acceptodds
Under review as a conference paper at ICLR 2027

Climbing the Causal Ladder: Twenty Language Models Share a Causal Variable

Abstract

Do different language models compute with the same variables? Universality is a founding hypothesis of mechanistic interpretability, yet cross-model evidence (similar representations, transferable steering vectors) shows that models are alike, not that they compute the same causal variable. We give a test that can establish sameness. It grades evidence by Pearl's ladder of causation and requires invariance across the behaviours that read a variable and across models, through maps that never see it. Applied to grammatical number in 20 language models from five families (70M–8B), it finds a causal variable shared across all of them, with evidence at every rung. Counterfactually, interchanging a direction estimated from labels alone at the subject noun mediates a median 0.85 of the full residual state's effect on five held-out agreement behaviours (0.86 in natural sentences), without moving gender. Interventionally, graded edits move every behaviour proportionally, and ablating the direction removes a median 0.53 of the number signal, against 0.01 for the gender direction. Across models, affine maps fitted only on generic text carry it with median efficiency 0.97 within and 0.95 across families (0.99 in natural sentences), while shuffled-pair maps, an untrained source and random directions carry nothing. The test also delimits and discriminates: the shared variable is the noun's own number, not number cued by a determiner (these sheep), and gender, causal in every model, is shared less reliably. Rankings of directions by behavioural strength, the field's usual evidence, reverse under calibration, recovery level and fitting cohort in our preregistered experiments. Invariance, not effect size, identifies the variables models share.

Primary area:
interpretability and explainable AI
Keywords:
mechanistic interpretability, universality, causal abstraction, interchange interventions, causal mediation, invariance, linear representations, cross-model transfer, model stitching, steering vectors, grammatical number, language models, interpretability, activation steering, activation patching, causal interpretability, distributed alignment search, evaluation methodology

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_odTWOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This model uses analogical reasoning (concretely, hypernymy and friends) to try and establish whether generative models 'compute' (really, activate) with the same 'variable' (trajectory/activation).

Critical strengths

Thorough experiments, and grammatically correct

Critical weaknesses

I think this paper has some critical weaknesses, divided in two major categories:

  1. The writing is unscientific and subjective
  2. Technically, the paper is good but must be considerably better to meet the bar.

I do appreciate the effort the authors put into this specific problem, which--at least to me--is compelling and has a few interesting branches. However, as it stands, the authors need to put in more effort in this work so it meets the bar. Here's how I would improve this paper:

In #1, regarding writing: broadly, you need to clearly separate your methods, results, and discussion sections. There is a reason for that: (1) it is easier to understand the setup; what you found (objectively); and how you interpreted these (subjectively). This paper does not do that, and hence every single aspect of the work is confirmatory to the thesis, not supportive. For example:

  1. The section titles are leading.
  2. Do not start an abstract with a question.It is not just style (which is unacademic), it simply does not fulfill the definition of an abstract: a condensed version of your paper. Clip out numbers and keep it readable.
  3. You should put the interpretation of the results in the discussion section, not in the results headers.
  4. Proofread your paper. In the abstract: 'Invariance, not effect size, identifies the variables models share'. What does this mean? It does appear that this does not contribute to the abstract at all.
  5. Captions are descriptive, which is good, but do not tell me what is the takeaway from it. For example, the medians over models plot seems like a clump without any interpretability--what am I looking at? Why am I looking at it?

In #2, regarding technical work:

  1. Maths need to be way more formal. L157-L160: 'recipient and donor activations at a declared site'. What is this? More importantly, what are hr,hdh_r, h_d really? Vectors? Tensors? What is a unit direction? Is this a basis or a metric tensor or something else? Also, this does not seem to contribute to the bulk of the paper.
  2. It is unclear to me what are your experiments. You have the models (without citations) and the 'stimuli and consumers'. But the latter seem to be entirely unrelated to the maps? I think this section is good, but needs to be clearer: what are we using the maps for? etc.
  3. Mechanistic interpretability's common pitfall is that the results are always confirmatory (hence my insistence on the earlier section). This lacks alternate interpretations of the results, in addition to better tests. Spearman is not sufficient to establish the conclusion since it is a correlation measure. Here, better tests would mean contrastive experiments.
  4. Terminology is unsound (compute, variable, etc). Fixing this is of crucial importance in any MI work.

Fundamentally, I think that this paper is not objective enough to be a scientific contribution. The problem is interesting but it really seems that it needed considerably more time to be a paper at the level of ICLR.

Questions for authors

What were, in a nutshell, the conclusions drawn from your work? I see a discussion section, but the conclusion is not in the work. I'd like to see that clearly stated so it can be tied to the thesis and properly evaluated.

Reviewer_KYBbOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The authors propose a framework based on causal tests and evaluate grammatical number across 20 models. The results provide evidence that all models share a causal variable representing grammatical number, supported by tests at different levels of the "causal ladder".

Critical strengths
  • Extensive evaluation with 20 LLMs from five different families
  • The paper reports unsuccessful experimental results, which adds transparency
  • The paper addresses an interesting research question
Critical weaknesses
  • The paper is very dense and difficult to follow. Many technical / abstract concepts and experimental details are introduced in a compressed way, without sufficient intuition or explanation. The paper frequently uses technical and abstract terminology without sufficient explanation.
  • The presentation of the results specifically is particularly difficult to follow. The paper reports many numerical values, evaluation criteria, and comparisons, but does not provide enough intuitive explanation of what these results mean. Additionally, the language is often abstract and not clearly phrased, making the results and their implications difficult to understand. Clearer explanations and more accessible language would significantly improve readability.
  • Although the authors report the unsuccessful experimental criteria in Table 1 and discuss them further in the appendices, their implications for the main conclusions could be more clearly summarised in the main text.
Questions for authors
  1. Several predefined experimental criteria are not satisfied, as reported in Table 1 and discussed further in the appendices. Could the authors clarify how these unsuccessful criteria affect the strength and generality of their main conclusions?
  2. To what extent do the authors expect their findings to generalise beyond grammatical number in English, particularly to more complex or abstract concepts?
Reviewer_TjgmOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper tests whether different language models compute with a shared causal variable, which is a stronger claim than having similar representations or steering effects that transfer. The authors build an evaluation framework on Pearl’s ladder of causation, with four tests: association, graded intervention, counterfactual interchange, and cross-model invariance. They apply it to grammatical number in 20 pretrained models from five families, ranging from about 70M to 8B parameters. At subject-noun positions, they find a number direction and measure how much it drives five held-out agreement behaviors. Interchange interventions on this direction recover a median 85% of the effect of patching the full residual state. Graded interventions, mean ablations, and gender controls add evidence that the direction matters for behavior and is specific to number. Affine maps between models, learned from generic text with no number labels, carry the direction over and keep high intervention efficiency both within and across families. Follow-up experiments test how far the result extends, using natural contexts, determiner-cued number, gender, negation, and training checkpoints. The authors conclude that models share a causal variable for noun-marked grammatical number specifically, not for phrasal number in general. They argue that invariance across contexts and models is stronger evidence of mechanistic universality than representational similarity or intervention strength alone. They also state plainly which pre-registered criteria were not met, including strict monotonicity across the full intervention range and some of the mapping and generalization conditions.

Critical strengths
  1. A rigorous framework for testing causal relevance, not just representational similarity. The paper combines four kinds of evidence: association, controlled interventions, counterfactual interchange, and cross-model transport. It tests on five held-out agreement behaviors, uses grammatical gender for double-dissociation controls, and compares against random and untrained-model directions. Together, these checks make it unlikely that the results come from surface-level similarity or from an intervention that works on one fitted behavior only. The framework is a useful step toward stronger evidence standards in mechanistic interpretability.

  2. Broad cross-model evaluation, with informative controls and boundary conditions. Twenty models from five families are tested. Number directions carried over by maps trained only on generic text come close to the effect of the target model’s own directions, which is strong evidence of functional correspondence across architectures. Shuffled-pair and untrained-source controls strengthen this result. The contrast between noun-marked and determiner-cued number shows that the framework can find where an abstraction stops holding, rather than confirming whatever direction is chosen. The authors also report the pre-registered criteria they did not meet, which adds to the paper’s credibility.

Critical weaknesses
  1. The central claim, that models share an identical causal variable, goes beyond what the tests establish. The paper frames its main contribution as showing that 20 models compute the same causal variable. What the experiments show is narrower: linearly transported directions produce similar causal effects on a selected set of agreement behaviors. That is an important result, but equivalence over a finite set of interventions and outputs does not uniquely identify an underlying causal variable. Models could implement different computations, for example with different distributed representations or redundant pathways, and still respond the same way to these interventions. The label-free mapping does not rule this out either. Generic text carries plenty of number information, so the mapping can align number structure without ever seeing explicit labels. The invariance criterion also checks only that transported directions have similar behavioral effects. It does not show a correspondence that preserves interventions across all relevant states, contexts, and downstream computations. The results therefore support cross-model functional correspondence rather than identity of a uniquely defined variable. This matters because causal identity is the paper’s main claimed conceptual advance.

  2. Reading normalized effects as mediation, and the operational definition of invariance, both need stronger justification. The paper divides the effect of a rank-one intervention by the effect of a full residual patch and calls the result the share of the effect mediated by the direction. The underlying computations are nonlinear, though, and effects need not add up. A ratio of two intervention effects is not automatically a natural indirect effect. That reading needs explicit assumptions about the mediator, what the intervention means, how the direction interacts with its orthogonal complement, and the outcome scale. Changing one coordinate while holding the others fixed can also produce activation states that never occur naturally. Because the argument leans heavily on statements like “the direction mediates 85% of the effect,” this gap matters. Likewise, the invariance tests rely on selected grammatical consumers and preset thresholds. Passing these criteria measures the strength of evidence under the protocol, but it does not amount to an identification result, and the thresholds are neither necessary nor sufficient for causal identity. The paper would be stronger if it stated formally which notion of causal equivalence it establishes, what that notion assumes, and which alternative mechanisms these tests cannot tell apart.

Questions for authors

Q1. What counts as the “same causal variable” across models? Different internal mechanisms could have equivalent effects on the tested behaviors. Is there a formal criterion or identifiability result that separates variable identity from functional equivalence over this restricted test set? If not, should the claim be stated as causal functional correspondence?

Q2. What justifies reading the normalized switch as a mediation proportion? In a nonlinear network, the direction can interact with complementary components. Under what assumptions does the ratio give the mediated share? Can the authors provide a derivation, or show the result is robust under other outcome scales and decompositions?

Q3. Are the mappings truly label-free? Generic text is full of number cues. Do transported directions keep their effect when mappings are fit on low-number-cue contexts or on very different corpora?

Q4. How sensitive are the results to site selection? Each model’s intervention site is tuned on the reference behavior. How much do the cross-model results change with nearby layers or token positions?

Q5. How should the failed pre-registered criteria weigh on the conclusion? Strict monotonicity, the standardized-mapping conditioning hypothesis, and the determiner-cued criterion all failed. Which tests are necessary for the universality claim, and which are only supporting diagnostics?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.