acceptodds
Under review as a conference paper at ICLR 2027

BIG: Bidirectional Information Gain for Visual Instruction Data Selection

Abstract

Existing visual instruction data selectors assess quality, coverage, or input contributions without jointly evaluating how the image and the particular instruction support the same reference answer. They can therefore retain examples whose answers require the image but are weakly constrained by the instruction, or follow from the instruction without requiring visual evidence. We propose BIG (Bidirectional Information Gain), which measures both conditional contributions using one frozen vision-language scorer. Each gain is the relative reduction in reference-answer loss from adding one input given the other. To align gain scales across sources and directions, BIG converts them to within-source percentiles and ranks examples by their minimum under fixed source quotas. This makes the weaker contribution decisive, preventing a strong direction from compensating for a weak one. BIG consistently outperforms random and single-direction selection in aggregate performance across two model–data settings, and also leads the comparison with established selectors under a shared training recipe. On Vision-Flan with Qwen2-VL-2B, BIG exceeds full-data aggregate performance across seven benchmarks using 50% of the data. Ablation analysis shows that removing calibration or allowing compensation shifts selection toward examples with weaker support in one direction. Selected-data analysis shows that BIG systematically recovers examples with stronger contributions in the direction overlooked by single-direction criteria.

Primary area:
foundation or frontier models, including LLMs
Keywords:
visual instruction data selection, visual instruction tuning, multimodal large language models, vision-language models, data selection, data-efficient fine-tuning, conditional dependence

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

This paper proposes BIG, a visual instruction data selection method that measures how much the image and instruction each contribute to predicting the same reference answer. Experiments on Vision-Flan and LLaVA-665K across seven benchmarks report improvements over random, single-direction, and existing selection methods.

Critical strengths

On Vision-Flan with Qwen2-VL-2B, BIG slightly exceeds full-data average performance using 50% of the data. Ablations and selected-data analyses examine the effects of percentile calibration and minimum aggregation.

Critical weaknesses
  • Experiments only consider 50% multimodal retention. So, evaluating different retention ratios and reporting total scoring and training costs would provide more stronger evidence of practical efficiency.
  • Only one training seed is reported, while improvements over the strongest comparators are approximately 0.8–1.1 average RP points. Multiple training runs are needed to establish whether these gains are robust.
  • It would be useful to examine whether the retained TextVQA performance is explained by OCR-text or answer coverage. Comparing these statistics with random selection and reporting results on answers unseen in the fine-tuning data would help clarify this.
Questions for authors

See the weakness

1 comment
AuthorsOpenReview ↗

Response to Reviewer YZro

W1: Only 50% retention is evaluated; total scoring + training cost is missing.

BIG's advantage is not limited to 50%. At the commonly used budgets of 25%, 50%, and 75%, BIG clearly outperforms same-budget random selection (Random), and it exceeds full-data training at 25% and 50%; at the extremely low budgets of 5% and 10%, BIG and Random are roughly on par. In terms of cost, a single run of BIG (scoring plus training) takes about as long as full-data training while performing better.

Results per retention ratio (Vision-Flan / Qwen2-VL-2B, Avg. RP, Full-data = 100):

Retention5%10%25%50% (Table 1)75%
BIG97.2998.29100.22100.5199.85
Random97.7498.6898.7798.7397.52
BIG − Random−0.45−0.39+1.45+1.78+2.33

The margin grows with the budget (+1.45 → +1.78 → +2.33). At extremely low budgets, random selection is a well-known strong baseline: PreSel (CVPR 2025) notes at a 15% budget that "random selection outperforms many baselines"; COINCIDE (EMNLP 2024) finds on LLaVA-1.5 at a 20% budget that "all baselines perform worse than the random sampling"; and on Vision-Flan at 5%, COINCIDE, EL2N, CLIP-Score, and PRISM also fall below random selection (Sun et al., 2026, Table 7).

Total scoring + training time (8×H20, full-parameter training):

SettingScoring (one-off)TrainingTotalAvg. RP
Full-data training—~46 min~46 min100
BIG 25%~40 min~12 min~52 min100.22
BIG 50%~40 min~23 min~63 min100.51

Scoring is done once, whereas training cost is paid again in every run. In workflows that involve repeated training, such as hyperparameter tuning, multiple seeds, comparing budgets, or upgrading the model, the scoring cost is amortized across all runs, and each subsequent run needs only a fraction r of the full-data training time (about 23 min at 50% vs. about 46 min for full data). The cost comparison with gradient-based methods is given in the General Response.

W2: Only one training seed; the margin over the strongest comparator is only ~0.8–1.1 points.

The margin holds across training seeds: BIG outperforms the strongest single-direction comparator, G_Q-only, in all five paired seeds, by about 1.54 points on average.

We added four training seeds for both BIG and G_Q-only (five in total: 40/41/42/43/44). The selected subsets are fixed; only the training seed changes:

Method (50%)Avg. RP (mean ± std, 5 seeds)
BIG100.74 ± 0.53
G_Q-only99.2 ± 0.5
Random98.09 ± 0.79

BIG also outperforms Random by about 2.65 points on average. At the 25% budget, five seeds give BIG 100.15 ± 0.61 vs. Random 98.14 ± 0.55, with BIG ahead in all five seeds.

W3: Is the TextVQA performance explained by OCR-text or answer coverage?

No. BIG's answer coverage is in fact lower than Random's and its OCR-token coverage is essentially the same, yet BIG still leads consistently on questions whose answers neither subset has seen.

  • Coverage statistics: answer coverage is 40.22% for BIG vs. 41.46% for Random; OCR-token coverage is 54.43% vs. 54.30%.
  • Unseen-answer subset: on the 2,764 TextVQA questions whose answers are covered by neither fine-tuning subset, BIG outperforms Random in all five seeds, by 3.97, 1.57, 1.40, 3.12, and 2.47 points (2.51 on average).

This is consistent with §5.5: BIG recovers examples with low image gain but high instruction gain, which require the model to read specific text in the image as directed by the question.

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

This paper proposes Bidirectional Information Gain (BIG), a data selection method for visual instruction tuning. BIG uses a frozen vision-language model to measure the conditional contributions of the image and instruction to predicting the same reference answer. The two scores are converted into within-source percentiles and combined using their minimum, followed by selection under fixed source quotas. Experiments on Vision-Flan and LLaVA-665K show improvements over random and single-direction selection, with 50% data retention slightly exceeding full-data aggregate performance in the Vision-Flan/Qwen2-VL-2B setting.

Critical strengths
  1. Clear presentation. The paper is well organized, and the motivation and selection procedure are clearly explained. The illustrative examples help readers understand the distinction between image and instruction contributions.
  2. Promising empirical results. BIG consistently improves aggregate performance over random selection and its single-direction variants in the two controlled settings. Achieving comparable or slightly better aggregate performance with substantially fewer training examples suggests practical potential.
  3. Interesting motivation. The observation that strong visual dependence does not necessarily imply meaningful instruction conditioning, and vice versa, is useful. Explicitly considering both contributions provides an intuitive perspective on visual instruction data selection.
Critical weaknesses
  1. Limited methodological novelty. Several central ingredients have close precedents: loss-based input contribution estimation, combining complementary contribution signals, and minimum-based aggregation. In particular, DICS already combines two information-gain signals through a minimum operator. BIG differs by aligning both directions to reference-answer prediction and introducing within-source percentile calibration, but the contribution appears incremental. Although the paper discusses these distinctions, it lacks direct comparisons with closely related methods such as DICS and mIFD, making the practical significance of the differences difficult to assess.

  2. Incomplete baseline coverage. The comparisons include several relevant selectors, but omit important gradient-based approaches such as TIVE and ICONS. Consequently, the results do not establish BIG’s competitiveness against methods that estimate training utility through gradients. Comparisons under matched data budgets and downstream training recipes, with differences in auxiliary data requirements and selection costs explicitly reported, would strengthen the evaluation. Moreover, the external baseline comparison is restricted to a single model–dataset setting and a 50% retention budget.

  3. Limited model and dataset coverage. The evaluation uses only Qwen2-VL-2B and LLaVA-1.5-7B, together with two established instruction datasets. There are no experiments on more recent model families, such as Qwen2.5-VL or later Qwen-VL models, or LLaVA-OneVision. Since the selection scores depend on the frozen model’s predictive behavior, it remains unclear whether the observed benefits extend to stronger models and larger, more diverse training mixtures.

  4. Insufficient efficiency and scaling analysis. BIG requires neither gradients nor additional scorer training, but it still performs three scoring passes per candidate example. Its computational advantage therefore needs to be measured rather than inferred. The paper should report selection time, GPU-hours, peak memory, and total selection-plus-training cost against representative baselines, particularly gradient-based methods. Experiments across larger candidate pools and multiple retention ratios would also clarify whether BIG maintains a favorable performance–cost trade-off and whether its gains become more substantial as data scale increases.

Questions for authors

See weakness above

6 comments
AuthorsOpenReview ↗

Response to Cfxg (1/3): Novelty and DICS

We thank the reviewer for the comments. We added direct comparisons with mIFD, TIVE, and ICONS, four newer models, five retention ratios, and complete cost measurements. For the remaining requests (reproducing the contemporaneous DICS, re-running every external method at multiple budgets and settings, and validating on larger candidate pools), we believe the current experiments already answer the underlying questions, as explained point by point below.

W1: Limited novelty; DICS already combines two signals with a minimum; no direct comparison with DICS or mIFD.

BIG's contribution does not lie in the minimum operator itself, but in aligning both directional contributions to the same reference answer, keeping the other input present when measuring each one, and using within-source percentile calibration so that the minimum can work. These design choices make a substantial difference: BIG outperforms mIFD by 5.91 points in a direct comparison, and DICS is contemporaneous work with no public implementation.

Taken individually, the components of any selection method have precedents. Table 3 shows that the design choices above determine whether the method works at all: removing calibration (94.34), or replacing the minimum with the mean (94.23) or the product (96.77), all fall below Random (98.73), whereas BIG reaches 100.51.

mIFD: we added the complete mIFD method in the paper's external-comparison setting (table in W2). It achieves 93.19 vs. 99.10 for BIG. mIFD's two terms predict different targets (the whole dialogue, and the answer without the image) and are combined by a product, which is precisely the kind of design that performs poorly in Table 3.

DICS: DICS first appeared on arXiv on 31 August 2026, about two weeks before the ICLR 2027 full-paper deadline (16 September). Under the ICLR 2027 Reviewer Guidelines, work published within two months is contemporaneous, and "the lack of such comparisons cannot be a basis for rejection." Nevertheless, §2.3 of our submission already cites DICS and discusses how it differs from BIG: DICS's second direction reconstructs the instruction from the answer, whereas BIG predicts the answer given the image, the opposite direction. DICS additionally relies on budget-adaptive diversity sampling. As of this response, DICS's public repository contains no scoring or selection code, so a faithful reproduction is not possible.

AuthorsOpenReview ↗

Response to Cfxg (2/3): Baselines and Fair Comparison

W2: Missing gradient-based baselines (TIVE, ICONS); matched budgets and recipes; auxiliary data and selection cost; the external comparison uses one setting and one budget.

Under the same budget and training recipe, BIG outperforms TIVE and ICONS (by 3.73 and 2.44 points, respectively), while its complete selection is faster and requires neither a warm-up model nor auxiliary data.

We added TIVE, ICONS, and mIFD in the paper's external-comparison setting (Vision-Flan / LLaVA-1.5-7B, 50%, shared LoRA recipe). The table below merges them with the original results of Table 2 (* marks the newly added methods):

MethodMMBenchTextVQAScienceQAMMEMMStarAI2DRealWorldQAAvg. RP
Random95.4695.9297.3698.2694.6395.3193.6695.80
Loss / Perplexity84.7296.7494.80100.7697.8588.7694.2593.98
CLIP-Score89.40103.6197.1499.2099.2391.7391.6395.99
COINCIDE95.76101.5599.0294.83104.5490.6590.0796.63
IFD93.3498.4898.1290.3896.8895.3586.1494.10
PreSel95.1699.4098.6498.5197.0595.5993.7396.87
ScalSelect99.2497.9198.20100.4498.8297.1390.8597.51
VisNec102.57104.43100.6896.2292.7595.8995.5598.30
mIFD*86.8197.6194.5399.8296.4094.7182.4393.19
TIVE*86.0997.6896.53100.0996.2193.1397.8495.37
ICONS*90.4397.8399.8596.86101.5289.06101.0896.66
BIG98.7897.69101.8297.32103.7497.5096.8899.10

Auxiliary-data requirements and per-stage costs are in the cost table of the General Response: TIVE and ICONS both require a warm-up model, and ICONS additionally requires target-task data; their complete selection takes about 143 and 240 minutes, respectively, versus about 118 minutes for BIG.

On a single setting and a single budget: the purpose of the external comparison is to position BIG against existing methods under a fully matched setting. Every method in the table above selects data with its own official scoring and selection procedure, from the same candidate pool, at the same 50% budget, and is trained with the same LoRA recipe. Running the main comparison at a single budget is also standard practice in this area: both main tables of PreSel (CVPR 2025) compare methods only at 15%, and TIVE's baseline comparison is likewise conducted at a single 15% budget. The table above contains ten external methods; each additional budget or setting would require re-running selection and training for all of them. The question such experiments would answer, namely whether BIG's effectiveness depends on the budget and the model, is already addressed directly: five retention ratios from 5% to 75% in the controlled setting (see our response to YZro W1) and four newer models (W3).

AuthorsOpenReview ↗

Response to Cfxg (3/3): Generalization and Efficiency

W3: Only two models and two datasets; no Qwen2.5-VL, newer Qwen-VL, or LLaVA-OneVision; no larger, more diverse mixtures.

Yes, BIG remains effective. On four models from the three families the reviewer named, BIG matches or exceeds full-data training with half of the data and outperforms same-budget Random.

The experiments use the same candidate pool, source quotas, and 50% budget as the Qwen2-VL-2B setting. Qwen2.5-VL-3B, Qwen3-VL-2B, and LLaVA-OneVision-0.5B follow the full-parameter recipe of the Qwen2-VL-2B setting; LLaVA-OneVision-7B is trained with LoRA (r = 128, learning rate 2e-4). BIG uses each model as its own scorer, and RP is relative to each model's own full-data training:

ModelBIG 50%Random 50%BIG − Random
Qwen2.5-VL-3B100.25100.02+0.23
Qwen3-VL-2B100.5497.60+2.94
LLaVA-OneVision-0.5B100.0995.55+4.54
LLaVA-OneVision-7B (LoRA)100.3799.79+0.58

Together with the two models in the paper, BIG is now validated on six models from three families: Qwen-VL, LLaVA-1.5, and LLaVA-OneVision.

On training data: LLaVA-665K (five sources, 665K examples) and Vision-Flan (191 tasks) are the standard candidate pools for data selection in this area. COINCIDE and PreSel in the table above are evaluated on both, and the newly added TIVE and ICONS use LLaVA-665K in their main experiments. Our experimental setup follows established practice in the field.

W4: Efficiency must be measured; performance–cost analysis across larger pools and multiple ratios; do gains grow with data scale?

The efficiency advantage is now measured rather than inferred, and it extends to larger candidate pools. Every item the reviewer listed (scoring time, GPU-hours, peak memory, selection time, and training time) has been measured; see the General Response. BIG's complete selection is faster than TIVE's and ICONS's and requires neither a warm-up model nor target-task data, and a single set of scores serves all retention ratios and can be reused across models of the same family.

On scale: in the multi-ratio experiments, BIG's margin over Random increases with the retention ratio (+1.45 at 25%, +1.78 at 50%, +2.33 at 75%). BIG's scoring cost grows linearly with the candidate-pool size, with no quadratic term such as pairwise similarity, so it applies directly to larger pools. How downstream gains change with candidate-pool size is a separate research question about data scaling and does not affect the paper's conclusions about the selection criterion itself.

Reviewer_p6QpOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces BIG (Bidirectional Information Gain), a training-free data selection method for visual instruction tuning. Using a frozen scorer, BIG measures how much the image and the instruction each improve prediction of the same reference answer, computes gains under image-absent and instruction-absent conditions, calibrates them into within-source percentiles, and ranks examples by the minimum of the two percentiles so that a strong contribution in one direction cannot compensate for a weak one. Selection is performed under fixed source quotas, preserving the source mixture. Experiments on Vision-Flan/Qwen2-VL-2B and LLaVA-665K/LLaVA-1.5-7B show that BIG improves aggregate relative performance over random and single-direction selection and compares favorably with existing selectors, while exchange and group-level analyses show it recovers complementary supervision but also shifts downstream strengths across contribution groups.

Critical strengths
  1. Measuring both image and instruction contribution against the same reference answer is intuitive and addresses a real gap: most selectors focus on one input or on quality/coverage, not on joint conditional contribution. The method requires no auxiliary scorer training and no tunable mixing weight.

  2. Within-source percentile calibration handles scale differences between image and instruction gains. The minimum aggregation is a principled non-compensatory rule: an example must be strong in both directions. The ablation in Table 3 supports both choices; removing calibration or using mean/product aggregation substantially lowers performance.

  3. The paper goes beyond aggregate scores by analyzing exchanged samples, matched comparisons on source/format/length/NLL, and group-specific TextVQA performance. This helps explain why joint selection works and shows it recovers complementary supervision rather than just selecting easier examples.

Critical weaknesses
  1. BIG requires three teacher-forced forward passes per example. For large pools this is expensive. The paper does not report wall-clock time, GPU memory, or scoring cost relative to baselines, even though “training-free” does not mean “cost-free.”

  2. Gains reflect the predictive behavior of the scorer. If the scorer differs from the target model or is weak, rankings may not transfer. The paper uses the same base model as scorer and target, but does not test cross-model scorer transfer or robustness to scorer quality.

  3. The image-absent pass zeroes visual embeddings and suppresses attention; the instruction-absent pass replaces the instruction with a generic prompt. These are reasonable approximations, but they may not perfectly isolate image or instruction contribution and can introduce artifacts. The paper does not validate these constructions beyond citing prior work.

Questions for authors
  1. No explicit diversity control within sources. Within each source, selection is purely contribution-based. This can select many similar high-gain examples and reduce coverage. Source quotas preserve coarse mixture but not fine-grained diversity. The paper does not analyze redundancy among selected examples.

  2. Trade-offs across contribution groups remain unresolved. The TextVQA diagnostic shows BIG gives up 3.15 points on high-image/low-instruction examples while gaining 3.41 points on low-image/high-instruction examples. This is an honest analysis, but it weakens the “consistent gains” framing: BIG redistributes downstream strengths rather than uniformly improving all groups.

2 comments
AuthorsOpenReview ↗

Response to Reviewer p6Qp

W1: Cost of three forward passes: wall-clock time, GPU memory, and cost relative to baselines.

The cost of the three forward passes is moderate and lower than that of gradient-based methods. Scoring the full pool of 186K candidates with Qwen2-VL-2B takes about 40 minutes, with a peak of about 5–6 GiB per scoring process. In the external-comparison setting (LLaVA-1.5-7B), BIG's complete selection takes about 118 minutes, versus about 143 minutes for TIVE and 240 minutes for ICONS; both of the latter require training a warm-up model, and ICONS additionally requires target-task data. The per-stage measurements are in the General Response. The scores can be reused for any retention ratio and across models (see W2).

W2: Can the rankings transfer when the scorer differs from the target or is weak?

Yes. A 50% subset selected with Qwen2-VL-2B scores, used directly to train the newer Qwen3-VL-2B and Qwen2.5-VL-3B, outperforms Random on both targets and is even slightly better than the targets' own BIG subsets (RP relative to each target's own full-data training):

TargetSubset from Qwen2-VL-2BBIG scored by the target itselfRandom
Qwen3-VL-2B100.96100.5497.60
Qwen2.5-VL-3B100.73100.25100.02

BIG's rankings therefore transfer across models within the same family: scoring once with an earlier-generation model can serve later, newer models. By default, BIG scores with the target itself and does not require an additional, stronger scorer.

W3: The input-absent constructions may introduce artifacts.

Replacing either construction leaves the scores and the selected subset largely unchanged, so BIG's selection does not hinge on a particular masking implementation or a particular prompt.

  • Image side: on Qwen2-VL, "zeroing + attention masking" and "padding + attention masking" give identical answer logits and NLL. When distinguishing samples with their original image from samples paired with a mismatched image, our construction achieves an AUROC of 0.730, comparable to directly removing the visual tokens (0.721).
  • Instruction side: on 4,101 examples covering all 191 sources, replacing the default construction with two generic prompts that contain none of the original question gives Spearman correlations of 0.925 / 0.931 for the G_Q ranking and 0.893 / 0.875 for the final BIG ranking; about 89% of the 50% subset is unchanged.

Q1: No explicit diversity control within sources.

We do not observe such redundancy. Under the same budget and source quotas, we compared BIG with matched random subsets on within-source image duplication, same-image-same-question duplication, near-duplicate QA text, and semantic coverage of the candidate pool by the subset. On all four metrics BIG differs from matched Random by less than 0.5 percentage points (the coverage score differs by about 0.003). Source quotas together with within-source percentiles already provide adequate coverage.

Q2: Trade-offs across contribution groups weaken the "consistent gains" framing.

By "consistent" we have always meant aggregate performance across settings, and the final version will make this wording explicit. The abstract states that BIG "consistently outperforms … in aggregate performance," and §5.5 and the conclusion explicitly state that BIG redistributes downstream strengths across contribution groups. To avoid misreading, in the final version we will change "consistent gains" in the §1 contribution list and the §5.1 heading to "consistent aggregate gains." We thank the reviewer for recognizing the candor of this analysis.

This redistribution also does not come at the expense of most examples. In a five-seed analysis with the contribution groups held fixed, BIG outperforms both Random and full-data training in all five seeds in the high-image/low-instruction, low-image/high-instruction, and high/high groups, by 2.13, 3.08, and 1.29 points over Random on average, respectively.

I appreciate your response. As some of my concerns have not been sufficiently resolved, I will retain my score.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We thank all three reviewers for their constructive feedback. To address the concerns raised, we added: more retention ratios (5%–75%) and five training seeds; direct comparisons with three external methods (mIFD, TIVE, and ICONS); four newer models (Qwen2.5-VL, Qwen3-VL, and LLaVA-OneVision at 0.5B and 7B) together with cross-model transfer; measured scoring and training costs; and mechanism analyses of the input-absent constructions, within-source redundancy, and TextVQA coverage. Point-by-point responses follow under each reviewer.

Shared concern: selection cost. All three reviewers asked about BIG's actual cost, so we report the measurements once here; the individual responses cite only the numbers relevant to each question.

  • Scoring cost. BIG's three conditions share the same frozen scorer and visual encoding. On 8×H20, scoring the full Vision-Flan pool (186K examples) with Qwen2-VL-2B takes about 40 minutes (about 5.3 GPU-hours), with a peak of about 5–6 GiB per scoring process; selection takes about 10 seconds.
  • Comparison with gradient-based methods (external-comparison setting: Vision-Flan / LLaVA-1.5-7B; wall-clock time, excluding downstream training):
MethodWarm-up modelTarget-task dataWarm-up trainingCandidate scoringOther stagesTotal
TIVEYesNo14 min (14,707 ex.)129 min—~143 min
ICONSYesYes (we use 2,906 independent support examples, deduplicated against the evaluation sets)9 min (9,302 ex.)220 minvalidation gradients 10 min; influence + voting ~1 min~240 min
BIGNoNo—~118 minselection ~10 s~118 min

Warm-up training was measured on 8×A800 and all other stages on 8×H20. Since the A800 has higher training throughput than the H20, measuring everything on H20 would only lengthen the TIVE and ICONS totals. BIG's complete selection pipeline is about one-sixth shorter than TIVE's and about half of ICONS's, and it requires neither a warm-up model nor any auxiliary data.

  • Reusable scores. Scoring is a one-off cost: the same scores can be used for any retention ratio and reused for other models of the same family (see our response to p6Qp W2). All multi-ratio, multi-seed, and cross-model experiments in this rebuttal were run from a single set of Qwen2-VL-2B scores, without re-scoring.

The paper's limitations section already acknowledges this preprocessing cost, and the final version will include the measurements above.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.