Symbolic and Connectionist Reasoning: Two Ways to Build "Think"

TheThe cognitive cycle names Think as a distinct stage without saying how it's actually implemented. This page covers the two fundamentally different answers computer science has produced, using two real, contrasting systems rather than describing the distinction only in the abstract.

Two philosophies of what "thinking" is

SymbolicSymbolic reasoning represents knowledge as explicit, discrete structures — rules, facts, logical relationships — and reasons by manipulating those structures according to formal inference procedures. If a symbolic system produces an answer, there is, in principle, a traceable chain of steps that led to it.

Connectionist reasoning represents knowledge as patterns of weighted connections between simple units, learned from data rather than written down by hand. A large language model is the current, most visible connectionist architecture: its "knowledge" is billions of numerical weights, adjusted during training, with no single weight corresponding to any human-readable fact.

Neither is a strictly better approach to Think — they trade different strengths for different weaknesses, and the clearest way to see the trade is against two real systems built to do genuinely similar work.

A worked symbolic case: inductive logic programming

PatLang,PatLang, an experimental programming language documented elsewhere on this site, includes a synthesis engine that induces new code from Given/When/Then examples using inductive logic programming (ILP) — searching a declared space of background facts and primitive operations for the smallest rule consistent with every example given^{1 2}. Every step of that search is inspectable: a failed search can report exactly which example couldn't be satisfied and why, because the system is manipulating explicit symbolic structures the whole way through³.

That inspectability comes at a real cost. The engine can only find rules built from primitives someone declared in advance — its own documented history records that its practical ceiling was consistently "the shape of its metarule template, not the difficulty of any one problem." Widening what it could express always required a person to widen the template first. A symbolic system's reasoning is traceable precisely because it can only reason inside a space a person explicitly built for it.

A worked connectionist case: large language models

AnAn LLM asked to solve the same kind of problem — infer a rule from a handful of examples — does something that looks superficially similar and works completely differently underneath. It has no declared space of primitives to search; it produces a continuation that statistically resembles patterns in its training data, shaped by billions of weights with no individual weight corresponding to "the concept of a precondition" or anything else a person could point to. This is why an LLM can solve problems nobody explicitly programmed it to solve — and also why, when it's wrong, there is no equivalent of PatLang's precise "this example is unsatisfiable because of X" diagnosis. There's no symbolic structure to interrogate, because none exists inside the system in the first place.

Why this distinction matters for cognitive architecture specifically

Lake,Lake, Ullman, Tenenbaum and Gershman's influential critique of purely connectionist approaches to intelligence argues that genuinely human-like cognition needs three properties together: compositionality (building complex representations from reusable simple parts), causality (modelling how things actually affect each other, not just correlating with each other), and the capacity to learn how to learn across genuinely new domains⁴. Symbolic systems like PatLang's ILP engine have strong compositionality — that's precisely what makes their search space describable at all — but no causal model of the domain and no capacity to widen their own search space unaided. Connectionist systems can approximate causal-looking behaviour statistically, and current large models show some capacity to generalise across domains, but neither property in an LLM comes with a symbolic system's traceability.

This is also why hybrid approaches are an active area of research rather than a settled question. Lake, Salakhutdinov and Tenenbaum's own account of human concept learning treats it as a form of program induction — search over a compositional space of primitives, much like PatLang's engine — while noting the search itself is guided by probabilistic, statistically-learned expectations, much closer to a connectionist system's own strengths⁵. Neither pure symbolic search nor pure statistical pattern-matching is currently a complete account of how humans actually do this. Building an agent's Think stage is, right now, a choice about which specific failure mode you can better afford: an LLM's untraceable confidence, or a symbolic system's inability to reach outside the space it was explicitly given.

Where this connects

• Synthesis, Human and Machine — the fuller academic treatment of this same comparison, developed through Bloom's taxonomy and the evaluation-before-creation argument.
• Inductive Synthesis: from BDD scenarios to PatLang code — the full technical account of the symbolic engine discussed above.
• Agent Archetypes: Five Tiers on One Diagram — where Think sits in the fuller cognitive cycle this page's distinction feeds into.

References