pith. sign in

arxiv: 2606.22430 · v1 · pith:RW5V6QNDnew · submitted 2026-06-21 · 💻 cs.CL · cs.AI· cs.LG

Words as Difference Makers: How Large Language Models Determine Causal Structure in Text

Wolfgang Pietsch This is my paper

Pith reviewed 2026-06-26 10:44 UTC · model grok-4.3

classification 💻 cs.CL cs.AIcs.LG
keywords large language modelscausal inferencevariational inductiondifference-makingtext dataself-attentiontoken embeddingsworld model
0
0 comments X

The pith

Large language models learn causal structure in text by identifying difference-makers through variational induction.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper argues that large language models determine causal relations within language using a difference-making logic called variational induction. This logic is put into practice during training by exposing the models to enormous quantities of text drawn from many different contexts, allowing them to detect which words or sequences change outcomes and which do not. Standard causal inference tools from Pearl and Rubin are said to fall short for explaining this process. Readers may find this useful because it offers an account of how these models build an implicit understanding of the world from language alone.

Core claim

LLMs employ a specific inductive approach based on a difference-making logic -- sometimes called variational induction -- realized during training where enormous amounts of text data from a wide range of contexts identify difference- and indifference-makers within word sequences. Central aspects of this logic are realized in training, and architectural features such as token embeddings and self-attention play roles in it. The difference-making logic of LLMs fundamentally parallels the experimental method, where causal relations are derived by systematically varying individual circumstances to determine their influence on a phenomenon.

What carries the argument

Variational induction, which identifies difference- and indifference-makers in word sequences to establish causal structure.

If this is right

  • LLMs need massive volumes of text from varied contexts to successfully extract causal relations.
  • Features like token embeddings and self-attention support the identification of difference-makers.
  • The approach in LLMs mirrors how experiments vary conditions to isolate causal effects.
  • This provides an explanation for why LLMs appear to possess a world model of causal and definitional structure.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Models trained on narrower datasets might show weaker grasp of causal connections in language.
  • This view suggests testing whether altering self-attention changes the model's sensitivity to contextual variations.
  • Human language learning may rely on similar difference-making observations in everyday speech.

Load-bearing premise

That the main formalisms of causal inference cannot explain how LLMs acquire causal knowledge from text data.

What would settle it

Training an LLM on text from only a narrow set of contexts and checking whether it fails to identify causal relations that a model trained on diverse contexts can detect.

Figures

Figures reproduced from arXiv: 2606.22430 by Wolfgang Pietsch.

Figure 1
Figure 1. Figure 1: determination of a value f(x) of a node in a subsequent layer given values x1, …, xn of nodes in the previous layer In a simplified reconstruction of a large language model, the input and output take the form of high-dimensional vectors each representing specific tokens in a sequence (see [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: on the left, a typical ‘one-hot’ input vector representing a token with ID number x (e.g. token ID for the word “induction” is 14355 using the GPT-2 tokenizer); on the right, a typical output vector representing a probability distribution over the entire vocabulary While large language models are essentially deep neural networks, some characteristics are tailored towards their principal aim: the processing… view at source ↗
Figure 3
Figure 3. Figure 3: typical trajectory of a gradient descent to a local minimum in a multidimensional [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
read the original abstract

Because large language models (LLMs) are impressively successful in predicting text, it appears that they must have access to a 'world model' representing causal and definitional structure. However, the dominant formalisms of modern causal inference -- Judea Pearl's interventionist approach and the Neyman-Rubin potential outcomes framework -- struggle to illuminate how LLMs learn causal structure. I resolve this puzzle by arguing that LLMs employ a specific inductive approach based on a difference-making logic -- sometimes called variational induction. I demonstrate how central aspects of this logic are realized during training, where LLMs require enormous amounts of text data from a wide range of contexts to identify difference- and indifference-makers within word sequences. Furthermore, I analyze specific architectural features of LLMs -- such as token embeddings and self-attention -- to determine their roles in variational induction. The difference-making logic of LLMs fundamentally parallels the experimental method, where causal relations are derived by systematically varying individual circumstances to determine their influence on a phenomenon.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper claims that LLMs acquire causal structure in text via a difference-making (variational induction) logic realized during training on diverse contexts, which identifies difference- and indifference-makers in word sequences; it further analyzes how token embeddings and self-attention implement this logic and draws a parallel to the experimental method, arguing that Pearl/Rubin frameworks are inadequate for explaining LLM causal learning.

Significance. If substantiated, the argument would supply a philosophically grounded alternative to interventionist causal formalisms for interpreting LLM world models, potentially informing both model interpretability and training data design. The manuscript credits the scale and diversity of training data as enabling the inductive process and offers an explicit architectural analysis, but these remain at a descriptive level without falsifiable mappings or tests.

major comments (2)
  1. [architectural analysis sections] The central claim that variational induction via difference-making isolates causal (rather than merely predictive) structure is load-bearing yet unsupported: the analysis of embeddings and self-attention (described in the sections on architectural features) provides no formal condition or divergence test showing when attention scores or embedding distances track causal influence versus any statistical association already captured by next-token prediction.
  2. [training data sections] § on training realization: the assertion that enormous varied text data necessarily identifies difference- and indifference-makers that track causal relations lacks a derivation or empirical demonstration; without it, the argument risks circular re-description of standard language modeling objectives.
minor comments (2)
  1. Notation for 'difference-makers' and 'indifference-makers' is introduced without a precise definition or contrast to related concepts such as interventions or counterfactuals, reducing clarity in the comparison to Pearl/Rubin frameworks.
  2. The parallel drawn to the experimental method would benefit from a short table contrasting variational induction steps with LLM training phases.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their insightful comments, which highlight areas where the manuscript can be strengthened. We address each major comment point by point below.

read point-by-point responses

  1. Referee: [architectural analysis sections] The central claim that variational induction via difference-making isolates causal (rather than merely predictive) structure is load-bearing yet unsupported: the analysis of embeddings and self-attention (described in the sections on architectural features) provides no formal condition or divergence test showing when attention scores or embedding distances track causal influence versus any statistical association already captured by next-token prediction.

    Authors: The architectural analysis is intended as an interpretive mapping of how the LLM components realize difference-making logic through context comparison in embeddings and attention mechanisms. The distinction from mere prediction is made by noting that only diverse training contexts allow identification of what makes a difference to the outcome. We agree that the lack of a formal condition or test leaves the claim somewhat unsupported empirically. We will revise the sections to propose a specific divergence test, for instance by examining attention score differences in synthetic causal vs. correlational text pairs, to make the claim more falsifiable. revision: yes

  2. Referee: [training data sections] § on training realization: the assertion that enormous varied text data necessarily identifies difference- and indifference-makers that track causal relations lacks a derivation or empirical demonstration; without it, the argument risks circular re-description of standard language modeling objectives.

    Authors: The derivation is conceptual, based on the requirement that to detect difference-makers, the model must encounter varied contexts where the same word sequence leads to different outcomes depending on other factors. This parallels the experimental method described in the paper. However, we acknowledge the risk of circularity and the absence of empirical demonstration. In the revised manuscript, we will expand the training realization section with a step-by-step logical derivation showing how the language modeling objective on diverse data leads to causal tracking, and note the limitations of the current argument. revision: partial

Circularity Check

0 steps flagged

No circularity; interpretive framework with no derivations or self-referential reductions

full rationale

The paper advances a philosophical interpretation that LLMs realize variational induction via difference-making logic during training on varied text, drawing parallels to experimental methods and analyzing embeddings/self-attention descriptively. No equations, parameter fittings, or formal mappings appear in the provided text. The argument does not invoke self-citations as load-bearing uniqueness theorems, nor does it rename known results or smuggle ansatzes. Central claims rest on conceptual parallels rather than any step that reduces by construction to its own inputs, making the derivation self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The claim rests on the domain assumption that successful text prediction requires a causal world model and that standard causal formalisms are inadequate, with no free parameters or invented entities introduced in the abstract.

axioms (2)
  • domain assumption LLMs must have access to a 'world model' representing causal and definitional structure because they successfully predict text.
    Opening premise of the abstract that sets up the puzzle to be resolved.
  • domain assumption Dominant causal inference formalisms (Pearl, Neyman-Rubin) are insufficient to explain LLM causal learning.
    Stated directly as the reason a new approach is needed.

pith-pipeline@v0.9.1-grok · 5697 in / 1281 out tokens · 29606 ms · 2026-06-26T10:44:54.569012+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

14 extracted references · 2 canonical work pages

  1. [1]

    Erkenntnis

    The Immortal Science of ML: Machine Learning and the Theory-Free Ideal. Erkenntnis. https://doi.org/10.1007/s10670-025-01010-x Bacon, Francis. 1620/1994. Novum Organum. Chicago, Il: Open Court. Baumgartner, Michael, and Gerd Graßhoff

    work page doi:10.1007/s10670-025-01010-x 1994

  2. [2]

    Synthese 203 (1):1-39

    Functional Concept Proxies and the Actually Smart Hans Problem: What’s Special About Deep Neural Networks in Science. Synthese 203 (1):1-39. Bolzano, Bernard. 1837/1972. Theory of Science (ed. R. George). Berkeley and Los Angeles: University of California Press. Brown, Tom B., Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Ar...

    1972

  3. [3]

    arxiv.org/abs/2005.14165 (downloaded on 25 March

    Language Models are Few-Shot Learners. arxiv.org/abs/2005.14165 (downloaded on 25 March

    Pith/arXiv arXiv 2005

  4. [4]

    https://arxiv.org/abs/1810.04805 Durán, Juan M

    BERT: Pre- training of Deep Bidirectional Transformers for Language Understanding. https://arxiv.org/abs/1810.04805 Durán, Juan M. and Giorgia Pozzi (eds.)

    Pith/arXiv arXiv

  5. [5]

    Dialectica 77(3): 319–333

    Causal Inference from Big Data? A Reply to Pietsch (2021). Dialectica 77(3): 319–333. Gillies, Donald

    2021

  6. [6]

    Olympiad-level formal mathematical reasoning with reinforcement learning. Nature. https://doi.org/10.1038/s41586- 025-09833-y Hübner, Dietmar, Mathias Frisch, and Uljana Feest

    work page doi:10.1038/s41586-

  7. [7]

    Big Data, Machine Learning – und diskriminierende Algorithmen? Unimagazin der Leibniz Universität Hannover 3/4 2021, 22-

    2021

  8. [8]

    arxiv.org/abs/2001.08361 Keynes, John M

    Scaling Laws for Neural Language Models. arxiv.org/abs/2001.08361 Keynes, John M

    Pith/arXiv arXiv 2001

  9. [9]

    University of Chicago Press

    Data-centric biology: A philosophical study. University of Chicago Press. Leonelli, Sabina. 2020/2025. Scientific Research and Big Data. In: Edward N. Zalta & Uri Nodelman (eds.), The Stanford Encyclopedia of Philosophy (Winter 2025 Edition), plato.stanford.edu/archives/win2025/entries/science-big-data/ Losee, John

    2020

  10. [10]

    Understanding AI

    “Understanding AI”: Semantic Grounding in Large Language Models. arxiv.org/abs/2402.10992 Mackie, John L

    arXiv

  11. [11]

    Erkenntnis 88(5), 1895–1909

    Can machine learning provide understanding? How cosmologists use machine learning to understand observations of the universe. Erkenntnis 88(5), 1895–1909. Mill, John S

    1909

  12. [12]

    arxiv.org/abs/2401.03910

    A Philosophical Introduction to Language Models – Part I: Continuity with Classic Debates. arxiv.org/abs/2401.03910. Mussgnug, Alexander M

    arXiv

  13. [13]

    In: Edward N

    The Analytic/Synthetic Distinction. In: Edward N. Zalta & Uri Nodelman (eds.), The Stanford Encyclopedia of Philosophy (Spring 2023 Edition), plato.stanford.edu/archives/spr2023/entries/analytic-synthetic/ Rubin, Donald B. and Guido W. Imbens

    2023

  14. [14]

    arxiv.org/abs/2409.11277 Trudel, Andre Curtis, Darrell P

    Machine Learning and Theory Ladenness – A Phenomenological Account. arxiv.org/abs/2409.11277 Trudel, Andre Curtis, Darrell P. Rowbottom, and David Barack (eds.). forthcoming. The Role of Artificial Intelligence in Science: Methodological and Epistemological Studies. Routledge. Vapnik, Vladimir

    arXiv