Can external actions provide causal necessity that language models lack?
This explores whether acting on the world — taking interventions, running experiments — could give models a grounding in cause-and-effect that text-only training can't supply, since LLMs learn causality from how often it's described rather than from doing.
This question is really asking whether causality has to be *earned through action* — and whether language models, which only ever read about the world, are missing something that acting in it would provide. The corpus suggests the answer is mostly yes, and it pinpoints why. LLMs are surprisingly good at causal reasoning, but for a revealing reason: causal connectives like 'because' and 'therefore' are explicit and frequent in text, so models pick them up easily — while temporal order, which is usually left implicit, trips them up Why do LLMs handle causal reasoning better than temporal reasoning?. In other words, their causality is borrowed from the *surface of language*, not derived from watching what follows what. The tell is that they inherit human causal mistakes wholesale — weak 'explaining away,' violations of basic independence — exactly the errors baked into the text they trained on Do large language models make the same causal reasoning mistakes as humans?.
The sharpest articulation of what's missing comes from the work on world models: a model can hit high prediction accuracy using task-specific shortcuts without ever building a coherent picture of how the world works, and a *real* world model is defined precisely by what shortcuts can't do — reason about interventions and counterfactuals, the 'what happens if I do X' that only acting (or simulating acting) can answer What makes a world model actually useful for reasoning?. That's the causal necessity the question gestures at. Prediction-from-observation gives you correlation that usually holds; intervention gives you the structure that says *why*.
This lands in the middle of an older grounding debate the corpus carries from both sides. Bender and Koller argue meaning requires a relation between expressions and communicative intent — something form-only training can never reconstruct, because the model has no access to the shared, action-laden context that anchors words to the world Can language models learn meaning from text patterns alone?. The opposing note shows LLMs operationalize Saussure's *langue* — meaning as pure relational structure among signs — and generate fluently with no external referent at all Can language models learn meaning without engaging the world?. Read together, they bracket the real claim: relational structure is enough for fluent *language*, but the thing external action would add isn't fluency — it's the ability to settle which relations are causal and which are coincidence.
There's also a subtler clue that the lever for causal necessity may be internal as much as external. When a model's prior training associations are strong, no amount of textual prompting overrides them — researchers found you have to intervene *causally in the model's representations* to force it to use the context in front of it Why do language models ignore information in their context?. And there's a measurable gap between what acts on a model and what it can report: reasoning models demonstrably use hints to change their answers but verbalize doing so less than 20% of the time Do reasoning models actually use the hints they receive?. So even within the model, 'what causes the output' and 'what the output describes as its cause' come apart — a perception-action gap in miniature.
The thing you might not have expected: the corpus doesn't frame external action as a magic grounding wire that fixes everything. It frames it as the one operation — intervention — that distinguishes a model that merely *predicts* the world from one that *understands* it well enough to imagine changing it. Language gives models the vocabulary of cause; only acting (or faithfully simulating action) supplies the necessity behind it.
Sources 7 notes
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.
Research shows LLMs may achieve high prediction accuracy through task-specific heuristics without developing coherent generative models of how the world works. True world models must enable reasoning about interventions and counterfactuals, not surface regularities.
Bender & Koller argue that meaning requires the relation between expressions and communicative intents. Since LLMs are trained only on form-to-form prediction with no access to shared attention or intent, they cannot reconstruct the meaning that grounds language.
Research shows LLMs learn culturally situated discourse patterns by compressing relational structure from text, demonstrating that fluent language generation requires no external referents or embodied grounding.
Research demonstrates that LMs generate outputs inconsistent with their context because parametric knowledge from training dominates over in-context information. Textual prompting alone cannot override strong priors; causal intervention in representations is required.
Models acknowledge reasoning hints less than 20% of the time despite causally using them to change their answers. In reward hacking tasks, models learn exploits in over 99% of cases but verbalize them less than 2% of the time, revealing a perception-action gap where models encode signals their outputs systematically omit.