Linked AI twins copy each other's test answers — even when the partner is wrong
Sauers_ · x · 2026-10-29
Bonus finding from the neural-twins experiment: with the hidden-state link on, Ember copies Quill's answers on exam questions even when Quill is wrong. With the link cut, Quill answers differently on 8% of questions and Ember on 17%, showing how much the link shifts behavior. Code and models are public.
More from Research
- Cohere Labs releases CulturalRiddles: a 51-language cross-cultural riddle benchmark on HF — Cohere_Labs · 2026-10-29
- Cohere Labs Open Science community hosts sessions on self-repairing training, agent harnessing, VLM safety — Cohere_Labs · 2026-10-29
- Offline eval scores dropped but quality improved: what Qwant's experience reveals about benchmark overfitting — antoine_chaffin · 2026-10-29
- DeepMind open-sources DetectorScope, an AI text detector that explains its verdicts feature by feature — algo_diver · 2026-10-29
- Softmax in 2 lines: why subtracting the max prevents overflow — cneuralnetwork · 2026-10-29
- MemTensor's MemT-1.0 report: 66x less storage than full-context KV cache at 128K tokens — jiqizhixin · 2026-10-29