acceptodds
Under review as a conference paper at ICLR 2027

Vocabulary-Relative Traceability: Representability Before Complexity

Abstract

Exact faithfulness in interpretable machine learning requires a declared explanation vocabulary that can express the frozen model decision. We test this requirement before minimizing explanation complexity. For binary decisions, the traceability budget is the minimum number of distinct named boundaries in an almost-surely exact Boolean transcript; if none exists. Under the stated regularity contract, an essential-boundary characterization yields three exhaustive obstructions: information loss, unavailable curved structure, and a missing flat direction. On finite samples, we compute maximum agreement over deterministic interface readouts, which also bounds every readout of every subinterface. A hard ECG bottleneck exposes a naming gap; PTB-XL supplies a separate post-hoc one-boundary witness resolving an eight-record sample obstruction. On public CUB concept-bottleneck checkpoints, an ordinal policy fixed before held-out endpoint outputs were opened refines only the displayed interface. Its primary confirmation-sample agreement ceiling rises from to , removing unavoidable disagreements and establishing exact sample factorization with a finite transcript. Nearly half the records retain shared codes; two fixed checkpoints also reach zero conflict. With 16 bins per concept, exact sample factorization holds for one checkpoint but fails for two: resolution alone does not determine sufficiency. The audit determines when exact explanation requires interface augmentation rather than further readout optimization.

Primary area:
interpretability and explainable AI
Keywords:
interpretable machine learning, explanation faithfulness, concept bottleneck models, decision sufficiency, representability, interface augmentation

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_a33qOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper studies the problem of constructing a global surrogate for a frozen model using logical rules defined over a set of features. These features may correspond to concepts in datasets such as CUB, concepts in CBMs, or features extracted by SAEs. The proposed framework applies thresholding to these features, resulting in binary indicators (selected/unselected), and then seeks logical rules over these indicators that capture the behavior of the frozen model. The authors refer to these logical rules as an "explanation". The paper identifies three fundamental representational limitations of such explanations: (1) information loss (B1), (2) unavailable curved structure (B2), and (3) missing flat direction (B3). These limitations characterize cases in which logical rules over thresholded features cannot faithfully represent the underlying model. The authors then conduct experiments across several benchmarks, including CUB, PTB-XL, and WikiText, using models such as GPT-2 and concept-bottleneck architectures for image and ECG classification. The experimental evaluation mainly focuses on demonstrating information loss (B1). Specifically, for each benchmark, the authors identify inputs that produce identical thresholded feature representations but different model predictions. Therefore, any logical rule defined only over these thresholded features must necessarily fail to represent the model correctly on at least one of these inputs, which is basically the limitation described by B1. The authors also show that refining the representation can remove this information loss. For example, on CUB, moving from binary concept thresholds to 32 levels per concept eliminates all such conflicting inputs on the evaluated sample.

Critical strengths
  1. The idea of constructing a global surrogate for a frozen model using logical rules over a set of interpretable features could be useful. However, I believe the paper would benefit from a clearer motivation and justification for why this particular form of surrogate is desirable (e.g., what other more standard forms are missing).

  2. The paper considers a relatively diverse set of settings, including CUB, CBMs, and SAEs, which helps illustrate applicability across different interpretability settings.

  3. I appreciate the inclusion of formal theoretical results, which I believe are important and still relatively uncommon in interpretability research. That said, the specific results presented here seem somewhat expected and, in my view, do not provide particularly significant new insights into interpretability (see weaknesses below).

Critical weaknesses
  1. Writing needs improvement. The paper is currently (very) difficult to follow. It took me several readings to fully understand the setting, the main objectives, and the contributions. Part of this difficulty is due to many paper-specific definitions and terms, such as "Joint1" (which, as far as I could tell, is never even formally defined), B1, B2, B3, Gate 1, and Gate 2. These make the overall framework unnecessarily complicated to follow. Additionally, several technical sections are written in a way that makes the arguments really difficult to follow. More fundamentally, the paper introduces the term "explanation" to refer to a logical rule intended to serve as a global surrogate for a frozen model over thresholded features. This is not necessarily the conventional meaning of an explanation in interpretability research, and the choice is not motivated. In particular, the paper should explain much earlier and more explicitly what setting it considers, why this setting is important, what problem it aims to solve, and how the proposed contributions address that problem. Overall, the motivation, problem formulation, and main contributions are presented very vaguely, making it difficult to understand what the paper actually does. I strongly encourage the authors to substantially revise the presentation and improve the accessibility of the paper.

  2. The “explanation” definition is not well motivated, and it seems to be the main source of the limitations the paper identifies. The paper shows that thresholding concepts into binary indicators discards information that the CBM head relies on, and that using finer thresholds recovers it. This is a fairly expected result, but my main concern is with the setup itself. The information loss comes from the paper’s own choice to define an explanation as a logical rule over thresholded features. Standard concept-based and mechanistic interpretability methods typically work with continuous activations or directions instead. A CBM head, for example, is a weighted combination of concept scores, so it keeps this information by construction. The paper does not explain why its more restrictive explanation definition should be preferred. In fact, under the paper’s own theory, this standard CBM explanation cannot be expressed at all: a head that is linear in continuous concept scores produces decision boundaries that are tilted or curved relative to per-concept thresholds, so the budget K is infinite no matter how many bins are used. In other words, the framework treats the most common way of explaining a CBM as inexpressible, and the 32-bin “repair” on CUB holds only for the finite set of evaluated images, not in the sense the theory defines. Finally, thresholding is not the common thread across the experiments either. In the ECG concept bottleneck, the concepts are already binary inside the model, and the failures arise only because three unnamed units that the head explicitly uses are left out of the explanation. Which diseases fail is therefore known in advance from the architecture, which makes this experiment close to trivial.

  3. Lack of motivation for using global rather than local surrogates. Interpretability methods often aim to explain individual predictions through local explanations or local surrogates, rather than attempting to capture the behavior of an entire model using a single global surrogate. Such a local formulation may be more realistic and could potentially avoid some of the representational limitations identified here. Overall, I believe the paper needs to provide a much stronger justification for its particular choice of explanation framework and clarify how it relates to existing interpretability approaches.

  4. Theoretical results are expected. While I appreciate the theoretical component of the paper, the main results seem relatively expected given the requirement that logical rules over thresholded features should perfectly represent the original model. In particular, the three limitations (B1, B2, and B3) appear to characterize fairly natural representational mismatches between the restricted logical-rule framework and the original model. Therefore, I am not convinced that these results offer important new insights into interpretability or reveal surprising limitations that would not already be anticipated from the formulation.

  5. Experiments are mainly on B1 and not on B2 and B3. The experimental evaluation does not seem to fully align the breadth of the theoretical analysis. While the theoretical section introduces three distinct representational limitations (B1, B2, and B3), the experiments focus mainly on demonstrating B1. In particular, the experiments show that two inputs may have identical thresholded feature representations while producing different model predictions. Although this is a valid demonstration of the way the authors defined information loss, it leaves the practical relevance of B2 and B3 largely unexplored.

  6. Mismatch between the theory and experiments definition. The theoretical framework requires the logical-rule representation to agree with the original model almost surely over the full data distribution. This seems like a rather strong requirement, particularly for a global surrogate intended to approximate a potentially complex model. Moreover, there appears to be a mismatch between the theoretical and experimental settings: while the theoretical analysis considers agreement over the full (continuous) data distribution, the experiments demonstrate the limitations using examples drawn from a finite sample of points.

  7. Missing related work. Concept leakage is a well established phenomena which seems to align closely with the ideas discussed here and is not discussed/compared to at all in the paper.

Questions for authors
  1. In settings such as CBMs and SAEs, explanations are often expressed through directions, continuous activations, or mappings in latent space, rather than binary thresholded features. Can the proposed framework accommodate such representations? It seems possible to me that a significant portion of the information loss identified in the paper arises specifically from thresholding, which is not necessarily the standard way these features are interpreted in mechanistic interpretability. How do the authors view this limitation?

  2. Would it make sense to consider a local version of the proposed explanation framework, where different logical rules are constructed for different inputs or local regions of the input space? This seems more closely aligned with several common interpretability approaches and could potentially improve some of the representational limitations discussed in the paper.

  3. Some of the assumptions in the theoretical analysis, particularly the requirement of exact (almost-sure) agreement between the logical-rule representation and the original model over the full data distribution, seem quite strong. Could these assumptions be relaxed, for instance, by considering approximate agreement? Would the main theoretical conclusions still hold under such relaxations?

  4. Is there a way to experimentally evaluate the limitations associated with B2 and B3, similarly to the evaluation of B1?

Reviewer_3FZDOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper is about model explainability. A frozen model takes decisions, and an explanation sees the inputs through a representation F (which can be the model's own features or separate measurements). The explanation can only ask yes/no questions of the form "is this named quantity above a threshold?"; the list of named quantities is the vocabulary. A transcript is a rule that combines the answers to these questions with AND/OR/NOT. The first research question (Gate I) is whether some transcript reproduces the model's decisions exactly. Then, supposing it does (Gate II), what is the minimal number of distinct thresholded questions such a transcript needs. On the theory side, the paper tries to show that when no transcript exists it is for one of three reasons: missing information, a curved boundary, or a missing direction. And it finally gives an experimental evaluation of that.

Critical strengths

Checking representability before minimising complexity seems a clean and reusable framing.

Critical weaknesses

To me the presentation is the main problem of the paper.

No contribution statement. No related work section and no limitations section.

Moreover, I list below, the terms and passages that are not understandable when reading the main text linearly. This is not exhaustive, as I stopped my reading early.

Abstract

"Exact faithfulness" (l. 11): never defined, here or later.
"Declared explanation vocabulary" (l. 11–12): vocabulary? declared by whom?
"Distinct named boundaries in an almost-surely exact Boolean transcript" (l. 14–15): "transcript" is defined only at l. 97.
"Deterministic interface readouts", "every readout of every subinterface" (l. 18–19)
"A hard ECG bottleneck exposes a naming gap" (l. 19–20)
"Post-hoc one-boundary witness resolving an eight-record sample obstruction" (l. 20–21)
"An ordinal policy fixed before held-out endpoint outputs were opened refines only the displayed interface" (l. 21–23)
"Primary confirmation-sample agreement ceiling", "exact sample factorization" (l. 23–24)
"Records retain shared codes", "zero conflict" (l. 24–25)
"16 bins per concept", "resolution alone does not determine sufficiency" (l. 25–27)
"Interface augmentation rather than further readout optimization" (l. 28–29)
ECG, PTB-XL and CUB are named without saying what they are.

1 Introduction

"Decision sufficiency of a learned representation" (l. 38).
"A binary display or named subset" (l. 40): first use of "display", never defined. How is it different from a representation or an interface?
"Restoring the activation order of displayed feature IDs" (l. 42–43): only understandable after §4.2 (l. 286–289).
This whole paragraph (l. 44–50): it uses FF, WW, B1–B3, "presented sample", "observed fiber", "regularity contract", "boundary patch" and "unnamed normal" before §2 defines any of them.
"Unavoidable disagreement" (l. 61): defined only in §2 (Eq. 1).
"Regular obstruction classification", "admissible augmentation class", "bounded selection-and-certification test" (l. 65–67)
This whole paragraph (l. 68–77): "distinct evidence roles", "previously unaudited Joint1 endpoint" (Joint1 is never explained, also at l. 245), "ordinal policy", "hard ECG bottleneck", "unnamed units", "sparse-feature display", "fixed-artifact example", "declared nine-scalar vector", "typed rate atom" (an atom is only defined at l. 96).
"Aggregation and tree-compression results instead approximate a base class or minimize an equivalent tree with inherited splits" (l. 80–82)
"A separately declared vocabulary" (l. 80): separate from what?
"Our certificate is indexed by the frozen RR, FF, named atom vocabulary WW, measure and equality contract" (l. 85–86): RR appears here for the first time, and none of these objects is defined yet.
"Display audit" (l. 88)
Figure 1
It took me a while to understand that the grey lines are the named cuts. Nothing in the figure says so.
The B1 panel is the least clear: the node "z" and the crossed orange arrow are not explained.
The B2 and B3 panels did not really help me understand B2 and B3. I get it but unsure that schemas are really helpful.
The caption refers to "the controlled study" (l. 125–126), which is introduced only in §5.

2 A vocabulary audit has two gates

"Named hyperplanes" (l. 102): only the half-spaces H(w,t)H(w,t) are defined, not the hyperplanes.
"Declared feature vector" (l. 135)
Proposition 2 (l. 136–138): "let ss be a score" without saying whose score, and it is not said that RτRτ​ is the audited decision. " "split", for a group, is not defined.
Equation 2: the maximum is over gg, but gg is not defined.
Questions for authors

None

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

The paper asks whether a frozen model's decision can be reproduced exactly (μ\mu-a.s.) by a Boolean formula over a declared vocabulary of named threshold atoms (Gate I), and only then how many distinct named hyperplanes the transcript needs (Gate II). For binary decisions the traceability budget KK is infinite when no transcript exists; Thm. 7/11 equates KK with the smallest named arrangement covering a support-based essential boundary, and Thm. 12 splits K=∞K=\infty into non-factorization (B1), a curved boundary patch (B2) or a flat patch with an unnamed normal (B3) under an o-minimal C2C^2 contract. On samples, the number of records outside their cell's majority label, UmcU_{mc} (Eq. 1, Prop. 3), gives the exact best agreement of any deterministic readout of an atom family and an upper bound for every subfamily, and the minimum-cut problem is NP-complete via set cover (Props. 5/14). Validation is by sample-level audits: CUB Joint1 bottleneck checkpoints, where a pre-fixed 32-level ordinal display lifts the confirmation-half ceiling from 98.83% to 100% (Table 3); an ECG hard bottleneck whose named subset fails at 13/66 cells; ordered versus unordered top-3 SAE IDs on GPT-2; a PTB-XL ResNet1D whose 30/66 failing operating points trace to one all-zero eight-record fiber, which a post-hoc RR atom separates at the LVH top-440 decision; and 3×100 synthetic worlds with planted obstructions (Table 6).

Critical strengths

A correct, carefully scoped formalization of exact vocabulary-relative representability. Defining KK on μ\mu-a.e. classes (Def. 9, l.887-898) makes it invariant to null edits, and Thm. 11 reduces it to covering ∂νeD\partial^e_\nu D by named hyperplanes; I went through Lemma 10, Thms. 11-12, Prop. 13 and the set-cover reduction of Prop. 14 (E.5.1-E.5.5, pp.23-25) and found no errors. The existence-before-complexity order separates "this vocabulary cannot express the decision" from "the search was weak", which a failed explanation search otherwise leaves ambiguous (l.049-050), and the paper says where its contract stops (concrete versus quotient classes, l.877-881; "rules outside this contract remain unclassified", l.189-190).

Bounded, checkable claims, reported with unusual candor. Prop. 3 gives an exact, optimizer-independent ceiling, and M.5 shows what it buys: at q2 a discovery-fitted ridge readout makes 198 confirmation errors, of which only U=34U=34 are forced by the interface (Table 20). The paper also reports analyses that cut against its own story: the GPT-2 occupancy reference that reproduces 96.5% of the rank gain (l.304-309, Table 23), "We therefore make no aggregate model-specific claim" for PTB-XL (l.2423), the failed q16 and compact-repair stages (l.2592-2596, l.2636-2645) and post-hoc labels on the RR atom (l.347-349). Every number I recomputed matches (98.83%, 47.53%, 3,472, Table 4, the Table 18 ceilings, 0.142 pp, the 96.3% Wilson bound, the Table 8 sign tests).

Critical weaknesses

On audited models, the essential-boundary theorem and B2/B3 are never exercised and B1 only on tied discrete or degenerate codes, while the sample tools that are exercised are classical, with Gate I unattributed. Thms. 7/11/12 assume ν∼λm\nu\sim\lambda_m, which implies (3); under (3) with independent draws no two records share a declared vector almost surely, so the tie test of Prop. 2 cannot fire for any rule and Prop. 6 yields only finite floors mn≤Km_n\le K. Every real-data Gate-I failure therefore sits on a finite or degenerate code outside the theorems (CUB Iq\mathcal I_q grids, the 19-bit ECG code, top-3 SAE ID sets, and the PTB-XL all-zero fiber, in a cohort for which "Theorem 11 is accordingly not asserted", l.1604-1605), and the CUB and GPT-2 decisions are multiclass, where the KK theory is not extended (l.140-141). What the trained models exercise is B1 (refine or extend a tied fiber, Table 1). Since these fibers carry positive probability (the condition l.130-133 sets for a population conclusion), the tie step in the proof of Prop. 6 (A.2) lifts each split to population non-factorization almost surely; on the near-continuous PTB-XL scalars, fold 10's one tied fiber is the all-zero one (8 records there, 96 in training, l.1601), which the ResNet1D splits at 30/66 operating points and whose mass comes from what F.3 calls an extraction degeneracy (l.1601-1602), on a cohort where (3) holds "only approximately" (l.1598). B2 and B3 appear only in synthetic worlds where the generator supplies the class (l.125-126) and MI and univariate baselines also recover 100/100 (Table 8; l.419-420 concedes these rates "do not distinguish informed selectors"). The paper scopes much of this (Table 2, l.130-133, Remark 1, and F.4's "with no tie group there is nothing for obstruction B1 to act on, whatever model is asked", l.1626-1627), yet the second and third contributions (l.064-067), Fig. 1 (whose flow runs from "Gate I fails" to "diagnose B1-B3; restrict augmentation") and the conclusion (l.457-461) present diagnosing B1-B3 as the step after a failed Gate I, although only B1 is ever diagnosed on an audited model. F.6 calls the PTB-XL census "obstruction B1 of Theorem 12 seen on a finite sample" (l.1692-1693), which is right about B1 but not about Theorem 12, whose contract does not hold on that cohort. The machinery the experiments do use is known: Umc/nU_{mc}/n and its monotonicity in the atom family (Eqs. 1-2, Prop. 3) are the inconsistency rate of consistency-based feature selection (e.g., Dash & Liu, AIJ 2003), Umc=0U_{mc}=0 is decision-table consistency in rough-set theory (Pawlak, 1982), and while Sec. 6 places Gate II in "the classical minimum-test-set/set-cover family" (l.431-434), Gate I gets no attribution even though the same passage names "the preceding vocabulary audit" as the contribution. The paper anticipates its CUB and ECG outcomes (l.039-041, l.240-241; N.4 ties the anchors to morphology such as the delta wave that "resisted expression as a threshold on the named measurement set during development", l.2767-2770), but it does not cite the soft-versus-hard concept-leakage work of which the binary-display result is a sample-level instance (Mahinpei et al., 2021; Margeloiu et al., 2021; Havasi, Parbhoo & Doshi-Velez, NeurIPS 2022), and beyond one-clause contrasts at l.082-085 it does not engage abductive or prime-implicant explanations (Ignatiev, Narodytska & Marques-Silva, AAAI 2019; Darwiche & Hirth, ECAI 2020) or the complexity of interpretability (Barceló et al., NeurIPS 2020); I could not find leakage, rough sets, inconsistency rates or any of these authors anywhere in the text or appendix ("rough" occurs only in the Chlebus & Nguyen venue name). Adding these citations would not by itself change my score. This weakness still weighs most in my rating for two reasons: l.433-434 names the vocabulary audit as the contribution, together with the obstruction-conditioned evidence chain, and on samples that audit is the inconsistency rate and decision-table consistency; and the new population part (the null-invariant budget, Thms. 7/11 and Thm. 12's classification) yields no finding about an audited model beyond the B1 lift, which needs only the elementary argument of Prop. 6 and, on the near-continuous PTB-XL scalars, rests on an extraction artifact. A bridge would resolve it, either conditions under which the growth of mnm_n with nn flags K=∞K=\infty under (3) or an analytic B2/B3 verdict for a trained head with the prescribed augmentation tested; failing that, the abstract, Fig. 1 and conclusion should be rescoped to tie-based failures on discrete or degenerate codes, with a related-work section that states what is added beyond these literatures.

Apart from GPT-2, the main-text readings of the real-data results are not separated from refinement, chance or construction. For CUB the grids are nested (l.2605-2606), so UmcU_{mc} cannot increase along them (Prop. 3), 1,520/2,897 = 52.47% of confirmation records are q32 singletons and hence conflict-free by definition, and the paper's own Prop. 18 (l.1120-1136) shows that plug-in optima are biased upward on sparse partitions and reach 1 once no two records share a cell. Fixing q32 before opening any Joint1 output removes grid-selection bias but does not supply this calibration. The success guard ("at least 5% of all confirmation records in non-singleton q32 fibers", l.2609-2611) is global and does not show whether the records from mixed q2 fibers, where the 34 conflicts sit, still share q32 codes, and neither the within-parent occupancy reference that reproduces 95.1-96.5% of the GPT-2 gain (Table 23) nor the fixed-count reference used for PTB-XL (Table 17) has a CUB counterpart (I searched App. M for occupancy, permutation, within-parent and reference analyses). What is held out is the grid, not a readout: the transcript with at most 3,472 atoms is compiled on the confirmation half (l.2631-2633, stated plainly), and a discovery-fitted ridge readout agrees on 93.17%, 92.82% (the execution-frozen primary comparison) and 91.30% of confirmation records at q2/q16/q32, with all 252 q32 errors on the 1,510 records whose codes are unseen in discovery (Table 20; l.2692-2696). The abstract's q16 sentence rests on confirmation counts of 2/1/0, yet Seed 1 moves from 0 on discovery to 2 on confirmation (Table 19), and because UmcU_{mc} over fixed thresholds is superadditive over disjoint samples, Seed 1's U16U_{16} on all 5,794 images is at least 2; on the denominator apparently used for the other recipe (l.2594-2595) the primary checkpoint therefore also retains conflict, which weakens the Sec. 6 contrast (l.438-440). For PTB-XL the paper itself reports what calibrates the count and accordingly makes "no aggregate model-specific claim" (l.2423): all 30/66 failures come from one fiber whose nine scalars are zero (l.316-317), which F.3 calls a retained "extraction degeneracy" (l.1601-1602), and O = 30 against E0 = 27.30 (z = 0.80, Table 17; my evaluation of Eq. 10 gives 0.85/3.71/6.28/9.16 expected failing decisions at top-22/110/220/440 against 1/4/7/9 observed). My complaint there is only that the main text leaves the match out: the "Resolution control" paragraph (l.324-327) points to the reference without reporting it. The remaining layers are labeled candidly but are weak or fixed by construction. The post-hoc 400 ms atom isolates the record with the smallest RR (376 vs 446 ms, Eq. 12), and the number of piloted candidates is not reported (l.2459-2461); for a scalar unrelated to the decision the lone positive is extreme with probability 2/8, so kk piloted scalars would yield such a cut with probability 1−0.75k1-0.75^k (0.76 at k=5k=5), and in F.7, 34 of 224 score-blind witnesses on the same fiber isolated a singleton (l.1729-1730). The Gate-II bracket [8,18] overlaps the permutation brackets, as l.355-356 states. Check C2 makes any failure at the eight non-anchor ECG endpoints a "setup defect" (l.2742-2744), so "all 13 failures occur for WPW, IRBBB and CLBBB" (l.278-280) is guaranteed once the mask check passes, leaving as non-forced content only that 13 of the 18 anchored cells fail. Only GPT-2 clears its reference (at layer 8 the excess is 170.33 disagreements, 0.142 pp, and every layer lies above the reference's 97.5th percentile). The paper scopes the CUB and ECG layers to sample facts ("a held-out fixed-policy repair of resolution", l.273-274; N.5, l.2776-2779), and I do not doubt those facts, but Sec. 4.1 tests the diagnosis "as an action rule" (l.245) and the abstract and Sec. 6 draw the q16 contrast from the CUB counts; for those readings the CUB layer as reported does not yet separate a model property from refinement. Running the O.3 within-parent reference and singleton-first decomposition for Joint1 q2 -> q16/q32, reporting the q32 fiber sizes of records from mixed q2 fibers and U16/U32U_{16}/U_{32} on the pooled set and re-drawn halves, and moving the PTB-XL and ECG calibrations (with the pilot count) into Sec. 4 could resolve it.

Questions for authors
  1. How many Seed-1 confirmation records from mixed q2 fibers share q32 cells, and what are pooled and O.3 within-parent-permuted U16/U32U_{16}/U_{32}? If permuted U32=0U_{32}=0 is rare, my CUB concern largely goes away.
  2. Under (3), can any finite-sample statistic (e.g., mn↑Km_n\uparrow K a.s.) flag a population obstruction? If the Joint1 head is linear, are its pairwise decisions a B3 instance closed by adding wa−wbw_a-w_b?
  3. Are the eight all-zero PTB-XL records failed extractions (l.1601-1602), and do failures persist with PTB-XL+ scalars? How many candidates preceded the 400 ms atom, and how accurate are the ResNet1D and ECG bottleneck?
  4. What do Gate I and UmcU_{mc} add beyond inconsistency rates and decision-table consistency, and Gate II beyond rough-set optimal discretization (Chlebus & Nguyen, 1998, already cited)? How do CUB/ECG findings relate to concept leakage?
  5. Stanley (2007) editors: "Ezra Miller, Bernd Sturmfels, and Victor Reiner (eds.)" should read "Ezra Miller, Victor Reiner and Bernd Sturmfels"; Merity et al. (2016, arXiv:1609.07843) should read ICLR 2017. Please also add PTB-XL+ and Breiman et al. (1984) for CART.
  6. Expand ECG abbreviations, define Joint1 and typed atoms at first use, resolve the W0W_0/W1W_1/A0A_0 clashes, caption H.1/H.4 tables, and bring the F.3, K.2, M.5 and N.2-N.3 caveats into Sec. 4.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.