NVIDIA ships Dory looped transformer model
The author argues that increasing a language model's thinking budget or looping transformer layers does not make it smarter when dealing with missing facts, because loops add computation rather than new information. Citing NVIDIA's four-billion-parameter looped model Dory and benchmark data from Fastino, the article demonstrates that extra reasoning passes increase latency and cost while sometimes degrading accuracy. The author recommends verifying whether an issue is an input assembly problem or a missing fact before increasing model thinking. Open article →