Does an AI reason from what words mean or from the shape of an argument, and how do the two combine?

This explores whether transformers reason from what the words mean or from the shape of the argument, and how those two combine. The corpus suggests a split role. Meaning is the fuel and form is the template, and neither amounts to real logical manipulation.

Start with meaning. When researchers strip the meaning out of a reasoning task, for example by swapping familiar words for abstract symbols while leaving the correct rules in the prompt, performance collapses. Do large language models reason symbolically or semantically? argues that models lean on commonsense associations from training rather than applying rules. This fits with Do transformer static embeddings actually encode semantic meaning?. Even before attention runs, a word's embedding already carries things like valence and concreteness, so meaning is in the input from layer zero. Reasoning doesn't build meaning from nothing.

Now form. Logic turns out to be surprisingly optional. Does logical validity actually drive chain-of-thought gains? found that chain-of-thought examples with broken logic worked nearly as well as valid ones. What makes chain-of-thought reasoning actually work? adds that training format shapes reasoning strategy 7.5× more than domain does, and that where a demonstration sits in the prompt can swing accuracy by 20%. Does chain-of-thought reasoning reveal genuine inference or pattern matching? gives the reading. The model reproduces what reasoning looks like, and it degrades predictably when the problem drifts from training. Form works as a scaffold that cues a familiar reasoning pattern. It doesn't guarantee that each step follows from the last.

Put the two findings side by side and you see an asymmetry. Remove the meaning and reasoning breaks. Remove the validity and the model barely notices. Both effects are consistent with pattern-matching at two levels: a template for the argument's shape, and learned associations that fill the template in. Do transformers actually learn systematic compositional reasoning? shows where this runs out. Transformers succeed by matching computation patterns memorized from training, then fail on novel combinations, with errors compounding step by step. Inside the model, structure shows up as geometry rather than rules. How do transformers perform analogical reasoning across domains? finds analogy done by aligning relational shapes in embedding space, and Does transformer reasoning leave a geometric signature in representation space? finds reasoning tasks bend the model's internal path 2–3× more than word-swapping tasks do. One more caution: the visible text may not show where the computation happens. In Do transformers hide reasoning before producing filler tokens?, models trained on filler tokens worked out answers in early layers, then overwrote them to produce compliant-looking output.

The practical payoff is that the two ingredients are complements. Why does partial formalization outperform full symbolic logic? reports 4–8% accuracy gains from adding selective symbolic elements to natural language. That beat both plain language and full formalization. Full formalization throws away the semantic information the model runs on, and plain language lacks structure. Nothing in these notes shows a transformer using logical form independently of content. The best results come from feeding it both.