acceptodds
Under review as a conference paper at ICLR 2027

DELTAGENOME: SPARSE-EDIT INFERENCE FOR GENOMIC FOUNDATION MODELS

Abstract

Genomic variant analysis repeatedly evaluates sequences that differ by one or a few nucleotides, yet genomic foundation-model inference usually treats reference and alternate alleles as independent dense inputs. DeltaGenome formulates this workload as sparse-edit inference: given a computed reference, determine which latent states and operations remain reusable after a sparse genomic edit. Across NT-v2, DNABERT-2, Caduceus, and HyenaDNA, single-nucleotide edits induce structured and predictable latent perturbations, but locality alone does not deter- mine computational value. The decisive systems property is reuse-frontier persis- tence: how long an architecture preserves exact reusable dependency structure af- ter the edit. On a frozen 384-variant benchmark, masked reconstruction preserves dense alternate perturbations at controlled selection-policy budgets, yet runtime diverges sharply. DNABERT-2 remains slower than dense inference in the tested executor, whereas HyenaDNA obtains a 1.54x warm-query speedup with a per- sistent full-stack causal reuse frontier and a larger reference cache. DeltaGenome identifies not a universal sparse executor, but a cross-architecture principle for serving genomic foundation models around nearby sequences.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
genomic foundation models, variant effect prediction, sparse-edit inference, incremental inference, computational reuse, genomic sequence modeling

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_7dTvOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes DeltaGenome, a framework designed to accelerate GFM inference by reusing computation between original sequences and slightly modified variants. The authors analyze how modifications affect the hidden representations and what parts can be reused in terms of inference. Results show that on four GFMs, those modifications are structured and predictable. But only HyenaDNA shows a 1.33x speedup, while others get little improvement or are even slower.

Critical strengths
  • The motivation of this paper is interesting and practically relevant. Reusing representations between slightly changed sequences is a promising direction for accelerating inference.
Critical weaknesses
  • Most figures in this paper are small in the text and have lots of dense visual elements, this makes them very difficult for readers to read and understand. Especially for Figure. 1, I can hardly read the text without zooming in a lot. Also, considering that key information in Figures. 2-5 is relatively limited, it would be better to use tables rather than those figures with tiny text and excessive whitespace to report the results. I would suggest the authors improve the readability of this paper.
  • This paper could be better organized. The key idea of this paper is to accelerate inference for genomic foundation models, so those results are expected to be presented earlier. But in the experimental section, 5.1-5.3 are talking about perturbation, prediction and reconstruction, and main acceleration results are only stated in Section 5.4 on page 8. I think the experimental narrative does not align well with the paper's primary objective.
  • For all four GFMs tested in the paper, only HyenaDNA shows a 1.33x speedup using the proposed method. The other three models NT-v2, DNABERT-2 and Caduceus get little improvement or are even slower. Considering that inference acceleration is the main motivation for this paper, the current results do not show practical significance or general applicability.
Questions for authors
  • Can the authors provide the results on the most recent large-scale GFMs like GenomeOcean[1] and Evo 2[2]? All four models tested in the paper are relatively old and may not represent SOTA GFMs. Those generative models can also do classification, and since they are large in parameters, the inference acceleration on them is potentially more valuable. I would suggest the authors add results on those models, especially Evo2.

[1] Zhou, et al. "Genomeocean: An efficient genome foundation model trained on large-scale metagenomic assemblies."

[2] Brixi, et al. "Genome modelling and design across all domains of life with Evo 2."

1 comment

Thank you for the practical motivation you identified and the specific suggestions on readability, organization and contemporary large-scale GFMs. I have substantially revised both the presentation and the experiments.

Presentation and organization. Figure 1 is redesigned as a simpler conceptual overview. The main text now has three figures and three tables: Table 1 summarizes mechanisms, Table 2 reports absolute latency, speedup and numerical fidelity, and Table 3 summarizes perturbation geometry. Dense diagnostic plots and detailed tables are moved to the appendix. The Results now begins with matched acceleration (§5.1), followed by amortization/state organization (§5.2) and contemporary-GFM execution (§5.3); perturbation prediction and online utility follow in §5.4, and downstream prediction preservation concludes §5.5. The abstract and introduction prioritize reference-aware execution across eight backbones, with the harmonized 384-SNV analysis clearly restricted to the original four. This addresses the original mismatch between motivation and experimental narrative.

Practical significance. The revised comparison separates full-system gains from incremental reuse. At K=128, HyenaDNA has 2.042× native-relative and 1.107× compiler-matched prepared-system speedups, NT-v2 1.003× and 1.017×, Caduceus 1.762× and 0.979×, and DNABERT-2 0.674× and 0.663×. The gains for HyenaDNA and NT-v2 are consistent across all eight references; their descriptive 95% intervals are [1.1033,1.1113] and [1.0164,1.0167]. Absolute timings, repeat variability, checkpoint identities, numerical errors and matched regional operators are reported in Table 2A and Appendices D–E. Caduceus's shared directional frontier reduces K=128 cache tensor storage by 79.2× relative to coordinate-specific caching, not peak GPU memory; it does not yield a matched latency gain. An independent HyenaDNA prefix comparison is also reported and remains competitive. The revision therefore does not attribute all native-relative acceleration to reuse.

GenomeOcean, Evo2 and other recent models. As requested, the study includes GenomeOcean-4B and Evo2-7B, as well as NT-v3 and Carbon-3B. At K=128, Evo2-7B achieves 1.292× native-relative and 1.016× matched setup-inclusive speedups, with wall totals of 98.022, 77.053 and 75.867 s for native, optimized dense and reuse. All 512 requests are bitwise exact. NT-v3 achieves 1.316× native-relative and 1.105× matched setup-inclusive CUDA speedups; a separate audit verifies bitwise equality on 512 requests. GenomeOcean-4B achieves 1.152× on 512 prepared eligible requests, with bitwise-exact outputs; a separate mixed workload with only 109/512 eligible requests yields 0.797× aggregate wall-clock speedup while preserving bitwise equality. Carbon-3B shows 1.022–1.053× warm gains across three edit-position strata, without setup-inclusive measurements. These results establish applicability of concrete reuse mechanisms beyond HyenaDNA, but not universal acceleration or controlled model-size scaling (Table 2B, Fig. 3, §5.3, Appendix F/Tables 15–19). Four of the six backbones evaluated against matched optimized dense retain positive setup-inclusive gains at K=128.

Downstream classification. To address whether reuse preserves biological predictions, the revision adds a HyenaDNA pathogenic/common variant classifier trained on frozen representations from 38,619 pairs, with an identical head across dense, independent prefix and DeltaGenome execution. Both reuse paths preserve all 1,018 held-out class decisions; DeltaGenome matches dense AUROC 0.644886, average precision 0.582140 and accuracy 0.592338. The natural test workload has 987 singleton groups among 1,002 references, yielding aggregate speedups of 0.725× (DeltaGenome) and 1.077× (prefix). In controlled unlabeled panels with greater sharing, DeltaGenome achieves 1.169× median paired speedup at K=128. These results validate prediction preservation and the importance of reference sharing, not improved classification accuracy. Classification on GenomeOcean/Evo2 is not claimed (§5.5, Appendix K, Fig. S3, Tables 26–28).

I appreciate the suggestions that led to a more readable execution-first paper and a substantially broader, more rigorous evaluation of modern genomic models.

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

DeltaGenome studies repeated genomic inference when alternate sequences differ from a shared reference by sparse nucleotide edits. It characterizes hidden-state perturbations across NT-v2, DNABERT-2, Caduceus, and HyenaDNA, trains predictors of positional perturbation importance, and evaluates masked reconstruction of dense reference–alternate differences. Separately, it benchmarks architecture-specific reuse implementations and distinguishes concentrated perturbations from dependencies that permit exact reuse. On a benchmark of 384 single-nucleotide variants in 2,047-bp contexts, the authors report a 1.54× warm-query speedup for HyenaDNA and a 1.33× setup-amortized speedup at 128 queries per reference, alongside a larger persistent cache. The tested DNABERT-2 implementation is slower than dense inference.

Critical strengths
  • S1: The paper addresses a useful computational genomics workload: repeated evaluation of variants that require alternate-conditioned representations. It clearly distinguishes small numerical perturbations from unchanged computational dependencies. This distinction prevents retained perturbation energy from being mistaken for executable savings and gives a useful way to reason about incremental inference.
  • S2: The study compares four different model architectures in nucleotide coordinates and reports both successful and unsuccessful reuse implementations. Separating warm-query speed, setup amortization, and persistent cache size helps interpret the scope and cost of the reported HyenaDNA benefit. The DNABERT-2 slowdown is also useful evidence that concentrated perturbations do not guarantee a practical gain in the tested implementation.
Critical weaknesses
  • W1: The demonstrated acceleration is limited in scope, and practical value remains insufficiently established. At 128 queries per reference, the reported setup-amortized speedups are 1.02×, 0.46×, 0.87×, and 1.33× for NT-v2, DNABERT-2, Caduceus, and HyenaDNA. Only HyenaDNA shows a clear gain, with a 198.06 MB reference cache and approximately no net gain at eight queries. Its successful reuse path follows causal prefix dependencies, while the learned predictor has no demonstrated acceleration benefit. Masking already computed dense differences establishes neither executable savings nor the predictor's advantage over simple positional selection. Exact backbone checkpoints and parameter counts are unspecified, and scaling across model sizes is not evaluated. The paper should also discuss how its reuse principles apply to newer models such as NTv3 [3] and Evo 2 [4], including how their different dependency structures affect reuse-frontier persistence. Missing execution-error measurements, absolute timings and variability, and query-workload details further limit assessment of practical applicability.
  • W2: Marginal contribution. The demonstrated advance is mainly a cross-architecture characterization of perturbations and reuse behavior, with limited conceptual or algorithmic novelty. fastISM [1] already caches reference activations and recomputes mutation-affected regions. Yuzu [2] exploits sparse intermediate differences using compressed sensing. Although these methods target different architectures, they establish relevant reuse principles. The successful HyenaDNA path also follows causal prefix reuse. The cross-model observations are informative, but the manuscript does not establish a substantial new reuse mechanism or generalizable insight beyond these principles. Neither fastISM nor Yuzu is cited or substantively compared, leaving this limited increment insufficiently justified.

References:

Questions for authors

Does the learned predictor improve the end-to-end speed–accuracy trade-off over a simple position-based selector after including prediction and execution overhead? If it is intended only as an analysis tool, please clarify its contribution and compare its selection quality against positional baselines on the held-out set.

1 comment

Thank you for recognizing the distinction between perturbation concentration and computational dependencies. The revision expands execution evaluation, reuse mechanisms and predictor validation.

W1 — Practical acceleration and reproducibility. The revised benchmark compares native dense, graph dense/reuse and regionally compiled graph dense/reuse. At K=128, prepared-system speedups are:

ModelNative-relativeCompiler-matched
HyenaDNA2.042×1.107×
NT-v21.003×1.017×
Caduceus1.762×0.979×
DNABERT-20.674×0.663×

HyenaDNA/NT-v2 improve on all eight references, with descriptive 95% bootstrap intervals [1.1033,1.1113] and [1.0164,1.0167]. All 1,024 distinct queries per backbone are validated: HyenaDNA logits and NT-v2/Caduceus final hidden states satisfy 1e-5 relative L2 tolerance. DNABERT-2 is approximate (median/max errors 1.014/15.496%). Execution uses eight references, nested K=1,8,32,128 distinct edits, B=1, five warmups and 20 paired repetitions. Table 2A and Appendices D–E report absolute latency, variability, checkpoints, parameter counts, numerical settings, matched operators and preparation costs.

Contemporary architectures. Evaluation now spans eight backbones (§5.3, Fig. 3, Appendix F).

  • NT-v3: Selective encoder recomputation achieves 1.316× native-relative and 1.105× matched setup-inclusive speedups. Eligible queries decline from 512 to 31 across six encoder blocks. A separate audit verifies bitwise equality on all 512 requests.
  • Evo2-7B: Compact projection achieves 4.734× local and 1.016× matched whole-model speedups (1.292× native-relative); all 512 outputs are bitwise exact.
  • GenomeOcean-4B: Causal suffix reuse achieves 1.152× on 512 eligible requests and 0.797× on a separate mixed workload (109/512 eligible), with bitwise equality in both.
  • Carbon-3B: Warm speedups increase from 1.022× to 1.053× across three reusable-prefix fractions.

Four of six backbones achieve positive matched setup-inclusive gains at K=128. Despite negative matched latency results, Caduceus achieves 79.2× smaller cache tensor storage than coordinate-specific caching. These findings demonstrate DeltaGenome's cross-architecture applicability and benefits in execution and storage efficiency.

W2 — Contribution beyond prior work. The revision credits fastISM, Yuzu and causal prefix reuse (§2). DeltaGenome introduces architecture-specific mechanisms, including NT-v2's algebraic first-layer attention correction and Caduceus's shared directional frontiers, alongside NT-v3 selective encoder recomputation and Evo2 compact projection.

An independent graph-enabled HyenaDNA absolute-prefix comparator yields prefix/DeltaGenome speedups of 1.022× warm, 0.949× prepared and 0.962× raw, with cache sizes of 54.23/263.84 MiB. Both preserve full logits within tolerance, revealing state-representation trade-offs (§3, §5.2, Appendices C,G).

Question — Predictor utility. The deepest-state predictor is diagnostic, not required for structural or algebraic reuse. Selection quality is evaluated on 56 held-out variants across four backbones against equal-budget positional controls.

Median energy recall at 30% nucleotide-position budget:

ModelFullPosition-onlyDistance-onlyFixed-window
NT-v20.79170.73930.74850.7485
DNABERT-20.99170.99180.99240.9924
Caduceus0.99700.99690.99690.9969
HyenaDNA0.99550.99530.98640.9864

NT-v2 also improves profile Spearman/top-k overlap over position-only (0.5229/0.5545 vs 0.3137/0.4390). At 40%, recall gains over the best positional median are 5.19 percentage points (5.35 after overlap exclusion). Reference-conditioned information adds predictive value in NT-v2, whereas positional structure largely explains selection quality in the other backbones.

A separate layer-wise DNABERT-2 controller tests online speed–accuracy trade-offs including selection and execution overhead:

Token budgetLearned/distance errorLearned/distance warm latency
30%1.033/0.879%45.69/24.70 ms
40%0.933/0.733%45.97/24.67 ms

Cold latencies are 59.50/38.51 and 59.78/38.49 ms. Learned selection does not improve the tested trade-off. These distinct experiments clarify the diagnostic contribution of learned prediction and its independence from dependency-valid reuse (§5.4, Appendices H–J).

Reviewer_fL44OpenReview ↗
Rating
3(+2)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

DELTAGENOME studies inference for genomic sequences that differ from a shared reference by a sparse edit. It compares perturbation geometry and importance prediction across NT-v2, DNABERT-2, Caduceus, and HyenaDNA. The benchmark contains 384 central single-nucleotide variants in 2,047-bp contexts. Masked reconstruction uses dense alternate states and is evaluated separately from execution. The reported HyenaDNA executor achieves 1.54x warm-query speedup and 1.33x setup-amortized speedup at 128 queries, with a 198.06-MB reference cache. NT-v2 and Caduceus show near-parity warm-query speedups, and DNABERT-2 is slower than dense inference (pages 4-8).

Critical strengths
  1. The harmonized analysis shows meaningful differences in perturbation concentration across four architectures. Held-out prediction results include variant-level IQRs. This provides useful evidence within the stated benchmark (pages 5-7, Figures 2-4; page 12, A.4).

  2. The reported HyenaDNA result provides a conditional example of useful executable reuse. The paper also exposes DNABERT-2's slowdown and separates incremental memory from persistent cache cost. These observations appropriately qualify the acceleration claim (page 8, Figure 5).

Critical weaknesses
  1. The systems evidence is not sufficiently specified. The paper omits absolute timings, repeat counts, timing uncertainty, batch size, precision, backbone checkpoint identifiers, and measured executor errors. It also does not establish matching timed outputs and baseline kernels. The K-query workload is unclear as well. Section 4 refers to repeated alternates sharing a reference. However, one fixed reference has only three distinct alternate bases at the benchmark's single edited coordinate, while K reaches 128. Identical repeated queries, edits at other coordinates, or a separate workload could explain this, but the paper must identify which. These gaps prevent a reliable assessment of the practical speedup (pages 5 and 8; page 12, A.6-A.7).

  2. The added contribution is not isolated from established reuse methods and simple positional structure. fastISM (Nair et al., Bioinformatics, 2022) caches reference activations and recomputes mutation-affected convolutional regions, including in an Enformer evaluation. It is not a direct replacement for the four tested backbones, but its reference-reuse principle is relevant and absent from the related work. The successful HyenaDNA path is explicitly prefix reuse followed by suffix recomputation. All edits are central, and the predictor receives six explicit edit-relative positional features. No position-only or fixed-window baseline is reported. The paper needs a clearer account of its architectural extension and evidence that reference-conditioned prediction adds value beyond these positional features (pages 3-4, 7-8, and 11-12).

Questions for authors
  1. Can you provide the exact executor algorithms, backbone checkpoints, output-error measurements, and matched absolute timing protocol? The paper distinguishes measured amortization from its idealized fixed-cost model. What changes across the K-query measurements, and how are timings aggregated in Figure 5C (pages 8 and 12-13)?

  2. What are the benchmark source and split rules? How are K=8,32,128 queries generated around one reference when the stated benchmark fixes a single edit coordinate? Are they identical repeats, edits at other coordinates, or drawn from another workload (pages 5 and 8)?

  3. What does the proposed execution add beyond fastISM's dependency-based reuse principle and a same-backbone causal prefix-reuse implementation? How much predictor performance remains after comparison with position-only or fixed-window controls (pages 3, 7-8, and 11-12)?

I am willing to revise my rating upward if the response makes the central acceleration result verifiable. The key evidence is executor correctness, comparable timing measurements, and a clear account of the workload and contribution beyond existing reuse methods.

7 comments

Thank you for identifying the missing systems evidence, and we do appreciate the constructive insights. The revised manuscript replaces the original evaluation with matched execution experiments.

W1/Q1 — Timing, correctness and reproducibility. All four backbones use A100-SXM4-80GB, FP32 (TF32 disabled), sequential B=1 queries, five warmups and 20 paired repetitions per reference–K condition. Native dense (A), graph dense/reuse (BG/CG), and compiled graph dense/reuse (MBG/MCG) use identical requests and numerical settings. Prepared-system timing includes reference/graph preparation; raw timing additionally includes input preparation. At K=128:

ModelA/MBG/MCG (ms)A/MCGMBG/MCG [95% interval]Max. rel. L2 error
HyenaDNA1032.91/559.39/505.412.0421.107 [1.1033,1.1113]3.32e-7
NT-v24064.28/4117.67/4050.851.0031.017 [1.0164,1.0167]6.19e-6
Caduceus2170.91/1205.97/1231.001.7620.979 [0.9787,0.9798]2.08e-7
DNABERT-21599.80/1685.00/2412.080.6740.663 [0.6461,0.6880]0.155

Latencies are marginal medians; speedups are paired reference-level medians. Intervals use 50,000 reference-bootstrap resamples. HyenaDNA/NT-v2 win on 8/8 references and 159/160 and 160/160 repeats; neither other model wins. All 1,024 distinct queries per backbone are validated: HyenaDNA logits and NT-v2/Caduceus final hidden states satisfy 1e-5 tolerance. DNABERT-2 is approximate (median/P95/max errors 1.014/2.488/15.496%); graph integration adds zero measured error against the selective target.

Both compiled arms share CUDA Graphs and regional operators (FFT epilogue, RoPE, reverse-add, scale-bias), verified by capture records (Table 5). Reuse-specific dispatches differ; identical whole-model kernel traces are not claimed. Exact checkpoint revisions and execution settings appear in Table 4/Appendix D. All 1,776 source-manifest entries match; the upstream HyenaDNA encoder remains unarchived. Table 2A, §5.1 and Tables 6–14 report absolute timings, errors, variability, preparation costs and ablations. All 240 CUDA/wall-clock comparisons agree on the side of parity.

W1/Q2 — Workload and split. The 384 central SNVs comprise 153 pathogenic/common variants and 231 QTL entries from Feng et al.'s benchmark (2,047 bp, edit index 1,023). The split is 270/58/56 (train/calibration/test; seed 20260821). Four train–test overlaps are excluded in sensitivity analysis without retraining; six train–calibration overlaps remain. Execution uses a separate workload: eight references, each with 128 distinct edit coordinates in [768,1278], center first, nested K=1,8,32,128. The 1,352 memberships represent 1,024 distinct requests per backbone, not identical ALTs or K-sequence batches. HyenaDNA's matched warm/prepared/raw speedups are 1.602/1.107/1.063 at K=128, the first tested K exceeding both dense comparators under prepared-system timing (§4, §5.2, Appendices A,D–E).

W2/Q3 — Novelty and controls. The revision credits fastISM, Yuzu and prefix caching (§2). Contributions include NT-v2 algebraic attention correction and Caduceus shared directional frontiers, reducing K=128 cache storage from 1,555.50 to 19.64 MiB (79.2×), without matched acceleration. An independent graph-enabled HyenaDNA prefix comparator preserves all 1,024 outputs within tolerance. Prefix/DeltaGenome speedups are 1.022× warm, 0.949× prepared and 0.962× raw; cache sizes are 54.23/263.84 MiB. Delta reuse is therefore not universally superior (§5.2, Appendix G).

Equal-budget positional controls establish NT-v2's reference-conditioned information advantage: 30% recall 0.7917 versus 0.7393/0.7485 (position/distance); overlap-excluded recall 0.7949 versus 0.7424/0.7580. At 40%, gains over the best positional median are 5.19/5.35 percentage points (full/overlap-excluded). Other models show negligible median gains at 30–40%; NT-v2 loses at 10%. Masked reconstruction is diagnostic, not executable savings. A separate DNABERT-2 online controller loses to distance selection on 54/56 variants at both token budgets; none of 336 conditions beats native dense (§5.4, Appendices H–J).

The revised evidence establishes where reference reuse is numerically valid and practically beneficial, while explicitly retaining negative results.

Thank you for the detailed response and the new execution experiments. They make the timing results much more transparent, and I appreciate your honest reporting of negative results. However, the new evidence also raises concerns about the core contribution. First, most of the reported speedup over native dense execution appears to come from graph compilation rather than from reference reuse. Against matched optimized dense baselines, the gain is small and limited to HyenaDNA. Second, on HyenaDNA, the proposed method does not outperform a simple prefix caching baseline and requires more cache memory. Third, the sparse edit prediction does not yet translate into practical savings on any backbone. Because the abstract still emphasizes speedups over native dense execution, I encourage you to revise it to reflect these findings. For these reasons, I will keep my current score. I hope these comments are helpful for future revisions.

Thank you for reviewing the new experiments and acknowledging the improved transparency. I would like to clarify one factual point concerning the revised results.

The matched gains are not limited to HyenaDNA. At K=128, four of six backbones evaluated against matched optimized dense baselines achieve positive setup-inclusive speedups: HyenaDNA (1.107×), NT-v3 (1.105×), NT-v2 (1.017×), and Evo2-7B (1.016×). In particular, NT-v3 achieves a matched gain comparable to HyenaDNA while preserving bitwise equality in a separate audit. These results are reported in Table 2 and Sections 5.1–5.3.

Furthermore, although Caduceus does not achieve a matched latency improvement, its shared directional frontier mechanism reduces cache tensor storage by 79.2-fold relative to coordinate-specific caching, representing a distinct systems contribution.

I agree that the independent prefix comparator remains competitive on HyenaDNA. However, the proposed framework extends beyond causal prefix reuse, including algebraic attention correction in NT-v2, shared directional frontiers in Caduceus, selective encoder recomputation in NT-v3, and compact projection in Evo2.

The deepest-state predictor is explicitly positioned as a diagnostic tool, with held-out positional comparisons across four backbones. Structural and algebraic reuse mechanisms do not depend on this predictor.

I appreciate the suggestion regarding the abstract and will prioritize the matched results. I hope the distinction between the HyenaDNA-specific comparison and the broader cross-architecture contributions can be considered in the final assessment.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.