ICML 2026 · Conference paper

Global Credit Assignment via Dynamical Criticality

Wentao Wang, Keren Gao, Guozhang Chen

Criticality-driven Online Local Alignment exploits long-range critical correlations so local recurrent updates approximate global backpropagation in constant activation memory.

The scientific question

Can dynamical criticality allow strictly local learning rules to achieve global credit assignment?

Backpropagation through time requires memory proportional to sequence length, whereas standard biological local rules suffer from severe gradient approximation bias.

Comparison of credit assignment paradigms and empirical memory and accuracy trade-offs.
Figure 1 from the original paper ↗

The core idea

Credit Propagation via Criticality

COLA places recurrent networks near dynamical criticality, using intrinsic long-range spatiotemporal correlations to propagate error information across time through a strictly local online update rule.

What we found

Sequence-Independent Memory and Stability

With only O(H) auxiliary state and constant activation memory independent of sequence length, COLA is competitive with backpropagation through time on standard benchmarks and demonstrates superior robustness on stability-sensitive tasks.

Continue the question

Related reading

Explore our research

Cite this work

Wentao Wang, Keren Gao, Guozhang Chen. Global Credit Assignment via Dynamical Criticality. ICML 2026.