Large Language Models as Statistical Optimisers

Large Language Models as Statistical Optimisers: Token Distribution Matching vs. Invariant Causal Modelling of Reality

2026-05-18 · llm-reasoning benchmarks-eval ai-architecture causal-representation-learning epistemology-of-ai · medium · source → · wiki →

key claims

Large Language Model pre-training is defined by autoregressive next-token prediction over observed text sequences, so the base optimisation target is predictive compression of token distributions rather than direct estimation of intervention-stable mechanismsBrown et al. (2020)Martin (2026)
Firthian distributional semantics, the view that words derive meaning from surrounding context, gives this training regime dense associational signal from co-occurrence and context, but that signal remains observational because text records patterns in language use rather than controlled interventions on the worldMartin (2026)LaViolette (2022)
Phase 2's Empirical Risk Minimisation and Pearl's causal hierarchy results imply that a model trained only on observational text cannot be assumed to recover Level 2 intervention semantics or Level 3 counterfactual structure merely by scaling the same observational objectiveCausalai (n.d.)Arxiv (n.d.)Research (2026)Research (2026)
Linear probes and other observational interpretability tools can show that causal or world-structured information is decodable from activations, but current reviews do not treat that as sufficient proof that the model is causally relying on an invariant internal world modelGavves (2024)Nanda et al. (2023)
Current causal benchmarks repeatedly show that Large Language Model performance degrades as tasks move from simpler causal association or familiar textual patterns toward harder intervention and counterfactual settings, which is the pattern expected from systems strongest at Level 1Chen et al. (2024)OpenCausalLab (2024)Wang et al. (2024)
Benchmark success on causal tasks cannot be taken at face value as proof of causal modelling because recent reviews argue that many tasks remain solvable through domain-knowledge retrieval or lexical regularities rather than through genuine intervention-grounded reasoningYang et al. (2024)Wang et al. (2024)
The best evidence for partial causal competence is that Generative Pre-trained Transformer family models can generate strong causal arguments on some tasks and generalise beyond training-cutoff datasets, but those same studies report unpredictable failure modes and explicitly recommend combining Large Language Models with external causal methodsKiciman et al. (2025)Research (2025)
Evidence that explicit causal constraints and sparsely interacting modules improve out-of-distribution performance suggests that default next-token language modelling does not already provide the invariant causal mechanisms that stronger causal robustness would requireGendron et al. (2024)

Research Question

To what extent do Large Language Models (LLMs) optimise strictly for linguistic form and statistical token distribution rather than constructing internal, invariant causal models of reality?

Findings

(Populated from §6 Synthesis above.)

Executive Summary

Present-day Large Language Models are systems trained to optimise predictive fit over textual token distributions, and current evidence does not show that they learn invariant causal models of reality.

Current evidence does not support the stronger claim that scale and architecture alone let text-only pre-training cross Pearl's observational ceiling, because Phase 2's causal-hierarchy argument still applies and current causal benchmarks remain fragile under harder intervention and counterfactual settings.

The strongest contrary evidence is narrower: some models can generate good causal arguments from text metadata, and some controlled interpretability studies recover structured internal algorithms, which suggests partial abstraction rather than mere bag-of-words shallowness.

But those results still fall short of proving invariant causal world models, because probe-based evidence is not the same as interventional proof, and robustness improves when explicit causal constraints are added.

The best-supported placement is therefore that present-day Large Language Models are predominantly Level 1 systems on Pearl's causal hierarchy, the framework that separates association, intervention, and counterfactual reasoning, with limited and uneven Level 2 or Level 3 competence reconstructed from textual regularities, stored world knowledge, and task scaffolding rather than from a verified invariant causal model.

Key Findings

Large Language Model pre-training is defined by autoregressive next-token prediction over observed text sequences, so the base optimisation target is predictive compression of token distributions rather than direct estimation of intervention-stable mechanisms.
Firthian distributional semantics, the view that words derive meaning from surrounding context, gives this training regime dense associational signal from co-occurrence and context, but that signal remains observational because text records patterns in language use rather than controlled interventions on the world.
Phase 2's Empirical Risk Minimisation and Pearl's causal hierarchy results imply that a model trained only on observational text cannot be assumed to recover Level 2 intervention semantics or Level 3 counterfactual structure merely by scaling the same observational objective.
Linear probes and other observational interpretability tools can show that causal or world-structured information is decodable from activations, but current reviews do not treat that as sufficient proof that the model is causally relying on an invariant internal world model.
Current causal benchmarks repeatedly show that Large Language Model performance degrades as tasks move from simpler causal association or familiar textual patterns toward harder intervention and counterfactual settings, which is the pattern expected from systems strongest at Level 1.
Benchmark success on causal tasks cannot be taken at face value as proof of causal modelling because recent reviews argue that many tasks remain solvable through domain-knowledge retrieval or lexical regularities rather than through genuine intervention-grounded reasoning.
The best evidence for partial causal competence is that Generative Pre-trained Transformer family models can generate strong causal arguments on some tasks and generalise beyond training-cutoff datasets, but those same studies report unpredictable failure modes and explicitly recommend combining Large Language Models with external causal methods.
Evidence that explicit causal constraints and sparsely interacting modules improve out-of-distribution performance suggests that default next-token language modelling does not already provide the invariant causal mechanisms that stronger causal robustness would require.
The strongest overall conclusion is that Large Language Models do learn abstractions, but present evidence supports statistical and text-mediated abstractions more strongly than verified invariant causal models of reality.

Assumptions

[assumption] Text corpora do not contain enough implicit intervention structure to overturn Pearl's generic observational ceiling without additional architectural or data assumptions, because no consulted source demonstrates such a collapse for present-day Large Language Models. [source: Research repo (2026-05-19) Research Question 2.4: Pearl's Causal Hierarchy and the formal limits of observational data for intervention and counterfactual reasoning aclanthology.org
[assumption] Benchmark performance is a usable but incomplete proxy for underlying causal competence, because latent-model structure cannot currently be read off directly at scale and benchmark reviews explicitly question surface-score interpretations. [source: Yang et al. (2024) A Critical Review of Causal Reasoning Benchmarks for Large Language Models arxiv.org/abs/2404.14082]

Analysis

The evidence is strongest where the question is about objective and information source, because both are directly specified.

The more speculative step is moving from "trained on token prediction" to "cannot learn any useful abstraction", and the consulted evidence does not justify that stronger negative claim.

The right synthesis is therefore asymmetric: the training objective and causal hierarchy justify scepticism about invariant causal modelling, while benchmark and interpretability evidence justify acknowledging partial abstraction and partial causal competence.

Alternative explanations remain live.

One alternative is that causal competence is mostly benchmark artefact or retrieval from text.

Another is that some causal abstractions do emerge, but only weakly or incompletely under the default objective and become more robust when explicit causal structure is added.

The consulted evidence supports the second explanation more strongly than the first, because there is repeated positive evidence for some causal-task competence, but not for full intervention-stable world modelling.

Risks, Gaps, and Uncertainties

Direct inspection of production-scale model internals remains limited, so the conclusion is partly constrained by what current mechanistic-interpretability tools can test rather than by a full latent-state ground truth.
Most empirical studies evaluate behaviour on tasks rather than intervention-grounded world interaction, which limits how strongly benchmark outcomes can be translated into claims about internal causal world models.

Open Questions

What changes when language models are trained with explicit interaction data, simulator interventions, or multimodal world grounding rather than text alone?
Can future interventional interpretability methods distinguish decodable causal variables from truly mechanism-governing internal representations at frontier scale?
Which observed causal-task gains come from stored textual world knowledge, which come from learned abstraction, and which require external tool scaffolding?