Can causal models be extended to include non-causal cognition?
This explores whether causal-reasoning models — the kind built to map cause-and-effect — can be stretched to cover the parts of human thinking that aren't causal at all: gut association, analogy, and emotion.
This explores whether causal-reasoning models can be stretched to cover the non-causal parts of cognition. The corpus has a direct answer to this, and it's refreshingly honest: causal models capture only a slice of how humans actually reason. The GenMinds work on causal belief networks is explicit that while these networks are excellent at representing cause-and-effect, they simply cannot represent associative links, analogical mappings, or emotion-driven shifts in belief — and crucially, its authors frame this not as a finished theory but as a tractable starting point Can causal models alone capture how humans actually reason?. So the question's premise is already conceded inside the research: causality alone leaves gaps, and the interesting work is in what you bolt on.
The reason this matters becomes clearer when you look at where causal reasoning is strong and where it isn't. In LLMs, causal reasoning consistently outperforms temporal reasoning — not because cause-and-effect is cognitively deeper, but because causal connectives ('because', 'therefore') appear explicitly and frequently in training text, while temporal order usually has to be inferred Why do LLMs handle causal reasoning better than temporal reasoning?. That hints at something important for your question: the 'causal' competence we observe may itself be a surface artifact of language statistics rather than a clean reasoning module. And models reproduce human causal *errors* exactly — weak explaining-away, Markov violations in collider networks — suggesting the same statistical substrate drives both the causal and the non-causal, messy parts of cognition Do large language models make the same causal reasoning mistakes as humans?. If causal and associative reasoning share one substrate, extending one to include the other may be less about adding a new module and more about acknowledging they were never fully separate.
There's a methodological cousin to this in interpretability research that reframes the whole question. To actually understand what a network is doing, you need *both* representational analysis (what's encoded, the correlational/associative picture) and causal analysis (what intervening actually changes). Neither alone is sufficient — representations show correlations without causation, causal tests show effects without explaining them Can we understand LLM mechanisms with only representational analysis?. That's essentially the same shape as your question one level up: a complete account of cognition needs the causal scaffolding *and* the associative/representational layer working together.
Where the corpus goes quiet is on a concrete unified architecture — nobody here has built the causal-plus-associative-plus-emotional model the GenMinds authors gesture toward. But the surrounding notes suggest what a non-causal extension would have to absorb: reasoning that draws on broad transferable procedural patterns rather than fact lookup Does procedural knowledge drive reasoning more than factual retrieval?, and a structural split where factual knowledge lives in lower network layers while reasoning adjustment happens higher up Why does reasoning training help math but hurt medical tasks?. The takeaway you didn't expect: the obstacle to extending causal models isn't that non-causal cognition is exotic — it's that causal reasoning itself may already *be* a special, well-labeled case of the same associative machinery, which means 'extension' might really mean dropping the assumption that causality was ever a clean, separable thing.
Sources 6 notes
Causal belief networks excel at modeling causal reasoning but cannot represent associative links, analogical mappings, or emotion-driven belief shifts. The GenMinds framework itself acknowledges this as a tractable starting point rather than a complete theory.
ChatGPT excels at causal relations but struggles with temporal ordering because causal connectives are explicit and frequent in training data, while temporal order is often implicit and must be inferred contextually.
LLMs show weak explaining away and Markov violations in collider networks, matching human error patterns exactly. This suggests shared mechanisms rooted in training data statistics rather than categorical reasoning inferiority.
Representational analysis alone identifies correlations without causation; causal analysis alone shows behavioral effects without explaining them. Only paired methods—locating candidate features representationally, then verifying causally—produce complete mechanistic claims.
Analysis of 5 million pretraining documents shows reasoning relies on broad, transferable procedural knowledge from diverse sources, unlike factual recall which depends on narrow, document-specific memorization of target facts.
Two-phase inference model shows knowledge retrieval operates in lower network layers while reasoning adjustment happens in higher layers. This separation explains why reasoning training improves math but can degrade knowledge-intensive domains like medicine.