acceptodds
Withdrawn from ICLR 2027

INTERVENTIONS AS ADDRESSES: CAUSAL QUERIES AS ABSTAINING MEMORY READS

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Estimators of causal structure whose parameters are shared across experimental conditions suffer from two structural vulnerabilities: catastrophic forgetting when experimental contexts arrive sequentially, and uncalibrated extrapolation when queried out of distribution or across counterfactual layers. We study an alternative, a memory of every episode stored under the experimental context that produced it, observational or which variable was set, with one rule added: it answers only where the stored evidence settles the query, and otherwise abstains. Held in a vector-symbolic memory with no trainable parameters, interventional distributions, conditioning and backdoor adjustment become reads and arithmetic. The abstention threshold is computed from the memory's own interference noise, so it needs no held-out data, and it is calibrated within a band of dimension. When experimental contexts arrive one after another, a gradient estimator forgets the earliest and the memory does not, since a sum does not depend on the order of its terms. An abstention comes for one of three reasons. A key, the address an episode is stored under, seen with two outcomes omits the variable that decides between them, and a search over the observables finds it. On Baba Is AI the search returns the rule text, and the memory then answers 0.977 of the held-out wall collisions it commits to while the rule is broken. On Alchemy it returns the displayed reward, and an agent given only the description language and the observation format reaches 0.91 of the ideal observer on held-out seeds. A key whose evidence several hypotheses still fit calls for an experiment, and expected free energy over a graded belief held beside the memory chooses it. A key naming two contexts at once can never be written, and the gate abstains from 60/60 such cross-world reads unprompted. Storing experience under its context thus gives an agent an interventional model that survives the order of its experiments, needs no retraining when a new variable is recorded, and says which of three things to do next when it cannot answer. This suggests that for an agent that learns by acting the useful form of causal knowledge may be a memory rather than a fit.

Primary area:
causal reasoning
Keywords:
causal inference, memory

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_PjPFOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper studies the use of memory to retain historical experience in sequential experimentation and causal querying. The central idea is to preserve queryable experimental records, thereby mitigating catastrophic forgetting that may arise when shared predictive models are updated sequentially, and to abstain when the available evidence does not adequately support a query. Through a series of small-scale experiments, the paper illustrates several roles of memory, such as directly estimating interventional probabilities or supplying evidence for inference and experiment selection.

Critical strengths
  1. The paper addresses a meaningful problem: when data from different experimental contexts arrive sequentially, a continually updated predictive model may fail to retain earlier experience. Maintaining historical records separately from the predictive model is a reasonable and intuitive design choice. The sequential learning experiments illustrate its potential to mitigate forgetting.
Critical weaknesses
  1. The main contribution is unclear. The analysis of the memory mechanism is relatively limited, while much of the subsequent discussion focuses on how to respond when memory is insufficient.
  2. The role of memory is not clearly established across experiments. Memory is used to directly estimate probabilities, provide evidence for inference or experiment selection, and supplement predictive models in different experiments. These settings involve fundamentally different roles of memory, making it difficult to understand what general conclusions about the proposed memory mechanism can be drawn from the experimental results.
Questions for authors
  1. The paper is not well-written. It introduces numerous technical details without sufficiently explaining their meanings or motivations. Both the methodology and experimental results are difficult to follow, and the main ideas are obscured by excessive technical details.
  2. The experimental section is poorly organized and difficult to follow. Experimental settings, methods, results, and analysis are often mixed together, making it difficult to identify the specific research questions addressed by each experiment and how the results support the paper's main claims.
  3. The final part of the appendix appears to contain an internal workflow or working-directory description unrelated to the paper’s technical content. The authors should check and ensure a clean submission.
Reviewer_XPXsOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper introduces a way to represent results of previous experiments as a vector-symbolic memory in which observations are indexed by their experimental context. Then, standard quantities of causal inference (empirical interventional distributions, conditional probabilities) are computed through approximate count retrievals. An abstention mechanism rejects queries for which the memory lacks sufficient evidence. The paper discloses substantial AI involvement at every step of the pipeline.

Critical strengths

The paper focuses on a potentially interesting general theme about the distinction between remembering every piece of past experimental information and compressing it into one shared representation.

Critical weaknesses

The paper is very hard to read, as little care was given to creating a logical flow of arguments that guides the reader through the paper's motivation and method. It often reads like the report of an AI coding agent to its user who is already familiar with the research. It is not even clear what problem the paper is trying to solve, what the approaches to this problem are, and what issues this paper is supposed to solve.

"Because an episode is an address, time does not enter a read. An interventional distribution read after up to 1000 interleaved unrelated episodes errs by at most 0.054, inside the 0.165 band" Where do these numbers come from? From which experiment?

Then looking at the content, the paper has a very minimal contribution. Storing the context of an experiment together with the result does not resolve causal discovery or causal inference; it is just a data storage mechanism. The paper speaks about causality and experiments but does not engage with the causal inference problem. In fact, by reading the paper, it seems that once the data is collected in the proposed structure, the causal machinery reduces to standard empirical estimation.

Overall, the paper is a collection of elementary ideas without a clear and well-motivated research purpose. The resulting write-up is just not fit for human consumption.

Questions for authors

What problem do you solve? What are the baselines? How was this done in the past? What is the problem with past approaches? How does your proposal solve these issues?

Reviewer_BB49OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a parameter-free, context-indexed episodic memory for causal reasoning, using vector-symbolic representations to retrieve interventional distributions and abstain when evidence is insufficient. The authors additionally study missing-variable discovery, active intervention selection, and counterfactual identifiability. The approach is primarily evaluated on synthetic causal systems and highly structured game environments.

Critical strengths
  • The paper is clearly written.
  • The paper presents an interesting perspective on causal reasoning through episodic memory rather than learned parametric models.
Critical weaknesses
  • Limited practical motivation and overly simplified benchmarks. Most experiments are done in highly structured, hand-designed environments with explicitly defined causal variables and intervention rules. It is unclear whether these tasks present meaningful challenges to modern large-scale models.

  • Unclear advantage of the proposed structured memory. The paper does not convincingly demonstrate why the proposed vector-symbolic memory is necessary compared to simpler alternatives.

  • Limited generalization beyond predefined structures. The approach relies heavily on predefined intervention contexts, discrete variables, and deterministic dynamics. It is unclear how the proposed method would scale to more realistic environments e..g, with continuous observations, stochastic transitions, or unknown causal variables.

Questions for authors

See my comments above.

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

Related papers

Discussion (0)

Sign in to comment.