acceptodds
Withdrawn from ICLR 2027

Causal Orbitals: Order-Sensitive Probing for Certifiable Test-Time Adaptation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Test-time adaptation (TTA) typically assumes that an unlabeled target batch provides sufficient evidence for deciding how a model should adapt. However, distinct causal shift mechanisms can produce nearly indistinguishable observations while inducing incompatible task-risk gradients, making adaptation fundamentally ambiguous. We study meta-trained, label-free TTA under causal mechanism uncertainty and ask whether additional test-time measurements can make a safe adaptation direction identifiable. We introduce Causal Orbitals, an operator-valued representation of unresolved causal mechanisms, and model self-supervised test-time probes as sequential measurement instruments that update this ambiguity state. Unlike a single-shot mechanism posterior, the proposed formulation captures order-sensitive probe responses, where applying the same probes in different sequences can yield different evidence about the underlying shift. We derive spectral certificates that determine when a candidate update guarantees local risk reduction and abstain when no common descent direction can be certified. Our analysis characterizes the classical boundary of the framework: it reduces to an ordinary posterior under commuting measurements, while selected noncommuting probe sequences can distinguish mechanisms that share identical single-probe marginals. We further develop certifiability-directed probing, which selects probe sequences according to their expected guaranteed adaptation value rather than generic uncertainty reduction. We evaluate the framework on controlled causal-shift benchmarks and realistic domain shifts, measuring sequential identifiability, certificate validity, negative adaptation, certified coverage, and probe efficiency.

Primary area:
learning theory
Keywords:
uncertainty

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 4 reviews, above 27% of the others.
Reviewer_tvgNOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces Causal Orbitals, a framework for mitigating negative adaptation in test-time adaptation through active, order-sensitive diagnostic probing. The method combines offline-calibrated mechanism gradient prototypes with a classical–operator state that tracks mechanism uncertainty and sequential probe responses. Its central geometric result characterizes when a robust common descent direction exists: the gradient uncertainty hull must exclude the origin; otherwise, the model abstains from updating. A Bellman planner selects probe sequences to maximize expected certifiable improvement, while split conformal calibration controls the marginal probability of harmful updates on empirical batch risk under exchangeability assumptions. Across synthetic causal shifts, ImageNet-C-style corruptions, and simulated MedMNIST environments, the method reports 81.7% certified adaptation coverage with a 2.1% joint unsafe rate, improving coverage by 2.9 percentage points over a matched GRU baseline.

Critical strengths
  1. A principled safety criterion for adaptation. The paper provides a clear geometric test for when an update is certifiably beneficial: the robust hull of plausible mechanism gradients, Γ_δ(π), must exclude the origin. When this condition fails, the method abstains rather than committing to an unsupported update. This matters because it directly targets negative adaptation, and the paper further supplies finite-sample, marginal risk control under explicitly stated exchangeability assumptions.

  2. A coherent link between active probing and coverage gains. The method treats probe order as informative and uses Bellman lookahead to select diagnostic sequences, thereby connecting evidence gathering to the decision to adapt. In the reported experiments, it attains 81.7% certified coverage at a 2.1% joint unsafe rate, and improves coverage by 2.9 percentage points (pp) over a matched GRU baseline. Matched comparisons and ablations support the claim that ordered evidence tracking and planning each contribute to these gains.

Critical weaknesses
  1. The safety guarantee is narrower than the deployment-facing claims suggest. The conformal result controls marginal empirical cross-entropy risk across exchangeable deployment episodes; it does not certify that any individual update is safe, nor does it bound population risk or classification error directly. Moreover, the paper itself notes that the conditional unsafe-risk bound can reach α/c\alpha/c, where cc denotes adaptation coverage—about 6.1% at the reported 5% nominal level and 81.7% coverage. This gap matters for safety-critical use: rare harmful updates, or failures under non-exchangeable shift, can be consequential yet are not excluded by the stated guarantee.

  2. The experiments provide limited evidence for generalization to realistic continual deployment. The evaluation relies on a finite set of candidate shift mechanisms, mechanism prototypes calibrated offline, and episode-level testing that resets the model to its source parameters. Although the benchmarks include ImageNet-C-style corruptions and simulated environments, they do not establish that the approach remains effective in an ongoing real-world stream subject to evolving or unmodeled shift. This matters because the paper motivates the method by open-world distribution change—a setting broader than the reported evaluation covers.

Questions for authors
  1. Clarification — Scope of the safety claim. The conformal guarantee is stated for marginal empirical cross-entropy risk across exchangeable deployment episodes, with the model reset between episodes. What evidence supports exchangeability in the evaluated settings, and how should the guarantee be interpreted under a continuously evolving, non-exchangeable stream? Please also clarify how the reported certified coverage and unsafe-adaptation rates are defined relative to that guarantee.

  2. Clarification — AI use statement. The official ICLR 2027 submission policy requires authors to explicitly disclose any use of AI in the research process and in manuscript preparation. The current manuscript, however, contains no AI usage statement. Please consult the official ICLR 2027 AI Policy for Authors and add a clear disclosure stating whether any AI tools were used. If they were, please specify the extent of their involvement—for example, in manuscript preparation, experimental design, data analysis, or other stages of the research.

  3. Suggestion for future work — Matched evaluation and attribution. A useful follow-up would compare Causal Orbitals directly against REM under a shared backbone, test stream, adaptation budget, and evaluation protocol, reporting both performance and safety/coverage measures. Ablations isolating the contributions of ordered probing, operator-state tracking, and Bellman lookahead would further clarify which components drive any observed gains.

Reviewer_rZ2UOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper addresses negative adaptation in test-time adaptation caused by ambiguous distribution shifts with conflicting risk gradients. It proposes Causal Orbitals, combining robust gradient geometry, order-sensitive operator representations, Bellman probing, and conformal safety calibration. Experiments are conducted on synthetic causal shifts, ImageNet-C-style corruptions, and simulated MedMNIST benchmarks. The method reports 81.7% certified coverage and a 2.1% joint unsafe adaptation rate.

Critical strengths
  1. The paper addresses an important problem in TTA by explicitly considering negative adaptation and introducing abstention when safe updates cannot be established.

  2. The framework combines several theoretically motivated components and includes comparisons with matched-supervision baselines and ablation studies.

Critical weaknesses
  1. Questionable theoretical soundness. Appendix D.5 incorrectly suggests that orthogonal gradients imply the origin lies in their convex hull, contradicting the proposed descent characterization. Furthermore, the finite conformal threshold reported for 15 calibration episodes at α = 0.05 (Appendix D.2) contradicts Theorem 7, which requires an infinite threshold in this setting. These issues raise serious concerns about the theoretical interpretation and safety claims.

  2. Inconsistent experimental reporting and insufficient evidence for the core contribution. The hyperparameters in Table 10 conflict with the main experimental setup, while Appendix D.2 reports accuracy and latency values inconsistent with Table 1. Moreover, the accuracy improvement over the matched GRU baseline is only 0.7 percentage points and is statistically insignificant (p = 0.080). Thus, the empirical contribution of the operator representation remains insufficiently established.

Questions for authors
  1. Could the authors clarify the gradient geometry in Appendix D.5, particularly why orthogonal gradients would imply origin containment in their convex hull?

  2. How is a finite conformal threshold obtained with 15 calibration episodes at α = 0.05 under Theorem 7?

  3. Could the authors reconcile the conflicting experimental configurations and results, and provide stronger evidence isolating the contribution of the non-commutative operator representation?

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

The paper proposes Causal Orbitals, a framework for preventing harmful test-time adaptation under distribution shift.

Critical strengths

The paper’s central technical contributions fall outside my expertise, and I cannot reliably assess their correctness or novelty.

Critical weaknesses
  • There is no AI use statement in the paper.

  • The paper’s central technical contributions fall outside my expertise, and I cannot reliably raise concerns about the paper's weaknesses.

Questions for authors

N.A.

Reviewer_Ah2KOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how to avoid harmful test-time adaptation when similar-looking shifts may require different updates. Causal Orbitals uses ordered diagnostic probes and offline mechanism knowledge to decide whether to adapt or retain the source model. It combines a gradient-based update rule with a calibrated safety gate and evaluates the approach across synthetic and image benchmarks.

Critical strengths
  1. The paper addresses an important problem: adaptation can hurt performance, and choosing not to update should be a valid option.

  2. The evaluation includes matched-supervision baselines and component comparisons, helping separate the effects of probe planning and the operator representation.

Critical weaknesses
  1. The method requires knowing the possible shifts in advance and using labeled data to learn their gradients and probe responses. This is a strong assumption. Its usefulness when deployment shifts fall outside the mechanism library remains unclear.

  2. “Certifiable safety” may give an overly strong impression. The theory assumes exchangeability between calibration and test environments and does not directly guarantee that accuracy will not decrease. The paper reports unsafe rates above 20% among deployed updates when this assumption is severely violated.

  3. The connection between the impossibility theorem and the proposed method is not sufficiently direct. The authors acknowledge that transforming the same inputs cannot distinguish environments with identical input distributions. The method instead relies on offline knowledge and distinguishable shifts, so the theorem does not establish why ordered probing is necessary.

  4. I am not fully convinced that the operator representation is needed. Order-dependent transformations do not necessarily require density operators; classical sequence models can also capture such effects. The comparisons support its effectiveness in the evaluated settings, but not broader claims about its representational advantages.

  5. The experiments remain fairly controlled. They mainly use four predefined mechanism families and short probing sequences, with the source model restored at every episode. Evidence for unknown shifts, mixed shifts, and accumulated updates in continuous deployment is still limited.

Questions for authors

See Weaknesses.

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

Related papers

Discussion (0)

Sign in to comment.