT and EDT, and every policy-FDT trial ended in cooperation — but action-FDT recommendations (as defined by equation 4 of the FDT paper) varied dramatically, with some models almost always cooperating and others almost always defecting. The author then examined whether requested reasoning effort explained part of the disagreement, concluding there's no consensus on interpreting action-based FDT.
More from Research
- Anthropic's Claude computes Hawking-Page transition order, solving 20-year question — rohanpaul_ai · 2026-11-10
- China's brain-reading device hits 94% mind-typing accuracy in day-long use — yogthos · 2026-11-10
- Marin 535B-A23B finishes 18T-token pretraining with zero instabilities, architecture detailed — ChengleiSi · 2026-11-10
- ICWM 2027 world models conference set for NYC, paper deadline December 22 — randall_balestr · 2026-11-10
- HuGo: Berkeley Team's LLM-Designed Humanoid Whole-Body Policy Paper Spotlit at CoRL 2026 — m_wulfmeier · 2026-11-10
- UniWALM: one diffusion video backbone for planning, memory and manipulation — drsrinathsridha · 2026-11-10