acceptodds
Withdrawn from ICLR 2027

DiSCO-SR: Residual-Guided Symbolic Selection under Predictive-Risk Constraints

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Predictively similar symbolic equations can imply different intervention responses. DiSCO-SR tests residual-guided equation selection within an explicit predictive-risk budget, separating equation choice from causal direction decisions. Separate data splits support fitting, selection, and direction decisions, with no refitting of the selected equations. We derive conditional guarantees for population risk and penalized finite-block residual dependence, and bound mean-intervention loss under correct orientation and bounded distribution shift. In independent confirmation on 1,920 synthetic cases from 240 mechanism clusters, the development-selected paired-stability rule achieves 95.21% direction accuracy and 0.1396 mean-intervention NRMSE, compared with 95.31% and 0.1385 for minimum-risk selection with the same direction rule. None of eight prespecified primary contrasts is significant after multiplicity correction. The stability rule changes 29.97% of directional equations, correcting four directions and corrupting six; its certificate counterpart always returns the minimum-risk reference across 2,916 confirmation cases. Matched regression controls attain stronger point estimates, and a separate bounded-score study exhibits large errors on measured physical interventions. These results provide an auditable framework for constrained symbolic selection, but do not establish an incremental mechanistic benefit from residual guidance.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
symbolic regression, causal direction, intervention prediction, predictive-risk constraints, equation discovery, residual-guided selection

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

This paper studies whether residual independence, the signature of a causal mechanism under an additive noise model, can be used to select a better symbolic equation among candidates that predict roughly equally well. The method, DiSCO-SR, fits a finite library of symbolic candidates and restricts attention to candidates whose selection-set risk lies within a small band of the minimum-risk candidate, and among those picks the one whose residuals are least dependent on the input, as measured by HSIC. They provide conditional guarantees on population risk, and evaluate across three pre-registered synthetic campaigns, the Tübingen pairs and a physical testbed (Causal Chamber).

Critical strengths

The experiments are quite thorough, and the authors matched-gate regression baselines separate the effect of equation selection from that of the direction rule. Proposition 2 (the wrong-direction intervention floor) and the direction/equation attribution identity are intuitive and nicely presented. The authors are also commendably honest that their method does not outperform the minimum-risk reference.

Critical weaknesses

I’m struggling with what the central claim of the paper is. The motivation is explicitly causal: equations that predict equally well can imply different intervention responses, so finding the equations with some sense of causal semantics will greatly aid generalisation. Yet the paper repeatedly states that its procedure provides no causal certificate and no evidence of a correct mechanism. If the method’s value is not causal, it’s hard to see why one would care about directionality at all, since predictively Y|X and X|Y are equally valid models of the joint. Can you help me understand who the "auditable interface" is for, and what decision it supports, if not causal mechanism recovery? My suggestion would be to state the causal claim once, conditionally on an identifiable two-node ANM without confounding, and evaluate against that.

My reading is that the null result is to be expected, and I would like to understand whether the authors agree. Under an ANM, a candidate’s residual is N + (f* − f)(X). Both its excess risk and its residual dependence on X are driven by the same term (f* − f)(X). Inside a tight risk band this term is constrained to be small in L^2, leaving HSIC little room to discriminate between candidates, particularly with only 50–125 selection rows. If this reading is correct, a discussion of it in the paper would help provide a clear explanation of the finding.

Relatedly, the finding I find most interesting seems to me to be one the authors underplay. Joint+gate, which applies residual dependence without the risk band, significantly improves OOD error in both V4 and V5. This suggests that residual guidance may help precisely where the risk constraint is removed, which is where extrapolation behavior differs. Could you comment on this, and on whether a design centered on OOD performance would be the more natural test of the hypothesis?

In the expriments, the gate never abstained on 160 nonidentifiable or 180 assumption-violating cases. Given that the causal interpretation rests on these assumptions, this deserves more than a single sentence.

If these points could be addressed adequately I would raise my score.

Questions for authors

The paper is quite hard to read. Almost every paragraph ends with a qualifier ("this is not a causal certificate"), and the density of these makes it difficult to find what is actually being claimed. The abstract is largely a list of numbers. The interleaving of three versioned campaigns (V4, V5, V6) with different selectors and locked parameters makes the paper tough to read at points, and I found myself constantly referring back to Table 4. A single clearly stated hypothesis, one primary campaign in the main text, and the remainder in the appendix would, in my view, improve readability considerably.

Notation is also overloaded in places: b denotes both the risk-reference index and the number of blocks, B both the loss bound and the block count, and L the log term, the loss and the number of bootstrap draws.

Reviewer_zzNPOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The writing of this work is so difficult to parse as to preclude the possibility of really understanding its goals. After reading it, I am not able to properly speak about the content of the work, despite having an extensive background in the related area of causality as well as on the differences between prediction and intervention. I strongly recommend a rewrite of the paper and resubmission after it has been restructured. I have detailed below several examples of where and how the writing lacks clarity and is too vague to follow.

Critical strengths

There appears to be a good foundation of ideas upon which to build.

Critical weaknesses
  • In my view, writing and presentation is a critical issue for this work, to the degree where it is often difficult to even parse the context (and more importantly, the significance) of what is being stated. Perhaps one way to describe it: the document is formatted as if it is to be read all at once (maybe by a language model?) rather than read in a linear fashion and understood along the way. There are several examples of this throughout, but I’ll document a couple in the subsequent bullets.

  • Much of the language is a terse summary of things that have not yet been defined or in any way contextualized. To make this concrete, I’ll highlight one example of this — the third paragraph of the paper — along with the entire list of questions I have specific to this one paragraph. (There is no need for the authors to answer these questions in their rebuttal — my point here is just to demonstrate the way in which the writing is too vague, leaving me with far too many questions after reading it.)

    • “DiSCO-SR makes these distinctions executable. Disjoint samples fit the finite library, select equations under an observed-risk cap, and decide direction from residual evidence with a regression-error fallback. Independent queries evaluate the returned coefficients without refitting. The risk safeguard holds on selection observations. Its population interpretation needs additional assumptions, and neither version certifies causal truth or extrapolation.”

      • What does it mean for samples to be "disjoint" in this context? Is this simply 'different' samples?

      • What is "the finite library"? I might deduce this is a library of symbolic equations, but this has not been stated. Are there cases where we might expect an infinite library, perhaps if we consider a continuous parameter space?

      • How is "residual evidence" interpreted? Is this just simply residuals themselves? What specifically is meant by a "regression-error fallback," and how does this differ from the aforementioned "residual evidence"?

      • What in this context is a "query" and what would be a dependent vs. independent query?

      • What precisely does it mean for a "risk safeguard" to "hold on selection observations"? 

      • There is a claim that "its population interpretation needs additional assumptions." What is this population interpretation? What are the additional assumptions being referred to here?

      • What is the significance of the statement "neither version certifies causal truth or extrapolation"? What specifically is meant by "causal truth" here? I presume this refers to a causal direction between a given pair of variables for which we are fitting/choosing a symbolic mechanism? With the term "extrapolation," what is it that is being extrapolated? Is this extrapolation of the functional form outside the observed range?

    • Parsing through the entire paper and then returning back to this paragraph, it is possible to find answers to some if not hopefully all of these questions. But the problem here is that this paragraph comes at the beginning of the paper — before much of anything has been introduced. And this same pattern of writing continues throughout the manuscript.

  • Many mathematical expressions are not properly defined before they are used (starting from Equation 1 onwards).

  • The field of causal discovery and/or structure learning is not discussed as a whole — which appears to be a big part of the entire context for this work. This should be added more thoroughly to the positioning and related work, and could help with the framing of the procedures themselves.

Questions for authors

N/A

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

The authors attempt to address the question of discovering the direction of causation between two variables as well as a symbolic representation of their relationship as a function of a set of fixed dictionaries of functions.

Critical strengths

Causal discovery is a well-established hard problem.

Critical weaknesses

The paper is unreadable as currently written; the question that the authors address is not clear what exactly it is; at the level that it is understandable, it is not clear why it is sufficiently motivated. The authors make particular choices that are un-substantiated (as outlined in the questions below) and hence the paper does not seem worthy of publication.

Questions for authors
  1. Section 3: what do you mean "executable equation"?
  2. Section 3: what do you mean by "A causal interpretation requires a two-node, causally sufficient ANM, an adequate retained representation, and directional identifiability."; what is "causal interpretation", what does it mean to "require", what is an "adequate retained representation"?
  3. What is a "score gap"?
  4. What is a "finite candidate selection"? What candidates are you talking about?
  5. Why do you define the constrained rule (4) a whole page after you use it at the beginning of section 3? What do you mean by "gate" in "gate below"? What is a "gate"?
  6. Why are you using this particular dictionary in equation (2)? What is special about these primitives?

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

Related papers

Discussion (0)

Sign in to comment.