acceptodds
Under review as a conference paper at ICLR 2027

Which Tasks Identify a Representation? Specificity Bounds and Task Acquisition

Abstract

Task structure can rule out latent mixing without making a representation easy to estimate. We study this distinction for finitely many outcomes with known overlapping supports and injective conditional signatures. Building on existing support-separation and partial-view principles, we characterize allowed mixing and derive a fractional-cover bound on block specificity. The bound combines task conditioning with signature error, avoids repeatedly charging the same excluded directions, and is optimal given its local variance constraints. A smooth signature example attains its constant and improves the coordinatewise bound by a factor 2(B-1) for (B) blocks. We then evaluate acquisition at a fixed sampled-label budget. In a Gaussian-signature comparison, a pilot-conditioned policy improves shifted accuracy over uniform candidate acquisition, and an information-matched inner-Lasso adaptation nearly matches the proposed spectral estimator. An invertible nonlinear model gives a more cautious result: a frozen 20-seed follow-up estimates a 2.52-point shifted-accuracy gain, with a 95% interval spanning zero. A separate 240-run composite-image benchmark exposes further failures outside the theorem. Together, the results distinguish structural separation, conditioning, and finite-budget utility, with released implementations, complete negative results, and explicit limits on practical certification.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Representation learning, identifiability, multitask learning, task acquisition, fractional cover, specificity bounds, partial observability

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_M3TDOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies which task-support patterns identify block-specific latent representations and whether acquiring more tasks helps under a fixed labeling budget. It characterizes the remaining mixing ambiguities and introduces a fractional-cover bound that combines task errors, conditioning, and support overlap. The paper also proposes linear and nonlinear estimators. Experiments show acquisition benefits in a controlled Gaussian setting, while nonlinear and image results are less conclusive.

Critical strengths
  1. The fractional-cover bound goes beyond a binary identifiability result. It combines overlapping task constraints without repeatedly charging shared nuisance directions. LP duality establishes optimality within the local-variance relaxation, and a constructed example shows a concrete improvement over the coordinatewise bound. This helps explain how support geometry and conditioning affect recovery after identifiability is established.

  2. The experimental comparisons are carefully controlled. The authors match labeling budgets and support information, include a baseline that nearly matches PATI, and report inconclusive and negative results. The nonlinear follow-up and image benchmark make the limits of the current method visible rather than presenting only favorable outcomes.

Critical weaknesses
  1. The experiments do not clearly show the added benefit of fractional-cover aggregation. The identifiability results build on prior work, so the significance of the paper depends heavily on this quantitative refinement. Yet its advantage is mainly illustrated by constructed examples, without a direct empirical comparison with the coordinatewise bound in Theorem 3. Figure 1(d) also reports a median bound-to-observed-variance ratio of 6.31, suggesting considerable slack. The acquisition experiments mostly show the value of separating tasks and choosing better-conditioned candidates. Beyond the controlled linear setting, the nonlinear gain has a confidence interval containing zero, and the image experiments do not establish consistent improvements.

  2. It is unclear how the bound could be used with real tasks. Theorem 4 requires population signature errors and inverse-Lipschitz constants, which are not directly available from labels. The synthetic experiments use access to ground truth to estimate these quantities. Although the paper acknowledges this limitation, it leaves unclear how a practitioner could compute a reliable bound or use it to guide task acquisition.

Questions for authors
  1. Can you compare the fractional-cover and coordinatewise bounds on trained representations across different overlap patterns, while controlling for identifiability and task conditioning? This would help isolate the contribution of fractional-cover aggregation from the benefits of support separation alone.

  2. Under what assumptions can the signature errors and inverse-Lipschitz constants be estimated or upper-bounded from pilot or validation data? Even a conservative data-dependent bound with a stated confidence level would help. If additional assumptions are essential, please explain what information is missing from labels alone.

  3. Do the acquisition gains persist when candidate tasks have similar conditioning or fewer pilot samples are available? In Section 6.1, the separating candidates have substantially different strengths (β=0.1\beta=0.1 and β=1\beta=1), and the policy always selects the stronger one. Tests with other support patterns and budget-to-dimension ratios, under matched total budgets, would clarify how broadly this result applies.

  4. What causes the shifted-accuracy decline in the six deteriorating nonlinear follow-up seeds: optimization failure, poor inverse conditioning, or remaining within-block ambiguity? Diagnostics for individual seeds would help explain the limits of PATI-Flow.

Reviewer_ASrrOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary
  • In this paper, the authors investigate representation identifiability in multitask learning, focusing on how overlapping task-support structures constrain latent-factor mixing and how these structural conditions relate to representation specificity and task acquisition. Building on existing support-separation and partial-view identifiability principles, the paper characterizes admissible representation ambiguities and establishes conditions for universal block identification.

  • I think the main contribution of this paper is a fractional-cover linear programming formulation that bounds nuisance variance using task-specific estimation errors and inverse-conditioning constants, together with an optimality result for the corresponding variance relaxation. The paper also presents spectral and nonlinear estimators and evaluates task acquisition under fixed labeling budgets through Gaussian, nonlinear normalizing-flow, and composite-image experiments.

  • While the results demonstrate the theoretical role of task-support structure and conditioning, the empirical findings show substantial limitations in bound tightness and the reliability of finite-budget acquisition improvements beyond controlled Gaussian settings.

Critical strengths
  • The paper provides a mathematically structured treatment of representation specificity by connecting overlapping task supports, inverse conditioning, and nuisance-variance control within a unified framework.
  • The fractional-cover formulation offers a useful quantitative refinement of coordinatewise bounds and establishes an optimal variance relaxation under explicitly stated assumptions.
  • The theoretical discussion carefully distinguishes structural identifiability from finite-sample recoverability, while the empirical evaluation includes information-matched baselines and controlled experiments across several settings.
Critical weaknesses

There are some weaknesses for which I need more clarification from the authors:

  • A major concern is the limited conceptual novelty of the proposed identifiability results. As explicitly acknowledged in the manuscript, the antichain support condition underlying Theorem 1 and Corollary 2 is equivalent to an existing assumption in Lachapelle et al. (2023), while the intersection-based identification principle builds on previous partial-view identifiability results. Consequently, the main original theoretical contribution is concentrated in the fractional-cover specificity bound and its optimal variance relaxation.
  • Although this refinement is mathematically meaningful, it primarily combines established variance-decomposition techniques and linear programming duality under restrictive structural assumptions. The paper does not sufficiently demonstrate that this quantitative extension represents a substantial conceptual advance over existing approaches to representation identifiability, particularly given the limited evidence of practical benefits.
  • The central theoretical guarantees rely on strong assumptions, including invertible observation mappings, perfectly known task supports, injective conditional signatures, controlled inverse Lipschitz constants, and independence or bounded density-ratio conditions. These requirements substantially limit the applicability of the framework to realistic representation-learning problems involving noisy observations, unknown factor-task dependencies, and potentially ill-conditioned neural mappings.
Questions for authors

I have some questions for the authors:

  • Given that the antichain identifiability condition and the underlying support-intersection principle are inherited from previous work, what is the fundamental conceptual advance of the proposed framework beyond applying fractional-cover optimization to existing variance constraints?
  • Under what practically verifiable conditions can the proposed fractional-cover bounds remain sufficiently tight to guide representation auditing or task acquisition? In particular, how would the framework handle estimated or partially incorrect task supports, unknown inverse-conditioning constants, and finite-sample signature errors?
Reviewer_9hTWOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies which overlapping tasks identify latent blocks and how task conditioning controls residual mixing. Its main addition is a fractional-cover specificity bound, followed by budget-matched task-acquisition experiments with positive linear results but inconclusive nonlinear benefits.

Critical strengths

The fractional-cover bound is a neat quantitative refinement: overlapping tasks can control nuisance variation without repeatedly charging the same directions. The matching variance construction gives the result some bite.

I appreciate the honest experiments. Matching support information largely closes the baseline gap, and the frozen nonlinear follow-up retains negative results rather than overselling identification as practical recovery.

Critical weaknesses

Practical use remains limited: the bound needs known supports, inverse-conditioning constants, and population signature errors. The acquisition policy uses a conditioning proxy, so the experiments do not directly establish the value of the new fractional-cover criterion.

Evidence beyond the controlled linear setting is weak. The nonlinear gain’s interval spans zero, image improvements are inconclusive, and the clinical results concern different models. These experiments establish boundaries more convincingly than practical utility.

Questions for authors

Can the fractional-cover criterion change acquisition decisions relative to the conditioning-only proxy, and improve outcomes at the same budget?

How robust are those decisions to errors in supplied supports or estimated conditioning constants?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.