acceptodds
Withdrawn from ICLR 2027

From Controlled Edits to Complex Bindings: Evaluating and Improving Structure-Aware Vision–Language Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

CLIP is effective at global image–text matching, and Structure-CLIP further improves discrimination on template-based attribute and relation perturbations. However, whether these gains extend to descriptions containing multiple entities, modifiers, and relations remains unexplored. A structural error in such descriptions may concern not only attribute or relation direction but also host assignment or spatial layout. We construct 666 image–text instances over six structural categories. Each instance pairs a natural image with a complex multi-entity caption P1, a meaning-preserving paraphrase P2, and a negative N that changes one visually verifiable binding. This P1/P2/N design evaluates whether a structural decision is stable across two valid expressions and also supplies two views of the same binding during training. An exploratory evaluation of pretrained checkpoints shows uneven transfer. Structure-CLIP remains strong on relation direction, but improves little or degrades on host, left–right, and depth bindings. We therefore introduce Complex-SCLIP, which applies the existing knowledge-enhanced encoder to both positives and jointly optimizes dual ranking, multi-positive contrastive learning, and positive consistency against the same structural negative. On the 240-instance test split, Complex-SCLIP reaches 69.2% Both-accuracy, outperforming Structure-CLIP (41.7%), five published compositional methods (32.1–39.6%), and an in-domain Structure-CLIP fine-tuning baseline trained with the released objective (48.3%). The improvement is consistent across three seeds and transfers to SUGARCREPE++. COCO 5K Mean Recall remains above CLIP (61.7% vs. 60.3%), though below the Structure-CLIP starting point.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Vision-Language Models, Compositionality, Structural Binding, Multi-Positive Learning, Structural Negatives

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_s4mrOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper studies whether Structure-CLIP, a CLIP-based model that incorporates scene-graph structure, remains effective when captions contain multiple entities, attributes, and relations rather than simple controlled edits. The authors construct a controlled dataset of 666 examples across six structural categories, split into 350 training, 76 validation, and 240 test examples. Each example contains an image, two semantically equivalent positive captions, and one negative caption that changes a single structural binding. They then propose Complex-SCLIP, which extends Structure-CLIP by training both positive captions to score above the same negative and encouraging the two positives to produce consistent representations. The method improves performance on the held-out test split and shows some transfer to SUGARCREPE++, while slightly reducing COCO retrieval performance.

Critical strengths
  1. The paper is clearly written and the setup is easy to follow.

  2. The authors provide a controlled comparison with several compositional baselines and include ablations, which makes it easier to understand where the gains come from.

  3. The paper is relatively transparent about some limitations, including the small dataset, reuse of the diagnosis set, and the drop in COCO retrieval performance.

Critical weaknesses
  1. I do not find the novelty strong enough. The main problem, whether VLMs can distinguish small but important semantic or structural changes, has already been studied in many prior works such as ARO, SugarCrepe/SUGARCREPE++, and other related works.

  2. The method itself is also quite limited. Most of the gain comes from adding the second positive caption during training. Table 3 shows that removing the additional consistency losses has almost no effect on the final Both accuracy. This makes it hard to see a substantial methodological contribution.

  3. The model is trained on only 350 examples. This is very small, especially since the main test set follows the same P1/P2/N format used during training. Because of this, the strong in-distribution improvement is not very convincing evidence that the model learned a more general structural capability.

  4. The out-of-distribution results are also not conclusive. SUGARCREPE++ uses a very similar positive/positive /negative setup, so it is not a very different evaluation setting. On COCO retrieval, the method performs worse than the Structure-CLIP initialization. Overall, I do not think the external evaluation is strong enough to show broad generalization.

Questions for authors

How was the semantic equivalence between P1 and P2 verified? Was there any independent human annotation or agreement measurement?

Since the model is trained on only 350 examples, how do the authors know the gains are not mainly due to fitting the specific P1/P2/N setup?

1 comment

Thank you for the careful assessment and for recognizing the clear setup, ablations, and stated limitations. We clarify the contribution and provide additional evidence on the training and generalization concerns.

W1: Relation to existing compositional benchmarks. Our benchmark tests whether valid multi-entity descriptions outrank descriptions containing a single edit to an attribute, relation, host assignment, or spatial binding, and whether this ordering survives meaning-preserving reformulation. The design retains non-target facts, while the six-family breakdown helps identify which binding types are difficult. We build on existing fine-grained and dual-positive evaluation, complementing ARO and SUGARCREPE++ through this controlled diagnostic focus rather than claiming a new general-purpose reasoning benchmark.

W2: Role of dual-positive supervision. A single-positive ranking objective does not explicitly constrain an equivalent reformulation. Supervising both valid descriptions against the same negative is therefore central to our adaptation design. The submitted ablations already note the small effects of the consistency regularizers. To match access to P2 while separating this design from the full objective's added value, we trained the released Structure-CLIP objective on (P1,N) and (P2,N), averaging losses before one update. This is an added dual-positive variant, not the unchanged published model. Initialization, image order, positive exposure, frozen visual encoder, optimizer, 160 updates, and three seeds were matched. Both use fixed step-160 checkpoints with test KEE disabled; the control retains original weighting, fusion, and warmup conventions.

With test KEE disabled, Test240 Both is 61.25% for the control and 60.69% for Complex-SCLIP, versus 41.67% for the unchanged starting checkpoint. Both schemes improve over this starting point, showing that the dual-positive adaptation design also works with the original objective, with numerically close Both scores. The full method does not exceed the matched control on Both here; on SUGARCREPE++, however, its ITT and TOT are higher than the control in all three seeds (see W4). We distinguish these findings from the contribution of any individual auxiliary loss.

W3 and the format-fitting question. We use the 350 examples for targeted adaptation of pretrained Structure-CLIP, keeping its visual encoder frozen. This builds on existing representations rather than learning vision-language competence from scratch. As described in Section 4.1, P1 and P2 express the same visual facts in different wording, while N changes one target binding. Supervising both positives against the same negative makes the structural distinction, rather than one particular wording, the intended learning target.

Using this format in both training and evaluation aligns supervision with the property being tested: whether the ranking remains correct under meaning-preserving reformulation. This motivates the design but does not by itself rule out format-specific adaptation. The 240 test records were excluded from fine-tuning, though they informed the earlier diagnosis and are not development-independent. We therefore interpret the results as adaptation within this controlled task; W4 discusses external transfer and further cross-format evaluation.

W4: External transfer and capability preservation. We selected SUGARCREPE++ to examine whether adaptation on our collection transfers to a separate dataset measuring a related compositional task, without further fine-tuning. It shares the P1/P2/N format with our benchmark, as you note.

For a fair comparison, both methods were evaluated with test-time KEE disabled. Compared with the training-matched dual-positive control, Complex-SCLIP improves mean ITT and TOT by 1.02 and 3.87 percentage points across three seeds, respectively, with both metrics higher in every seed. These results support an additional benefit from the full method on related compositional tasks, but alone do not establish broad, format-independent generalization.

COCO retrieval performance is lower than the Structure-CLIP initialization. We therefore regard incomplete preservation of retrieval capability as a limitation of the current adaptation, rather than interpreting gains on a related task as improvements across all tasks.

In future work, we plan to evaluate benchmarks with different task formats to assess transfer beyond the current setting.

Question on P1/P2 equivalence. Two raters independently examined 100 test instances. Both confirmed the intended equivalent positive pair for 97/100 instances; agreement on equivalence was 99% (Cohen's kappa=0.795).

Reviewer_GJnyOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper benchmarks vision–text alignment, focusing on detecting factual inaccuracies in textual descriptions containing multiple facts. To this end, the authors construct a benchmark comprising 666 image–text instances across six structural categories. Their empirical results show that existing CLIP-based image–text alignment models struggle with this task, while their proposed approach, fine-tuned on the collected data, outperforms the baseline models.

Critical strengths
  1. Evaluating the detection of factual inaccuracies in detailed textual descriptions offers some novelty.
  2. The empirical results suggest that the proposed approach is effective on the authors’ benchmark.
Critical weaknesses
  1. Limited benchmark novelty. The main distinguishing feature is the use of sentences containing multiple facts. However, decomposing each sentence into individual factual statements would make the task closely resemble existing evaluation settings. The paper should clarify what additional challenges arise from evaluating multiple facts jointly.

  2. Limited benchmark scale and coverage. The benchmark is small, raising concerns about whether it adequately covers diverse textual and visual domains.

  3. Unclear technical novelty. The paper provides limited discussion of how the proposed method differs from existing techniques. Although the experiments show performance improvements, the novelty of the proposed module remains unclear.

  4. Potentially confounded comparison. The training and test splits are constructed using the same procedure. Thus, the proposed model’s advantage may reflect adaptation to the benchmark’s construction process, particularly if the baselines are not fine-tuned on the same training split. Comparisons under matched training conditions would help establish the method’s effectiveness.

  5. Unclear negative-example generation. The paper does not clearly explain how negative examples are generated.

Questions for authors

My rating is based on all 5 weaknesses above.

1 comment

Thank you for recognizing the relevance of factual discrimination in detailed descriptions and for identifying where the contribution and controls need clarification. We address the five concerns below.

W1: Benchmark contribution and fact decomposition. Decomposing captions into facts is a reasonable approach. Our question is whether the model ranks captions correctly when a target binding is embedded among other facts in a full description and expressed in two valid wordings. Decomposition must preserve entity identity, attribute ownership, and relation direction. In the C-Host example in Appendix A, swapping clothing between two people preserves the inventory of people and clothes; object recognition alone cannot resolve ownership. A decomposition retaining ownership could address this example. We do not assert that multi-fact sentences are indivisible or claim superiority over a validated decomposition-and-aggregation baseline. The contribution is the controlled diagnostic setting and its binding-specific analysis.

W2: Scale and coverage. The 666 instances cover six targeted binding families, with 350/76/240 training/validation/test records. The design changes one target binding while retaining other facts, allowing family-specific diagnosis alongside paraphrase testing. This controlled scope helps locate failure patterns, but the 34-50 test instances per family limit precision and coverage. We restrict conclusions to the evaluated families and tasks.

W3: Technical contribution. A single-positive ranking objective does not explicitly constrain an equivalent reformulation. Our design targets this gap: both valid descriptions must outrank the same structural negative. Complex-SCLIP extends Structure-CLIP's knowledge pathway to both positives, coupled through multi-positive alignment and consistency terms. The contribution combines complex-binding diagnosis with this adaptation design, not the invention of multiple-positive learning or KEE. Section 2 discusses related work including CLIC; W4 separates the training design from the full method's additional benefits.

W4: Training controls and construction bias. Beyond the paper's in-domain P1/N baseline, we trained a dual-positive control to match access to P2: we applied the released Structure-CLIP objective to (P1,N) and (P2,N), averaging their losses before one update. This extends how the original objective is applied, rather than testing the unchanged published model. Initialization, image order, positive exposure, frozen visual encoder, optimizer settings, 160 updates, and three seeds were matched. Both use fixed step-160 checkpoints and no test KEE; the control retains original weighting, fusion, and warmup conventions.

With test KEE disabled, Test240 Both is 61.25% for the control and 60.69% for Complex-SCLIP, versus 41.67% for the starting checkpoint. Both dual-positive schemes improve over the starting model. The control shows that the adaptation design also works with the original objective, yielding numerically close Both scores. The full method does not exceed the matched control on Both here; on SUGARCREPE++, however, its ITT and TOT are higher than the control in all three seeds.

We also freshly evaluated both methods on all 4,757 SUGARCREPE++ instances under one evaluator and environment, without test KEE or additional external fine-tuning. The control obtains 69.59 ± 0.48% ITT and 65.90 ± 1.03% TOT; Complex-SCLIP obtains 70.60 ± 0.94% and 69.78 ± 1.55% (three-seed means ± sample SD). Both metrics favor Complex-SCLIP in every seed, supporting an additional external benefit over dual-positive training with the original objective. Removing four identified training-overlap records leaves ITT at 70.63% for Complex-SCLIP versus 69.60% for the control.

Test records were excluded from fine-tuning but used in the earlier diagnosis. The external evaluation complements this controlled test, while the lack of a newly collected test set independent of method development remains a limitation.

W5: Negative generation. As described in Appendix A, we modify one target binding in P1 while preserving non-target facts. Unambiguous attribute swaps or left-right reversals are generated deterministically. When string replacement could affect the wrong phrase, a vision-language model drafts the edit and a second agent checks it, followed by author spot checks. The retention criterion is image support for both positives and a visually contradicted binding in N. In the additional independent two-person audit of 100 test instances, both raters confirmed the single-binding edit in 96/100 cases.

Reviewer_5TrCOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates whether Structure-CLIP compositional improvements generalize to complex multi entity captions. The authors construct a 666-instance benchmark spanning six structural categories, where each image is paired with an original positive (P1), a meaning-preserving paraphrase (P2), and a structural negative (N). The primary metric requires both positives to outrank the negative. The proposed Complex-SCLIP extends Structure-CLIP with dual-positive ranking, knowledge enhanced representations, multi positive contrastive alignment, and consistency regularization. On the 240 instance test set, it achieves 69.2% Both accuracy compared with 41.7% for Structure-CLIP and 48.3% for an in-domain fine-tuning baseline. Improvements also transfer to SUGARCREPE++, although text-only discrimination and COCO retrieval performance decline.

Critical strengths

[S1]. Meaningful evaluation of complex compositional reasoning. The P1/P2/N benchmark extends template-based structural evaluation to multi-entity captions and tests whether binding decisions remain consistent under valid paraphrasing. The six categories reveal informative limitations of Structure-CLIP, particularly for host assignment and spatial relationships.

[S2]. Substantial improvements with useful ablations. Complex-SCLIP improves both accuracy by 20.9 pp over the in-domain Structure-CLIP fine-tuning baseline. Three seed experiments, paired significance tests, category wise analysis, and component ablations support the effectiveness of dual positive supervision. Improvements on SUGARCREPE++ provide additional evidence of transfer.

Critical weaknesses

[W1]. Unmatched baselines and unclear test-time structural supervision. The main improvements over frozen checkpoints conflate the proposed method with in-domain adaptation. The in-domain Structure-CLIP baseline also differs in positive caption supervision, visual encoder freezing, training exposure, and checkpoint selection. A matched dual positive baseline is needed to isolate the proposed objectives contribution. More importantly, Section 4.2 applies shared KEE features derived from target triples to positives but not negatives, while the controlled benchmarks test time scoring protocol is unclear. If scoring requires ground truth structural information or treats known positives differently, the main evaluation would be substantially confounded.

[W2]. Limited benchmark independence and generalization. The 240 instance test set was drawn from the same collection used for the initial diagnosis, with only 34 to 50 examples per category. Benchmark construction relies partly on AI assisted validation without reported human agreement or language only shortcut controls. Furthermore, SUGARCREPE++ text only accuracy drops from 69.4% to 58.0%, while COCO Mean Recall decreases from 64.9% to 61.7% relative to Structure-CLIP. These limitations weaken claims about generalizable compositional improvements and preservation of existing capabilities.

Questions for authors

[Q1]. Test-time KEE: Are KEE features applied during controlled benchmark evaluation? If so, which captions receive them, and are the triples obtained from ground truth annotations or independently parsed? Please report results without test time KEE or with uniform, label independent candidate processing.

[Q2]. Matched baseline: Can the authors train Structure-CLIP using both (P1, N) and (P2, N), matching Complex-SCLIP data exposure, frozen visual encoder, optimization settings, and checkpoint selection? Please clarify Complex-SCLIP effective batch size.

[Q3]. Benchmark validity: Can the authors provide human annotation audit statistics and text-only or image shuffling controls to exclude linguistic shortcuts? Given that the test collection informed method development, can performance be verified on independently collected complex-binding examples?

[Q4]. External generalization: Can the authors report the in-domain Structure-CLIP baseline on SUGARCREPE++, including per-category results, and clarify whether source image overlaps with external evaluation datasets were excluded?

[Q5]. Capability preservation: What explains the substantial degradation in SUGARCREPE++ text-only discrimination and COCO retrieval? Can the effects of KEE, multi-positive training, and small-data adaptation be isolated?

2 comments

Thank you for recognizing the diagnostic setting and empirical analysis. We address your concerns with additional controls and an external comparison.

Q1 / W1: Test-time KEE. P1 and P2 express the same visual facts, so we use shared structural knowledge to support consistent representations across wordings. Training extends Structure-CLIP's positive-side KEE fusion to both captions. The original evaluation retained this policy, enhancing P1/P2 with construction-time target triples but not N. We additionally evaluated all 240 test instances under two uniform candidate-processing protocols: no test KEE, and KEE using independently parsed triples for every candidate. Each parser request used one caption, without images, labels, paired captions, or existing graphs; all candidates used the same six-triple selection rule.

Test-time protocolMatched control (Q2), Both (%)Complex-SCLIP, Both (%)
Original positive-only KEE60.14 ± 1.5869.17 ± 0.72
KEE disabled for all candidates61.25 ± 1.5060.69 ± 1.34
KEE using independently parsed triples for all candidates61.11 ± 1.3460.42 ± 2.20

Both methods use the same three seeds, each with its own fixed step-160 checkpoints; values are means ± sample SD. The full model leads under the original protocol; the methods have numerically close Both scores under both uniform protocols. Trained weights and training-time knowledge supervision are unchanged. External comparisons below use no test KEE.

Q2 / W1: Training-matched dual-positive control. Beyond the submitted P1/N fine-tuning baseline, we trained an additional dual-positive control by applying the released Structure-CLIP objective to (P1,N) and (P2,N) within the same image batch and averaging the losses before one update. This extends the original objective to both positives; it is not the unchanged published model. We matched initialization, image order, positive-caption exposure, frozen visual encoder, AdamW, learning rate (1e-6), weight decay, 160 updates, and three seeds. Both methods use fixed step-160 checkpoints and no test KEE. The effective image batch size is four, not eight: each image has two positive captions. Both methods use KEE only for positives during training, without N triples. The control retains the original objective's weighting, fusion, and warmup conventions.

With test KEE disabled, Test240 Both is 61.25% for the control and 60.69% for Complex-SCLIP, versus 41.67% for the unchanged starting checkpoint. Both dual-positive schemes improve over this starting point, supporting the adaptation design with either objective. The full method does not exceed the matched control on Both here; on SUGARCREPE++, however, its ITT and TOT are higher than the control in all three seeds.

After training the matched dual-positive control, we evaluated its checkpoints alongside Complex-SCLIP on all 4,757 SUGARCREPE++ instances using the same evaluator, environment, precision, and inference batch size, with test KEE disabled. Neither method was additionally fine-tuned on this external benchmark.

MethodITT (%)TOT (%)
Matched dual-positive control69.59 ± 0.4865.90 ± 1.03
Complex-SCLIP70.60 ± 0.9469.78 ± 1.55

With training data, positive-caption exposure, and training budget matched, both metrics favor Complex-SCLIP in each of the three seeds. Mean gains are 1.02 ITT and 3.87 TOT percentage points, calculated before rounding. These results support an additional benefit on SUGARCREPE++ from the full method over dual-positive fine-tuning with the original objective. The matched dual-positive control is distinct from the original P1/N continual baseline in Part 2.

Q5 / W2: TOT correction and capability preservation. The manuscript's 58.0% was mistakenly taken from mean-image-feature control accuracy, not text-only TOT. We apologize for this reporting error. TOT requires both sim(P1,P2) > sim(P1,N) and sim(P1,P2) > sim(P2,N). Recounting the saved predictions gives 75.24% under the original asymmetric KEE protocol. Separately, disabling test KEE gives 69.78 ± 1.55%, close to the starting checkpoint's 69.39%; the fresh external evaluation above reproduces this result. The reporting-field correction and the evaluation-protocol change are separate steps.

Under uniform no-test-KEE evaluation, text-only performance is therefore close to the starting model, rather than showing the large decline reported in the manuscript. COCO Mean Recall, however, decreases from 64.87% to 61.66 ± 1.36%. A possible explanation is a trade-off between small-data adaptation for structural discrimination and general image-text retrieval. Part 2 reports category results and source-image overlap checks.

Below we address Q3 and Q4.

Q3 / W2: Human verification, shortcuts, and independence. We completed an independent two-person audit of 100 test instances across all six families, with randomized instance and candidate orders. The raters checked each caption's image support, P1/P2 equivalence, and single-binding edits. All five judgments agreed on 99/100 instances; 96/100 met every criterion for both raters. These results provide direct human evidence supporting annotation quality on the audited subset.

We evaluated image controls on all 240 test instances with test KEE disabled. Across three training seeds, Both is 60.69% with the paired image, 55.72% averaged over 20 image-shuffling assignments, and 61.11% with the mean image feature. Paired images outperform shuffled images, but the mean-image result means we cannot claim to have excluded linguistic shortcuts. Mean-image accuracy is not text-only TOT. Due to time and other constraints, we have not completed further diagnostics to disentangle visual and textual cues, and retain this unresolved issue as a limitation of the current evaluation.

The test instances were excluded from the 350-example supervised fine-tuning split, although they participated in the earlier exploratory diagnosis. We therefore distinguish a test split held out from fine-tuning from a newly collected test set independent of method development. The external comparisons below complement the controlled diagnosis; the human audit addresses annotation quality, while the lack of development-independent testing remains a limitation.

Q4: In-domain Structure-CLIP baseline on SUGARCREPE++. We evaluated the original P1/N continual-fine-tuning baseline using the original configuration and three seeds. Validation selected step-160 for seeds 20260714 and 20260716, and step-120 for seed 20260715 (a retrained reproduction under the original configuration). The matched dual-positive control is a separate baseline using both positives and fixed step-160 checkpoints.

All columns report results with test KEE disabled on all 4,757 instances. P1/N FT columns report single-seed results; the last two are three-seed means ± sample SD. Entries are ITT (%).

Edit typeP1/N FT 20260714P1/N FT 20260715P1/N FT 20260716Matched dual-positiveComplex-SCLIP
replace_att64.0963.7163.2069.80 ± 0.2570.18 ± 1.11
replace_obj79.0079.7279.4883.39 ± 0.9185.21 ± 1.44
replace_rel49.5749.0048.0854.46 ± 0.5054.93 ± 0.88
swap_att68.4768.9268.1769.87 ± 0.3871.37 ± 0.38
swap_obj60.0063.2760.4161.90 ± 0.6261.36 ± 0.85
Overall ITT (%)65.3865.6364.9469.59 ± 0.4870.60 ± 0.94

Within seeds 20260714/15/16, Complex-SCLIP improves ITT over P1/N FT by 5.05, 4.14, and 6.68 percentage points, respectively, calculated before rounding. For the reproduced seed, TOT is 67.37% for P1/N FT versus 68.47% for Complex-SCLIP. Thus, the additional seed also supports the external benefit over single-positive fine-tuning. The starting checkpoint obtains 69.56% ITT. Against the stronger matched dual-positive control, Complex-SCLIP retains an average ITT gain of 1.02 percentage points, with higher mean ITT in four of five categories; swap_obj decreases by 0.54 percentage points.

Seeds 20260714/16 reuse verified prior evaluations; the reproduced seed is newly evaluated. The matched-control and full-model columns use the same evaluator and environment.

Q4: Source-image overlap. To address this concern, we checked exact image hashes and documented source identities and evaluated after excluding the identified overlaps. The audit identified four SUGARCREPE++ records and 16 COCO evaluation images overlapping our fine-tuning images. Removing the four SUGARCREPE++ records leaves 4,753 instances: ITT is 70.63 ± 0.93% for Complex-SCLIP and 69.60 ± 0.48% for the matched dual-positive control. TOT is 69.82 ± 1.56% versus 65.94 ± 1.03%, respectively. Both methods' ITT and TOT means remain nearly unchanged from the full evaluation set, and Complex-SCLIP retains its advantage over the matched control on both metrics after these exclusions.

For COCO, we removed the 16 images and their captions and recomputed retrieval ranks over the reduced gallery. Mean Recall is 61.63 ± 1.36% for the full model versus 64.84% for the starting checkpoint, again nearly unchanged from the full gallery. Thus, excluding the identified image overlaps leaves the external comparison conclusions unchanged.

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

Related papers

Discussion (0)

Sign in to comment.