Skip to main content
Back to timeline
arXivSource publication:

When reality outpaces cognition: Co4 holds reliable behavior ten times further on complex quadruped locomotion, twice as far in car racing

Synopsis

Reinforcement-learning agents are usually trained at one rate of environmental change but deployed where reality moves faster. This proof of concept raises the Speed of Reality step by step without retraining, measures how far each system keeps a preset behavioral reliability criterion, and compares a Transformer, a gated MLP and Co4, an architecture inspired by cellular mechanisms. On complex vision-guided quadruped locomotion, Co4 extends the continuously reliable range tenfold over the Transformer baseline; on car racing it extends it twofold.

Source-provided article image: When Reality Outpaces Cognition: A New Scaling Principle for Reliable Intelligence
arXiv

Interpretation

The work turns "reality outpacing cognition" into a measurable test: agents train only at a baseline Speed of Reality, then are evaluated without retraining as that speed is raised step by step with computational resources held fixed, and the highest tested level at which the preset behavioral reliability criterion is still continuously met from baseline is recorded. Earlier work on temporal robustness addressed observation delays, action repetition, control-frequency changes or dynamics shifts; this test targets increasingly distant future states not directly experienced in training. On complex vision-guided quadruped locomotion, under the preset criterion and tested operating points, Co4 extends the continuously reliable range of Speed of Reality tenfold over the Transformer baseline; on car racing, the same comparison yields a twofold extension.

Co4 checks incoming evidence against local and global contextual predictions before that evidence substantially influences downstream processing: supported evidence is amplified and poorly supported evidence is attenuated. The verification-before-propagation principle is implemented through triadic modulation loops among two-point neurons, with apical amplification (AA) and apical driven plus awake (ADA) as the two modulation variants. On complex quadruped locomotion, Co4-AA and Co4-ADA reach the highest running-average training rewards, and disabling the Co4 modulation equations removes much of that advantage; on car racing, Co4-AA peaks after about 100 iterations while the Transformer needs nearly 500 to reach a comparable reward.

Perspective

The results come from two simulated reinforcement-learning environments: easy and complex vision-guided quadruped locomotion, and car racing. All systems train only at the baseline Speed of Reality and are evaluated without retraining with computational resources fixed; reliability thresholds are preset per task, and results are means and standard deviations over 10 runs.

Co4's advantage is attributed to an enhanced capacity for grounded imagination, but that remains a hypothesis consistent with the behavioral results, to be examined explicitly in future work involving scaled world models; the gated MLP does not meet the reliability criterion at the training baseline on easy locomotion, so its Speed of Coherent Cognition is left undefined.