acceptodds
Under review as a conference paper at ICLR 2027

Decision-Sufficient Epistemic Evaluation: When to Redesign and When to Acquire

Abstract

AI evaluation often optimizes additional measurement before asking whether measurement is a valid repair: when a representation erases an action-relevant distinction, more data through that representation cannot recover it. We formalize this prior repair question by distinguishing semantic erasure from statistical insufficiency through two repair gates. Semantic aliasing induces unavoidable intervention regret, while preserved distinctions may remain statistically unrecoverable. Locally, decision recoverability requires and, when estimable, is governed by . After preservation, classical directional design targets acquisition along the downstream decision direction. Across 2,400 held-out controlled interfaces with hidden failure mechanism and encoder matrix, a blind gate reaches 92.6% diagnosis accuracy and lowers post-routing regret to 0.1175 versus 0.2085 for Always Acquire and 0.1710 for Always Redesign. We further train 1,800 representations without assigning semantic/statistical failure labels. Using only independent calibration separability, the gate attains regret 0.1442, below Always Acquire (0.1480) and Always Redesign (0.1648). Post-hoc decision-subspace overlap tracks the repairability crossover. Results support a contract-specific criterion for deciding whether additional measurement is a valid repair under controlled populations, rather than a universal natural-failure classifier.

Primary area:
learning theory
Keywords:
Epistemic Evaluation, Decision Sufficiency, Repairability, Semantic Aliasing, Decision-Aware Acquisition

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_SKpUOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This work investigates whether unreliable downstream decisions should be addressed by improving the representation or collecting additional measurements. For example, a representation may fail to preserve action-relevant information, or it may preserve the relevant information but suffer from insufficient measurement precision. The authors propose a Preserve–Recover–Acquire framework based on decision-relevant uncertainty and targeted acquisition. The empirical validation relies primarily on controlled simulations with synthetic data, including linear representations, where a calibration-based router reduces regret compared to strategies that always acquire additional measurements or redesign the representation.

Critical strengths
  • The paper addresses an interesting and practically relevant question: whether downstream decision failures can be mitigated by collecting additional measurements or whether the underlying representations need to be improved. This remains an important challenge for the efficient allocation of resources in practical AI systems.

  • The experiments with blind and learned representations use separate calibration and evaluation settings and measure regret after routing. This provides evidence that the proposed diagnostic can identify useful interventions in controlled settings.

Critical weaknesses
  • Significance and practical applicability. While the underlying idea is interesting, I am not sufficiently convinced that the contribution goes substantially beyond combining established concepts. The acquisition objective corresponds to classical c-optimal design, and the strongest empirical validation relies on controlled linear representations and predefined repair operators. Moreover, the improvement over an always-acquire strategy is modest. Although the paper acknowledges these limitations, I remain uncertain about the framework's broader applicability.

  • Justification for prioritizing redesign. The aliasing theorem establishes the irreducible limitations associated with discarding information. However, additional measurements could still improve estimates of the retained information and thereby reduce regret. Under a limited budget, acquisition might therefore be the preferable initial intervention, even when representation loss is present. It remains unclear under which conditions redesign should be prioritized over acquisition.

  • Clarity of presentation. A secondary concern is the presentation of the framework. Connecting the theoretical formulation to the different experimental settings requires substantial effort from the reader. A worked example explicitly linking the underlying state, representation, downstream decision, and available repair strategies would help clarify the approach.

Questions for authors
  1. Under which conditions should redesign take priority over acquisition, particularly when representation loss and measurement noise coexist under a limited budget?

  2. How would the proposed framework be applied to an existing system? What calibration data would be required, and how could the redesign intervention be specified without knowing the true decision-relevant direction?

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

The paper proposes a Preserve–Recover–Acquire framework to distinguish representation failures that require redesign from statistical insufficiency that can be addressed by acquiring more data. It develops theoretical results on decision sufficiency and decision-relevant identifiability, and uses directional optimal design to guide additional measurement. Experiments on controlled interfaces and learned representations suggest that the proposed diagnostic gate can reduce intervention regret relative to fixed repair policies. However, the evidence remains largely based on controlled settings, and the practical effectiveness of the framework on real-world AI evaluation tasks is not yet convincingly established.

Critical strengths
  1. The paper addresses a meaningful problem: collecting more data is not necessarily useful when the representation has already discarded information relevant to the downstream decision. The distinction between representation failure and statistical insufficiency provides a clear conceptual framework.
  2. The theoretical analysis connects decision-relevant identifiability with directional optimal design, and the experiments compare the proposed approach with several alternative acquisition and repair policies.
Critical weaknesses

Major concerns

  1. The main results rely heavily on controlled synthetic settings. Even the learned-representation experiments use a restricted linear representation family and a prescribed data-generating process. The real-model studies provide limited evidence that the proposed framework can reliably diagnose and repair naturally occurring AI failures.
  2. The paper primarily evaluates diagnostic accuracy and intervention regret under constructed evaluation contracts. However, the selected-repair experiments report no task recovery in either the in-domain or transfer setting (0/20 and 0/30, respectively). This weakens the claim that improved diagnosis translates into meaningful improvements in actual AI task performance.
  3. The theoretical framework assumes a specified downstream decision boundary, while the proposed acquisition rule depends on estimating its decision-relevant direction. The reported sensitivity to direction misspecification raises concerns about practical applicability when the decision rule is uncertain or changes across tasks. The paper does not sufficiently establish how these requirements can be met in realistic settings. Minor concerns
  4. The FASGL measurement framework needs stronger construct-validity evidence. The observed coupling between Support and Attribution makes it difficult to interpret the five dimensions as distinct measurement constructs.
  5. The paper is technically detailed but sometimes difficult to follow. The relationship between the general theoretical framework, the FASGL instantiation, and the different experimental protocols could be explained more clearly, with a sharper distinction between theoretical results and empirical findings.
Questions for authors
  1. How would the proposed gate operate when the downstream decision boundary is unknown or misspecified? Can the authors provide evidence of robustness beyond the controlled perturbations of the decision direction?
  2. Given that the selected-repair experiments show no task recovery, what evidence supports the claim that the proposed diagnostic framework improves practical AI performance rather than only reducing regret under the specified evaluation loss?
  3. How sensitive are the main findings to the choice of synthetic representation family, calibration sample size, and diagnostic threshold? Would the reported advantages persist under nonlinear representations and naturally occurring model failures?
Reviewer_M1jNOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

When an LLM evaluation (or "AI evaluation" as the paper refers to it) gives an unreliable result, the usual response is to collect more measurements. This paper asks whether more measurement can fix the failure at all. It separates a 'semantic ' failure and 'statistical failure'. The paper motivates this with FASGL apprach, a five-role instrument for scoring LLM claims, but the main results come from synthetic Gaussian problems where a gate routes each case to redesign or acquire, first with the encoder known, then blind, then on learned linear representations. This is a classic design optimization setup on separating aleatoric uncertainty (irreducible wth sampling for a fixed representation) and epistemic uncertainty (reducible).

Critical strengths

The idea is interesting. Checking whether measurement can repair a failure before optimizing what to measure, with FASGL's roles as a guide, reads like a sensible blueprint for designing an LLM evaluation harness.

Critical weaknesses
  1. The motivating example and the claims are not tested sufficiently. The introduction is about LLM evaluation pipelines and FASGL, but every positive result comes from a toy example. The studies on real models (Gemma, Qwen, Mistral) in the appendix and are also not really testing the main claim, they are more on understanding if FASGL is even capable of breaking down the errors. I appreciate that these are reported, but it means the framework has not been shown working on an LLM. They are mostly negative results, the FASGL scoring doesnt actually work, the diagnosis localization is marginally better but scores 0% on clean recognition etc. Im not very sure what to make of the results.

  2. The paper is very unclear, the writing is very poor, confusing, and vague. It needs significant work before it can be published. The paper does a very poor job of defining terms, that are vague and imprecise-- representation, state, family, etc. The introduction never gives a concrete example of a semantic or a statistical failure, and FASGL's five roles are only defined in an appendix.. "Epistemic" refers to the epistemic state of a claim, not epistemic uncertainty, which is extremely confusing considering this is about design optimization, where epistemic uncertainty is a core concept. I still dont have a clear understanding of what the "representation" is. In the first gate it seems to be what the evaluator records about a single LLM answer. The second gate is instead about estimating the model's overall skills from many test questions. The experiments use neither, just a random vector multiplied by a matrix.

  3. As far as i can tell the theory is not novel, the redesign-versus-acquire split is the representation-relative aleatoric/epistemic view already set out by Hullermeier & Waegeman (https://arxiv.org/pdf/1910.09457). The aleatoric/epistemic literature is never cited which is a very glaring omission.

The idea of a harness that first asks whether measurement can repair a failure, and only then what to measure, is nice and would be useful to the community. But the paper is very hard to follow, the implementation is a linear toy that does not match the LLM motivation in the introduction, and the real-model evidence is effectively null. A concrete worked example, positioning against the aleatoric/epistemic literature, and one real LLM evaluation where the gate picks the repair that helps would make this much stronger.

Questions for authors

N/A

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.