Can an AI chain facts it already knows into a brand-new multi-step chain of reasoning it's never actually seen before?

This explores how transformers link several reasoning steps together, and whether that ability carries over to new territory. The short answer from the corpus: transformers compose steps well when they've seen similar combinations before. Getting them to compose new ones may depend less on more training data and more on how the architecture reuses its own computation.

Controlled training experiments show the skill doesn't arrive all at once. Models first memorize facts, then learn to chain them on familiar examples, and only later handle chains they haven't seen. Success shows up as related entities clustering together inside the model, and the second step of a chain only generalizes if training explicitly included composed examples How do transformers learn to reason across multiple steps?. A more skeptical line of work argues that what looks like composition is often matching against fragments of computation memorized during training. That works on familiar problems but breaks on new combinations, and errors pile up with each additional step Do transformers actually learn systematic compositional reasoning?. A related finding: when the meaning is stripped out of a logic problem and only the rules are left, performance collapses even with the correct rules in the prompt. Models lean on what concepts mean rather than on formal logic Do large language models reason symbolically or semantically?.

The cross-domain part of the question has a surprisingly geometric answer. When transformers solve analogies, they first line up the relational structure of one domain with another inside their internal representation space, then apply a learned mapping between them. The same signature appears in toy tasks and in pretrained LLMs How do transformers perform analogical reasoning across domains?. Reasoning also leaves a visible shape: as a problem passes through the model's layers, reasoning and analogy tasks trace paths that bend two to three times more sharply than simple word-variation tasks Does transformer reasoning leave a geometric signature in representation space?. On this view, carrying reasoning across domains means aligning shapes, not applying abstract rules.

The architecture sets a ceiling. A standard transformer has a fixed number of layers. Each reasoning step uses up depth, so long chains eventually run out of room, and writing out chain-of-thought is a costly patch that moves the model's running notes into tokens Why do transformers need explicit chain-of-thought reasoning?. Looped transformers, which run the same layers repeatedly, generalize to unseen combinations and to longer chains than they were trained on, while vanilla transformers don't Can looped transformers generalize to unseen knowledge combinations?. A small model that pairs slow high-level planning with fast detailed computation solves Sudoku and maze tasks where chain-of-thought fails completely Can recurrent hierarchies achieve reasoning that transformers cannot?.

One consequence you might not expect: the reasoning steps a model writes out are not necessarily where the composition happens. Models trained to output filler compute the correct answer in their first few layers and then overwrite it Do transformers hide reasoning before producing filler tokens?. Fine-tuning makes the written steps matter less to the final answer Does fine-tuning disconnect reasoning steps from final answers?, and chain-of-thought often copies the form of familiar reasoning rather than doing new inference Does chain-of-thought reasoning reveal genuine inference or pattern matching?. That is why some researchers argue multi-step reasoning should be studied as movement through the model's hidden states, with the visible text treated as only a partial window onto it Where does LLM reasoning actually happen during generation?.