acceptodds
Under review as a conference paper at ICLR 2027

How Causality Bridges the Semantic Gap

Abstract

Numerical measurements capture how a system behaves, but often leave the meanings of its variables unspecified. Some variables are measured but never labeled, and others are never measured at all. Existing methods assign semantics to such variables by consulting general human knowledge, but this inherits its biases where that knowledge exists and offers nothing where it does not. We bridge this gap between measurements and their meanings with causal structure instead, reading a variable's semantics from how it acts on other variables. We formalize this as structure-constrained semantic alignment, in which the embedding of each unnamed variable is solved under the dependence relations implied by the causal graph, with the embeddings of a few known names as anchors. Accordingly, we build **CausalBridge**, a framework that discovers the causal graph from the measurements, latent variables included, solves for the embeddings under those relations, and expresses them as names through a language model. The causal structure reflects the mechanism that generated the measurements and is recovered from the measurements alone, which may make it the one source of information free of bias from human knowledge. We evaluate CausalBridge on five questionnaires and three robotics scenarios, with 20 to 90% of the variable names masked. It recovers the semantics of observed and latent variables more accurately than existing methods that rely on association, and its lead widens as less of the system is documented. The graph it discovers names variables as accurately as the documented one, and a new system is named in minutes and at a fraction of the cost of sampling methods. Once the semantic gap is bridged faithfully, machines can understand the world and take actions causally.

Primary area:
interpretability and explainable AI
Keywords:
interpretability, semantic alignment

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_6LhZOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies semantic naming of unnamed observed and latent variables from numerical measurements and a small set of known names. CAUSALBRIDGE combines causal discovery, structure-constrained embedding optimization, and prefix-based language-model adaptation. Experiments on questionnaires and robotic/vehicle data show improvements in name retrieval.

Critical strengths
  1. Jointly constraining unknown semantic embeddings using relational information and named anchors is an appealing alternative to independent naming.

  2. The paper considers multiple domains, masking levels, graph sources, and language-model backbones.

Critical weaknesses
  1. The practical motivation for naming observed variables is not fully convincing. In many scientific applications, the names of observed variables are already determined, and the primary goal is to discover their causal relationships. Moreover, variable names may simply be identifiers, such as gene A, gene B, and gene C, rather than meaningful semantic descriptions. The authors should clarify when recovering a name constitutes a substantive scientific contribution and distinguish naming a variable from understanding its function or causal role.

  2. Statistical dependence is represented through cosine similarity between semantic embeddings. I'm not sure whether this implementation makes sense: causally related variables need not have similar meanings.

  3. In Table 29, removing the prefix improves latent-variable naming quality at all four masking levels under discovered graphs. These results require reconciliation with the claimed central role of the prefix.

As a researcher in causal inference, I find that this paper does not contribute new knowledge to the causal inference community (which is, of course, not necessarily a weakness). Instead, it applies existing causal inference methods to extract information from data to determine the semantics of variables. I find this task somewhat unusual, and I am not convinced that it presents substantial challenges or has significant practical or theoretical value. However, I do not hold this view strongly, and I would like to consider the opinions of other reviewers and the authors' rebuttal before deciding whether to revise my assessment.

Questions for authors

I find semantic interpretation of latent variables to be the most compelling application of this framework. I therefore encourage the authors to engage more deeply with related work on causal discovery among latent variables [1-8].

[1] Learning linear bayesian networks with latent variables. ICML 2013.
[2] Triad constraints for learning causal structure of latent variables. NeurIPS 2019.
[3] Latent hierarchical causal structure discovery with rank constraints. NeurIPS 2022.
[4] Structural estimation of partially observed linear non-gaussian acyclic model: A practical approach with identifiability. ICLR 2024.
[5] Score-based causal discovery of latent variable causal models. ICML 2024.
[6] Learning Discrete Latent Variable Structures with Tensor Rank Conditions. NeurIPS 2024.
[7] Recovery of causal graph involving latent variables via homologous surrogates. ICLR 2025.
[8] Efficient and trustworthy causal discovery with latent variables and complex relations. ICLR 2025.

8 comments
AuthorsOpenReview ↗

We sincerely appreciate your thoughtful review and helpful references (1/3)

Dear Reviewer 6LhZ,

We sincerely appreciate your time and the thoughtful review. We are glad that you found the semantic interpretation of latent variables the most compelling application of our framework. Following your suggestions, we have tested the link between dependence and names, shown that the names of latent variables come from the causal structure, and added a discussion of causal discovery with latent variables [1-8]. All section, table and appendix numbers below refer to the updated manuscript, where every change is marked in blue. Please find our point-by-point responses below.


W1. When does naming an observed variable matter, and how does a name differ from a function or causal role?

Thanks for raising this question. Naming is substantive whenever the meaning of a variable is unknown. We would like to highlight the difference between a semantic meaning and an identifier for observed variables. The semantic meaning here describes what a variable means with richer semantic information, such as “rear left wheel speed”, while an observed identifier such as “gene A” does not provide anything beyond that identifier. Recovering the semantic meaning, like a description, for both observed and latent variables is our task. This need arises when a vehicle CAN bus records channels that its maker does not document, which engineers decode by hand and our vehicle experiments reproduce. It also arises when shared robot corpora such as Open X-Embodiment record observation and action spaces that differ across robots. It arises most clearly for latent variables, which causal discovery returns with only an identifier.

As for the difference between a name and a causal role, the two play different parts in the method. The functional/causal role is the input of the method, namely what a variable causes, what causes it and what it is independent of. The name is the output, a description in words that agrees with that role and with the named variables.

AuthorsOpenReview ↗

We sincerely appreciate your thoughtful review and helpful references (2/3)

W2. Choice of cosine similarity for the link between dependence and names.

Thanks for the question. We state the link between dependence and names as a modeling assumption (Section 2, L187-189), and the method asks only that a cause and its effect share part of their meaning, because the two can mean different things. The embedding of an effect is generated from the embeddings of its causes by a learned operator and keeps a residual of its own (Section 3.2). For strongly dependent pairs, the similarity only has to reach a fraction of their dependence. On a negatively correlated edge, the cause contributes its opposite meaning through a learned negation map.

The true names support this assumption on every data set. We take the pairs of a cause and its effect from the published keys and the supplied graphs, and measure how similar their true names are by the cosine similarity in the space that scores the names (Section 4). We then compare this similarity with the median similarity over all pairs of the same kind, such as all pairs of a latent variable and an item. If causal relations said nothing about names, half of the cause-effect pairs would fall below this median. Over all data sets, about 16% do (Appendix N).

Moreover, the link is stated for a general measure of dependence between embeddings (Section 3.2, L255-257). The conditional mutual information (CMI) is a standard measure of conditional dependence that is defined for any distribution and any form of relation. With the CMI, the embeddings are asked to share as much information as the variables. The cosine similarity corresponds to the CMI under a Gaussian model of the embeddings (Appendix C, L1486-1501), and the main method uses it for its simplicity and scalability. The two versions write essentially the same names, and on the discovered graph their NRR@1 differs by at most .05 at every masking level (Appendix M, L3568-3576). The measure is therefore a choice of the user, who can take the one that fits the scenario best.


W3. The prefix and naming quality in Table 29.

Thank you for this careful observation. The gap in naming quality comes from the form of some names. With the prefix, the name of a discovered latent variable is sometimes written as a sentence in the style of a questionnaire item. For example, the full method gives “I felt I was not good enough.” for the depression scale of DASS, and the variant without the prefix gives the short name “worthlessness”. The judge recognizes the sentence as depression, but the form check of naming quality rejects it, since it is not a name. The table compares the full method with the variant without the prefix on latent variables at 20 / 50 / 75 / 90% masking. All rows except the fourth are under the discovered graph.

Latent variablesWith the prefixWithout the prefix
Identification.531/.511/.441/.254.549/.474/.443/.229
Recognized, but written as a sentence.086/.086/.052/.037.011/.000/.006/.003
Naming quality.446/.426/.390/.217.537/.474/.437/.226
Naming quality under the supplied graph.596/.518/.500/.332.496/.465/.462/.273
NRR@1.540/.486/.386/.240.414/.386/.406/.186

Naming quality is identification minus the second row. Identification is about the same with and without the prefix, so the gap in naming quality comes from these sentences. The sentences concentrate on discovered latent variables with only one or two items. The model then writes the name in the style of those few items. Under the supplied graph, each latent variable has more items, and the prefix raises naming quality at every masking level (fourth row).

In the paper, the central role of the prefix is measured in NRR@1, which scores how precisely a name places its meaning among all names of the system. Removing the prefix lowers NRR@1 on all variables at almost every masking level (Tables 27 to 29). The last row shows this on the same latent variables, where the full method is clearly ahead while many names remain. We have added this analysis and the examples to the updated manuscript (Appendix Q).

AuthorsOpenReview ↗

We sincerely appreciate your thoughtful review and helpful references (3/3)

W4. Contribution to causal inference and the value of the task.

Thank you for sharing this view. Our contribution to causal inference is the step that follows discovery. Methods such as [1-8] return variables as nodes with only an identifier such as L1, and their real-data examples name them by hand (Appendix S). A perfectly identified latent variable may be well disentangled from the others, but no semantic meaning is recovered in the procedure. CausalBridge gives it a meaning from its causal relations to the named variables, and Proposition 3 characterizes the assignments that satisfy these relations. The output of discovery thus becomes hypotheses that a person can read and test.

A new experiment shows that these names come from the causal structure. We replaced a third, two thirds or all of the edges of the discovered graph with random edges of the same kind and kept everything else fixed, including the language model, the prompt template and the decoding. We then scored each latent variable against its name under the unchanged graph. The table gives NRR@1 at 20 / 50 / 75 / 90% masking, with the best value at each level in bold.

Edges replacedNRR@1 on latent variables
None.540/.486/.386/.240
1/3.374/.322/.275/.203
2/3.223/.234/.209/.149
All (random graph).133/.107/.130/.151

Accuracy falls as more edges are replaced, so the name comes from the causal relations of the latent variable. Under the supplied graph the conclusion is the same, and we have added these experiments to the updated manuscript (Appendix N).

The task is hard because the meaning must come from the structure. A latent variable has no text of its own, and an unnamed channel has only an identifier, so a language prior alone cannot supply their meaning. Standard interpretability tools ask a language model what the data suggest, and with no structure supplied they stay well below CausalBridge on the vehicle CAN at every masking level (Table 9). Beyond the settings of W1, naming the features of neural networks is a practical use, and CausalBridge writes each name from one prompt, with 2.6 to 56.6 times fewer tokens per latent variable than the sampling methods (Section 4).


Q1. Related work on causal discovery with latent variables [1-8].

Thank you for these recommendations. These works discover the causal structure among latent variables and state when it is identifiable, and the updated manuscript now cites them in the introduction. Appendix S discusses them in detail, including the data type each of them fits and how their latent variables are named when they are applied to real data. They return the latent variables as nodes with only an identifier, such as L1. CausalBridge takes such a graph as input to name them, so whichever of these methods fits the data can supply the graph.


Thanks again for the time and care you invested in reviewing our work. If any part of our responses remains unclear, please feel free to let us know, and we will be glad to clarify further.

Reviewer_zwjcOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a method for naming observed and latent variables using numerical measurements and a subset of known variable names. It first learns a causal graph, then optimizes semantic embeddings using the graph and the names of other RVs (that have known names) as anchors. A trained adapter passes these embeddings to a language model to generate names. The method is evaluated on robot and vehicle measurements and five questionnaires, with 20–90% of names hidden. The authors report improved name retrieval over several baselines.

Critical strengths
  1. This is an interesting problem. Naming undocumented measurements and interpreting discovered latent variables are useful tasks. Combining graph-guided embedding optimization with a language model seems to offer a practical way to address both.

  2. The experiments cover simulated and real systems, several masking levels, and both supplied and discovered graphs. They include component ablations and a language model given the graph as text. The reported improvements, particularly on vehicle channels, provide encouraging evidence of practical value.

Critical weaknesses
  1. The claims about recovering semantics/names from the causal structures are stronger than the presented justification. Statistical dependence does not necessarily imply semantic similarity, so the embedding constraints are modeling assumptions rather than consequences of causal theory. The method also uses named anchors, supervised training, a concept bank, and language-model knowledge. Proposition 3 characterizes when the loss is zero; it does not establish recovery of the correct names. The results support predicting missing names more clearly than discovering meaning independently of existing knowledge.

  2. The experiments do not isolate the main claimed contribution. The component ablations do not show whether causal relationships improve naming over correlations under comparable training and decoding. This distinction is central to the paper’s claim.

  3. TLVD, which the paper cites as a related approach to naming latent variables from a recovered graph, is also missing from the comparisons, while TLVD seems to be the most relevant baseline and has with an open source code.

These gaps make the advantage of the proposed causal formulation difficult to assess.

Questions for authors
  1. How does performance change when causal relationships are replaced by correlation-based or undirected relationships, with comparable training, model capacity, and decoding?

  2. What supports the assumption that statistical dependence should correspond to semantic similarity? Is the intended claim supervised prediction of missing names, or identification of meanings from causal structure? How does this account for the ambiguity acknowledged in Appendix C?

  3. Are comparisons with TLVD available on the same naming tasks, using the same graph and known-name information? This would help establish what the embedding optimization and adapter add.

6 comments
AuthorsOpenReview ↗

We sincerely appreciate your clear and constructive review (1/3)

Dear Reviewer zwjc,

We sincerely appreciate your time and the clear and constructive review. We are glad that you found the problem interesting and the combination of graph-guided embedding optimization with a language model practical. We also appreciate that you found the improvements encouraging, particularly on the vehicle channels. Following your suggestions, we have compared causal relations with correlation-based and undirected ones inside CausalBridge, run TLVD on the same task, and tested the assumption that links the graph to the names. All section, table and appendix numbers below refer to the updated manuscript, where every change is marked in blue. Please find our point-by-point responses below.


W1. More discussion on the modeling assumption behind the embedding constraint.

Thanks for the comment. We have revised our claim so that it matches the evidence, and the updated introduction now states it as the prediction of missing names from the causal structure (Section 1). The structure narrows the meanings that an unnamed variable can take, and the anchors and existing knowledge select and phrase one of them. The experiments under W2 show that this narrowing is what carries the meaning of the latent variables.

We also state the link between the graph and the embeddings as a modeling assumption (Section 2, L187-189), and new analyses support it. Causal theory determines which variables are dependent and which are independent, and we assume that the embeddings of their names follow the same relations. The method asks only that a cause and its effect share part of their meaning. The embedding of an effect is generated from the embeddings of its causes and keeps a residual of its own, and for strongly dependent pairs the similarity only has to reach a fraction of their dependence (Section 3.2).

First, the true names agree with this assumption on every data set. We take the pairs of a cause and its effect from the published keys and the supplied graphs, and measure how similar their true names are by the cosine similarity in the space that scores the names (Section 4). We then compare this similarity with the median similarity over all pairs of the same kind, such as all pairs of a latent variable and an item. If causal relations said nothing about names, half of the cause-effect pairs would fall below this median. Over all data sets, about 16% do (Appendix N).

Second, the link is stated for a general measure of dependence between embeddings (Section 3.2, L255-257). The conditional mutual information (CMI) is a standard measure of conditional dependence that is defined for any distribution and any form of relation. With the CMI, the embeddings are asked to share as much information as the variables. The cosine similarity corresponds to the CMI under a Gaussian model of the embeddings (Appendix C, L1486-1501), and the main method uses it for its simplicity and scalability. The two versions write essentially the same names, and on the discovered graph their NRR@1 differs by at most .05 at every masking level (Appendix M, L3568-3576). The measure is therefore a choice of the user, who can take the one that fits the scenario best.

Proposition 3 characterizes the set of embeddings that satisfy the relations of the graph. Within this set, the anchors and existing knowledge select the name, and our experiments measure how often the selected name is correct. Existing knowledge enters through the concept bank and the language model. The trained parts are fitted on separate data and never see the evaluated systems (Appendix D.5).

AuthorsOpenReview ↗

We sincerely appreciate your clear and constructive review (2/3)

W2. Causal relations against correlations under the same training.

Thanks for the suggestion. In light of it, we conducted new experiments that change only the graph inside CausalBridge, and they show that causal relations improve naming over correlations. The trained parts, the model capacity, the prompt template, the decoding and the judge stay as in the paper.

A causal graph differs from a correlation graph in two ways, since it contains latent variables and its edges have directions. A latent variable is a common cause of its child variables, so neither a partial-correlation graph nor an undirected graph can name it. Its name also depends on its causal relations. We replaced a third, two-thirds, or all of the edges of the discovered graph with random edges of the same kind. The table gives NRR@1 on the latent variables at 20 / 50 / 75 / 90% masking, with the best value at each level in bold.

Edges replacedNRR@1 on latent variables
None.540/.486/.386/.240
1/3.374/.322/.275/.203
2/3.223/.234/.209/.149
All (random graph).133/.107/.130/.151

Accuracy falls as more edges are replaced.

The directions also matter on the robot channels where every variable is observed, and the three graphs can be compared directly. The partial-correlation graph connects the channels with the strongest partial correlations and has as many edges as the discovered graph. The undirected graph keeps the edges of the discovered graph and gives each of them a random direction, since the method needs a direction to find the parents of a variable. We compare them in type accuracy, since the structure determines the type of a channel (Section 4). Type accuracy measures whether a name identifies the right kind of channel, such as a joint angle or a wheel speed.

When few names remain, the causal graph gives the highest type accuracy on all three sets of robot channels. Removing the directions then lowers type accuracy, so the direction of each causal relation carries information that an undirected graph lacks. The paper also shows the converse. When the discovery step admits the directions that the physics of a platform fixes, type accuracy on the simulated arms rises at every masking level (Table 15). This agrees with the ablation in Section 4. While many names remain, the evidence line already indicates the type of a channel, and from 75% masking on, the prefix that carries the structure matters most. The same conclusion holds under the supplied graph, and we have added these results to the updated manuscript (Appendix N).

AuthorsOpenReview ↗

We sincerely appreciate your clear and constructive review (3/3)

W3. Comparison with TLVD.

Thanks for pointing to TLVD. We compared CausalBridge with TLVD on the same naming task, and CausalBridge names the latent variables more precisely at every masking level, even when TLVD runs on a much larger model.

TLVD names the latent variables of a graph recovered by RLCD, which is also our discovery method for the questionnaires. Two language-model agents propose names from the neighbors of a latent variable, and a coordinator settles on one. We ran its released code with its released prompts, number of agents and decoding, and gave TLVD the same discovered graph and the same named items as CausalBridge for every latent variable. The release also searches the web for evidence about each edge, but its code does not pass this evidence to the coordinator, so we ran the naming loop alone. To separate the method from the model, we ran TLVD with two models. Kimi-K2 is the model of its release and has about 1T parameters, and Qwen3-4B is the model of CausalBridge. The table gives NRR@1 at 20 / 50 / 75 / 90% masking, with the best value at each level in bold. The TLVD values are means over three seeds.

MethodNRR@1
CausalBridge (Qwen3-4B, 1 call).540/.486/.386/.240
TLVD (Qwen3-4B, 3 calls).460/.428/.330/.228
TLVD (Kimi-K2, 3 calls).488/.476/.375/.211

With the same model, CausalBridge names the latent variables more precisely than TLVD at every masking level, and this gap is what the embedding optimization and the adapter add. Against Kimi-K2, CausalBridge also stays more precise at every masking level while using about a quarter of the tokens (Table 34). We have added this comparison and the details of how we ran TLVD to the updated manuscript (Appendix O).


Q1. How does performance change when causal relationships are replaced by correlation-based or undirected ones?

Thanks. The experiments under W2 answer this question, with the training, the model capacity and the decoding fixed. Neither the partial-correlation graph nor the undirected graph can represent a latent variable, so neither can name one. On the robot channels, the causal graph gives the highest type accuracy of the three graphs when few names remain.


Q2. What supports the assumption, what is the intended claim, and how does it account for the ambiguity of Appendix C?

Thanks for these questions. The assumption rests on a simple reason, and the data support it. Causally related variables usually belong to the same mechanism, so their names share part of their meaning. Examples include a latent trait with the items that measure it, and the wheel speeds with the vehicle speed computed from them. We now state this link as a modeling assumption (Section 2, L187-189), and the analysis of the true names under W1 supports it on every data set (Appendix N).

The updated introduction now states our claim (Section 1). We predict the missing names of a system from its causal structure. The structure constrains the meaning of each unnamed variable, and a language model phrases this meaning as a name. The few named variables of the system serve as anchors. The structure fixes how each unnamed variable relates to them, and the masked names are never exposed during training and inference.

As for the ambiguity of Appendix C, the causal structure recovered from the measurements decides where each unnamed embedding lies relative to the anchors. Existing knowledge only resolves the freedom that these relations leave. It enters through the concept bank and the language model, and our experiments measure how often the resulting name is correct.


Q3. Are comparisons with TLVD available with the same graph and known-name information?

Thanks. Yes, the comparison under W3 gives TLVD and CausalBridge the same naming task, the same discovered graph and the same named items. With the same model, the gap between the two methods is what the embedding optimization and the adapter add.


Thanks again for the time and care you invested in reviewing our work. If any part of our responses remains unclear, please feel free to let us know, and we will be glad to clarify further.

Reviewer_faYcOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Many systems record numbers whose semantic meanings are unknown to us, producing unlabeled columns. Other variables are hidden causes that are never measured at all (latent variables). Existing methods name these mostly by consulting human knowledge, e.g., by asking a language model to guess from the data or from neighboring names; the paper argues that this struggles when little has been documented (e.g., the vehicle CAN example in the Introduction, where two baselines give one name to several channels).

This paper proposes CausalBridge, which infers a variable's meaning from its causal relationships to the few variables that are known. It works in three main steps:

  1. Discovers a cause-and-effect graph from the data alone (BOSS over two consecutive time steps for robot and vehicle channels, and rank-based RLCD for questionnaires with latent variables).
  2. Places each unnamed variable in a text-embedding space so that its position agrees with the graph, with the named variables acting as fixed anchors.
  3. Feeds that position to a frozen language model as a prefix, which gives the unnamed variable a name.

The authors test the method on five questionnaires, simulated and real robot arms, and car sensor (CAN) data, masking 20–90% of the true variable names. They report that CausalBridge names both unlabeled and hidden variables more accurately than existing methods. Its advantage is largest where labels carry little meaning (robot and vehicle channels) and grows on latent variables as more is masked, while it narrows at 90% masking, where few anchors remain. The graph it discovers works about as well as a documented one and the authors report using 2.6–56.6 times fewer tokens than the sampling-based baselines. The authors conclude that causal structure can help connect raw measurements to meaningful names.

Critical strengths
  • Broad evaluation. The method is tested in multiple different domains and a variety of data (5 questionnaires, 6 simulated arms, 3 real-robot logs, and 10 vehicle drives), at 4 levels of masking. It is compared against several existing approaches and outperforms them in most settings (Fig. 4).

  • Efficient. The method names each variable with only one prompt to an LLM. The authors also report much smaller token usage than competing methods (2.6-56.6 times fewer tokens per latent variable; 10 drives named in about 7 minutes), plus the ability to work with either a discovered or a supplied graph and with 3 different LLMs (3B Llama; 4B and 8B Qwen models).

Critical weaknesses
  • Unfair comparison; no isolation of the causal structure. In the main setting (no structure supplied), Causal-bridge discovers a causal graph while all baseline lack any structure (Sec 4). The ablation (RQ2) removes the prefix, the evidence line, and the low-rank update. It could also replace the causal graph inside Causal-bridge (e.g., by a correlation graph, an undirected skeleton, or a random graph of the same density). The gain due to causal, as opposed to associational, information is not established as far as I can tell. The method needs its own structure-aware baselines and an extensive ablation. The association-graph baselines (correlation kNN, Feature Propagation over a correlation graph) appear only in the appendix.
  • Claims exceed scope. The idea is good, but the implementation is limited and the claims are wider than what the experimental results suggest. The abstract and introduction make wide claims about machines coming to "understand the world" and "take actions causally," with the main example (the rover in Fig 1) describing an environment that no human has named at all. The experiments however were run on variables whose names are already human-known and understandable (only masked), so the setting the paper motivates is not really tested; the authors acknowledge this as a main limitation (Sec 5).
  • Limited data. The vehicle results come from 10 drives of the same car, and the real-robot results from only 3 robot logs. It is not clear to me whether these findings would generalize to another car, different robots, or other sources.
  • Human knowledge is still there. The paper argues that causal structure should free the naming from the bias of human knowledge; nevertheless, the method also relies on patterns learned from human language (the pre-trained text encoder that defines the anchors, a regularizer toward english phrases, the prompt with the nearest phrases, and the LLM). The appendix states that which solution "carries the semantics of a variable is left to existing knowledge". The ablation shows that removing the evidence line costs more than removing the prefix on the arms and logs up to 50% masking. This makes it difficult to separate the contribution of the causal structure from that of existing language knowledge.
  • Scoring. Results are scored by an embedding model and an LLM judge (Gemma 3, 12B) and not human evaluators. They only ask whether a name is closer to the right answer but not to the other names in the same system, and the right answer is always on that list (the judge picks among 5 candidates), making it closer to multiple-choice testing instead of an actual naming. A human evaluator (like a domain expert) could check several things like meaning over wording, usefulness, and partially correct answers (e.g. the right type of sensor, but on the wrong side).
  • Untested core assumptions. The method treats cosine similarity of names as a surrogate for statistical dependence ("causally related variables should be semantically related"), but a cause and its effect can mean different things. It also relies on anchors to constrain each unnamed variable, but there is no analysis of how naming accuracy changes when the nearest named variable is several edges away in the graph.
Questions for authors
  1. How much of the accuracy comes from the causal structure, and how much from the prompt text and the language model?
  2. Has the method been tried on a system where the true names were never known, like the rover example in the introduction (Figure 1)?
  3. Did a human evaluator check whether the generated names actually make sense?
  4. How does the method tell apart variables that behave almost the same, such as the speed channels of different wheels? In appendix, Fig 8 shows 2 of 3 unnamed speed channels receiving the same name on a straight drive.
  5. How often does CausalBridge give the same name to several variables?
  6. How well does CausalBridge work when the nearest named variable is several hops away from the target in the graph? It seems the signal from an anchor would weaken with each edge it passes through, so it would help to see results broken down by distance to the nearest anchor.
  7. The method assumes that if two variables are strongly dependent, their names should mean similar things, but a cause and its effect can mean different things. How often does this assumption fail in the data, and how does CausalBridge account for it?
7 comments
AuthorsOpenReview ↗

We sincerely appreciate your detailed review and constructive suggestions (1/4)

Dear Reviewer faYc,

We sincerely appreciate your time and the detailed review. We are glad that you found the evaluation broad and the method efficient. In light of your suggestions, we have added new experiments that change only the graph inside CausalBridge, a human evaluation of the names, and analyses of the core assumption and of the distance to the anchors. All section, table and appendix numbers below refer to the updated manuscript, where every change is marked in blue. Please find our point-by-point responses below.


W1. Comparison with correlation, undirected and random graphs inside CausalBridge.

Thank you very much for this great suggestion. We conducted new experiments that change only the graph inside CausalBridge, and they show that the causal structure carries the meaning of the variables and the direction of its edges matters as well.

A causal graph differs from a correlation graph in two ways, since it contains latent variables and its edges have directions. We therefore compared the discovered graph with the three suggested graphs and kept everything else fixed. The partial-correlation graph connects the observed variables with the strongest partial correlations and has as many edges as the discovered graph. The undirected graph keeps the edges of the discovered graph between observed variables and removes their directions. Neither of them contains latent variables, because a latent variable is a common cause of its child variables and needs directed edges. Since the method needs a direction to find the parents of a variable, each of their edges receives a random direction, which carries no information. The random graph replaces every edge with a random edge of the same kind and therefore has the same density as the discovered graph. We also replaced only a third or two thirds of the edges in the same way.

The results show that the causal structure carries the meaning of the latent variables. Neither the partial-correlation graph nor the undirected graph can name them, because neither contains them. The table gives NRR@1 on the latent variables at 20 / 50 / 75 / 90% masking, with the best value at each level in bold.

Edges replacedNRR@1 on latent variables
None.540/.486/.386/.240
1/3.374/.322/.275/.203
2/3.223/.234/.209/.149
All (random graph).133/.107/.130/.151

Accuracy falls as more edges are replaced, so the names of the latent variables come from their causal relations.

The direction of each causal relation also carries information, which we test on the robot channels. There we measure type accuracy, since the structure determines the type of a channel (Section 4). Type accuracy measures whether a name identifies the right kind of channel, such as a joint angle or a wheel speed. When only a few names remain, the discovered graph gives the highest type accuracy on all three sets of robot channels, and removing the directions of its edges lowers it. The paper also shows the converse. When the discovery step admits the directions that the physics of a platform fixes, type accuracy on the simulated arms rises at every masking level (Table 15). This agrees with the ablation in Section 4. While many names remain, the evidence line already indicates the type of a channel, and from 75% masking on, the prefix that carries the structure matters most. The same conclusion holds under the supplied graph, and we have added all these results to the updated manuscript (Appendix N).

Regarding the fairness of the comparison, we would also like to note that the original paper includes structure-aware baselines. In Tables 12 to 14, every method that can use a graph receives exactly the same graph as CausalBridge, including the one that CausalBridge discovers. These methods include GraphMAE and the base model given the graph. Even then, CausalBridge leads in NRR@1 in almost every setting. Its advantage therefore does not come from having a graph that the baselines lack. As for correlation kNN and Feature Propagation, they build a name from the known names alone. Where their output is a known name, it can never rank the true name first, which is why the paper reports them separately in Appendices I and J.

AuthorsOpenReview ↗

We sincerely appreciate your detailed review and constructive suggestions (2/4)

W2. Scope of the claims and the evaluation setting.

Thanks for the comment. Accordingly, we removed or weakened some claims. For example, the updated manuscript now phrases the closing sentence of the abstract as the motivation of the task. We used masking in our experiments mainly for evaluation, since we need to have the ground truth in order to rigorously validate our results. However, it is worth noting that these masked names are never exposed during training and inference, mimicking the scenario where we do not have knowledge about them. We would also like to note that Figure 1 contains known names as anchors, which is consistent with our method and experiments.


W3. Generalization to other cars, robots and data sources.

Thank you for raising this point. Our evaluation is designed to test generalization to new systems, since CausalBridge never sees an evaluated system during training. Its trained parts are fitted on separate questionnaire scales and on generated arms and vehicles (Appendix D.5), and each evaluated system is named only from its own measurements and its own discovered graph. Under this design, CausalBridge gives the most accurate names on every set of variables (Section 4).

For different robots, the results cover nine platforms, with six simulated arms of six different robots and three real logs from three other platforms (Appendix D). For the car, CausalBridge is trained only on generated vehicles, so the car of the evaluation is already a new car to it. The ten drives of this car range from stop-and-go traffic to highway speed, and they also differ in the relations their data show. On average, the discovered graphs of two drives share only 16% of their edges, and no edge appears in all ten drives (Appendix E), so each drive poses a different structural problem. For other sources, the five questionnaires come from a completely different domain, with scales of 21 to 240 items that range from clinical symptoms to social values.


W4. Roles of the causal structure and of language knowledge.

Thank you for this thoughtful comment. The contribution of the causal structure can be separated from that of language knowledge. The experiments under W1 change only the graph and keep the language side fixed, and the accuracy drops once edges of the discovered graph are replaced with random ones. Conversely, the base model given the graph keeps the same prompt, evidence line and language model as CausalBridge but receives no embedding solved under the structure. It stays below CausalBridge in NRR@1 in almost every setting of Tables 12 to 14, so the prompt and the language model alone cannot explain the results. On the robot channels in particular, it names few channels correctly. It comes closest on latent variables, because its prompt lists the items that each latent variable causes, and this list itself comes from the discovered structure.

Human knowledge plays a different role. It supplies the anchors and the words, while the causal structure recovered from the measurements decides where each unnamed embedding lies relative to the anchors. Human knowledge only resolves the freedom that these relations leave, as Appendix C states. The claim on bias in the abstract refers to the structure, and the updated introduction now states this division explicitly (Section 1).

The ablation in Section 4 does not show that language knowledge outweighs the structure, since the evidence line is built from the measurements and from the solved embedding. It lists the named channels most correlated with the unnamed one, together with the phrases nearest to its solved embedding. This information helps most while many names remain. However, the structure becomes more important as names become scarce. From 75% masking on, removing the prefix costs more than removing the evidence line, since the prefix carries the structure (Section 4, RQ2).

AuthorsOpenReview ↗

We sincerely appreciate your detailed review and constructive suggestions (3/4)

W5. Human evaluation of the names.

Thank you for this suggestion. A new human evaluation shows that NRR@1 reflects what a person sees, so the lead of CausalBridge in Table 9 holds. Specifically, we invited three human evaluators to grade the names that CausalBridge wrote for a random sample of masked robot channels without knowing which method wrote a name. The share graded correct stays within .09 of NRR@1 at every masking level. A name was graded correct if it named the same channel in any wording, and partially correct if it named the right type of channel with a wrong index, axis or side. This partial grade covers the example of the right sensor on the wrong side. It also reflects the usefulness of a name, since the right type of channel already tells a reader what kind of signal a channel carries.

We have added the setup, the agreement between the evaluators and the partially correct grades to the updated manuscript (Appendix P).

We would also like to clarify that NRR@1 rewards a name for getting the meaning right and gives no credit for copying a known name. It ranks the written name against every name of the same system, including the known ones (Appendix H). A name therefore earns credit only when it is closer to its own true name than to the name of any other variable. For instance, if a method copies the name of a channel that is already named, the known name itself becomes the closest name to its output, so the true name cannot rank first. With a paraphrase, the true name can still rank first, because the paraphrase carries the same meaning. For example, “longitudinal acceleration” can rank first for the written name “forward acceleration”. As for the concern about multiple-choice testing, identification is the only metric in which a judge chooses among five candidates. In addition, type accuracy and coordinate accuracy give credit for partially correct names.


W6. More analysis of the core assumption and of the distance to the anchors.

Thank you for raising these two points, which we address in turn.

Cause and effect with different meanings. We state the link between dependence and names as a modeling assumption (Section 2, L187-189), and the method asks only that a cause and its effect share part of their meaning, because the two can mean different things. The embedding of an effect is generated from the embeddings of its causes by a learned operator and keeps a residual of its own (Section 3.2). For strongly dependent pairs, the similarity only has to reach a fraction of their dependence. On a negatively correlated edge, the cause contributes its opposite meaning through a learned negation map.

The true names support this assumption on every data set. We take the pairs of a cause and its effect from the published keys and the supplied graphs, and measure how similar their true names are by the cosine similarity in the space that scores the names (Section 4). We then compare this similarity with the median similarity over all pairs of the same kind, such as all pairs of a latent variable and an item. If causal relations said nothing about names, half of the cause-effect pairs would fall below this median. Over all data sets, about 16% do.

Moreover, the link is stated for a general measure of dependence between embeddings (Section 3.2, L255-257). The conditional mutual information (CMI) is a standard measure of conditional dependence that is defined for any distribution and any form of relation. With the CMI, the embeddings are asked to share as much information as the variables. The cosine similarity corresponds to the CMI under a Gaussian model of the embeddings (Appendix C, L1486-1501), and the main method uses it for its simplicity and scalability. The two versions use essentially the same names, and on the discovered graph their NRR@1 differs by at most .05 at every masking level (Appendix M, L3568-3576). The measure is therefore a choice of the user, who can take the one that fits the scenario best.

Distance to the anchors. Naming accuracy decreases as the nearest named variable gets farther away, but most masked variables lie close to a named one, so this affects only a small share of them. Following the suggestion, we added this analysis, which breaks identification down by the distance in the discovered graph. Identification is highest within one hop of a named variable at every masking level, and most masked variables lie within two hops. We have added both analyses to the updated manuscript (Appendices N and Q).

Reviewer_UaANOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The authors are interested in the problem of naming variables in datasets or latent variables produced by models.

Through either a discovered or pre-supplied causal graph relating the variables, the authors use a textual encoder to represent the existing labels of variables and use the connectivity of the causal graph to learn textual embeddings for the unlabeled variables. Next, using a base LLM, and for each unlabeled variable, the learned embeddings are transformed into prefixes to be added to the input of each transformer layer and a textual prompt is generated using the graph structure and the closest words to the learned embedding. The output of the LLM is the label of that variable. This base LLM is adapted through the network that produces the prefixes, the keys and value matrices for the prefixes, and a LoRA adapter.

The method is evaluated in two main types of data: time-series of sensor data from robots / vehicles and latent-variable models that explain questionary results. The authors evaluate both the case when the causal graph is supplied and not supplied. The results are evaluated in terms of the correct identification of ground-truth labels, an ablation of the presence of the prefix tokens, test of transferability across different base LLM, and token cost.

Critical strengths

Methodology is backbone agnostic and attempts to inject information into the LLM backbone through multiple channels.

Important ablation studies were done, explaining architectural choices.

Critical weaknesses

The method relies on multiple stages: causal discovery, embedding optimization, and LLM adaptation. Although the paper evaluates the discovered graph against a ground-truth graph, it is unclear how the method degrades when intermediate stages produce imperfect outputs. In particular, it is not clear how errors in the causal graph or embedding optimization propagate into the final naming accuracy, nor if such failures can be detected.

The introduction and baseline descriptions cover several related approaches, but a more systematic comparison of their assumptions, limitations, and differences from the proposed method would help readers understand the methodological contribution more clearly.

Questions for authors

Can the authors elaborate on the failure modes of their proposal if any of the steps of their pipeline fail? Such as the causal discovery returning incorrect graphs or the embedding vector discovery not converging? More specifically, is it possible to empirically quantify how the magnitude of errors in the previous steps can propagate into the final result?

What is the biggest bottleneck of the method currently? How does run-time scale in terms of the total number of variables, for a fixed size of unlabeled variables? For example, if one would attempt the example of neural network interpretability, ignoring that most nodes are unlabeled, would this methodology scale for large causal networks?

4 comments
AuthorsOpenReview ↗

We sincerely appreciate your positive assessment and insightful questions (1/3)

Dear Reviewer UaAN,

We sincerely appreciate your time and the positive assessment. We are glad that you found the method backbone agnostic and the ablation studies important for explaining its design choices. Your questions helped us examine the method more closely. In light of them, we have quantified how errors in the graph and in the embedding carry into the names, compared the related approaches systematically, and timed each step on systems of growing size. To measure naming accuracy in these analyses, we mainly use NRR@1, the main metric of the paper. It counts a name as correct when its true name ranks first among all names of the system (Appendix H). All section, table and appendix numbers below refer to the updated manuscript, where every change is marked in blue. Please find our point-by-point responses below.


W1. Robustness to errors in the intermediate stages, and their detection.

Thanks for this important question. We quantified how errors in the graph and in the embedding carry into the names. In both cases, naming accuracy decreases as the error grows, and the variables at risk can be identified before naming.

Propagation from the graph. Naming accuracy decreases as more edges of the graph are replaced with random edges. With a third of the edges replaced at random, CausalBridge still names variables more accurately than the best baseline at every masking level, in NRR@1 on the questionnaires and in type accuracy on the robot channels (Tables 8 and 9). Type accuracy measures whether a name identifies the right kind of channel, such as a joint angle or a wheel speed. In this experiment, we replaced a third, two thirds or all of the edges of the discovered graph with random edges of the same kind and kept every other step unchanged. The table gives NRR@1 on the latent variables at 20 / 50 / 75 / 90% masking, with the best value at each level in bold.

Edges replacedNRR@1 on latent variables
None.540/.486/.386/.240
1/3.374/.322/.275/.203
2/3.223/.234/.209/.149
All (random graph).133/.107/.130/.151

Latent variables are affected most, since they take their meaning only from their relations. On the robot channels, type accuracy decreases mainly when few names remain (Appendix N). While many names remain, the named channels most correlated with an unnamed one already indicate its type. The same holds under the supplied graph.

Propagation from the embedding. A name is much more often correct when its embedding lies closest to its true name than when many other candidates lie closer. We measured the error of a solved embedding as the share of the other candidates of its system that lie closer to it than its true name does. For latent variables, NRR@1 is more than twice as high when the embedding lies closest to the true name as when more than 30% of the candidates lie closer, at every masking level. The supplied graph shows the same pattern (Appendix Q). These errors are already part of the results in the paper, where CausalBridge gives the most accurate names on every set of variables (Section 4).

Detecting failures. The variables at risk of such failures can be identified before naming. The distance from an unnamed variable to its nearest named variable is known once the graph is built. Identification is highest for the variables directly connected to a named variable, at every masking level (Appendix Q).

AuthorsOpenReview ↗

We sincerely appreciate your positive assessment and insightful questions (2/3)

W2. A systematic comparison with related approaches.

Thanks for the suggestion. We have added a systematic comparison to the updated manuscript (Appendix O). It shows that CausalBridge is the only approach of the evaluation that uses the discovered causal structure to constrain the meaning of every unnamed variable before naming it in one model call, whether observed or latent. The table compares what each approach reads, which graph it uses, how it treats latent variables, how it decides a name and what it assumes.

MethodReadsGraphLatent variablesName decided byAssumesModel calls per variable
CLIP-Dissect-e5Activation profileNoneFactor analysisBest-matching concept in a fixed bankThe bank holds a concept that matches the activation pattern0
AutoInterp, DelphiSampled casesNoneFactor analysisA language model explains the cases and scores the explanationThe sampled cases show what the variable responds toSeveral
SASC-styleResponse profileNoneFactor analysisA language model proposes explanations scored on the dataA short explanation fits the response profileSeveral
DeViSE-styleCorrelations with named variablesNoneFactor analysisRegression into the embedding space of namesCorrelations map linearly to name vectors, at least two named variables0
Correlation kNN, Feature PropagationCorrelationsCorrelation graph (Feature Propagation)Factor analysis (Feature Propagation)A copied or diffused known nameA variable shares the name of its correlated neighbors0
GraphMAENode featuresSupplied or discoveredFrom the graphReconstruction decoded to the nearest known nameA graph, a name close to a known name0
Base model given the graphGraph as text, evidence lineSupplied or discoveredFrom the graphThe language prior over the listed neighborsThe names of the neighbors suffice1
TLVDNeighbors of a latent variableRecovered by RLCDLatent variables onlyLanguage-model agents agree on a nameA graph from RLCD3
CausalBridgeMeasurements, known namesDiscovered or suppliedDiscovered with the graphGraph relations constrain an embedding, which a language model phrasesA few named variables, a discovery method for the data type1

The limitations of the other approaches follow from these assumptions. Without a structure, a method can name a latent variable only through a factor-analysis front end, since a latent variable has no data column of its own. GraphMAE and correlation kNN can only return a name that is already known, the sampling methods need several model calls per variable, and TLVD names latent variables only. In turn, CausalBridge needs a few named variables and a discovery method that suits the data type.

AuthorsOpenReview ↗

We sincerely appreciate your positive assessment and insightful questions (3/3)

Q1. What are the failure modes when a step fails, and can their effect be quantified?

Thanks for the question. Yes, the analyses under W1 quantify how the size of the errors in the graph and in the embedding carries into the names. Each step of the pipeline has one main failure mode, which we describe in turn.

If discovery returns an incorrect graph, naming accuracy decreases gradually with the share of wrong edges, and with a third of the edges replaced at random, CausalBridge still outperforms the best baseline (W1). The embedding optimization always returns an embedding, since it starts from a closed-form solution and runs a fixed number of steps (Appendix F). Its failure mode is therefore an inaccurate embedding, and the accuracy of the name follows that of the embedding (W1). Finally, the naming step can miss even when the embedding is accurate, in two cases. First, channels that carry the same signal cannot be told apart from the measurements alone, as two wheel speeds on a straight drive show (Figure 8). Second, a latent variable with only one or two items is sometimes named with a sentence in the style of its items. We have added these analyses to the updated manuscript (Appendix Q).


Q2. What is the main bottleneck, and how does the run time scale with the number of variables?

Thanks for the question. For the systems in the paper, the main bottleneck is the embedding optimization, and a whole system is still named in minutes. Causal discovery becomes the bottleneck only for large systems whose causal order is unknown. The table gives the run time of each step for different data types and sizes, with 30 unnamed variables in the embedding optimization (Appendix R).

StepData type (example)SizeTime
Embedding optimizationOrdinal, i.i.d., linear, latent causes (questionnaires)150/600/2,400 variables175/256/1,039 s
Embedding optimizationContinuous, linear Gaussian, every variable observed (robots, vehicles)100/500/1,000 variables84/180/306 s
Discovery: RLCDOrdinal, i.i.d., linear, latent causes (questionnaires)150/600/2,400 variables10/33/371 s
Discovery: BOSSContinuous, linear Gaussian, every variable observed (robots, vehicles)100/500/1,000 variables0.8/53/400 s
Discovery: sparse regression of each layer on the previous layerContinuous, linear non-Gaussian, every unit observed, causal order known (network activations)19,456 units on 1 GPU/350,208 units on 8 GPUs52/268 s
Discovery: BOSS/DirectLiNGAMThe same, causal order not used250/100 units247/566 s

As for the scaling, the embedding optimization grows about linearly with the total number of variables when the number of unnamed variables is fixed. Its growth comes mainly from one residual fit per variable, and these fits can run in parallel. In contrast, discovery grows faster, at a rate set by the data type. With every variable observed, it therefore overtakes the embedding optimization at the largest size. Finally, naming takes one prompt per unnamed variable, at 18.5 variables per second (Appendix K).

Moreover, the method scales to neural networks. A network has a known causal order given by its layers. With this order, discovery reduces to the second stage of DirectLiNGAM, which regresses each layer on the previous one with a sparsity penalty. This covers every MLP unit of Qwen3-4B in under five minutes. In contrast, BOSS and the full DirectLiNGAM search for the order and already need several minutes for a few hundred units or fewer. The embedding optimization then remains the main cost. It grows about linearly and can run in parallel across units.


Thanks again for the time and care you invested in reviewing our work. If any part of our responses remains unclear, please feel free to let us know, and we will be glad to clarify further.

AuthorsOpenReview ↗

We would greatly appreciate your feedback on our responses

Dear Reviewer UaAN,

Thank you again for your positive assessment and your insightful questions. Since the discussion period will close soon, we would like to kindly check whether our responses have addressed your questions. For your convenience, the main points are summarized below.

  • Robustness of the pipeline (W1 and Q1). We measured how the quality of the graph and of the embedding carries into the names. With a third of the discovered edges replaced at random, CausalBridge still names variables more accurately than the best baseline (Appendices N and Q).
  • Comparison with related approaches (W2). A new table lists what each approach reads, uses and assumes (Appendix O).
  • Scalability (Q2). We timed each step on systems of growing size. With a known layer order, discovery covers every MLP unit of Qwen3-4B in under five minutes (Appendix R).

If any part remains unclear, we would be glad to clarify it right away. Thank you again for your time and care.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary of the changes in the updated manuscript

A list of the main changes can be found below. For your convenience, every change is marked in blue in the updated manuscript.

  • The link between dependence and names is stated as a modeling assumption for a general measure of dependence (Section 2, L187-189; Section 3.2, L255-257).
  • Theory for the conditional mutual information (CMI) as the dependence measure, with the new Proposition 4 (Appendix C, L1486-1501) and the objective with the CMI (Appendix F).
  • Results with the CMI as the dependence measure (Appendix M, Tables 35 and 36).
  • Experiments that change only the graph inside CausalBridge, with correlation, undirected and random graphs (Appendix N, Tables 37 and 38).
  • Analysis of how similar the true names of a cause and its effect are (Appendix N, Table 39).
  • Comparison with TLVD and a table of the related approaches side by side (Appendix O, Tables 40 and 41).
  • Human evaluation of the names by three evaluators (Appendix P, Table 42).
  • Further analyses of naming accuracy, covering the quality of the embedding, the distance to the nearest named variable, repeated names and the prefix in Table 29 (Appendix Q, Tables 43 to 46).
  • Runtime by data type and size (Appendix R, Tables 47 and 48).
  • Further discussion of causal discovery with latent variables (Section 1, Appendix S).
  • Refined claims in the abstract and the introduction, and clarifications to the text as suggested by the reviewers.
AuthorsOpenReview ↗

Common response to all reviewers

We sincerely appreciate the time and care of all reviewers. Four common questions from the reviewers were:

(i) how much the names depend on the causal structure (faYc, zwjc, UaAN),

(ii) what supports the link between statistical dependence and the meaning of names, and whether the cosine similarity is a suitable measure (faYc, zwjc, 6LhZ),

(iii) what the paper claims and what role existing knowledge plays (faYc, zwjc),

(iv) how CausalBridge compares with TLVD and related approaches (zwjc, UaAN).

We answer them in turn below.


(i) The role of the causal structure (Reviewers faYc, zwjc, UaAN)

The names come from the causal structure. To isolate the contribution of the structure, we followed the suggestions and changed only the graph inside CausalBridge, with the training, the model capacity and the decoding fixed (Appendix N). We replaced the causal graph with a partial-correlation graph, an undirected graph, or random graphs in which 1/3, 2/3, or all of the edges are replaced.

The latent variables show this most clearly. A latent variable is a common cause of its child variables, so only a causal graph contains it. Therefore, neither the partial-correlation graph nor the undirected graph can name any latent variable. When the causal edges are replaced with random ones, the accuracy on the latent variables also falls as more edges are replaced, so their names come from their causal relations. The directions matter as well, since removing them lowers type accuracy on the robot channels when few names remain.


(ii) The link between dependence and names (Reviewers faYc, zwjc, 6LhZ)

We state this link as a modeling assumption (Section 2, L187-189) and support it with data and theory. The assumption asks only that a cause and its effect share part of their meaning, because the two can mean different things. Accordingly, the embedding of an effect is generated from the embeddings of its causes and keeps a residual of its own (Section 3.2).

Evidence from the true names. We analyzed the true names of the cause-effect pairs in our data (Appendix N). If causal relations said nothing about names, half of the pairs would have names less similar than the median over all pairs of the same kind. Over all datasets, about 16% do, so the assumption holds for most cause-effect pairs.

Theory for a general measure. The link is stated for a general measure of dependence between embeddings, and the cosine similarity is one choice (Section 3.2, L255-257). We added the theory for the conditional mutual information (CMI). The CMI is a standard measure of conditional dependence that is defined for any distribution and any form of relation. New Proposition 4 shows that the cosine similarity corresponds to the CMI under a Gaussian model of the embeddings (Appendix C, L1486-1501), and Appendix F gives the objective with the CMI. With the CMI as the measure, CausalBridge writes essentially the same names (Appendix M, L3568-3576).

Overall, the measure is a choice of the user, who can take the one that fits the scenario best. The main method uses the cosine similarity for its simplicity and scalability.


(iii) The claim and the role of existing knowledge (Reviewers faYc, zwjc)

We refined the claim in the abstract and the introduction, which now state it as the prediction of the missing names of a system from its causal structure (Section 1).

The structure constrains the meaning of each unnamed variable, and a language model phrases this meaning as a name. The updated introduction states that existing knowledge enters through the named anchors, a preference for embeddings near English phrases and the language model. The experiments under (i) support this division, since the accuracy changes with the graph while the prompt and the language model stay fixed.


(iv) Comparison with TLVD and related approaches (Reviewers zwjc, UaAN)

CausalBridge names the latent variables more precisely than TLVD at every masking level (Appendix O).

We ran the released code of TLVD on the same discovered graph and the same named items as CausalBridge, and CausalBridge stays ahead even when TLVD runs on a much larger model. As suggested by Reviewer UaAN, a new table also sets the related approaches side by side in what each of them reads, uses and assumes.


The remaining questions are answered in each reviewer's thread, including the human evaluation (faYc), the run time (UaAN), the prefix in Table 29 (6LhZ) and the related work on causal discovery with latent variables (6LhZ). We would be glad to clarify any point that remains unclear.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.