Work

Think While You Speak: Adaptive Computational Effort in Neural Dialogue Systems

Year: 2026
Type: preprint
Abstract: Adaptive Cognitive Processing (ACP) is a minimal two-pathway architecture in which inference-time computation is input-dependent. A linear fast pathway answers familiar inputs by default, and a monitor checks the input against a small episodic memory, computing a novelty score that gates a deliberate MLP pathway. The system is implemented in pure NumPy on CPU, with analytical gradients verified against finite differences (max relative error 1.7e-9). On a 5,000-example synthetic task across three difficulty levels, ACP learns a clear routing gradient: it deliberates on 2.8% of easy inputs, 99.7% of medium inputs, and 99.2% of hard inputs, reaching 93.5% accuracy (mean of 10 seeds) versus 93.7% for an always-deliberate baseline while deliberating on only 67.2% of inputs. A 3-seed ablation removing the episodic novelty feature drops task accuracy from 94.3% to 73.7%, showing that the memory readout does real work. This is a mechanism study on a synthetic task; dialogue is the motivating application, not the experimental setting.
Language: English
Open Access status: green
OpenAlex ID: W7236002993
Cites:
Cited by:
Related to: 10