acceptodds
Under review as a conference paper at ICLR 2027

Semantic-Frame Identifiability and Recovery in Geometric Representation Learning

Abstract

Geometric representations are often used not only to predict relations, but also to infer hierarchy, abstraction, or uncertainty from an embedding’s origin and direction. However, accurate distance-based prediction does not establish that these semantic choices are identifiable. We formalize this distinction as semantic-frame identifiability: observationally equivalent representations can support different semantic targets despite identical intrinsic predictions. We establish a minimax lower bound determined by the semantic ambiguity within each observational equivalence class and characterize how asymmetric evidence resolves decision-relevant ambiguity. Building on this analysis, we introduce FramePosterior, a semantic decision layer that leaves the encoder and native predictor frozen, conditions candidate frames on sparse, noisy anchors, and aggregates task-specific decisions across rooted charts. Exact-isometry audits distinguish unchanged native predictions, altered semantic readouts, and measurable downstream loss. On ImageNet-100, a radial-depth readout loses 12.68 percentage points in within-one-level accuracy, while 32 noisy landmarks recover 89.88% of its depth-error gap. On the full WordNet noun DAG, posterior voting outperforms matched MAP inference by 1.48 percentage points with 16 anchors under 20% corruption. A topology-informed routing atlas improves delivery from 69.90% to 77.92% over a matched no-revisit control. Stable visual readouts provide an important boundary: frame ambiguity does not necessarily degrade every task. These results clarify when metric prediction is insufficient for semantic decisions and how explicit evidence enables targeted recovery without changing native predictions.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
geometric deep learning, representation identifiability, hyperbolic representation learning, hierarchical representation learning, symmetry and invariance, graph representation learning, uncertainty and abstention

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_n1wHOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper distinguishes metric prediction from decisions requiring an external semantic frame. Isometric observations can preserve native predictions without identifying a root, orientation, or fixed-origin radial interpretation. It provides orbit-risk bounds, noisy-anchor recovery conditions, and decision-resolving conditions. FramePosterior freezes the native predictor and aggregates frame-dependent decisions using additional evidence. Experiments cover graphs, visual hierarchies, radial uncertainty, and routing, including stable-readout counterexamples.

Critical strengths
  1. The observation law, semantic target, and decision equivalence are explicit, avoiding an immediate identification of coordinate sensitivity with task failure. The main orbit, recovery, and posterior-approximation bounds appear sound under their assumptions (Section 2; Appendices A–B).
  2. Exact isometries and transported-reference controls separate prediction preservation, readout sensitivity, and actual loss. Weak or null CIFAR and public-code HIE results are retained.
  3. Information budgets are transparent, including WordNet chart misspecification, source calibration radii in ImageNet, and extra topology in routing. Recovery is not presented as obtaining semantics automatically from distances alone.
Critical weaknesses
  1. A gap remains between the identifiability setting and natural deployment failures. The theorem concerns a specified invariant observation law. Full training systems with centering or hierarchy constraints contain additional information. A post-training isometry preserves the native distance predictor without necessarily preserving all training constraints. Since transporting the root or center resolves coordinate changes, the paper should establish when real systems lose that reference and require new semantic evidence. Some strong results use deliberately constructed radial tasks.
  2. The WordNet gains do not fully isolate the value of posterior inference. At 16 anchors and 20% corruption, MAP root recovery is already 100%; errors arise from undirected charts approximating directed depth. MAP predicts ties on 2.36% of queries versus .007% for posterior voting. This improvement is useful, but comparisons with direct depth regression, reasonable MAP tie-breaking, and simple neighboring-root ensembles under the same anchors are needed to separate uncertainty marginalization from structural approximation repair.
  3. ImageNet mainly supports sparse calibration, not ensemble risk control. Single-isotonic-chart MAE is 1.343 versus 1.340 for the ensemble, and disagreement does not produce a monotone risk–coverage curve (Appendix D.8). RDC also uses stored source calibration radii in addition to the 32 noisy values. This is disclosed, but comparisons should match the full information budget and test realistically queryable depth/order anchors.
  4. Conditional risk bounds lack corresponding empirical calibration. The T=1 Huber Gibbs update is not the exact likelihood of the ±1-hop corruption process, and the WordNet target may lie outside the root-chart family. Retained top-K posterior mass is not reported, preventing an operational interpretation of the truncation TV bound. Posterior confidence, abstention, and actual risk under model misspecification need independent validation.
  5. Routing gains need clearer attribution to additional topology. The main score is an ALT lower bound from landmark shortest paths, plus a small learned-distance term and no revisit. The no-revisit control is useful, but a signatures-only control without learned geometry and topology-based methods with the same storage budget are still needed. This application does not directly establish an advantage of frame-posterior inference.
Questions for authors

See weakness.

3 comments
AuthorsOpenReview ↗

Response on calibration, matched controls and routing

1. Observation law and natural failures. The orbit theorem requires equal complete observation laws for admissible states with different semantic answers, including centering constraints and hierarchy supervision. HIE/HypStructure/UNCHA interventions test fixed-reference readouts while preserving checked native outputs; transforming points and their reference together preserves relative distances. Additional untransformed checks found no natural reference error repaired by anchors. These are separate theorem and intervention claims. (Table 7, p. 21; Appendix E.3, p. 23)

2. WordNet: matched regression, ties and neighbors. With 16 anchors and 20% corruption, MAP/posterior/uniform-top-64 tie accuracies are 96.56/98.04/98.17%; non-tied MAP/posterior are both 98.89%. Structure-only ridge reaches 93.38%. The tie rule votes over the same anchor-loss-ranked top-64 pool only on MAP-tied queries. Root recovery is already 100%; the differences concern chart predictions and ties. A correct root plus the directed DAG yields exact depth using richer information. (Table 2, p. 9)

The neighbor ensemble includes MAP and all one-hop graph neighbors, uniformly votes ordinal signs and counts unresolved ties as errors. It matches anchors, queries and 50 paired trials per noise level, without tuning the neighborhood. Neighbor/posterior/tie accuracies are 96.81/98.04/98.17% at 20% corruption and 96.54/97.74/97.94% at 40%. (Appendix F.1.1, p. 28)

3. ImageNet calibration. Single/ensemble MAEs are 1.34330/1.33984 with the same 32 noisy values, stored source radii and decoder. The ten-seed paired MAE reduction is 0.00345 levels (95% CI [0.00065, 0.00626]). Sparse calibration supplies most recovery; disagreement produces a nonmonotone empirical risk-coverage curve. Recovery observes noisy source radii, and taxonomy/calibration targets are obtained offline. Realistically queried depth/order-anchor quality and cost remain unmeasured. (Appendix E.3.1, p. 23)

4. Working weights, truncation and risk. T=1 Huber Gibbs weights are a loss-based working model, not the exact clipped-hop likelihood; Gaussian recovery bounds retain their noise assumptions. In a separate truncation diagnostic, WordNet top-64 retains 99.9959/99.9787% mean working mass at 20/40% corruption; Disease top-128 retains 22.08%. Retained mass measures approximation within that working model, not semantic calibration. (Appendix F, p. 27; Appendix F.3.1, p. 32)

Complete-candidate Disease disagreement occurs on 64.70% of queries; full/minimum-loss uniform voting reach 77.08/77.13% on this hard subset. The 95%-working-mass set gives 63.60% disagreement and includes the true root in 96/100 trials. These are descriptive model-set results. (Appendix F.3.2, p. 32)

External calibration adds separately counted clean pair labels. At 2% target risk and 8192 labels, WordNet posterior coverage is 97.50/97.82% at 20/40% anchor corruption, versus tie handling's 97.60/97.92%; no certified rule exceeds target population risk in these primary conditions. Across 7200 configurations, one certified rule exceeds its 1% target with 2048 labels, reaching 1.0193% population error. At 5% risk with 2048 labels, warm32 HPO posterior/landmark-64 ridge mean coverage is 7.42/75.24%. (Appendix G.2, p. 36; Appendix F.4.3, p. 35)

The corrected selector separates unordered pairs/reversals across selection, calibration and test. Known independent symmetric-noise correction is evaluated separately from conservative contamination correction, which mostly abstains. An additional audit uses 1024 possibly corrupted selection labels plus 1024 independent clean verification labels beyond depth anchors. Across 240 Disease cases sharing 30 roots, selected-rule/fixed-posterior mean coverage is 73.17/65.14% at 5% target risk; no accepted test set exceeds 5% empirical error. Selection chooses the rule; independent verification selects its certified threshold. The result applies to simple rules too and requires correct verification labels and the stated finite query population; verification-label corruption and future query shift remain untested. (Appendices G.4-G.5, p. 38)

5. Routing attribution. With matched 64 landmarks and no-revisit settings, delivery is 25.68% for ALT alone, 21.48% with an L1-signature score, 52.44% with degree and 77.91% with HGCN; HGCN without ALT gives 69.90%. The stronger BFS-tree delivers 100% at 1.119 stretch, versus hybrid's 77.912%. Tree/hybrid persistent-plus-packet state is 169,599/693,827 serialized bytes, including common CSR and visited state but excluding scratch, evaluator/Python memory and distributed metadata. (Appendices H.2, H.4, p. 41)

I appreciate the authors' efforts. The author’s response has indeed given me a better understanding of the paper; however, these revisions are not sufficient to warrant a higher score, so I am maintaining my original rating.

AuthorsOpenReview ↗

Follow-up on the matched controls and remaining concerns

Thank you for the follow-up and for confirming that the revision clarifies the paper. We would appreciate one further clarification about the remaining basis for your assessment.

Requested matched controls. The MAP-plus-all-one-hop-neighbors comparison now uses the same anchors, queries and paired trials. With 16 anchors and 20% corruption, its accuracy is 96.81%, versus 98.04% for posterior voting and 98.17% for uniform-top-64 tie handling. We retain the stronger tie rule explicitly; this comparison does not establish a general posterior-aggregation advantage. (Table 2, p. 9; Appendix F.1.1, p. 28)

The revision also reports retained working mass, distinguishes the Huber Gibbs model from the actual corruption likelihood, and evaluates reliability with additional calibration and verification labels under the stated sampling and label-quality assumptions. Routing separates learned scores from landmark topology and includes the stronger BFS-tree comparator. These additions make the information budgets and attribution tests explicit. (Appendices F.3, G and H)

Remaining concern. The absence of a naturally occurring, anchor-repairable reference failure and the limited incremental benefit over strong simple rules remain substantive limitations. We state them alongside the results.

Could you indicate whether the added comparisons address your original requests for matched controls, and which remaining concern is most decisive for your recommendation? This would help the AC distinguish the evaluation questions now addressed from the scientific limitations that remain. We would appreciate that distinction being reflected in your final assessment.

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

This paper studies semantic-frame identifiability and recovery in geometric representation learning. The authors analyze how symmetric geometric observations, such as pairwise distances, can leave certain semantic orientations unidentified, and propose FramePosterior to incorporate external semantic anchors into inference over candidate reference frames. The framework keeps the original geometric predictor frozen, assigns posterior weights to candidate frames, and aggregates their task-specific predictions. The paper also distinguishes recovering a unique reference frame from recovering the decisions required by a downstream task. Experiments on WordNet, Disease, ImageNet, and related settings examine the effects of isometric transformations on semantic readouts and evaluate recovery using anchors and posterior aggregation.

Critical strengths

1. The distinction between geometric prediction and coordinate-dependent semantic interpretation is relevant. The paper draws attention to a useful evaluation issue: preserving geometric relationships and supporting a particular semantic interpretation are separate properties of a representation. This distinction matters when quantities such as radius or coordinate position are interpreted as hierarchical depth. The controlled isometric transformations provide a concrete way to examine the sensitivity of these readouts to the choice of reference frame.

2. Task-level recovery is a useful evaluation target. The paper considers whether candidate reference frames support the same downstream decisions, allowing recovery to be evaluated according to the information a task actually requires. This provides a useful way to organize experiments under limited semantic observations and to distinguish uncertainty about a reference frame from uncertainty that affects the task outcome.

Critical weaknesses

1. The methodological contribution of FramePosterior beyond existing anchor-based inference remains unclear. I accept the central conclusion of Theorem 1: symmetric geometric distances alone can leave semantic orientation unidentified. Introducing asymmetric information through roots or anchors is also a reasonable response. My question is what FramePosterior adds once this information is available. The paper moves from a setting with only symmetric observations to reference-frame inference conditioned on additional semantic evidence. The orientation information comes from those external observations; candidate roots are then weighted using distance residuals and contribute to task predictions. What is the fundamental methodological difference between this procedure and conventional anchor-based root inference or semantic calibration using the same information? More directly, what makes “geometry + asymmetric evidence + posterior voting” a new research contribution in this setting? The manuscript explains the motivation for adding semantic information more clearly than it establishes the technical increment in the frame representation, posterior construction, decision rule, or recovery guarantees.

2. The experiments establish the value of anchors more clearly than the independent value of FramePosterior. My main experimental question is whether the results demonstrate that anchors supply useful information or that FramePosterior uses that information more effectively than simpler methods. These possibilities require comparisons with matched anchor observations, structural access, and calibration budgets. The WordNet inference procedure especially needs explanation: if the correct root has already been identified, why aggregate undirected-distance predictions from multiple candidate roots? If directed relations are available to the method, directly computing depth from the correct root and those relations is an important baseline. If access is restricted to undirected structure, that restriction should be explicit. The authors should also separate improvements caused by uncertainty over roots from improvements that compensate for disagreement between undirected distances and semantic depths, or that arise from resolving tied predictions. In particular, I would like to see MAP inference with appropriate tie handling. ImageNet raises a similar question: how much performance can a single calibration recover using the same source-frame semantic information, and how much does multi-frame or resampled aggregation add? Overall recovery and the incremental benefit of aggregation should be reported separately.

3. The recovery experiments rely heavily on constructed settings that closely match the inference assumptions. In Disease, semantic roots and anchor depths are constructed within a known graph, and recovery evaluates candidate roots using distances from that same graph. When observations are generated according to the structural rule subsequently used for recovery, the experiment primarily tests recoverability under matched assumptions. I would like the authors to clarify which existing experiments use independently defined semantic targets or examine a mismatch between geometric distances and semantic relationships. Aggregate ordinal accuracy also needs to be interpreted alongside the difficulty of the evaluated queries. If many candidate roots give the same answer to a query, that query provides a weak test of the claimed decision-recovery capability. What proportion of queries produce disagreement among plausible candidate roots, and how well does FramePosterior perform on that subset? The isometry experiments raise a related scope question: transforming an embedding while keeping a semantic reference fixed constructs a reference mismatch, but the paper should explain how comparable mismatches arise in trained models and identify the evidence for them.

4. The theoretical contributions need clearer separation from established identifiability and information-acquisition results. Several central components have direct precedents. Identifying graph nodes through distances to landmarks is studied by Khuller et al. (1996), and updating candidate weights through exponentiated negative losses follows the generalized Bayesian framework of Bissiri et al. (2016). Identifying only the class or region needed for a decision is central to equivalence-class determination and decision-region determination in Golovin et al. (2010) and Javdani et al. (2014). I would like a precise account of what Theorem 1 adds to a standard observational-indistinguishability argument, and what new conditions or guarantees the noisy-anchor recovery and active-query results introduce. The manuscript already acknowledges some reductions to existing frameworks. The remaining question is what is contributed beyond those reductions, and how the new results guide the method or yield additional testable predictions. At present, the theoretical section mainly appears to organize established principles around semantic-frame recovery, while the size of the theoretical increment remains difficult to assess.

Questions for authors

1. What does FramePosterior add when competing methods receive exactly the same root or anchor information? Please identify the most substantive difference from conventional anchor-based root inference, root-conditioned geometric inference, or semantic calibration. Which specific algorithmic step or theoretical capability is new, and which experiment directly tests that increment?

2. Which improvements come from the anchors, and which come from FramePosterior? If the correct root has already been identified, what additional problem is solved by voting over undirected-distance predictions from multiple candidate roots? When directed relations are available, how does direct depth inference from the correct root perform? Under an undirected-only restriction, how does MAP inference perform with appropriate tie handling? For ImageNet, please compare single calibration and multi-frame aggregation using the same semantic observations, and report the additional gain from aggregation together with its variability.

3. How far do the recovery results extend beyond the assumptions used to construct the experiments? Which evaluations use semantic targets obtained independently of the graph or reference information used by the recovery procedure? In Disease, what fraction of queries elicit different predictions from plausible candidate roots, and what is the performance on that subset? What observations in trained models correspond to the reference mismatches introduced by the controlled isometry experiments?

4. Which theoretical results are new, and what do they add? Please distinguish new results from applications of existing principles and guarantees inherited through reductions. In particular, what is added by Theorem 1, the noisy-anchor recovery analysis, and the active-query guarantees? How do those additions affect the design of FramePosterior or explain its experimental behavior?

1 comment
AuthorsOpenReview ↗

Response on contribution and matched comparisons

Q1/Q4. Contribution and theoretical provenance. We define semantic-frame identification jointly through the complete observation law and requested query family: frames need separation only when their answers differ. Theorem 1 applies the standard indistinguishability/triangle-inequality argument to orbit-separated semantic targets and decision disagreement. Noisy-anchor bounds express recovery through signature separation: decision error sums only over answer-disagreeing frames. Independent sub-Gaussian noise and the stated Gaussian/canonical-family/packing assumptions delimit the corresponding guarantees. Under bounded noise, cross-decision signatures must differ by more than twice the noise bound. For finite DAG frames, noiseless decision identification is equivalent to covering answer-disagreeing frame pairs. Set-cover and EC2 guarantees follow through the stated reductions; the implemented entropy heuristic has no EC2 guarantee.

Graph resolving sets, generalized Bayes and decision-region acquisition are explicitly credited. FramePosterior implements established Gibbs weighting and voting for reference-conditioned charts while keeping the native predictor fixed. The resulting analysis separates reference recovery, depth-map adequacy and inference under matched evidence. (Table 5, p. 17; Appendix C.3, p. 19)

Q2. Matched information and voting after root recovery. With the same undirected structure, 16 anchors and 20% corruption, MAP/posterior/uniform-top-64 tie handling attain 96.56/98.04/98.17% ordinal accuracy. MAP root recovery is already 100%, and non-tied MAP/posterior accuracy is 98.89%. These differences concern chart-based predictions and ties. The tie rule votes over the same anchor-loss-ranked top-64 pool only when MAP ties on the query; the pool is not restricted to equal-loss roots. Structure-only ridge attains 93.38%. Correct root plus the full directed DAG gives exact depth with additional directional information. (Table 2, p. 9)

MAP plus all one-hop neighbors attains 96.81/96.54% at 20/40% corruption, versus posterior's 98.04/97.74% and tie handling's 98.17/97.94%, without neighborhood tuning or extra labels. (Appendix F.1.1, p. 28)

ImageNet single/resampled-chart MAEs are 1.34330/1.33984 using the same 32 noisy values, stored source radii and decoder. The paired mean reduction is 0.00345 levels, 95% CI [0.00065, 0.00626], across ten seeds after averaging paired draws within seed. Sparse calibration provides most recovery. (Appendix E.3.1, p. 23)

A separate warm32 diagnostic raises native AUC from 0.8706 to 0.9253, yet posterior recovery is 84.80-85.27%, versus landmark-64 ridge's 90.92-92.65%, with the same 32 labels and equal per-node numeric-table bytes. Improved encoding therefore does not establish a recovery-method advantage. (Appendix F.4.3, p. 35)

Q3. Target independence and query difficulty. Disease tests structural recoverability under matched assumptions. WordNet targets are directed taxonomy depths, while recovery receives undirected distances, creating chart mismatch. ImageNet targets are taxonomy-defined, but hierarchy also supervises the encoder. Additional untransformed model checks found no natural reference error repaired by anchors. (Table 7, p. 21; Appendix E.3, p. 23)

Across all 2665 Disease roots, working-Gibbs label Gini disagreement exceeds 0.1 on 64.70% of queries (95% trial-level half-width 6.24 percentage points), versus 10.37% for the earlier top-128 proxy. Compensated full/minimum-loss uniform voting attain 77.08/77.13% on the all-root hard subset. The 95%-working-mass set gives 63.60% disagreement and 76.58/76.64% hard-subset accuracy. These are model-defined sets, rather than externally calibrated plausible-root sets. (Appendices F.3.1-F.3.2, p. 32)

The main text now includes matched-method results, anchor-budget sensitivity and visual calibration recovery; Algorithm 1 specifies the implemented Gibbs/top-K procedure. (Tables 2-3, pp. 9-10; Figures 3, 5-6, pp. 7, 9-10; Algorithm 1, p. 6)

Reviewer_77knOpenReview ↗
Rating
2
Confidence
1
Informativeness
3
Presentation
1
Soundness
2
Read review
Paper summary

The paper argues that models trained only to predict distances (e.g., whether two nodes are linked) cannot reliably establish the direction of hierarchical depth: they capture how far apart items are, but not which is “higher” in the hierarchy. In hyperbolic embeddings, depth is commonly read as distance from the center of the embedding, but the whole embedding can be moved without changing any pairwise distance, which changes these depth readouts.

The authors call the missing information a “semantic frame” (e.g., which node is the root, or which direction is more general). They argue that existing approaches don’t provide it: Plain hyperbolic models never learn it, centering methods such as HIE and HypStructure give a stable but not necessarily semantic reference, and entailment cones learn local order but not global depth. They provide theoretical bounds on this ambiguity, and propose FramePosterior, which uses a small number of nodes with known depth (“anchors”) to infer the frame without retraining the model. Experiments audit several published models and test recovery on Disease, WordNet and ImageNet-100

Critical strengths
  • The authors identify an under-appreciated issue with a common interpretation in hyperbolic models (distance from the origin as depth or uncertainty), and explains it with a simple example: a three-node path where either endpoint could be the root.

  • The audits apply to published methods and are carefully controlled.

  • The proposed method leaves the trained model unchanged, so the original model's performance is unaffected and there is no additional training cost. Finally, it needs only a few nodes with known depth to work.

Critical weaknesses
  • The paper does not give a concrete example of an application where this ambiguity causes a real failure. Most failures are shown by deliberately applying a transformation to a trained model, which would not happen in practice. A clear real-world case where a practitioner relying on radius-based depth would reach a wrong conclusion would make the problem much more convincing.

  • The paper is quite difficult to follow for a non-specialist. The abstract and introduction are dense with terminology, and the main message is spread across many loosely connected experiments and a long appendix. A clearer separation between (1) the problem, (2) the theory, (3) practical failure modes of existing methods, and (4) the proposed implementation, with one running example carried through the paper, would help.

  • FramePosterior is given extra information (a few nodes with known depth) that the models it is compared against do not have. It is therefore unclear how much of its advantage comes from the method itself, and how much simply comes from having this extra information. A comparison against simpler methods given the same information would help clarify this.

Questions for authors
  • Can the authors give a concrete scenario where this ambiguity leads to a wrong decision in practice, without an artificial transformation being applied?

  • Do models trained with different random seeds actually end up with different frames? Would this indciate that the problem originates from the training stage itself?

  • How many anchors are needed before the method becomes reliable, and how does performance change as the number of anchors decreases?

  • What happens if the anchors themselves are wrong or poorly chosen? Is there a way to tell when the method’s answer can’t be trusted?

3 comments
AuthorsOpenReview ↗

Response to Reviewer 77kn

We address the practical-failure, seed, anchor-budget and reliability questions separately, with equal-information comparisons to identify what comes from the anchors and what comes from inference.

Q1. A natural failure without an artificial transformation? The isometry experiment establishes controlled fixed-reference sensitivity. Additional audits of untransformed centered models and checkpoint transfers found no natural reference error that anchors repaired. Centering constraints and hierarchy supervision are included in the complete observation law and can supply orientation information. The theorem diagnoses ambiguity under a specified invariant law; the intervention tests a frozen readout. We now distinguish these claims throughout. (Table 7, p. 21) (Appendix E.3, p. 23)

Q2. Do seeds produce different frames, indicating a training-stage cause? Across all 45 pairs of ten full HypStructure checkpoints, radius rankings on the same 60 held-out taxonomy nodes have a mean Spearman correlation of 0.9951 and ordinal-decision disagreement 2.168%. This quantifies readout variation. Distances and learned content also change between seeds, so these observations alone cannot attribute variation specifically to frame drift. (Appendix E.3.1, p. 24)

Q3. How many anchors are needed, especially at smaller budgets? Under the fixed WordNet protocol (20% corruption, farthest anchors, T=1, top-64), 4/8/16/32/64 anchors yield 91.548/91.789/98.225/98.202/98.314% ordinal accuracy. Respective 95% half-widths across 50 corruption trials are 0.559/0.611/0.047/0.006/0.009 percentage points. The largest gain occurs when the budget increases from eight to sixteen anchors. This ablation and the matched-comparator table use separate trial streams. The required count depends on the queries, chart adequacy, signature separation and noise. (Table 3, p. 10)

Q4. Wrong or poorly selected anchors, and when to distrust an answer? Sixteen clustered Disease anchors yield 84.39%, almost identical to eight (84.38%); simple offset correction handles common +2 bias better than the original posterior. HPO stress tests also show root-dependent rankings. Performance therefore depends on where anchors are placed, whether their errors share a bias, and which semantic depth is specified. (Appendix F.2.1, p. 28) (Appendix G.6, p. 39) (Appendix G.7, p. 40)

Internal vote concentration measures agreement among charts. Actual errors are evaluated separately with additional semantic-pair labels. At 2% target risk, WordNet coverage with 8192 clean calibration labels is 97.50/97.82% across two corruption conditions, versus 9.56/5.65% with 512 labels; a simple tie rule is equally effective. At 5% target risk and 2048 labels, warm32 HPO mean posterior coverage is 7.42%, versus landmark-64 ridge's 75.24%. Reliability therefore depends on both the model and the external information budget. (Appendix G.2, p. 36) (Appendix F.4.3, p. 35)

For corrupted selection labels, the corrected protocol isolates unordered pairs and reversals across selection, calibration and test. Known independent symmetric flip-rate correction is evaluated separately from conservative contamination correction, which mostly abstains. (Appendix G.4, p. 38)

An additional option uses 1024 possibly corrupted selection labels plus 1024 independent clean verification labels, beyond the depth anchors. Across 240 saved Disease cases sharing 30 roots, mean selected-rule coverage is 73.17% at 5% target risk, versus fixed posterior's 65.14%. No accepted test set exceeded 5% empirical error in these runs. Selection-label accuracy first chooses a rule; independent clean verification then selects its highest-coverage certified threshold. The five selection-label conditions give coverage means that are equal only after rounding; some selected rules change. The finite-population audit requires correct offline oracle verification labels and the stated sampling scheme. It applies to simple rules as well; unknown corruption of verification labels remains outside this result. (Appendix G.5, p. 38)

Equal information and presentation. With the same WordNet anchors, posterior/query-wise uniform tie handling attain 98.04/98.17%. ImageNet single/ensemble calibration share the same 32 noisy values and stored source radii/decoder; aggregation reduces MAE by 0.00345 levels. The abstract, introduction, Figures 1–6 and Algorithm 1 now separate evidence, inference and reliability; the appendix groups protocols and controls by scientific question. (Table 2, p. 9) (Appendix E.3.1, p. 23) (Figure 5, p. 9) (Algorithm 1, p. 6)

Thank you for the response and for answering my questions directly.

On a real-world failure (Q1), the authors state that they “found no natural reference error that anchors repaired” in untransformed models. On seeds (Q2), independently trained models give nearly identical depth rankings (Spearman 0.995). On equal information, a simple tie-breaking rule with the same anchors (98.17%) performs on par with FramePosterior (98.04%).

These answers clarify the scope of the paper, but they also suggest that the problem does not commonly arise in practice, and that the benefit comes mainly from the anchors rather than the method itself. As I am not able to assess the theoretical contribution, I will maintain my score and defer to the other reviewers on that aspect.

AuthorsOpenReview ↗

Follow-up on stability and semantic-readout validity

Thank you for the clear follow-up. We agree that the matched WordNet comparison does not establish a posterior-aggregation advantage over the strongest same-anchor rule, and that a naturally occurring, anchor-repairable reference failure remains missing.

A distinction in the seed evidence. In the same untransformed ten-checkpoint audit, mean cross-seed radial Spearman correlation is 0.9951, while mean radial depth-order accuracy is 74.82% on 1,518 unequal-depth pairs among the same 60 held-out taxonomy nodes. The checkpoints were trained with centering and hierarchy supervision. Thus, the correlation measures agreement across checkpoints, not accuracy against the taxonomy. These errors do not establish that the reference is wrong or that anchors can repair it; they show why reproducibility and semantic-readout validity need separate evaluation. (Appendix E.3.1, p. 24)

The contribution beyond aggregation. The identification question is which requested answers the available observations determine. Different roots need not be distinguished when they give the same answers; conversely, even a known root can leave a distance-based depth map inaccurate. The evidence conditions and matched controls separate reference recovery, depth-map adequacy and the effect of the decision rule. These tests guide the choice between collecting reference labels, changing the depth map and using a simpler rule with the same labels. (Section 2; Table 2, p. 9)

We would appreciate a reassessment of the paper's identification/evaluation contribution and of whether the revised explanation and figures resolve your original presentation concern. The absence of a natural, anchor-repairable reference failure remains an explicit limitation.

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

The paper claims that in geometric embeddings, just fitting pairwise distances is not enough as it doesn't determine which point is the root and which direction is depth as an isometry preserves every distance, but does not necessarily preserve the origin. The authors call this the semantic frame, formalize when it is identifiable, give a lower bound when it is not, and bound how many noisy anchors can recover the frame. The authors propose FramePosterior which gives each candidate root a score based on agreement with the anchors. An answer to a query is then a weighted sum of each root's answer. If the sum is close to zero, the roots disagree and their method needs another anchor.

There are three sets of experiments. The first applies isometries to trained models (their method has no training) and shows that the origin-based depth / uncertainty readouts do indeed change, showing that the problem exists. The second gives noisy anchors for three datasets and show that their method can recover the depth order. The third is a real-world scenario on an airport graph. They show that their method raises the success rate over a greedy approach.

Critical strengths

(1) I like the choice of "decision-level identifiability". Instead of trying to just identify the root, the paper doesn't care if two roots are different as long as they give the same answer to the query. This seems like the right way to attack the problem as I don't see why we need uniqueness. For example with DAGs, you can have multiple roots and as long as they give the same answer to a query, the paper's set-up doesn't need to tell them apart. I see this show up in Table 1 where on Disease, even though the root is identifiable only a small amount of the time, the pair accuracy is very high.

Critical weaknesses

(1) I don't think that the experiments show that the problem exists in practice. In figure 4b, as the authors increase the strength of the slide by the isometry, the accuracy degrades, showing that the models do suffer degradation as the origin changes. However, and this may be coming from my lack of knowledge of the area, I can't tell why you would want to do this. If the point is that you don't know the frame, then I can agree, but in every setting the authors consider, they have access to the frame (or mask it). In practice, it seems like you know the frame so this wouldn't be an issue. In the methods they test on, they push towards on origin. HypStructure has the centering loss, HIE tries to estimate a center, and UNCHA uses a norm. So if these do choose an origin, then I think the claim might be that this is the wrong origin and an isometry needs to be applied. However, I do not see evidence that convinces me of this either.

In D.2, the claim is made that HIE doesn't find the right origin (although it is consistent), but in D.3, the authors say that they trained HGCN and did three readouts (distance from origin, from HIE center, and from correct root), and all did poorly. Since the correct root fails, I don't think the problem is that the frame is incorrect, it's the depth information is not carried in any frame. This is a different problem than the one the paper says.

(2) I don't exactly understand what the second and third sets of experiments (4.2 and 4.3) are actually testing. 4.2 says that it tests whether "limited evidence restore useful decisions", and the third tests the case where the root changes at every query. However, it seems like neither are actually using the learned embeddings, but rather just using the graph with the metric being the shortest path (4.2 says it uses the graph for WordNet and Disease and for the routing in 4.3.). I understand that the authors say they use this when the learned metric doesn't give the direction (line 975), but this still seems like we're no longer working with geometric representations, which is what the paper is supposed to be about.

As someone not familiar with the field, I found this a dense paper to read. I think including more background would make the paper more accessible.

Questions for authors

Can you address my questions / confusion in the weaknesses section?

1 comment
AuthorsOpenReview ↗

Response to Reviewer sKvq

We have reorganized the paper around the distinction you highlighted: recovering the requested decisions can require less information than identifying a unique root. The revision also separates reference ambiguity from a representation's ability to express semantic depth.

1. What does the isometry experiment establish, and do existing centers already solve the problem? An isometry is a controlled test of a fixed-reference readout: it preserves the checked native outputs while changing the coordinates used for semantic interpretation. Applying the same isometry to the points and their reference preserves their relative distances. The orbit theorem has a stronger premise—equal complete observation laws for two admissible states—which must include centering constraints and hierarchy supervision. Preserving a frozen distance decoder alone does not establish that premise for the complete trained system. We now state these scopes explicitly. (Table 7, p. 21)

The additional untransformed checks found no natural reference error repaired by anchors. Across ten full HypStructure seeds, nested center selection consistently preferred the origin over the calibration midpoint and all-leaf midpoint; checkpoint-transfer tests likewise found no reference error that anchors corrected under the audit criteria. The intervention establishes readout sensitivity, while the prevalence of naturally lost or incorrect references remains an empirical question. (Appendix E.3, p. 23)

2. Why does supplying the correct root sometimes fail? Supplying the correct root fixes the reference, but the tested distance-to-root readout can still fail to reproduce the target depth. We retain the oracle-root negative control and identify chart adequacy as a separate requirement. A supplementary joint-training diagnostic improves native link AUC from 0.8706 to 0.9253, but changes the encoder and is reported separately from frozen-model FramePosterior. Even with these improved learned charts, posterior recovery reaches 84.80–85.27%, versus 90.92–92.65% for landmark-64 ridge with the same 32 labels and equal per-node numeric-table bytes. Encoder improvement therefore does not establish a semantic-recovery advantage. (Appendix F.4.3, p. 35)

3. What do graph recovery and routing test? WordNet and Disease recovery use supplied undirected shortest-path distances and depth anchors. They test recovery from structural geometry. The metric source is now explicit in the method definition and protocol. On WordNet, posterior voting reaches 98.04%, versus 98.17% for same-anchor query-wise uniform tie handling and 96.81% for MAP plus all one-hop neighbors. These comparisons measure how posterior weighting changes decisions when all methods receive the same anchors. (Table 2, p. 9) (Appendix F.1.1, p. 28)

Complete-candidate Disease analysis finds disagreement on 64.70% of queries; full voting and minimum-loss uniform voting attain 77.08/77.13% on that hard subset. Figure 3c now reports the all-candidate hard subset; the earlier top-128 proxy remains in the appendix. (Appendix F.3.2, p. 32)

Routing has a known destination and additional landmark topology. It tests topology and learned-distance tie breaking. The stronger state-accounted BFS-tree comparator delivers 100%, versus hybrid's 77.912%, using less serialized persistent-plus-packet state. Routing consequently supports application attribution rather than an advantage of unknown-frame inference. (Appendix H.4, p. 41)

4. Presentation. The abstract and introduction explain how hierarchical depth, radial uncertainty and known-destination forwarding use a task reference, how their decisions are evaluated, and which evidence supports them. Figures 1–2 carry a running example; Figures 3–6 separate structural recovery, controlled interventions and matched increments. Algorithm 1 states the implemented Gibbs/top-K ordinal procedure. The appendix is organized by scientific question, with a reader guide to the theory, proofs and controls. (Figure 3, p. 7) (Figure 5, p. 9) (Algorithm 1, p. 6)

The main text now includes the matched WordNet and anchor-budget tables and the visual recovery curve. (Table 2, p. 9) (Table 3, p. 10) (Figure 6, p. 10)

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of revisions and additional controls

We have revised the manuscript around three contributions: decision-level semantic-frame identification, evidence requirements for noisy anchors and finite DAG frames, and frozen-predictor inference evaluated through matched controls.

Problem and scope. The identifiability analysis uses the complete observation law, including centering constraints and directed supervision. Controlled isometries test fixed-reference readouts while preserving checked native outputs. Untransformed-model and checkpoint-transfer checks found no natural reference error repaired by anchors. (Table 7, p. 21)

Matched comparisons. With the same undirected structure, 16 anchors and 20% corruption, WordNet posterior voting and uniform-top-64 tie handling attain 98.04% and 98.17% ordinal accuracy. Neighboring-root, regression and complete-candidate difficulty controls distinguish anchor information from the effect of inference. (Table 2, p. 9) ImageNet single-chart and ensemble calibration share 32 noisy values, stored source radii and the decoder; their paired MAE difference is 0.00345 levels. (Appendix E.3.1, p. 23) Known-destination routing tests learned-distance and topology contributions; the stronger BFS-tree comparator exceeds the hybrid in delivery and accounted state. (Appendix H.4, p. 41)

Presentation and materials. The revision clarifies references as graph roots or embedding centers, explains why a correct root can still give inaccurate depth readouts, and reorganizes the figures, algorithm and appendix. The prepared supplementary archive contains the marked manuscript and the original experimental code release, with a machine-specific plotting path removed. Implementations of the additional discussion-stage analyses are outside this preserved code release. Detailed responses address each review separately.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.