acceptodds
Under review as a conference paper at ICLR 2027

CALLOSUM: One Path between Seeing, Search, and Speaking

Abstract

Vision-language models (VLMs) increasingly serve as retrieval encoders in deployment systems, where their embedding space becomes a fixed interface for lookup collections, rerankers, and downstream applications. Yet naive generative fine-tuning can silently alter that interface: rank lists drift, thresholds shift, and candidates change despite stable aggregate retrieval scores. Can a unified VLM add generation while keeping retrieval drift within a measured budget? We propose Callosum, a one-path adaptation method that treats the original retrieval representation as an anchor. CAL–LO–SUM captures our method's three principles: preserving a cache-side anchor, adapting through low-rank operators, and unifying multimodal retrieval with generation. We show that Callosum shifts the retrieval–generation frontier beyond architecture-matched baselines and approaches the generation quality of a frozen-anchor-plus-generator reference. To examine system stability, we compare full rankings, top-10 candidate sets, and threshold decisions, finding Kendall , 93% top-10 overlap, and merely 3.2% threshold crossings. Thus, Callosum enables VLMs to see, search, and speak along one path while preserving the retrieval infrastructure.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Vision-language Models, Multimodal Retrieval, Retrieval-to-Generation Adaptation

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

This paper investigates how to equip retrieval-oriented vision-language models with generative capabilities while preserving existing multimodal retrieval indices and minimizing retrieval representation drift caused by generative fine-tuning. The authors propose CALLOSUM, a three-stage framework for integrating retrieval and generation capabilities. First, it constructs a retrieval anchor surrogate within the generative model through representation distillation. Second, it uses singular value decomposition (SVD) to decompose generative parameter updates into an anchor-aligned shared component and a relatively independent private component, thereby mitigating interference between the two tasks. Finally, an input-adaptive, layer-wise gating mechanism dynamically adjusts the fusion weights of retrieval and generative updates. The experiments cover image-text retrieval, visual question answering, hallucination evaluation, and retrieval stability analysis. The results show that CALLOSUM improves generation performance while maintaining the original retrieval model's aggregate Recall@1, and reduces GPU memory usage and inference overhead compared with deploying two full models.

Critical strengths
  1. The paper investigates the impact of generative fine-tuning on existing retrieval systems, rather than focusing solely on retrieval and generation performance. The authors point out that stable aggregate Recall@1 does not necessarily imply unchanged retrieval results. Accordingly, they further evaluate ranking correlation, Top-K candidate overlap, and changes in threshold-based decisions. This system-compatibility-oriented perspective is practically meaningful.

  2. The authors compare various parameter-efficient fine-tuning and model merging methods, and conduct ablation studies to examine the contributions of individual components. CALLOSUM improves visual question answering performance while maintaining the original retrieval model's aggregate retrieval performance. Furthermore, the paper includes comparisons with FR-Head and Two-Path, reporting GPU memory usage, latency, and end-to-end task performance. These evaluations extend beyond isolated retrieval or generation metrics.

Critical weaknesses
  1. I have some reservations about the motivation for adopting a unified model. I do not fully understand why retrieval and generation must be integrated into a single model, rather than simply using two independent models for the respective tasks. I am not entirely convinced that there is a compelling need for a unified model in real-world applications. In particular, considering the advantages of a two-model approach, such as task independence, easier maintenance, and avoiding interference with the existing retrieval system, I believe that the resource savings offered by model unification may not be sufficient to justify the additional complexity and loss of retrieval stability.

  2. Second, the paper argues that generative fine-tuning changes retrieval representations and that rebuilding a large-scale index is costly, thereby motivating the integration of retrieval and generation capabilities while preserving the existing index. However, the authors do not adequately quantify the actual cost of re-encoding the corpus and rebuilding the index, nor do they evaluate retrieval performance when index rebuilding is allowed. For many retrieval systems, re-encoding the entire corpus can be performed as an offline batch process and may not necessarily incur prohibitive costs. More importantly, a model after generative fine-tuning may simply become incompatible with the original embedding space, while its actual retrieval capability may not have deteriorated and could even have improved. Without further analysis, the paper does not sufficiently justify why a complex anchor-preservation mechanism is necessary instead of directly updating the model and rebuilding the index.

  3. Even if we accept the assumption that the existing retrieval index must remain unchanged, the paper still does not adequately explain why retrieval and generation capabilities need to be integrated through such a complex fusion mechanism. Since retrieval and generation tasks can usually be explicitly distinguished during inference, a natural and simpler alternative would be to maintain separate retrieval and generation LoRA adapters on a shared backbone and switch between them according to the current task. This approach could also avoid deploying two full models while reducing interference between task-specific updates. In particular, CALLOSUM's gating network uses prompt-format and modality information, while retrieval and visual question answering tasks in the experiments have clearly distinguishable input formats. Therefore, the gating mechanism may primarily be learning to identify task types rather than achieving deeper dynamic collaboration between the two capabilities. Without comparison against this simple yet competitive alternative, it is difficult to determine how much additional value the proposed complex fusion mechanism actually provides.

Questions for authors

Please refer to the concerns raised in the Critical Weaknesses section. I would particularly appreciate clarification on the practical necessity of a unified model, the justification for preserving the existing retrieval index, and the advantages of the proposed fusion mechanism over simpler task-specific LoRA switching.

2 comments
AuthorsOpenReview ↗

Response to jDck

We thank the reviewer for the comments. We follow the three points in your Critical Weaknesses section.

Weakness 1. Why use one model rather than separate retrieval and generation models?

From early vision-language foundation models to today's LLMs and VLMs, researchers have repeatedly joined retrieval with generation. CoCa (Yu et al., TMLR 2022) jointly trains retrieval embeddings and captioning, and BLIP (Li et al., ICML 2022) supports retrieval, captioning and VQA in one vision-language pretraining framework. GritLM (Muennighoff et al., 2024) brings the same idea to instruction-conditioned LLM embeddings and generation for text RAG. Recent VLM work now studies this as one serving workflow: SV-RAG (Chen et al., ICLR 2025) retrieves evidence from multi-page documents for question answering, URaG (Shi et al., AAAI 2026 Oral) uses early and later Qwen2.5-VL layers for retrieval and answering, and Hydra (Georgiou, 2026) combines ColBERT retrieval with autoregressive generation in one Qwen3.5 VLM. In Hydra's 4B setting, one-model serving reduces peak memory from 28.85 GB to 10.77 GB. This progression establishes a concrete need: visual RAG requires evidence retrieval and answer generation in the same request, and a unified model avoids keeping a second full model resident. Callosum adds generation to an existing service while retaining its corpus embeddings.

Our joint request searches an image collection, answers using the evidence and searches again using the interaction. In 2,000 such requests, Callosum reaches 79.9% answer accuracy and 62.9% complete-request success. Shuffling or removing the evidence lowers answer accuracy to 52.3% or 47.1% (supplementary Table 39).

SystemVQAv2 (%)Memory (GB)Linked latency (ms)
FR-Head84.123.9130
Two-Path85.146.1189
Callosum85.623.8138

Against Two-Path, we improve VQAv2 while using 48.4% less memory and reducing linked latency by 27.0%. Against FR-Head, we gain 1.5 VQAv2 points and raise complete-request success from 61.4% to 62.9% at an 8 ms latency increase.

Weakness 2. Why keep the existing index, and what happens if we rebuild it?

We test rebuilt galleries directly. Callosum retains its retrieval advantage.

Let A be the deployed anchor, M ordinary generation fine-tuning, C Callosum and S task-switched LoRA.

Query encoderGallery encoderCOCO R@1 (%)
AA64.2
CA64.2
MA60.0
SA62.8
MM62.5
CC63.5
SS63.5

Rebuilding raises M from 60.0 to 62.5 and S from 62.8 to 63.5. Thus changing the embedding space explains part of the original loss. Callosum reaches 64.2 using the existing gallery, exceeding both rebuilt alternatives. Keeping the existing gallery delivers a quality benefit as well as avoiding corpus re-encoding.

The retrieval-only reference at 66.7 uses its own encoder for both queries and gallery. It defines a different index from the deployed anchor at 64.2. Callosum keeps the existing corpus vectors and reaches 85.6 VQAv2 at 64.2 COCO R@1.

Re-encoding cost. We will add the rebuilt-gallery results, the cited unified-model research and a corpus re-encoding cost breakdown to the camera-ready version. The cost breakdown will state corpus size, hardware, encoding time, index-building time and storage, and distinguish measured values from scale estimates.

Weakness 3. What do we gain over switching between task-specific LoRA adapters?

Callosum improves both task quality and retrieval stability over task switching.

We run the proposed two-LoRA alternative on one backbone. We use the same initialization, data, training steps, number of trainable parameters, cache access and validation selection rule. An explicit task label chooses the adapter.

MeasureTask-switched LoRACallosum
COCO R@1, existing gallery (%)62.864.2
VQAv2 (%)84.085.6
GQA (%)66.568.1
POPE-Adversarial (%)81.082.5
Threshold crossings (%)4.53.2
Complete-request success (%)61.7562.9

Thus Callosum outperforms task switching on every reported measure.

We also compare the branch designs under the same gate. With task-specific layer weights, our shared and private branches add 1.0 COCO point and 0.5 VQAv2 points over standard adapters. With input features and a task label, they add 1.1 and 0.4 points. Input-dependent weights separately add 0.4 or 0.5 VQAv2 points, with COCO changes of -0.2 or -0.3 points. The branch gain holds under both routing choices. We will add these comparisons and settings to the camera-ready version.

We hope the above addresses the concerns, and thus, the reviewer would consider raising their score. We are happy to answer any further question!

Reviewer_NGVLOpenReview ↗
Rating
2
Confidence
1(−1)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes CALLOSUM, a practical VLM adaptation framework that facilitates retrieval VLMs with generative capabilities while minimizing the impact on their retrieval behavior. It has multiple low-rank parameter sets to respectively handle retrieval and generation, and integrates gating modules into the models to adaptively balance the updated parameters in different tasks. Experimental results show that CALLOSUM manages to improve generation performance of VLMs while preserving their retrieval capabilities.

Critical strengths
  • Clear method design: the method clearly separates retrieval and generation capabilities, and proposes different modules and a gating mechanism to handle the two tasks respectively.
  • Extensive experiments with strong empirical results: experimental results show that the method outperforms multiple baselines in the tradeoff between generation and retrieval. The results are obtained under multiple settings against multiple baselines.
Critical weaknesses
  • Insufficient problem motivation: in practical applications, there are not many cases where a model is required to do both retrieval and generation, and there are even fewer cases where such versatile models are built upon already deployed frozen retrievers with high performance. The authors do not establish why the target setting is preferable.
  • Over-sensitive to prompt format: the method suffers from performance decline under prompt format shifts. This also results in restricted generalization capabilities.
Questions for authors
  • For some VLMs, the last few layers are not involved in embedding calculation, so only training the last few layers for generation does not affect their retrieval performance. Can the authors provide additional comparisons with such baselines to further validate the effectiveness of the method?
  • Can the authors compare the method with backward-compatible representation learning methods?
  • How to set proper prompt formats in specific tasks? Can the authors provide additional guidance?
5 comments
AuthorsOpenReview ↗

Response to NGVL

We answer the motivation and three questions with experiments and practical guidance.

Weakness 1. Why combine retrieval and generation while keeping an existing index?

This is an established research direction, and our joint requests demonstrate its value.

From CoCa (Yu et al., TMLR 2022) and BLIP (Li et al., ICML 2022) to GritLM (Muennighoff et al., 2024), research has repeatedly joined retrieval and generation in vision-language models and LLM-based RAG. Recent VLMs make this joint workflow explicit: SV-RAG (Chen et al., ICLR 2025) retrieves evidence from multi-page documents for question answering, URaG (Shi et al., AAAI 2026 Oral) retrieves and answers within one Qwen2.5-VL, and Hydra (Georgiou, 2026) combines ColBERT retrieval and autoregressive generation in one Qwen3.5 VLM. Hydra reports 10.77 GB peak memory for one 4B model versus 28.85 GB for two resident models. Callosum brings this established direction to an existing service while retaining its stored corpus vectors.

In 2,000 retrieve-answer-retrieve requests, Callosum reaches 79.9% answer accuracy and 62.9% complete-request success. Shuffling or removing evidence lowers answer accuracy to 52.3% or 47.1%. Against Two-Path, Callosum improves VQAv2 from 85.1 to 85.6, uses 48.4% less memory and reduces linked latency by 27.0%. We will foreground this workload in the camera-ready introduction.

Q1. Does training only the last generation layers achieve the same result?

We freeze both retrieval embedding paths and train the last four generation layers after those outputs. We freeze the vision encoder, retrieval projection heads and shared upstream modules. We test both LoRA and full fine-tuning of these four layers.

Both versions score 0.5 points below the same backbone's instruction baseline on VQAv2, with zero change in retrieval embeddings. Training these layers preserves the retrieval mapping but does not provide Callosum's generation gain. Callosum also beats FR-Head on VQAv2 (85.6 vs 84.1), GQA (68.1 vs 66.0) and POPE-Adversarial (82.5 vs 80.3) (supplementary Table 32). We will include the last-four-layer results and exact architecture details in the camera-ready version.

Q2. How do we compare with BCT for reusing an older model's embeddings?

We use BCT (Shen et al., CVPR 2020) together with the same generation objective. We freeze the old model and gallery and train the new representations for retrieval against the old corpus vectors.

MethodCOCO R@1, old gallery (%)COCO R@1, own gallery (%)VQAv2 (%)
BCT with generation63.063.584.0
Callosum64.263.585.6

Both methods reach 63.5 with their own rebuilt galleries. Callosum raises old-gallery retrieval by 1.2 points and VQAv2 by 1.6 points. Our advantage covers both retrieval against the existing index and answering quality. We will add this comparison and the BCT implementation details to the camera-ready version.

Weakness 2 and Q3. How do we handle prompt changes, and which format should users choose?

Practical prompt guidance (Q3). Keep the prompt wrapper and format code consistent for each operation. Retrieval takes a caption or query description. VQA takes an image and an explicit question. For illustration, use "a red bus parked beside a building" for retrieval and "What color is the bus?" with the image for VQA. Keep the task identity unchanged when paraphrasing, and assign the format code using the template mapping.

Results under new prompt formats (Weakness 2). For a new prompt format, we train only the gate and freeze the backbone and adapters. Answering examples use answer labels. Retrieval examples use positive pairs or frozen-anchor targets. We separate calibration, model selection and final testing by image identity and template family. We use the same final test set at all four sample counts.

Calibration examplesVQAv2, held-out templates (%)COCO R@1 (%)
082.563.9
10082.7563.9
50083.563.9
2,00084.063.9

Calibration adds 1.5 VQAv2 points while COCO R@1 remains at 63.9. In the paper's separate experiments, mixed-template training reaches 83.9 VQAv2 and 64.1 COCO R@1, and multilingual calibration reaches 84.2 and 64.1 with 2,000 prompts (supplementary Tables 21 and 37). We will add the calibration curve, prompt examples and format-code mapping to the camera-ready version.

I appreciate the authors' detailed response and additional experiments. However,

1. Problem formulation

From my understanding, none of the provided related work shares an identical problem formulation to CALLOSUM:

  • CoCa, BLIP, and GritLM aim to support both retrieval and generation during training, instead of developing one capability on an already-deployed model with the other capability.

  • SVRAG and Hydra use task-specific adapters in a capability-decoupled manner. Adapters for one capability do not impact the inference path of the other capability.

  • URaG addresses a different task setting.

I am still not convinced of the motivation. Unifying retrieval and generation is reasonable. However, one can develop a unified model early in the training process or use separate models/adapters for different tasks. What are the unique advantages of "modifying the retrieval path of a high-performance, already-deployed industrial model to enable generation, and maintaining its retrievel embeddings at the same time"?

2. Comparisons with backward-compatible methods

The comparison with BCT is appreciated, but there are multiple more recent works with higher performance and possibly closer relations to CALLOSUM (e.g., XBT (ICCV25) aims to upgrade a retriever while maintaining the original embeddings). Though not sharing the same problem formulation, specific modules in these methods can be adopted as baseline, and the results will make the claim much more convincing, if applicable.

In summary, the statements and experimental results address some of my concerns, but I still think the motivation is rather unclear.

AuthorsOpenReview ↗

Clarification on Misunderstanding

On Problem Formulation and Potential Baselines that Do not Share the Same Problem Formulation As the Reviewer Flagged.

  • For a service whose archive is already indexed, the simplest alternative is to freeze retrieval and train a generation head.
  • We tested exactly that.
  • At essentially the same memory footprint (23.8 vs. 23.9 GB), adapting the shared path raised VQAv2 from 84.1% to 85.6% and complete-request success from 61.4% to 62.9% across 2,000 requests. The measured cost was that 3.2% of retrieval threshold decisions changed. Serving two full models required 46.1 GB. These results give an operator a concrete basis for deciding whether the answer-quality gain warrants the drift.
  • XBT’s projector could be incorporated into a generation-trained model. The distinction matters because XBT’s Eq. (2) evaluates whether a newer retriever improves aggregate retrieval against an old gallery. For a query with one relevant image ranked first, replacing the other nine top-10 images leaves every Recall@K unchanged. Consequently, even strong results under XBT’s criterion leave open which evidence a downstream generator receives and how well it answers.
  • Our existing comparisons isolate the deployment constraint directly. CALLOSUM, BCT trained with the same generation objective, and task-switched adapters all score 63.5% COCO Recall@1 when each uses its own rebuilt gallery. The tie breaks when all three must query the same deployed gallery. CALLOSUM reaches 64.2%, compared with 63.0% for BCT with generation and 62.8% for switched adapters. CALLOSUM also reaches 85.6% VQAv2, compared with 84.0% for both alternatives.
  • The retrieval advantage appears precisely when the service needs to reuse the index it already owns.
  • The unique advantage is stated in the paper clearly. We want to give an real-world example here. Users want to ask questions about retrieved images and use the answers in a follow-up search. In our 2,000 such requests, removing the retrieved evidence reduced answer accuracy from 79.9% to 47.1%. Reusing the index requires new queries to remain comparable with its stored vectors. Joint pretraining alone gives no such guarantee. Even a rotation of the embedding space preserves all of a replacement model’s internal similarities while potentially changing which items it retrieves from the old index.
Reviewer_FYVhOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Callosum adapts an instruction-tuned VLM to support retrieval and generation against frozen corpus embeddings. It combines retrieval-anchor distillation, shared–private low-rank updates, and input-dependent gating. The reported model matches the anchor’s rounded aggregate retrieval scores while improving generation. Query-level audits show that some ranking and threshold drift remains.

Critical strengths
  1. The fixed-index setting addresses a practical compatibility problem. Measuring candidate overlap, rank correlation, and threshold crossings adds useful information beyond aggregate recall. These measurements directly reflect what can change in an existing retrieval pipeline.

  2. The reported comparison includes a meaningful exact-interface control. Callosum improves VQAv2 from 84.1 to 85.6 relative to parameter-matched FR-Head. The ablations also indicate contributions from anchor distillation, private capacity, and adaptive fusion.

Critical weaknesses
  1. The benefit of the proposed decomposition is not clearly separated from task-conditioned routing. Removing prompt-format features reduces VQAv2 from 85.6 to 81.8. This makes a simpler task-conditioned adapter mixture an important control. The available evidence does not establish how much the shared–private construction adds beyond routing and anchor distillation.

  2. Key evidence cannot be assessed from the provided manuscript. The exact initialization checkpoints, training and cache composition, gate-feature implementation, and several matched comparisons are deferred to supplementary material unavailable here. These details matter for evaluating training-data overlap, resource matching, and reproducibility. This is a limitation of the available evidence, rather than a claim that the authors did not conduct these experiments.

Questions for authors
  1. How does Callosum compare with anchor and generation adapters combined by a simple task-conditioned gate, using the same cache access, features, parameter budget, and training schedule?

  2. Please identify the exact anchor and instruction checkpoints, training datasets, cache construction, and split boundaries. How is cache uncertainty computed at inference without running the external anchor?

  3. What is the incremental benefit over FR-Head under a fixed retrieval-drift budget? FR-Head preserves the interface exactly and has lower reported linked-request latency. A matched comparison would clarify when Callosum’s generation gain justifies its additional drift.

2 comments
AuthorsOpenReview ↗

Response to FYVh

We thank the reviewer for recognizing the value of keeping the existing index and measuring retrieval changes for individual queries. Our additional experiments separate the contribution of the branches from task routing and quantify the gain over FR-Head.

Q1 and Weakness 1. What does our branch design add beyond a simple task-based adapter mixture?

Our shared and private branches improve both tasks under either gate.

We use the same backbone initialization, training data, batch composition, training steps and validation selection rule for all four settings. We use the same number of trainable parameters, counting every adapter and gate, and give each method the same access to the cache and features.

BranchesGateCOCO R@1 (%)VQAv2 (%)
Shared and privateTask-specific layer weights63.884.5
Shared and privateInput features and task label63.684.9
Standard adaptersTask-specific layer weights62.884.0
Standard adaptersInput features and task label62.584.5

With task-specific weights, our branches add 1.0 COCO point and 0.5 VQAv2 points. With input-dependent weights, they add 1.1 and 0.4 points. All four settings have 92.0% Top-10 overlap, 3.0% changes in the top-ranked result, 4.5% threshold crossings and a 1.0-point nDCG loss with the frozen reranker. The branch advantage holds under both routing choices and the same measured threshold changes.

Within each branch design, input-dependent weights add 0.4 or 0.5 VQAv2 points. COCO R@1 changes by -0.2 or -0.3 points. This separates the gain from the branch design from the effect of routing within a task. The independent task-switched LoRA baseline scores 62.8 COCO and 84.0 VQAv2. Full Callosum reaches 64.2 and 85.6. We will add the full four-setting comparison and its training details to the camera-ready version.

Q2 and Weakness 2. Which models, data and gate features do we use, and how do we compute uncertainty online?

Models and data. The retrieval anchor defines the stored embedding space. We fit its surrogate inside the instruction model.

We build the cache from 50,000 COCO training image-caption pairs after excluding evaluation image and caption identifiers. We sample generation data from VQAv2, GQA and ChartQA in proportions of 0.50, 0.30 and 0.20. We use batch size 32 and 2,000, 6,000 and 2,000 training steps for the three stages. We split the 5,000 gate-development examples into 4,000 for training and 1,000 for selection. GQA generation uses train, gate development uses separate validation subsets, and test-dev is for final evaluation. We will include the exact model names, data versions, split definitions and gate-feature implementation in the camera-ready version.

Online uncertainty. For each new query, this internal surrogate produces the normalized vector qsq_s. We compare it with the 32 nearest normalized cached vectors cic_i and compute

u=-\frac{\sum_{i=1}^{32}p_i\log p_i}{\log 32}.$$ The external anchor provides offline training targets and cached vectors. We do not run it for incoming queries. We will make this distinction explicit in the formula and implementation description. The gate also uses prefix states, vision tokens, an eight-way prompt-format code and a modality mask. **Q3. What do we gain over FR-Head under the same limit on retrieval changes?** For each limit, we use the same predefined model set. We keep models whose validation crossing rate meets the limit and choose the one with the highest validation VQAv2. We then report its test results. | Validation limit | Test crossings | Callosum VQAv2 (%) | FR-Head VQAv2 (%) | |---|---:|---:|---:| | 1% | 1.0% | 83.75 | 84.1 | | 3% | 3.0% | 84.5 | 84.1 | | 5% | 5.0% | 84.6 | 84.1 | Callosum gains 0.4 and 0.5 points at the 3% and 5% limits. FR-Head has zero crossings. These rows follow validation selection. The paper's main Callosum model separately reaches 85.6 with 3.2% crossings, a 1.5-point VQAv2 gain. We will provide the full selection settings in the camera-ready version. Callosum also raises GQA from 66.0 to 68.1 and POPE-Adversarial from 80.3 to 82.5 over FR-Head. In 2,000 linked requests, it raises success from 61.4% to 62.9%. Memory is 23.8 versus 23.9 GB, and latency is 138 versus 130 ms. These figures quantify the generation and request-level gains alongside the retrieval measurements.

We hope the above addresses the concerns, and thus, the reviewer would consider raising their score. We are happy to answer any further question!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.