Can neural networks achieve genuine reasoning, or are they fundamentally limited to statistical pattern matching? Recent work by DataAlchemy (Zhao et al., 2025) demonstrated that Transformer chain-of-thought reasoning collapses entirely under distribution shift—length generalization drops to near zero. This has led to a growing belief that reasoning fragility is an inevitable property of all neural networks. We challenge this conclusion. We present Cognitio, a brain-inspired cognitive architecture based on multi-minicolumn group collaboration. Its core innovation is establishing internal language as a planning causal bottleneck through a role-content disentangled codebook, and achieving complex reasoning via a hierarchical integration-differentiation loop. We conducted six progressive experiments on digit-to-word translation and ROT-k displacement tasks. Experiments 1-4 validated the causal efficacy of the discrete bottleneck, the feasibility of group collaboration, and the necessity of hierarchical architecture. Experiment 5 achieved 100% zero-shot generalization across task, length, and format dimensions—the model was trained only on single-step samples, yet correctly solved multi-step sequences of arbitrary length by reusing the learned operation at each position. Experiment 6, in direct comparison with DataAlchemy under fully comparable conditions, achieved 100% accuracy on 3/5/7/10-step length generalization where Transformer accuracy collapsed to near zero. This stark contrast demonstrates that reasoning fragility is not an inevitable destiny of neural networks, but a structural defect of homogeneous continuous architectures. Cognitio achieves genuine rule understanding by separating operation types (roles) from operation parameters (content), enabling learned rules to be flexibly reused under novel conditions—a capability fundamentally absent in Transformer-based reasoning. We provide complete runnable code for the B3 model.
Huanran Xu (Tue,) studied this question.