acceptodds
Under review as a conference paper at ICLR 2027

ESCROW: Learning Groups and Hierarchies from Streaming Categorical Data

Abstract

Categorical records arrive over time without supplied groups or parent relationships. A system must learn useful organization while avoiding groups that merely fit noise. ESCROW addresses this task through a data-source and industry-agnostic interface. Its central contribution is a complete conditional code that charges for groups, links, memberships, and cell owners. Each cell is encoded once, even when groups overlap. The code bounds unsupported selection under a stated no-association null; a separate result controls evidence accumulated from later blocks. Applied to COBWEB, selection raises mean membership F1 from 0.415 to 0.936 at depth one and from 0.622 to 1.000 at depth two, recovering all six reference groups and their links. It improves shallow recovery but lowers recovery at depths five through seven, and selects no groups on matched null streams. Recursive search separately achieves F1 above 0.999 in a 20-stream two-level confirmation. A semantic head shares sparse counts, and a memory head adapts trust in predictors. Matched experiments on Amazon and Wikidata compare classical, zero-shot, and few-shot LLM proposals. On Amazon, selected parent-based ESCROW saves 1.434 conditional bits per record, versus 1.032 for COBWEB and 0.705 for FISHDBC, with positive savings in all five orders. We measure structural recovery, conditional compression, and conventional prediction separately. Component studies isolate the two heads and ownership costs.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
streaming clustering, hierarchical clustering, minimum description length, categorical data, structure learning, model selection, online learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_1c2BOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper learns overlapping groups and parent links from categorical records that arrive over time. Its core is a conditional MDL code that pays for the graph, the memberships and one owner per record-field cell, with column counts fixed. Theorem 1 turns this code into a compression test with false selection probability at most 2−b2^{-b} under a column-permutation null, and Theorem 2 gives a Ville-type bound for evidence collected on fresh blocks. Two prediction heads add co-occurrence smoothing and fixed-share expert weights. Experiments use synthetic hierarchies of depth 1 to 7 and 600-record samples of Amazon Berkeley Objects and Wikidata, against COBWEB, FISHDBC, classical clusterers and prompted LLMs.

Critical strengths

Cell ownership is a clean fix for double counting when groups overlap, and the Kraft argument behind Lemma 2 and Theorem 1 is simple and correct as far as I checked. The certifier also works well as a filter on other methods' proposals. In Figure 3 it lifts COBWEB from 0.415 to 0.936 F1 at depth one and returns zero groups on every null stream.

I also like that native and pruned graphs are reported side by side and losses are stated openly, such as k-means beating ESCROW at s=0.65s = 0.65 (0.755 vs 0.511 ARI).

Critical weaknesses

The streaming framing is not supported by the setup. Prefixes are 300 or 600 records, recursive search revisits the stored prefix at each repair, and recursive proposals take up to 4,923.8 seconds on 600 Wikidata records (Section 8). There is no drift and no memory bound, so this is batch structure learning on small samples. Categorical stream and drift work such as LIMBO (Andritsos et al., 2004), COOLCAT (Barbará et al., 2002) and Cao et al. (2010) is missing. MDL clustering of categorical data (van Leeuwen et al., 2009) and hierarchical latent class models (Zhang, 2004) would be closer baselines than DP-means or BIRCH on multi-hot vectors.

The guarantee is weaker than the abstract suggests. The null means no association at all, and it says nothing about spurious groups once real structure exists (Section 3). All 345 pruned decisions are unchanged for bb from 2 to 30 (Appendix D.10), so the margin does little work. Gains are mixed too. In Table 2 selected COBWEB has the best mean F1 (0.781), and on Wikidata selection drops parent-based ESCROW from +0.236+0.236 to −0.845-0.845 bits saved. Internal labels like EHR05 and R27 make the paper hard to follow.

Questions for authors
  1. What happens when the hierarchy changes mid-stream, for example a group splits? Can repairs remove groups that lose support?
  2. How do time and memory per record grow with n, and can certification avoid revisiting the full prefix?
  3. Could you add an MDL categorical clustering method or a latent class model as a proposer under the same certifier?
  4. Is there a weaker null than full column independence under which false selection of single groups can be bounded?
4 comments

Thank you for your careful review, for checking the Kraft argument behind Lemma 2 and Theorem 1, and for the pointers to related work. We are glad you found cell ownership a clean fix for double counting, and that you liked seeing native and pruned graphs side by side with the losses stated openly. We answer each point below; changes in the revised PDF are in blue.

1. Streaming. Section 4 now defines streaming for each search. Parent-based insertion is one pass: one descent per record, never revisited. Recursive search repairs over the stored prefix on a schedule, so its memory grows with n. The theory is built for arriving records: Theorem 1 certifies a graph at a declared prefix, and Theorem 2 adds evidence from each later block.

2. Mid-stream change (Q1). Full-prefix repairs keep a group that stops at record 600, since its old records still pay for it, and do not find a split's new children. Run fresh on the last 300 records, the same search and certifier track the change: the dead group is gone in 10 of 10 streams by record 900, both new children are found under their parent in 7 of 10 by 900 and 10 of 10 by 1,200, and no null stream keeps a group (App. K.12).

3. Time and memory (Q2). The 4,923.8 s is recursive search; certifying its graph took 0.41 s. On all 3,362 Wikidata records, parent-based insertion takes 110 ms per record in 107 MB, and certifying its 1,013-group pool from scratch takes 5,255 s. Without duplicate nodes, a leaf edit is priced from count tables and the leaf's rows alone (new Proposition 4, App. J): at most 1.6 ms for every n from 1,000 to 100,000 on synthetic 40-node graphs, against 0.04 to 6.4 s to recompute, with equal results (App. K.13).

4. Closest baselines (Q3). Thank you for these references. We added latent class analysis (LCA), COOLCAT, Squeezer, k-modes, LIMBO, ROCK and Bayesian hierarchical clustering, each checked first, native and certified on the 82 inputs (App. K.14). With BIC, a two-part description length, LCA picks one class (no group) on all 40 real inputs. Certified, none keeps a group on any null stream it finished. Section 9 now discusses COOLCAT, LIMBO, Cao et al. (2010), van Leeuwen et al. (2009) and Zhang (2004).

5. The guarantee (Q4). We would like to clarify that the abstract states the null, and Section 3 adds the repeated-check result. For single groups inside a parent given by a rule fixed from the past, yes: new Proposition 3 (App. J) tests a child there on fresh iid records at level 2^-b, with false-discovery control across nodes. The tested fields may depend on the fields that define the parent; the null is that, inside the parent, they carry no further association. App. J.4 states the scope. It kept no child in 593 true leaves. With real structure present, the certifier kept the best searched split of a true leaf 0 of 600 times (App. K.11).

6. "The margin does little work." We read this result as a strength of the code: kept graphs score at least 35.3 bits and rejected non-empty graphs at most -16.59, so every b between gives the same 345 decisions (App. K.8).

7. Mixed gains. As your review notes, the certifier works well as a filter on other methods' proposals, and the paper's claim is the code, not one best proposer. Table 2's best mean F1 is that filter at work: certifying COBWEB lifts it from 0.676 to 0.781, above every native graph (best: FISHDBC, 0.761). On Wikidata, the selected value is negative against a reference that knows each block's counts; against background with the same heads, selected graphs of both ESCROW searches, COBWEB and FISHDBC gain on every variant, also on 2,000 unseen records per order (App. K.1). Selection tests support, not prediction; the paper reports both.

8. Labels. We agree. EHR05, R27 and the other run labels are removed; each method and selection rule now has one name.

Summary of your points and the evidence

  • Streaming: one-pass insertion, and prefix and block guarantees; Section 4 defines both searches.
  • Mid-stream change: a 300-record window drops a dead group and finds a split (K.12).
  • Time and memory: 110 ms per record in 107 MB on 3,362 records; a leaf edit is priced in at most 1.6 ms up to n = 100,000 (K.13).
  • Closest baselines: seven added, native and certified; LCA picks one class on all 40 real inputs (K.14).
  • Single groups: Proposition 3 tests a child inside a past-fixed parent; 0 of 600 searched splits kept inside true leaves (App. J, K.11).
  • The margin: any b between -16.59 and 35.3 gives the same 345 decisions (K.8).
  • Mixed gains: the certifier gives Table 2's best F1; Wikidata gains against background (K.1).
  • Labels: one name per method.

Thank you again for your time and your careful review. If any remaining concern is keeping your score at 2, we would welcome the chance to address it before the discussion ends.

Thank you for the detailed response and the new experiments. The drift study, the added categorical baselines and Proposition 3 address my main questions, so I am raising my score to weak accept. I still see the streaming evidence as small scale, and I encourage the authors to state this limit clearly in the final version.

Thank you very much for your thoughtful re-evaluation and for raising your score to weak accept. We appreciate the time you took to review the new experiments, the drift study, the added categorical baselines, and Proposition 3.

Our work is primarily theoretical, and the experiments were designed to validate the theoretical results rather than to serve as a large-scale empirical benchmark. We agree with your point on the scale of the streaming evidence, and we will state this limitation clearly in the "Scope and next steps" section of the final version.

Thank you again for your constructive feedback, which has helped strengthen the clarity and rigor of the paper.

Reviewer_kU2rOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies learning useful organization without supplied groups or parent relationships. It proposes a method, ESCROW, to provide a complete conditional code for groups, links, memberships, and cell owners. Each cell is encoded once, even when groups overlap. The code bounds unsupported selection under a stated no-association null, and a separate result controls evidence accumulated from later blocks. The paper measures structural recovery, conditional compression, and conventional prediction separately. Component studies isolate the two heads and ownership costs.

Critical strengths

S1. The complete conditional code charges for overlapping memberships and cell ownership.

S2. On three synthetic depth-two streams, conditional pruning recovers all six reference groups from COBWEB proposals.

Critical weaknesses

W1. Hierarchy recovery is still limited for deep hierarchies and short streams. In Section 6, recursive ESCROW reaches native depth, i.e., seven membership F1 of 0.553. COBWEB and FISHDBC reach 0.686 and 0.676 on the same comparison. A separate 20-stream confirmation raises recursive ESCROW to 0.690 with 1,000 records, but this remains modest. Section 5 uses 300-record prefixes with six informative and six irrelevant fields. All three selected methods return no groups through signal 0.90. At full signal, parent-based ESCROW reaches only F1 0.112, while COBWEB reaches 0.545 under the same conditional certifier. These results weaken the evidence for reliable streaming hierarchy recovery. Better proposals or repairs need matched validation.

W2. Conditional selection has inconsistent predictive effects. Section 7 compares parent-based ESCROW before and after selection. Conditional bits saved decrease on all four variants. On Wikidata, saving falls from +0.236 to -0.845 bits per record. Conventional loss improves on both Amazon variants. It worsens on both Wikidata variants. Selection tests structural support, not future prediction. These results limit selection's demonstrated value for prediction. The selection and prediction design needs substantive validation on fresh records.

Questions for authors

Q1. The implementation floors numerical probabilities. Does the implemented fresh-block conditional distribution still have total probability mass at most one? Theorem 2 requires this.

Q2. The semantic similarity scale is 0.25. The memory head sharing rate is 0.01. How were these values chosen? Were evaluation outcomes used in that choice?

Q3. The paper reports method rankings in conventional loss. Do these rankings remain the same under present-cell loss? Do they remain the same under seen-symbol loss?

5 comments

Thank you for your careful review and your clear summary of the method. We appreciate your note that the complete code prevents the same saving from being counted twice, and that conditional pruning recovers the reference groups and links from COBWEB proposals. Your three questions were precise, and each led to a new analysis. We answer each point below; changes in the revised PDF are in blue.

W1. Deep hierarchies. We would like to clarify that 0.686 and 0.676 come from the native pools, before certification; on null streams, native COBWEB returns 103.7 groups. Under the same pruning, recursive ESCROW, COBWEB and FISHDBC score 0.553, 0.554 and 0.586 at depth 7, where pruning removes nothing from recursive graphs (precision 1.0). F1 weights all seven levels equally; recursive search reaches three (a perfect three-level answer scores 0.600). At 1,000 records recursive search reaches levels four and five (15.3 of 16, 13.3 of 32 groups), pruned scores are 0.686, 0.705 and 0.712, and native COBWEB's 438 groups (254 true) prune to 56 (App. K.4).

W1. Improved proposals, matched. We agree that improved proposals need matched validation. Since the paper's claim is the code, not one best proposer, K.14 gives that validation for proposals beyond our own: same inputs, seeds and certifier. Certified, LIMBO's full tree scores 0.789 at depth 7 and BHC's 0.700, against 0.553 for recursive search, and from these two trees the certifier keeps 48.0 and 29.7 groups, against 12.3 from recursive search. Native LIMBO scores 0.887 but groups all 21 depth null streams.

W1. Short streams. On the 300-record ladder the planted graph itself fails the margin in all 27 zero cells: a sample-size limit, not a search failure. At signal 0.9 recursive search finds the groups (native F1 0.99), and the true graph clears the margin from signal 0.8 with 600 records and at 0.65 with 2,000. Section 5's 0.112 comes from the one-pass parent-based proposer. On the same inputs at full signal, recursive search with the same certifier scores 0.787, 1.000 and 1.000 on the three seeds, against a mean of 0.545 for COBWEB (App. K.5).

W2. Selection and prediction. Thank you for this point. The negative Wikidata values are measured against the type-class reference, which knows each block's counts. Against background with the same heads, selected graphs of both ESCROW searches, COBWEB and FISHDBC gain on every variant, also on 2,000 unseen records per order, every interval above zero (Wikidata: recursive +0.833, parent-based +0.657 bits per record; App. K.1). As you note, selection tests structural support, and the paper reports support and prediction separately.

Q1. Flooring. Yes, up to rounding, and Theorem 2 still holds. No floor is active: every probability is at least 4e-6 in reported settings. Next-cell mass is within 6e-16 of one in float64 and 1.8e-7 in the float32 shared-count expert; Lemma 5 (App. J) keeps Theorem 2 at a cost of 2.6e-7 bits per scored cell. Rescoring in float64 changes no reported number (App. K.9).

Q2. Settings 0.25 and 0.01. From the development record: 0.25 is the code's default and was not varied during development; 0.01 is the sharing rate that saved the most bits on ABO records in a development comparison of memory variants (on a synthetic fixture, no forgetting saved the most). Neither enters selection. To choose each corpus's values without its own outcomes, we scored a 5 x 6 grid (scale 1/16 to 1, rate 0.0003 to 0.1) and gave each corpus the setting that is best on the other corpus: Amazon scale 1/16 and rate 0.003, Wikidata scale 1/8 and rate 0.1. Under these choices all ten tests of the abstract and Section 7 hold (Amazon: selected parent-based ESCROW 1.483 bits per record, COBWEB 1.018, FISHDBC 0.620). The submitted values rank 13th and 12th of 30 on the two corpora. Varying one value at a time, every top-ranked method stays first and each Section 8 head gain keeps its sign in at least seven of eight means (App. K.10).

Q3. Rankings by measure. The rankings do change with the measure, and the revision now reports each one. Among the original methods, selected parent-based and recursive ESCROW rank first and second on all-cell loss on both Amazon variants; 10-NN (no structure) is first on present-cell loss everywhere; background leads the four core graph methods on Wikidata seen-symbol loss (App. K.3).

Summary of your points and the evidence

  • Deep hierarchies: within 0.033 of COBWEB and FISHDBC under the same pruning; 0.789 from LIMBO's tree through the same certifier (K.4, K.14).
  • Short streams: every zero cell is a sample-size limit; 0.787 to 1.000 at full signal (K.5).
  • Selection on fresh records: gains against background on 2,000 unseen records per order (K.1).
  • Flooring: mass is at most one up to rounding, and Lemma 5 keeps Theorem 2 (App. J, K.9).
  • 0.25 and 0.01: set as the development record shows; chosen on the other corpus instead, all ten tests of the abstract and Section 7 hold (K.10).
  • Rankings: reported per measure (K.3).

Thank you again for your time and your careful review. If any remaining concern is keeping your score at 2, we would welcome the chance to address it before the discussion ends.

Thank you again for your precise questions. In short: next-cell mass stays at most one up to rounding, and Lemma 5 keeps Theorem 2 (Q1); the two settings were fixed during development, and choosing them on the other corpus keeps all ten claim tests (Q2); rankings are now reported per measure (Q3). On deep hierarchies, every group recursive ESCROW keeps at depth seven is a true group (precision 1.000), and no selected graph keeps any group on a null stream. The paper's contribution is the code and its guarantees; the experiments check them. If any concern is keeping your score at 2, we would be grateful to hear it before the discussion ends.

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

ESCROW constructs a complete conditional MDL code for overlapping categorical groups and hierarchies, explicitly charging for graph structure, memberships, cell ownership, and residual types. This supports a fixed-prefix compression test and a separate fresh-block e-process.

Critical strengths

The ownership-aware code is a substantive and technically interesting contribution. The Kraft accounting appears careful, and the paper is unusually transparent about implementation details, computational cost, ablations, and negative results.

Critical weaknesses

The central issue is a mismatch between the proved guarantee and the system being evaluated. Theorem 1 controls one fixed prefix under column-permutation invariance. The repair pipeline repeatedly tests overlapping prefixes using the same b=10b=10 threshold without alpha-spending. Theorem 2’s fresh-block e-process is a separate alarm and does not certify the individual nodes, links, or repeated repair decisions. Hence, the paper’s reliability story does not currently extend to its actual structure-recovery procedure; the null also excludes drift and serial dependence. The empirical story is fragmented. Flat+split supplies the noise ladder, recursive search supplies hierarchy recovery, while parent insertion produces much of the real-data prediction result. Parent recovery deteriorates sharply beyond shallow depths, recursive search takes roughly 4,924 seconds on only 600 records, and both Wikidata variants have negative conditional savings. The comparison also omits the closest categorical streaming and hierarchical methods, including COOLCAT, Squeezer/SCLOPE, ROCK, LIMBO, and Bayesian/rose-tree approaches under matched memory and runtime.

Questions for authors

Can sequential repair be rerun with explicit global error control? How does ESCROW compare with the closest categorical methods under equal memory and compute? Can the gain from recovered hierarchy be separated from the semantic and adaptive-memory heads? Since generative AI was used to create synthetic data, please give the exact generator provenance and confirm that the generators were frozen before inspecting experimental outcomes.

4 comments

Thank you for your careful and detailed review. We are glad you see the ownership-aware code as "a substantive and technically interesting contribution", and that you found the paper transparent about its costs, ablations, and negative results. Your review led to new experiments and substantial improvement in the paper. We answer each point below; changes in the revised PDF are in blue.

1. Error control for repeated repairs. We would like to clarify that each reported decision is one certified test at one declared prefix. Repairs, splits and pruning only choose the graph tested, and Theorem 1 bounds the chance that any graph clears b, so it covers all search. To make this explicit, new Corollary 1 (App. J) sets margin b + log2(j(j+1)) at repeated check j. Rerun with it, both fresh-block studies change 0 of 320 checkpoint decisions, recursive search certified at every repair changes 0 of 24, and no null graph is kept (App. K.7).

2. Single nodes. You are right that Theorem 2 is a global alarm; it is one by design. New Proposition 3 (App. J) adds a node-level test for a child inside a parent given by a rule fixed from the past: the child is scored on fresh iid records at level 2^-b, with false-discovery control across nodes. It kept no child in 593 true leaves. With real structure present, the certifier kept the best searched split of a true leaf 0 of 600 times (App. K.11).

3. Drift and serial dependence. As you note, drift or serial dependence inside a block is outside the null (Section 3 and App. J). Between blocks, the null lets the counts change: Theorem 2 conditions on the past and on each block's counts (Section 3). K.12 tests drift directly: one group stops, or splits, at record 600. Run on the last 300 records, the same search and certifier drop the dead group in 10 of 10 streams by record 900, find both new children in 10 of 10 by 1,200, and keep no group on any column-shuffled null stream.

4. One code, two declared searches. The paper's claim is the code and its guarantees, not one best search. Both ESCROW searches, COBWEB and FISHDBC run through that code on identical depth and real inputs: recursive search proposes hierarchies, parent-based insertion is the one-pass predictor, and a claim map ties every number to its run (App. K.15). The 4,923.8 s is search; certifying that graph took 0.41 s. On 20 unseen noise-ladder seeds, flat search with split moves reaches ARI 0.997 at full signal with no group on any null stream; every transferred baseline groups all 20 nulls (App. K.6).

5. Categorical methods and compute. Added COOLCAT, Squeezer, k-modes, latent class analysis (LCA), LIMBO, ROCK and Bayesian hierarchical clustering (BHC), each checked first and run on the 82 inputs, native and certified, on one host, with CPU time and memory (App. K.14). Both selected ESCROW searches beat the first four on every real variant (Amazon: 1.434 and 1.036 bits per record against at most 0.260), also when COOLCAT and k-modes restart until they use parent-based ESCROW's same-host generation time. Natively, all but LCA and ROCK group every real null stream they finish; certified, none keeps a group on any null stream it finished. LIMBO, ROCK and BHC return trees of up to 598 groups; certified, they take 9 to 48 times parent-based ESCROW's CPU time on Amazon, and 11 of 20 Wikidata-variant LIMBO jobs passed one hour. Selected parent-based ESCROW leads all three on both Amazon variants; selected LIMBO leads on Wikidata and selected ROCK on Wikidata*.

6. Groups versus heads. With heads fixed, we replaced each native graph's groups by random groups of equal count, sizes and nesting. On 2,000 unseen records per order, with the full predictor, the real groups win for both ESCROW searches, COBWEB and FISHDBC on every variant, by 0.22 to 1.81 bits per record, every interval above zero (App. K.2).

7. Depth. Parent-based insertion is the one-pass predictor; hierarchy is recursive search's job. At depth 7, under the same pruning, recursive ESCROW, COBWEB and FISHDBC score 0.553, 0.554 and 0.586 in membership F1, with precision 1.000 for recursive search (App. K.4). From LIMBO's and BHC's trees the certifier keeps deeper structure: certified, they score 0.789 and 0.700 at depth 7 and keep 48.0 and 29.7 groups, against 12.3 for recursive search (App. K.14). Native LIMBO scores 0.887 but groups all 21 depth null streams.

8. Negative Wikidata savings. We would like to clarify that these values are measured against the type-class reference, which knows each block's counts. Against background with the same heads, selected graphs of both ESCROW searches, COBWEB and FISHDBC gain on every variant, also on 2,000 unseen records per order, every interval above zero (Wikidata: recursive +0.833, parent-based +0.657 bits per record; App. K.1).

9. Generator provenance. Thank you for the chance to state this clearly. All synthetic streams come from seeded Python generators in the supplement. No language model produced any record. The protocols in evidence/protocols record the seeds, settings and input hashes, and the confirmation protocols record their freeze times.

Summary of your points and the evidence

  • Repeated repair tests: Corollary 1; 0 of 320 and 0 of 24 decisions change (App. J, K.7).
  • Single nodes: Proposition 3, for a child inside a past-fixed parent; no child kept in 593 true leaves (App. J, K.11).
  • Drift: block counts may change under Theorem 2; a 300-record window tracks a dead group and a split (Section 3, K.12).
  • Fragmented story: one code, identical inputs, a claim map (K.15).
  • Closest categorical methods: seven added, native and certified, with CPU time and memory on one host (K.14).
  • Hierarchy versus heads: with the full predictor, the real groups of both ESCROW searches, COBWEB and FISHDBC beat random ones by 0.22 to 1.81 bits per record (K.2).
  • Depth: from LIMBO's and BHC's trees the certifier keeps deeper structure (K.4, K.14).
  • Wikidata: all four methods gain against background with the same heads (K.1).
  • Provenance: seeded Python generators (evidence/protocols).

Thank you again for your time and your careful review. If any remaining concern is keeping your score at 2, we would welcome the chance to address it before the discussion ends.

Thank you again for your careful review. We know this is a busy period. In short, the paper's contribution is theoretical: a complete conditional code for overlapping groups and hierarchies, with three guarantees. Theorem 1 bounds false selection for any search at a declared prefix; Corollary 1 extends it to repeated checks with error spending (applied to the reported runs, it changes none of the 320 certified checkpoint decisions); Proposition 3 gives a node-level test inside a parent fixed from the past. The experiments check these results; our reply above lists each of your points with its evidence. If any concern is keeping your score at 2, we would be grateful to hear it before the discussion ends.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.