Keynotes | K&Ts | GACs | Talks | Posters | Search

Poster D in Poster Session D: Wednesday, August 5, 2:00 – 3:45 pm, Kimmel Center, Shorin & Rosenthal Rooms

Slow reasoning can reflect pressure to minimize information processing costs

Nicholas Fagan¹, Carlos G. Correa¹, Noga Zaslavsky¹, Marcelo G Mattar¹; ¹New York University

Presenter: Nicholas Fagan

Human cognition is strikingly slow: across tasks, people process only about 10 bits per second, far below the throughput of sensory systems and modern AI. Why would a system that must behave flexibly operate so slowly? One possibility is that slow, serial processing is not a limitation but an adaptation -- an efficient strategy when there is pressure to minimize the amount of information processed at each moment. Here we show that slow reasoning emerges naturally in agents trained to solve a logical reasoning task when they are penalized for both processing time and representational precision. We trained recurrent neural networks on a boolean evaluation task with variable difficulty, rewarding accuracy while penalizing the number of processing steps and the bits encoded per step. We found a systematic trade-off: agents under stronger precision pressure learn coarser representations and compensate by using more processing time, yet achieve higher accuracy than those that encode inputs more precisely. Agents also dynamically allocate more steps to harder problems, mirroring patterns seen in human deliberation. These results demonstrate that slow, serial cognition can be computationally advantageous, minimizing information-processing costs at the expense of latency, and offer a normative explanation for why biological cognition may have evolved to operate at such a surprisingly low bandwidth.

Topic Area: Decision-Making, Cognitive Control & Event Cognition