acceptodds
Withdrawn from ICLR 2027

Task-Completed Identifiability: Recovering Task-Relevant Factors for Selective Use

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Representations can preserve all information needed for prediction while still mixing the latent factors that a task should use separately. A particularly important case arises when environmental variation alone cannot distinguish task-relevant factors, even though task responses provide complementary constraints on the same latent structure. We propose Task-Completed Identifiability (TCI), which formalizes when these two information sources jointly remove such relevant-factor ambiguities within a shared generative model. We derive the Task-Relevant Rank Condition (TRC), a projected rank condition on the first- and second-order derivatives of environmental and task-response contrasts. Under the stated assumptions, TRC identifies task-relevant factors and the common task–factor correspondence up to permutation and factor-wise invertible transformations, without requiring individual identification of factors irrelevant to all tasks. We further construct settings in which either information source alone admits genuine mixing among relevant factors, whereas their combination resolves it. To estimate the shared model, we introduce TC-iVAE, which couples a response-free inference branch with task-conditioned variational branches through shared latent coordinates. A composite-ELBO decomposition connects exact observable-law fit and posterior approximation to the population identification result. Experiments show that task responses can resolve relevant-factor mixtures left unresolved by environmental variation, while preserving predictive information. The resulting representations support more selective factor-wise editing, with substantially reduced changes in non-target factors. Together, these results distinguish preserving task information from identifying the factors needed to use that information selectively.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Representation learning, Identifiability, Disentanglement

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_L1oYOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary
  • In nonlinear ICA settings, it is not always possible to decompose and extract latent factors per task even if all task-relevant information is preserved in latent representations. This paper shows that a new condition is required for the decomposition and extraction.
  • Proposed Task-Relevant Rank Condition (TRC): a sufficient condition to decompose and extract task-relevant factors (up to invertible transformations) for a model whose task responses are formulated as an exponential family dist. with factor-additive natural parameters. After collecting the log density ratios (contrasts) of environmental variation and task responses, compute the ranks of their derivatives. The derivative directions of task-relevant factors exhibit no rank degeneracy and no intersection with the task-irrelevant factors.
  • Theory is presented in Sec. 3. Sec. 4 presents how to conduct an analysis via a composite ELBO optimization.
  • Experiment: Synthetic experiments effectively show that the existing baseline (iVAE(U,Y) etc.) cannot separate and extract the factors required for tasks even if the information for the factors is preserved (rel. R^2≈1). On the other hand, the latent model learned by the TC-iVAE successfully achieves well-separated factor extractions.
Critical strengths
  1. Clear claim and supporting result
  • Figure 1 clearly depicts the difference between related model families and the model assumed in this work.
  • Experimental results clearly show the importance of distinguishing the information preservation and the factor reconstructions.
  1. Reasonable compromise
  • Leaving task-irrelevant factors unidentified is practically unproblematic. Thus, I think this partial-identification design is a good compromise.
Critical weaknesses
  1. The experiments are not strong.
  • In 5.1, only the mixing matrix is unknown. This amounts to a sanity check that the model behaves as the theory assumes, and the other methods naturally fail exactly as designed. In fact, experiments in 5.2 (the model does not know g) observe that the gap in editing leak shrinks substantially, and the variance of the irrelevant R^2 becomes larger.
  • It is natural to assume that real-world data violate the background model assumptions. And one cannot verify whether TRC holds in his/her own data. Therefore, I think it is important to present practical limitations (how robustly the method works in that regime) for motivating the readers to try it.
  • The experimental systems are small-scale in factor counts, dimensions, and number of tasks. The authors explicitly state the computational non-scalability as a limitation, but this leaves serious doubt about operability in large-scale systems where people want to perform selective factor extractions.
  1. In Sec. 4.3, too many regularizers are introduced.
  • Note that these terms are not needed by theory, and they are not derived from TRC. These are implementation choices. This causes two problems. First, it is non-trivial to choose hyperparameters for these regularizers. Second, the current manuscript does not provide evaluation concerning the sensitivity of these regularization terms to lambda. So one cannot disentangle the effect of TC-iVAE itself from the effects of regularizers.
Questions for authors
  • Is there a procedure to verify TRC from data (even approximately)?
  • Is it possible to provide stability results characterizing how much the recovery quality degrades when TRC is not exactly satisfied (approximate TRC)?
  • What is the individual contribution of each regularization term, and how sensitive is the performance to lambda? How difficult is tuning lambda in practice?
  • What happens if we cannot fully observe task responses due to random missingness or structurally partial observability.
Reviewer_WbL4OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper addresses the limitation that existing representations may perform well on prediction tasks but fail to support selective control over individual factors. To distinguish predictive information from separately identifiable factors, the paper constructs two types of log-density contrasts. Intuitively, after eliminating directions associated purely with nuisance factors, each task-relevant factor must contribute distinguishable derivative directions, allowing these factors to be identified separately.

Critical strengths

The paper provides a clear and mathematically well-motivated formulation of the distinction between information preservation and selective factor identification

Critical weaknesses

Lack of clarity on practical applications: The paper would benefit from a more detailed discussion of how the proposed framework can be applied in practice. In particular, the authors should provide concrete examples of scenarios where identifying task-relevant factors separately offers advantages over conventional predictive representations.

Insufficient distinction from prior work: The motivating framework and learning paradigm substantially overlap with prior work, particularly C-disentanglement (https://arxiv.org/abs/2310.17325), although the theoretical results appear distinct. C-disentanglement already addresses latent non-identifiability, introduces observed labels as inductive information, and develops a conditional VAE-based framework for disentangled representation learning. The authors should clarify how TCI differs from this framework and, more importantly, what new mathematical or algorithmic capabilities it provides. Specifically, what would happen if the confounding label (C) in C-disentanglement were replaced with TCI's environment variable (U) and task response (Y)? What identifiability guarantees or capabilities would TCI provide beyond those already formulated within weakly supervised conditional disentanglement? A more rigorous comparison is necessary to establish the novelty of the proposed framework.

Limited experimental evaluation: The experimental design is insufficient to demonstrate the practical value and effectiveness of the proposed framework.

  1. Lack of realistic benchmarks and downstream applications: The paper does not evaluate the method on real-world datasets, image benchmarks, or natural downstream tasks. Consequently, it is difficult to assess whether the theoretical advantages translate into meaningful practical improvements or identify scenarios where the proposed method would be useful.

  2. Outdated and incomplete baseline comparisons: The experimental evaluation primarily compares against relatively old baseline methods, without including more recent identifiable representation learning or disentanglement approaches from 2025–2026. Although the related work section cites several more recent studies, these methods are not included in the experimental comparisons. The authors should explain these omissions and provide comparisons with more competitive baselines to better establish the advantages of the proposed approach.

Questions for authors

See above

Reviewer_ctasOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
3
Read review
Paper summary

In this paper, the authors aim to identify task-relevant latent factors by combining complementary information from environmental variation and task supervision, in settings where neither source alone is sufficient to identify these factors.

Critical strengths

The authors address an interesting question concerning the identifiability of task-relevant latent factors when environmental variation and task supervision provide complementary information. However, I found the paper very difficult to follow.

Critical weaknesses
  • Clarity and presentation: The main issue with the paper is that it is extremely difficult to follow. The text is overly compressed, and even relatively simple ideas are expressed in an unnecessarily convoluted way. For example, consider the following passages:

    “A central question is what happens when one information source leaves mixing among task-relevant factors unresolved, but another source constrains different directions.”

    “TRC requires relevant factor directions to remain distinguishable after quotienting out the column space of factors irrelevant to all tasks.”

    “Thus task information need only eliminate the environment-permitted directions that still involve relevant coefficients; ambiguity confined entirely to nuisance coefficients may remain.”

    These passages require multiple readings to grasp, even though the underlying ideas could be expressed much more simply. While the terminology may be legitimate, its makes the paper unnecessarily difficult to read rather than helping the reader understand the underlying concepts.

  • Positioning in the literature: I found the paper to be poorly positioned relative to existing work. In the second paragraph of the introduction, the authors mention several relevant studies without discussing their limitations, then jump directly to their research question without clearly explaining how it emerges from the existing literature or what gap it aims to address. In addition, there is no dedicated related work section, which would have been particularly necessary given the insufficient discussion of prior work in the introduction.

  • Experimental evaluation and practical relevance: Although the experiments verify several intended properties of the proposed method, including through a negative control, I find the evaluation rather weak. The experiments are conducted in synthetic settings specifically designed around the method's assumptions. Yet, even under these favorable conditions, the results in the vector-factor setting are not entirely convincing. The improvements over the strongest baseline are modest relative to the variability across seeds (factor R²: 0.740 ± 0.184 vs. 0.637 ± 0.141 for bilevel sparsity; target error: 0.496 ± 0.345 vs. 0.707 ± 0.263). Even if we overlook the high variability in the results, the practical relevance of the proposed framework remains unclear. This issue is closely related to the paper's insufficient motivation and positioning. I would have liked to see at least a concrete example or realistic setup illustrating where the proposed framework would be relevant and useful in practice.

Questions for authors
  • Main objective and positioning: Given how difficult I found the paper to follow, I may not have fully grasped its scope. Could the authors clarify the main goal of the paper, which specific gap in the existing literature they aim to address, and how their contribution differs from prior work?

  • Practical relevance: Could the authors explain why they do not provide a concrete real-world example or a realistic experimental setting illustrating where the proposed framework would be relevant and useful?

Table 2 presents the results of the vector-factor experiment discussed in Section 5.2, but it is not explicitly referenced in that section.

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

Related papers

Discussion (0)

Sign in to comment.