acceptodds
Under review as a conference paper at ICLR 2027

ECRL: Explainable Causal Representation Learning via Semantic Anchoring of Exogenous

Abstract

Understanding dynamical systems requires not only accurate prediction but also representations that reveal which external mechanism drives each state change, enabling mechanism-level explainability. Yet predictively equivalent causal representations may assign different latent blocks to the same named mechanism, leaving the semantic meaning of block-level interventions unresolved. We study this semantic orientation ambiguity using anchor environments that reveal which exogenous block changed, but not its realized value. Under affine residual ambiguity, known block dimensions, and block-spanning mean shifts, we show that off-target response bounds cross-block mixing, with zero response identifying named blocks up to within-block invertible affine transformations. We derive a finite-sample certificate and extend the result to exact temporal innovations. Motivated by this theory, ECRL localizes anchor responses in transition innovations while sharing coordinates with the state, decoder, and predictor. Known-affine calibration validates the certificate, while full-span neural experiments evaluate factor-value-free intervention-effect prediction. Across five controlled SCM families, 40 seeds, and eight held-out mechanism banks, ECRL reduces effect MSE from 1.1442 to 1.0856 versus a capacity-matched CITRIS-style model, a 5.12% reduction with a two-way cluster-bootstrap interval of [0.0438,0.0740]. ECRL thus links conditional semantic orientation to mechanism-specific effect prediction.

Primary area:
interpretability and explainable AI
Keywords:
Explainable Causal Representation Learning, Exogenous Semantic Anchoring, Named-Block Identifiability, Temporal Causal Representation Learning, Intervention-Effect Prediction

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 3 reviews, above 27% of the others.
Reviewer_8nziOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies the semantic ambiguity of learned causal representations in dynamical systems, where different latent blocks may mix information from multiple underlying exogenous mechanisms. Under an affine representation assumption and sufficiently informative anchor interventions, the authors derive bounds on cross-block mixing and propose ECRL to improve intervention-effect prediction. Experiments are conducted on synthetic dynamical systems to evaluate the proposed method.

Critical strengths
  1. The question studied in this paper is very interesting: learned latent blocks may mix information from multiple underlying mechanisms and when doing intervention of the mechanisms, the falsely matched latent variables may lead to suboptimal performance of modeling the changing exogenous mechanisms and predicting the intervention.
  2. Under proper assumptions, the authors build a theoretical bound showing how the mismatched cross-block mixing is controlled. And motivated by this analysis model, they proposed the method ECRL and experiments on controlled synthetic datasets show its effectiveness.
Critical weaknesses
  1. The methodology part remains unclear: the motivation like why introducing orthogonal head QwQ_w or adding the four different loss terms remains unclear, some notations like μ^j,eϵ\hat\mu_{j,e}^{\epsilon} are not introduced, the numerical values like 0.18, 0.16, 0.06 in Lmargin\mathcal{L}_{\text{margin}} and total loss L\mathcal{L} remains heuristic and authors didn't explain how they are obtained. When introducing training, why adopting three-stage training also remains unclear and seems heuristic.
  2. The theory contribution seems to be marginal: Theorem 1 simply applies Gk⪰γkIG_k\succeq \gamma_kI and Proposition 2 applies Hoeffding's inequality for concentration. The final bound still relies on an unknown parameter γk\gamma_k which seems critical to how the cross-block mixing diminishes. Authors didn't analyze this part. Also, for experiment part, Figure 2 adopts Gk∝IG_k\propto I so the inequality of Eq. (3) reaches the equality condition so this is tight. The experiment says nothing when GkG_k is non-isotropic.
Questions for authors
  1. The paper introduces both theory and experiment method. However, the connection between them is not sufficiently established, particularly regarding how the theoretical guarantees motivate the actual learning objective.
  2. The paper is generally well-written. However, some key concepts for understanding the method are not well-explained and might hinder the reading.
  3. "AI Use Statement" Section is missing in the paper.
1 comment

We thank the reviewer for the positive assessment of the research question and the theoretical connection between anchor responses and semantic mixing. We address the methodological clarity, theoretical contribution, and reporting concerns individually below.

Weaknesses

W1. The motivation for the orthogonal head, individual loss terms, coefficient choices, and three-stage training schedule is insufficiently explained. Some notation is unclear.

We clarify the mathematical and empirical roles of these design choices and complete a 480-run controlled audit across 24 configurations and 20 paired seeds.

D1 establishes the orthogonal head's metric-preservation property; general invertible transformations preserve zero-intervention prediction when the correct inverse is used, but can change innovation-response scale.

D2 establishes independent roles for Localization and Margin. Removing Localization increases OOD effect MSE by 0.11056, while removing Margin increases it by 0.04501. We also report coefficient and threshold sensitivity.

D3 shows that Joint-24 improves Selectivity, ID rollout MSE, and OOD effect MSE over the original 12/8/4 schedule. Thus, the original staged schedule is a documented design choice rather than a necessary condition.

We also clarify empirical innovation-mean notation and distinguish the margin threshold m=0.18m=0.18 from the three auxiliary loss coefficients.

Please see Common Questions, Q4.

W2. The theoretical contribution appears marginal, the mixing bound depends critically on the anchor coverage constant γk\gamma_k, and the original calibration uses isotropic Gram matrices.

The contribution concerns named-block semantic identification and quantitative residual-mixing control, not the individual concentration inequalities.

We extend the theoretical analysis to rank-deficient anchors, partial identifiability, sharpness, and optimal coverage.

C1 explicitly varies Gram condition numbers from 1 to 100. Under a fixed excitation budget, γk\gamma_k decreases from 0.3200 to 0.00634, while the median finite-sample bound ratio increases from 8.77 to 60.86 at N=10,000N=10{,}000.

C2 demonstrates that covariance contrasts can restore a direction invisible to mean shifts, increasing moment-visible dimension from 1 to 2.

These experiments directly address anisotropic coverage and the distinction between certificate validity and tightness.

Please see Common Questions, Q2.

Questions

Q1. How do the theoretical guarantees motivate the actual learning objective?

The approximate-affine extension links absolute off-target innovation responses to residual cross-block mixing.

The Localization objective directly encourages target-specific response allocation, and E3 tests the connection empirically. Adding Localization raises Selectivity from 0.5442 to 0.8410 and reduces absolute off-target response from 0.016280 to 0.002601.

The theoretical bound and learning loss use related but distinct response quantities; the audit measures both rather than treating them as interchangeable.

Please see Common Questions, Q1, and the objective analysis in Q4.

Q2. Several key concepts and methodological choices require clearer explanations.

We expand the definitions of named exogenous blocks, matched anchor contrasts, learned innovation means, normalized response energies, the orthogonal orientation head, and the three auxiliary loss terms.

We also explain the training stages and report direct architectural, coefficient-sensitivity, and schedule comparisons.

The mathematical identities and implementation details are organized in Sections 5.1–5.3 and Appendices K.1–K.3, with controlled design audits in Appendix K.7.

Please see Common Questions, Q4.

Q3. The AI Use Statement is missing.

Thank you for identifying this omission. We will add a dedicated AI Use Statement in accordance with the submission requirements, accurately disclosing any AI-assisted activities and their scope. The authors remain responsible for verifying the scientific claims, derivations, experimental results, references, and final manuscript.

Reviewer_R6C5OpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies causal representation learning, aiming to align different latent blocks with named external mechanisms. It proposes an anchor-based method and theoretical guarantees connecting off-target responses to cross-block mixing.

Critical strengths

1.The theoretical argument appears sound to me under the stated assumptions, and the finite-sample bound makes the analysis more concrete. 2. The intervention-effect improvement is consistent across most seeds and all evaluated mechanism banks.

Critical weaknesses

1.The affine assumption is restrictive. The theory assumes an affine relationship between learned representations and true sources, but the neural method does not guarantee this relationship. 2.The experiments do not fully convince me. They focus on low-dimensional controlled SCMs, with limited comparisons against other causal representation learning methods. 3. The predictive benefit is mixed. Intervention-effect MSE improves by 5.12%, but in-distribution prediction MSE worsens by about 1.91%. It is unclear whether this trade-off is worthwhile beyond the controlled settings.

Questions for authors

1.Can the affine assumption be relaxed, or can the analysis quantify errors when this assumption only approximately holds? 2.How can users assess whether the theoretical assumptions hold without observing the true latent sources? 3.Can the authors demonstrate the benefits on more realistic or higher-dimensional data, with broader baseline comparisons?

1 comment

We thank the reviewer for recognizing the soundness of the theoretical argument under its stated assumptions and the consistency of the intervention-effect improvements. We address the concerns regarding affine applicability, experimental generality, and the ID–OOD trade-off below.

Weaknesses

W1. The affine assumption is restrictive, and the neural method does not guarantee an exact affine relationship between learned representations and true exogenous sources.

We extend Theorem 1 to approximate-affine innovation representations, incorporating an explicit residual-contrast correction into the cross-block mixing bound.

E1 validates the extended bound against known mixing matrices: the residual-corrected population bound achieves 100% coverage, compared with 0% when the residual is omitted in the controlled cancellation test.

E2 independently estimates affine components of frozen neural representations, while E3 demonstrates that Localization improves both relative response concentration and absolute off-target response.

Please see Common Questions, Q1.

W2. The experimental evidence is limited to relatively low-dimensional controlled SCMs and a narrow range of causal representation baselines.

We extend the evaluation through B1 official-source CITRIS/iCITRIS adaptations and B2-A higher-dimensional controlled SCMs.

B1 compares ECRL with adapted source implementations under a common task, while preserving the original matched CITRIS-style comparison as the cleaner test of semantic anchoring.

B2-A evaluates five configurations with 20–256 observation dimensions and latent blocks up to (8,8,4)(8,8,4). ECRL maintains a fixed-order Selectivity advantage over a prediction-only GRU in all five settings.

Please see Common Questions, Q6.

W3. Although OOD intervention-effect MSE improves by 5.12%, ID prediction MSE worsens by approximately 1.91%, leaving the value of the trade-off unclear.

These results concern different endpoints: OOD intervention-effect prediction and ordinary ID rollout prediction.

The original experiment establishes a 5.12% OOD benefit with a 1.91% ID cost. Our additional D3 audit shows that the trade-off is not fixed by the original training schedule: Joint-24 reduces ID rollout MSE from 0.82594 to 0.75014 and OOD effect MSE from 1.08113 to 1.01885, while improving Selectivity.

This is a within-ECRL schedule comparison; we report it separately from the historical CITRIS-style result.

Please see Common Questions, Q7.

Questions

Q1. Can the affine assumption be relaxed, or can the analysis quantify errors when it only approximately holds?

We introduce an approximation-aware response–mixing bound for

ϵ=Aξ+η,\epsilon=A\xi+\eta,

where the residual term η\eta enters through anchor-dependent response contrasts.

The extension is verified in E1 using known mixing matrices and evaluated on frozen neural representations in E2. The corrected population bound achieves 100% coverage in the controlled cancellation experiment, and the neural audit quantifies both fitted mixing and approximation behavior.

Please see Common Questions, Q1.

Q2. How can users assess whether the theoretical assumptions hold without observing the true latent sources?

We clarify which quantities are observable from learned responses and which require independent source-space calibration.

P1 evaluates a source-free rank diagnostic across 60 frozen checkpoints and finds 98.5% repeatability for ECRL, while detecting only 20.65% of truly collinear rank-deficient anchors. Thus, observable response stability provides a diagnostic, not a certificate of true source rank.

P2 then quantifies the effects of incorrect target labels, block partitions, and missing anchors on semantic correspondence and intervention prediction.

Please see Common Questions, Q3.

Q3. Can the benefits be demonstrated on more realistic or higher-dimensional data with broader baseline comparisons?

We complete both requested controlled extensions.

B1: Official-source CITRIS-VAE and iCITRIS-VAE adaptations are evaluated alongside ECRL, CITRIS-style, and a prediction-only GRU under a shared vector protocol.

B2-A: Five higher-dimensional SCM settings cover 20–256 observation dimensions, including larger latent blocks. ECRL retains a Selectivity advantage across all settings, with effect-NMSE improvements supported in the 128-dimensional, (2,2,1)(2,2,1) configuration.

The results explicitly distinguish semantic localization gains from task- and scale-dependent predictive performance.

Please see Common Questions, Q6.

Reviewer_L8XhOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents a method for exploiting external information about sets of exogenous variables in the process of causal representation leartning. This is motivated by the fact that most modern ML methods learn models that entangle the variables humans might like to use to describe what is going on, thus limiting their interpretability. The paper is one of a long line of papers seeking to uncover explainable causal representations, and is primarily differentiated by the way it uses external information. The model applied is that the learned representation is an affine function of the sources, so it is applicable only in this case.

Critical strengths

Th epaper tackles one of the ongoing problems in the field

The experiments are statistically well thought out, which is nice to see.

SOme effort has been made to list the relevant literature, which seems to be often skipped in causal representation learning papers. This doesn't go as far as is required to explain how teh paper builds upon its comparators, but it's a good start.

The statements as to the limitatios of the approach are candid, which is great.

Critical weaknesses

The core challenge is that the theory doesn’t apply to the model evaluated. The theory is based on affine embeddings but the model used in the experiments is a GRU, which isn't affine. The affine fit statistic on line 455 is 0.54–0.59 which would seem to indicate that the model really doesn't apply.

The setting requires knowing quite a lot about the scenario, and no case is made as to why this might be justifiable.

The paper doesn't make it clear what is really novel about the approach. Theorum 1 is a relatively minor step, and proposition 2 is similarly not all that novel. The contribution seems marginal as compared to Ahuja et al. 2023, and int eh absence of a clear argument as to why it's important it is hard to identify the required level of novelty.

The baselines in teh experiments are also relatively weak. There's a “CITRIS-style” reimplementation rather than CITRIS itself or iCITRIS

Questions for authors

If explainability is defined as block correspondence, why is the metric effect MSE?

Given that teh data is synthetic why didn't you report cross-block mixing against ground truth?

1 comment

We thank the reviewer for the constructive assessment of our work, particularly the recognition of its relevance to causal representation learning, the statistical evaluation, and the discussion of limitations.

Weaknesses

W1. The theoretical guarantees assume affine representations, whereas the evaluated GRU is nonlinear. The reported affine-fit R2R^2 of 0.54–0.59 raises concerns about the applicability of the theory.

We extend the analysis to approximate-affine innovations, ϵ=Aξ+η\epsilon=A\xi+\eta, deriving a response–mixing bound that explicitly incorporates anchor-dependent approximation residuals. The new E1 experiments validate the correction against known mixing: population coverage increases from 0% to 100% after residual correction. E2 further evaluates frozen GRU representations using independently fitted affine components, and E3 connects the bound to the Localization objective.

These results distinguish the validity of the approximation-aware bound, the empirical affine approximation quality of neural representations, and the semantic effects of the learning objective. Please see Common Questions, Q1.

W2. The method requires substantial prior knowledge about the scenario, and the justification for this information is unclear.

The required information is protocol-level supervision: named external mechanisms, their block dimensions, intervention-target labels, and matched reference–anchor environments. Realized exogenous values and true mixing matrices are not used during training or intervention prediction.

P1 distinguishes observable learned-response stability from source-space coverage, while P2 measures the consequences of incorrect targets, partitions, and missing anchors. For example, 20% target-label noise reduces Selectivity by 0.1614, and removing all anchors for one block reduces that block's Selectivity by 0.6208. Please see Common Questions, Q3.

W3. The theoretical novelty relative to existing identifiable CRL methods, including Ahuja et al. (2023), is insufficiently clear.

Our contribution is a quantitative characterization of fixed-order, named-block semantic correspondence, rather than the use of spectral or concentration inequalities alone. Theorem 1 connects observable off-target anchor responses to residual cross-block mixing. We further analyze identifiable subspaces under rank deficiency, bound sharpness, optimal anchor coverage, and covariance-based recovery of otherwise unexcited directions.

The C1/C2 experiments demonstrate how anchor geometry affects both identification and certificate tightness, making the additional contribution explicit. Please see Common Questions, Q2.

W4. The experimental baselines are relatively weak because the main comparison uses a CITRIS-style reimplementation rather than canonical CITRIS/iCITRIS.

The original CITRIS-style recurrent comparator was deliberately constructed to isolate semantic anchoring under matched prediction architecture, capacity, block structure, intervention information, and optimization budget.

We now complement this controlled comparison with official-source CITRIS-VAE/iCITRIS-VAE adaptations under a shared experiment vector protocol. Across 20 paired seeds, ECRL achieves fixed-order Selectivity of 0.82797, compared with 0.80425 for CITRIS-style and 0.40859/0.30956 for the official-source adaptations. Please see Common Questions, Q6.

Questions:

Q1. If explainability is defined as correspondence between learned blocks and named mechanisms, why is intervention-effect MSE the main evaluation metric?

Intervention-effect MSE measures the downstream predictive utility of mechanism-specific representations; it is not our sole measure of explainability.

We additionally evaluate fixed-order Selectivity, absolute target/off-target responses, and mechanism-to-block attribution under simulator-controlled interventions. In X1, ECRL improves Selectivity from 0.56931 to 0.87014 over w/o Localization and reduces standardized off-target response from 0.30812 to 0.19746. These results directly evaluate mechanism-level semantic alignment independently of effect-prediction accuracy. Please see Common Questions, Q5.

Q2. Given that the data are synthetic, why is cross-block mixing not directly evaluated against ground-truth mechanisms?

We conduct an oracle-controlled mechanism-response audit and directly measure how changes in each true exogenous block affect the corresponding learned blocks, without fitting an affine mixing matrix or using true exogenous values during inference.

We separate no-coupling, first-step, lagged, and instantaneous responses to distinguish mechanism-specific localization from dynamical propagation.

True affine mixing is evaluated separately in the known-map certificate experiments, while the neural audit directly evaluates source-controlled responses. Please see Common Questions, Q5, with the approximation diagnostics in Q1.

Comments on the paper

8 comments

Common Questions and Responses

Q1. How Do Our Theoretical Guarantees Apply to Nonlinear Neural Representations? (L8Xh: W1; R6C5: W1, Q1; 8nzi: Q1)

The affine assumption concerns the relationship between learned representations and exogenous sources, rather than the encoder architecture itself. We extend Theorem 1 to approximate-affine innovations,

ϵ=Aξ+η,\epsilon=A\xi+\eta,

where η\eta captures the non-affine residual. The resulting response–mixing bound includes an additional anchor-dependent residual term, quantifying how approximation error affects cross-block mixing.

For source block kk and learned block jj, the corrected bound takes the form

\leq \frac{\sqrt{R^\epsilon_{jk}}+\delta_{jk}} {\sqrt{\gamma_k}}$$ where $\gamma_k$ is the source anchor-coverage constant and $\delta_{jk}$ quantifies the corresponding residual contrast. We validate this extension and its connection to ECRL through three complementary experiments. - **E1: Approximation-aware certificate validation.** Even with an affine-fit $R^2=0.999807$, ignoring residual contrasts produces only **0% population coverage**, with an average bound-to-true-mixing ratio of **0.2163**. The residual-corrected population bound achieves **100% coverage** and a ratio of **1.0014**. At $N=10{,}000$ samples per environment, the corrected finite-sample bound has a ratio of approximately **5.20**. These results establish both the importance and the effectiveness of the approximation correction, while distinguishing population tightness from finite-sample tightness. - **E2: Neural cross-block mixing.** Using independently fitted affine components of frozen neural representations, ECRL Full achieves a fitted mixing ratio of **0.4053**, compared with **0.4336** for CITRIS-style and **0.6479** without Localization. The paired reductions have 95% confidence intervals of **[0.0170, 0.0397]** and **[0.2152, 0.2694]**, respectively. The residual-corrected bound achieves **100% empirical consistency with the fitted affine component**. This evaluates the neural representation's approximation behavior; it is distinct from statistical coverage of an unknown true neural mixing matrix. - **E3: Theory–objective alignment.** The theoretical bound depends on **absolute off-target responses**, whereas Localization optimizes **relative response concentration**. Adding Localization increases Selectivity from **0.5442 to 0.8410** and reduces absolute off-target response from **0.016280 to 0.002601**, an **84.0% reduction**. Thus, the learning objective improves both semantic concentration and the response quantity directly associated with the theoretical mixing bound. **Revisions.** We extend the theoretical analysis in **Section 4 and Appendix A.11**, clarify the relationship between the theoretical response quantity and the learning objective in **Section 5.2**, and report E1–E3 in **Sections 6.4–6.5 and Appendix L.4**. **The approximation-aware theorem quantitatively connects non-affine residuals to cross-block mixing. E1 validates the corrected guarantee against known mixing, E2 characterizes fitted mixing in neural representations, and E3 demonstrates the effect of Localization on the response quantities underlying the theory.**

Q2. What Is the Theoretical Novelty, and How Does Anchor Coverage Affect the Guarantee? (L8Xh: W3; 8nzi: W2)

Our theoretical contribution is a quantitative characterization of named-block semantic correspondence, rather than the use of spectral or concentration inequalities in isolation. Building on causal representation identifiability research, we examine how observable anchor responses constrain residual cross-block mixing in a fixed mechanism-named order.

Specifically, Theorem 1 establishes the response–mixing relationship

Rjkμ=tr⁡(AjkGkAjk⊤)≥γk∥Ajk∥F2,γk=λmin⁡(Gk)>0.R_{jk}^{\mu} = \operatorname{tr}(A_{jk}G_kA_{jk}^{\top}) \geq \gamma_k \lVert A_{jk} \rVert_F^2, \qquad \gamma_k = \lambda_{\min}(G_k) > 0.

This quantifies the remaining mixing between named blocks while allowing invertible affine transformations within each block. We further characterize the role of anchor geometry through theoretical extensions and controlled experiments.

  • Identification conditions and theoretical sharpness. Theorem 1 links observable off-target responses to cross-block mixing. Appendices A.9–A.12 develop moment-based identification, finite-sample certification, and constructive failures under insufficient anchor coverage. We further characterize identifiable subspaces under rank deficiency, the sharpness of the mixing bound, and optimal anchor coverage under a fixed excitation budget.

  • C1: Anisotropic coverage and certificate tightness. With fixed total anchor energy, increasing the Gram condition number from 1 to 100 reduces the minimum eigenvalue from 0.3200 to 0.00634. At condition number 100, population bound-to-true-mixing ratios are 1.00, 10.00, and 6.45 for weak-direction, strong-direction, and random-direction mixing, respectively. At N=10,000N=10{,}000, the median finite-sample ratio increases from 8.77 to 60.86. Across all 28 full-rank condition-number/sample-size settings, simultaneous coverage of the six off-diagonal blocks is observed for 20/20 affine maps. These results distinguish empirical coverage from certificate tightness and directly address the anisotropic geometry absent from the original isotropic illustration.

  • C2: Recovering directions through covariance contrasts. When the mean-shift rank remains one, introducing covariance contrasts along the unexcited direction increases the moment-visible dimension from 1 to 2 and the squared separation constant from 0 to 0.2048. The population mean response in that direction remains zero, while the population moment bound reaches a 1.00 bound-to-true-mixing ratio. However, during our experiments, we observed that at N=10,000N=10{,}000, the finite-sample moment bound remains relatively conservative, reaching a bound-to-true-mixing ratio of 387.15. This highlights the distinction between population identifiability and finite-sample statistical precision.

Revisions. We strengthen the theoretical characterization of named-block identification and anchor coverage in Section 4 and Appendix A.10 and incorporate the C1/C2 anisotropic-coverage and covariance-rescue experiments into Section 6.2 and Appendix L.2. The original calibration and neural prediction results remain unchanged.

In summary, our contribution is a response-based account of which named mechanism directions are identifiable, how anchor geometry controls mixing-bound sharpness, and when additional moment information restores otherwise invisible directions.

Q3. Are the Required Prior Assumptions Justified and Verifiable?

(L8Xh: W2; R6C5: Q2)

Our assumptions correspond to a target-labelled intervention protocol in which mechanism identities and intervention targets are recorded, while realized exogenous values remain unobserved. We clarify the required information and provide two additional audits of prior reliability.

  • Justification of the priors. ECRL requires named exogenous mechanism blocks, their dimensions, intervention-target labels, and matched reference–anchor metadata. Such information can be specified in controlled diagnostic or intervention experiments with recorded perturbation targets. True exogenous realizations and mixing matrices are not used for training, checkpoint selection, or intervention-effect prediction (Section 3; Appendix A.7).

  • P1: Source-free assumption audit. We evaluate 60 frozen checkpoints over 20 paired seeds using 21,000 observational diagnostic records. Learned-response diagnostics exhibit high repeatability: ECRL achieves 98.5% agreement between independent rank assessments. However, only 20.65% of genuinely rank-deficient collinear anchors are detected. Thus, the audit establishes an important distinction: stable learned responses characterize observable representation behavior but do not certify source-space rank. Absolute mixing certificates require independently calibrated source coverage, whereas response repeatability and localization can be evaluated without observing exogenous values.

  • P2: Sensitivity to prior violations. We train 110 matched checkpoints across 11 conditions with identical architecture and optimization budgets. At 20% target-label noise, Selectivity decreases by 0.1614 , while fitted cross-block mixing increases by 0.1172 . Removing all anchors for one mechanism decreases that block's Selectivity by 0.6208 and increases its target-specific OOD effect MSE by 0.3415 . Incorrect block partitions also impair semantic alignment, whereas anchor contamination shows condition-dependent rather than uniformly degrading effects.

Revisions. We clarify the required information and its interpretation in Section 3 and Appendix A.7 and report the P1/P2 assumption audits in Section 6.5 and Appendix L.5.

The required priors consist of protocol-level intervention metadata rather than observed exogenous factor values. P1 establishes what can be assessed from learned responses and what requires independent source calibration; P2 quantifies the consequences of incorrect labels, partitions, and missing anchors.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.