Paper deep dive
From Generalist to Specialist: A Context-Fusion Framework for Endoscopic Polyp Reporting with a Frozen VLM
Ruijie Yang, Yan Zhu, Peiyao Fu, Siyuan Li, Te Luo, Zhihua Wang, Quanlin Li, Pinghong Zhou, Xian Yang, Shuo Wang
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 93%
Last extracted: 8/22/2026, 2:32:15 AM
Summary
The paper introduces a context-fusion framework to specialize a frozen general-purpose Vision-Language Model (VLM) for endoscopic polyp reporting without modifying its pretrained weights. The framework combines explicit transduction context (retrieved image-report pairs via a self-supervised polyp encoder) and implicit instruction context (learned continuous specialist tokens). Experiments on 2,056 expert-annotated images show the framework outperforms general-purpose VLMs, task-specific predictors, and weight-adaptation methods, adding only 0.006% trainable parameters.
Entities (7)
Relation Signals (6)
Context-Fusion Framework โ uses โ Frozen VLM
confidence 95% ยท We introduce a context-fusion framework that specializes a frozen general-purpose VLM... without modifying its pretrained weights.
Context-Fusion Framework โ incorporates โ Implicit Instruction Context
confidence 94% ยท The framework combines two forms of specialist context... Implicit instruction context is represented by learned continuous specialist tokens
Context-Fusion Framework โ incorporates โ Explicit Transduction Context
confidence 94% ยท Explicit transduction context consists of retrieved imageโreport pairs that supply query-specific clinical evidence.
Context-Fusion Framework โ evaluatedon โ EndoVL
confidence 93% ยท Experiments were conducted on 2,056 expert-annotated public endoscopic images... derived from the public EndoVL release
Explicit Transduction Context โ generatedby โ EndoFinder
confidence 92% ยท EndoFinder [36], a self-supervised polyp encoder trained on polyp images is utilized in this study... to retrieve related image-report pairs
Context-Fusion Framework โ targets โ Paris Classification
confidence 90% ยท Reliable endoscopic polyp reporting requires integrating... standardized Paris classification... The framework... prompts it to estimate diameter, classify Paris type
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Reliable endoscopic polyp reporting requires integrating quantitative lesion sizing, standardized Paris classification, and clinically meaningful morphological description within a single record. General-purpose vision-language models (VLMs) offer a unified interface for image understanding and report generation. Existing specialization strategies, however, typically rely on task-specific models or model-weight adaptation, leaving unresolved how to introduce reliable specialist knowledge while preserving both this unified interface and the VLM's pretrained capabilities. We introduce a context-fusion framework that specializes a frozen general-purpose VLM through both implicit instruction context and explicit transduction context without modifying its pretrained weights. Specifically, a self-supervised polyp encoder retrieves related image-report pairs as explicit, query-specific evidence, while learned continuous specialist tokens provide implicit instruction context shared across cases. Experiments were conducted on 2,056 expert-annotated public endoscopic images. We compared the framework with general-purpose VLMs, task-specific predictors, and weight-adaptation methods to assess specialist performance, unified reporting, and adaptation efficiency. Across numerical, categorical, and report-generation metrics, the proposed framework substantially improved direct frozen-VLM inference and achieved the strongest overall performance among the evaluated methods. It added trainable parameters equal to only 0.006% of the frozen VLM's parameter count. When the top-1 retrieved case carried the correct target category, our framework corrected 70.5% of the errors made by a weight-adaptation baseline. These findings support the context-fusion framework as a lightweight and effective strategy for specialist adaptation of a frozen VLM.
Tags
Links
- Source: https://arxiv.org/abs/2608.15580v1
- Canonical: https://arxiv.org/abs/2608.15580v1
Trouble viewing inline? Open PDF directly โ
Full Text
70,942 characters extracted from source content.
Expand or collapse full text
From Generalist to Specialist: A Context-Fusion Framework for Endoscopic Polyp Reporting with a Frozen VLM Ruijie Yang a,b,c,d,1 , Yan Zhu e,f,1 , Peiyao Fu e,f , Siyuan Li c,d , Te Luo c,d , Zhihua Wang a,b , Quanlin Li e,f , Pinghong Zhou e,f,โ , Xian Yang g,โ and Shuo Wang c,d,โ a Zhejiang University, Hangzhou, China b Shanghai Institute for Advanced Study, Zhejiang University, Shanghai, China c Shanghai Key Laboratory of MICCAI, Shanghai, China d Digital Medical Research Center, School of Basic Medical Sciences, Fudan University, Shanghai, China e Endoscopy Center and Endoscopy Research Institute, Zhongshan Hospital, Fudan University, Shanghai, China f Shanghai Collaborative Innovation Center of Endoscopy, Shanghai, China g Alliance Manchester Business School, The University of Manchester, Manchester, United Kingdom A R T I C L E I N F O Keywords: Colonoscopy report generation Context fusion Multimodal retrieval Vision-language model A B S T R A C T Reliable endoscopic polyp reporting requires integrating quantitative lesion sizing, standardized Paris classification, and clinically meaningful morphological description within a single record. General- purpose vision-language models (VLMs) offer a unified interface for image understanding and report generation. Existing specialization strategies, however, typically rely on task-specific models or model-weight adaptation, leaving unresolved how to introduce reliable specialist knowledge while preserving both this unified interface and the VLMโs pretrained capabilities. We introduce a context- fusion framework that specializes a frozen general purpose VLM through both implicit instruction context and explicit transduction context without modifying its pretrained weights. Specifically, a self-supervised polyp encoder retrieves related imageโreport pairs as explicit, query-specific evidence, while learned continuous specialist tokens provide implicit instruction context shared across cases. Experiments were conducted on 2,056 expert-annotated public endoscopic images. We compared the framework with general-purpose VLMs, task-specific predictors, and weight-adaptation methods to assess specialist performance, unified reporting, and adaptation efficiency. Across numerical, categorical, and report-generation metrics, the proposed framework substantially improved direct frozen-VLM inference and achieved the strongest overall performance among the evaluated methods. It added trainable parameters equal to only 0.006% of the frozen VLMโs parameter count. When the top-1 retrieved case carried the correct target category, our framework corrected 70.5% of the errors made by a weight-adaptation baseline. These findings support the context-fusion framework as a lightweight and effective strategy for specialist adaptation of a frozen VLM. 1. Introduction Colonoscopy plays a central role in colorectal-cancer prevention by enabling polyp detection and the removal of premalignant lesions [1]. Within this clinical work- flow, reliable polyp reporting supports lesion management, post-polypectomy surveillance, clinical communication, and quality review [2, 3]. Reporting standards require explicit documentation of polyp size and morphology [4], while the Paris classification standardizes macroscopic lesion type [5]. Automated polyp reporting is challenging because several clinically distinct judgments must be made from a single monocular image. The system must estimate lesion size without an intrinsic scale, distinguish subtle morphological patterns for Paris classification, and describe surface, base, border, and associated findings in clinically meaningful language [6, 7]. โ Corresponding authors zhou.pinghong@zs-hospital.sh.cn (P. Zhou); xian.yang@manchester.ac.uk (X. Yang); shuowang@fudan.edu.cn (S. Wang) ORCID(s): 1 These authors contributed equally to this work. Existing approaches to automated polyp reporting broadly follow two paradigms. Modular pipelines [8] predict prede- fined clinical attributes and assemble them into a template- based report, making selected findings directly assessable but limiting the report to a fixed set of outputs. End-to- end image-to-text systems [9, 10] generate more flexible descriptions through a unified interface. However, errors in size or type can remain hidden within otherwise fluent reports [11]. Neither paradigm therefore combines unified multi-task reporting with direct verification of quantitative, categorical, and descriptive outputs. Recent advances in general-purpose VLMs [12, 13] of- fer a promising foundation for solving medical reporting tasks. Direct prompting generalist VLMs, however, does not reliably provide the specialist knowledge needed to inter- pret subtle endoscopic findings, and hallucinations remain a concern [14]. On the other hand, conventional special- ization adapts model weights to domain data, which may alter pretrained reasoning capacity and response diversity [15]. The unresolved challenge is therefore to introduce reliable specialist knowledge into a general-purpose VLM without modifying its pretrained weights, while preserving its pretrained capabilities and unified generative interface. Yang et al.: Preprint submitted to ElsevierPage 1 of 16 arXiv:2608.15580v1 [cs.AI] 16 Aug 2026 Context fusion for endoscopic polyp reporting To address this challenge, we introduce a context-fusion framework that equips a frozen general-purpose VLM with specialist context rather than adapting its pretrained weights. The framework combines two forms of specialist context. Explicit transduction context consists of retrieved imageโ report pairs that supply query-specific clinical evidence. Implicit instruction context is represented by learned con- tinuous specialist tokens that provide persistent reporting guidance across cases. At inference, the query image and both forms of specialist context are jointly presented to the frozen VLM under a structured instruction, which prompts it to estimate diameter, classify Paris type, and generate a morphological description in a single JSON report. The framework thereby supports multiple specialized reporting tasks through one generative interface. Because specializa- tion is confined to the context, it requires no report-specific changes to the VLM architecture and leaves all pretrained VLM weights unchanged. Figure 1 summarizes the report- ing task, the limitations of the two existing paradigms, and the proposed context-based alternative. We instantiate this hypothesis using latest open-source VLM [16] and 2,056 public endoscopic images with in- house expert annotations. We compare the proposed frame- work with directly prompted general-purpose VLMs, weight- adaptation methods and modular designs with task-specific predictors. Across the validation and test sets, the framework achieves the strongest overall performance among the evalu- ated methods while adding only 0.5 M trainable parameters. Ablation studies further show that the two contexts provide complementary gains and that retrieval benefits depend on evidence relevance rather than input length. Our contributions are threefold: โข We formulate multi-task endoscopic polyp reporting as unified generation of continuous diameter, cat- egorical Paris type, and open-ended morphological description within one structured schema. We further construct and release a 2,056-image benchmark with expert annotations. โข We introduce a context-fusion framework that com- bines explicit transduction context with implicit in- struction context. The framework confines specialist adaptation to the input context and leaves all pre- trained VLM weights unchanged. โข We systematically compare different strategies adapt- ing general-purpose VLMs. The proposed context- fusion framework achieves the best overall perfor- mance while maintaining high efficiency. 2. Related work 2.1. Automated endoscopic and polyp reporting Prior work on automated endoscopic reporting broadly follows two design patterns: assembling reports from task- specific predictions or generating report text directly from images. Qu et al. [8] combined multiple recognition models with rule-based report assembly and directly evaluated le- sion classification and size error, but the generated reports were confined to predefined recognizers and report fields. In contrast, Fonollร et al. [9] used an image encoder and a BERT language module to generate BASIC-based polyp descriptions. As automated report generation has advanced, rigorous assessment of report quality has emerged as a distinct methodological challenge. Metrics such as BLEU, ROUGE-L, and METEOR were used to measure resem- blance to reference text, but did not separately expose errors in quantitative measurements or categorical findings. SAM- ColonPolypGen [10] supplemented language metrics with two factual checks: a generated size was correct when it differed from the reference by no more than 0.1 cm, and morphology required an exact label match. Because these values were embedded in free text, the metric definitions assumed that a single numeric value and morphology label could first be recovered from the report. Omissions, mal- formed values, or ambiguous mentions would still require an additional matching policy. Report-Angel [17] turned to expert subjective assessment of accuracy and completeness. Experts judged whether a report was clinically acceptable and rated sentence coherence, feature completeness, and feature accuracy, while quantitative lesion sizing or categor- ical morphology classification were not evaluated. Existing systems therefore either make selected findings directly ver- ifiable within a restricted schema or generate richer prose whose key facts require additional extraction. A shared pub- lic benchmark that jointly evaluates regression, standardized classification, and narrative description remains absent. 2.2. Medical specialization of VLMs Adapting general-purpose VLMs to the medical domain has generally relied on large-scale, domain-specific multi- modal data, particularly for continued pretraining and mul- timodal instruction tuning. Medical VLMs such as LLaVA- Med [18], BiomedGPT [13], and Hulu-Med [19] acquire broad medical capability from large, heterogeneous imageโ text corpora. When the target is a narrower reporting do- main, adaptation is usually restricted to part of the model. Flamingo-CXR [20] trains its visual and connective compo- nents while freezing the language model, whereas PeFoMed [21] learns a visual projection and applies LoRA without up- dating the vision or language backbones. Prompt tuning [22] and visual prompt tuning [23] reduce the trainable compo- nent further to continuous tokens. These parameter-efficient strategies lower optimization cost, but they do not remove the need for representative specialist supervision. The effect of limited data is evident in prior studies: Flamingo-CXR [20] observed overfitting when its language component was updated, and ConceptVLM [15] found that direct specialist fine-tuning reduced the general analytical capability of the underlying VLM. This constraint is particularly relevant to endoscopy, where expert imageโreport pairs are less abun- dant than in radiology. Report-Angel [17], one of the few endoscopy-specific reporting systems, was trained on 20,617 imageโtext pairs and combined a fine-tuned MLLM with Yang et al.: Preprint submitted to ElsevierPage 2 of 16 Context fusion for endoscopic polyp reporting Figure 1: Workflow and motivation for unified endoscopic polyp report generation. (A) Conventional image-to-report generation embeds quantitative, categorical, and descriptive findings within a natural clinical sentence. We reorganize the same clinical content as a structured multi-task JSON report: diameter and Paris type can be evaluated directly, while the complete output is converted into a fixed clinical sentence for report-generation evaluation. (B) Direct general-VLM prompting lacks specialist case context, whereas separate task-specific models fragment reporting across independent predictors. (C) The proposed framework combines explicit transduction context and implicit instruction context around a frozen VLM, enabling all three reporting tasks through one unified interface. conventional models; meanwhile, GI-Bench [24] identifies substantial gaps in visual grounding and factual correctness among general-purpose models. Reliable specialist adapta- tion from smaller endoscopic reporting datasets therefore remains insufficiently studied. 2.3. Context engineering for large models Context engineering [25] organizes instructions, demon- strations, retrieved evidence, and learned context around a pretrained model. Existing methods instantiate this princi- ple at different levels, including instruction optimization, demonstration selection, and external knowledge retrieval. At the instruction level, Automatic Prompt Engineer [26] generates and selects candidate instructions, while OPRO [27] iteratively improves instructions using the scores of earlier candidates. Their final prompts provide instruction guidance shared across queries rather than query-specific evidence. At the demonstration level, LLM-R [28] trains a dense retriever from LLM feedback to select inputโ output examples, whereas visual in-context learning [29] retrieves, summarizes, and composes visual demonstra- tions. Beyond instructions and demonstrations, retrieval can supply a model with external knowledge relevant to the current query. RAG [30] retrieves external text passages, whereas GraphRAG [31] organizes document collections Yang et al.: Preprint submitted to ElsevierPage 3 of 16 Context fusion for endoscopic polyp reporting Figure 2: Context-fusion framework for a frozen VLM. (A) Offline, a fixed self-supervised polyp encoder maps the training images to L2-normalized CLS representations ํ 1 , ...,ํ ํ , which are stored with their reports in the case bank; online, the same encoder retrieves the top-ํพ imageโreport pairs for a query. (B) The retrieved pairs provide explicit transduction context, while learned continuous specialist tokens provide implicit instruction context within the multimodal prompt. (C) The composed context is supplied to the frozen VLM to generate one structured report. into entity graphs and community summaries. BiomedKAI [32] extends text retrieval to biomedical knowledge graphs with query-aware retrieval and specialized agents; these sys- tems retrieve textual knowledge rather than paired clinical images and reports. Medical multimodal RAG has moved closer to case-based evidence. CXR-RePaiR [33] retrieves reference reports for chest radiographs, MMed-RAG [34] uses domain-aware retrieval and adaptive context selection on radiology, ophthalmology, and pathology datasets for medical VQA and report generation, and FactMM-RAG [35] trains a fact-aware retriever for radiology report gen- eration. In endoscopy, EndoFinder [36] performs image- to-image retrieval of visually similar colorectal polyps to support explainable diagnosis. Taken together, prior work has explored shared instruction guidance, query-specific demonstrations, and multimodal retrieval largely as separate forms of context. Their integration within a frozen VLM remains underexplored for endoscopic reporting. 3. Method 3.1. Problem formulation Let ๎ฐ = (ํฅ ํ ,ํฆ ํ ) ํ ํ=1 denote the reporting dataset, where ํฅ ํ is an endoscopic image andํฆ ํ = (ํ ํ ,ํ ํ ,ํ ํ ) is its structured report. The report contains three fields corresponding to three tasks: diameter estimation ํ ํ , Paris-type classification ํ ํ , and free-text morphology description ํ ํ . Given a query image ํฅ ํ , the objective is to generate ฬํฆ ํ = ( ฬ ํ ํ , ฬํ ํ , ฬํ ํ ) in one autoregressive response. Let ํบ ํ denote a pretrained VLM with parameters ํ. Conventional adaptation uses ๎ฐ to update ํ for the target domain. In contrast, we keep ํ fixed and formulate specialist reporting as context-conditioned generation: ฬํฆ ํ = ํบ ํ (ํฅ ํ ,ํถ ํ ),(1) where ํถ ํ denotes the additional context supplied with the query. Adaptation therefore changes the information avail- able to ํบ ํ rather than its pretrained parameters. 3.2. Framework overview Our framework specializes frozen VLMs through two complementary forms of context (Fig. 2). The first is query- specific: explicit transduction context retrieves prior imageโ report pairs that demonstrate how visually related lesions were described and assessed. The second is shared across the dataset: implicit instruction context uses learned con- tinuous specialist tokens to condition the frozen VLM to- ward the reporting problem. These sources are combined with the query image, textual reporting rules, and output schema in one multimodal input. The frozen VLM receives Yang et al.: Preprint submitted to ElsevierPage 4 of 16 Context fusion for endoscopic polyp reporting this composed context and generates all three report fields in a single response. The framework therefore separates specialist adaptation from the VLM weights: case-specific information is supplied through retrieval, whereas shared reporting information is represented by a compact trainable context. 3.3. Explicit transduction context We first re-index the ํ training imageโreport pairs as (ํฅ ํ ,ํฆ ํ ) ํ ํ=1 . For each pair, a fixed image encoder ํธ(โ ) produces an L2-normalized CLS representation ํ ํ = ํธ(ํฅ ํ ). EndoFinder [36], a self-supervised polyp encoder trained on polyp images is utilized in this study. We compute these representations once and construct the case bank ๎ฎ = (ํฅ ํ ,ํฆ ํ ,ํ ํ ) ํ ํ=1 . For a query image ํฅ ํ , we obtain ํ ํ = ํธ(ํฅ ํ ) and calculate its similarity to the ํth bank image as ํ ํ = ํ โค ํ ํ ํ ,(2) which is cosine similarity because both representations are normalized. We select ํพ imageโreport pairs and retain their scores as the query-specific retrieval set ๎พ ํ = [ (ํฅ ํ ํ ,ํฆ ํ ํ ,ํ ํ ํ ) ] ํพ ํ=1 .(3) During training, ํพ cases are randomly sampled from the ten highest-scoring candidates. This prevents the same evidence set from being used for a query in every epoch. At inference, ๎พ ํ contains the ํพ highest-scoring cases in descending similarity order. For each retrieved pair, letํฏ ํ ํ denote the visual-token sequence of image ํฅ ํ ํ and letํญ ํ ํ denote the text-token sequence obtained from reportํฆ ํ ํ . The pair and its similarity score form one multimodal evidence block: ํต ํ = [ ํฏ ํ ํ ,ํญ ํ ํ ,ํ ํ ํ ] , ๎ฑ ํ = [ํต 1 , ...,ํต ํพ ].(4) Each block retains the association between the retrieved image, its report, and its similarity score. In the implemented prompt, the retrieved cases are numbered by rank so that each image is explicitly paired with its report and score. A retrieved image supplies a visual analogue of the query, while its paired report provides the corresponding diame- ter, Paris type, and clinical description. The pair therefore demonstrates an observed mapping from endoscopic appear- ance to specialist report rather than supplying isolated labels or ungrounded text. We refer to these inspectable, query- dependent cases as explicit transduction context. 3.4. Implicit instruction context We employ prompt tuning technique to parameterize the implicit instruction context. It is a sequence of continuous specialist tokensํ = [ํ 1 , ...,ํ ํ ] โโ ํรโ , whereํ is the prompt length and โ is the hidden dimension of the frozen VLM. The sameํ is shared across all cases and is optimized from the report-generation objective whileํบ ํ remains fixed. Because these tokens are learned from all training reports, they provide shared conditioning for producing the required numerical, categorical, and descriptive fields. Unlike a tex- tual instruction, these continuous vectors are learned directly in the VLM embedding space; unlike retrieved cases, they contain no identifiable image, report, or query-specific fact. The specialist tokens are inserted immediately after the query-image tokens. Under causal attention, their hidden states can depend on the current image before the later instruction and retrieval tokens are processed. Thus, al- though their learned embeddings are shared across cases, their contextualized representations are conditioned on the query. The parameterization follows prompt tuning, but its role in the framework is specifically to provide case-shared implicit context that complements the case-specific evidence supplied by retrieval. It is trained under the same multimodal evidence composition used for report generation rather than as an isolated textual prompt. 3.5. Context composition and fusion Letํฏ ํ denote the visual-token sequence of query image ํฅ ํ , and let ํ rule and ํ schema denote the token sequences of the textual reporting rules and output schema, respectively. The context for each query is composed as ํถ ํ = [ํฏ ํ ,ํ,ํ rule ,๎ฑ ํ ,ํ schema ].(5) The query-image tokensํฏ ํ appear first because they repre- sent the case to be reported. The specialist tokens ํ follow immediately, allowing their hidden states to be conditioned on the query image. The textual-rule tokens ํ rule then define the three reporting tasks before the retrieved evidence ๎ฑ ํ is introduced. Each evidence block ํต ํ โ ๎ฑ ํ contributes one retrieved image, its paired report, and its similarity score. Finally, the output-schema tokens ํ schema specify the required keys and close the user prompt before generation begins. In implementation, the multi-image processor seri- alizes all image placeholders before the textual prompt body; rank labels preserve the one-to-one correspondence between every retrieved image and its report. This ordering creates a single context in which each source has a distinct role. The query image supplies the primary observation;ํ provides shared specialist condition- ing; the retrieved pairs provide query-specific precedents; and the textual rules constrain the reporting task and re- sponse structure. We refer to their joint use within the same multimodal input as context fusion. During autoregressive generation, the frozen VLM conditions each report token on this composed context and produces one structured multi- task report. 3.6. Training objective The ground-truth structured report is appended to ํถ ํ as the assistant response and tokenized as ํฆ 1โถํ . Prompt positions are masked, and the autoregressive loss is applied only to report tokens: ๎ธ = โ ํ โ ํก=1 logํ ํ,ํ ( ํฆ ํก โฃ ํฆ <ํก ,ํถ ํ ) .(6) Yang et al.: Preprint submitted to ElsevierPage 5 of 16 Context fusion for endoscopic polyp reporting Figure 3: Dataset characteristics of the structured polyp-reporting subset. Each image is paired with diameter, Paris type, and a morphology description. These fields are evaluated with numerical, categorical, and language-generation metrics, respectively. The distribution panel summarizes Paris-type and diameter-bin counts across the training, validation, and test splits. The VLM parameters ํ, EndoFinder encoder, and case bank remain fixed; only ํ is optimized. Consequently, gradi- ents from the reporting loss update the implicit instruction context without altering the pretrained visual or language representations. At inference, the same sequence structure is retained, the learned ํ is reused for every query, and the ํพ highest-scoring cases provide explicit transduction con- text. Prompt templates, optimization settings, and decoding details are provided in the supplementary material. 4. Experiments The experiments compare directly prompted general- purpose VLMs, task-specific predictors, and weight-adaptation methods under the same multi-task reporting setting. We then examine the individual and combined contributions of the two context sources, how retrieval relevance and depth affect performance, whether requesting the three tasks jointly changes their individual performance, and how the two contexts correct errors or handle conflicting evidence in representative cases. 4.1. Dataset The study subset is derived from the public EndoVL release [37], which aggregates 2,709 gastrointestinal endo- scopic images from nine public datasets. We retained polyp- like lesions with a determinate diameter, a Paris-type label, and a non-empty morphology description. The resulting 2,056 imageโreport pairs were divided into 1,439 training, 206 validation, and 411 test cases. The split was stratified by the joint label of Paris type and four diameter intervals (< 0.5, 0.5โ< 1.0, 1.0โ< 2.0, and โฅ 2.0 cm), preserving comparable categorical and size distributions across the three subsets (Fig. 3). Each imageโreport pair was assigned to exactly one subset. Model optimization and case-bank construction used only the training subset; validation and test cases were used solely as held-out queries, and their reports were never available as retrieval evidence. This separation prevents image- and report-level leakage across the splits. Source-level and label-level split statistics are provided in Supplementary Section 2. Each report contains diameter, Paris type, and mor- phological description and is serialized as one JSON re- sponse. Figure 3 summarizes the reporting task and label distributions. Diameter is evaluated by mean absolute error (MAE) and threshold agreement at 0.5 and 1.0 cm. These boundaries correspond to commonly used clinical size cat- egories [38, 3]. ACC>0.5 records whether the prediction and reference fall on the same side of 0.5 cm; ACC>1.0 is defined analogously. Paris type is evaluated by exact-match accuracy and macro precision, recall, and F1, with multi- label combinations treated as complete category values. For whole-report evaluation, recoverable formatting vari- ants in the predicted JSON objects are first normalized. Each prediction and its reference are then converted into clinical sentences using the same fixed verbalization rule. The resulting sentences contain all three report components: lesion diameter, Paris type, and morphology description. We compute BLEU-4 [39], ROUGE-L [40], and CIDEr [41] on these complete sentences, complementing the task-specific metrics with sentence-level agreement for the entire report. CIDEr is divided by 10 for display on a 0โ1-oriented scale. The normalized JSON objects and verbalized sentences are retained as per-case records. For the general-VLM and Yang et al.: Preprint submitted to ElsevierPage 6 of 16 Context fusion for endoscopic polyp reporting Table 1 Comparison of general-purpose VLMs and the proposed method on the validation and test sets, with bootstrap 95% confidence intervals. Bold and underlined values denote the best and second-best distinct point estimates within each split and metric, respectively. Section MethodDiameterParis typeReport generation MAEโACC>0.5โACC>1.0โACCโPrecisionโRecallโF1โBLEU-4โROUGE-LโCIDErโ Val GPT-5.5 0.354 (0.304โ0.404) 0.684 (0.621โ0.748) 0.665 (0.602โ0.728) 0.296 (0.238โ0.359) 0.448 (0.219โ0.505) 0.265 (0.203โ0.333) 0.237 (0.163โ0.313) 0.386 (0.368โ0.404) 0.651 (0.638โ0.665) 0.063 (0.050โ0.077) Claude Sonnet 5 0.356 (0.304โ0.408) 0.767 (0.704โ0.821) 0.728 (0.670โ0.786) 0.388 (0.320โ0.451) 0.358 (0.181โ0.392) 0.177 (0.142โ0.226) 0.172 (0.113โ0.231) 0.420 (0.401โ0.439) 0.686 (0.673โ0.700) 0.080 (0.064โ0.097) Gemini 3.5 Flash 0.314 (0.269โ0.364) 0.816 (0.757โ0.864) 0.728 (0.670โ0.787) 0.403 (0.335โ0.466) 0.280 (0.238โ0.320) 0.235 (0.192โ0.275) 0.227 (0.183โ0.267) 0.443 (0.422โ0.464) 0.689 (0.674โ0.704) 0.086 (0.073โ0.102) Qwen3.6 Plus 0.374 (0.335โ0.417) 0.748 (0.689โ0.806) 0.670 (0.602โ0.733) 0.325 (0.262โ0.388) 0.198 (0.179โ0.221) 0.138 (0.112โ0.169) 0.121 (0.095โ0.149) 0.389 (0.374โ0.406) 0.660 (0.649โ0.671) 0.060 (0.051โ0.069) Frozen VLM 0.569 (0.509โ0.628) 0.534 (0.471โ0.602) 0.408 (0.345โ0.476) 0.214 (0.160โ0.267) 0.240 (0.175โ0.301) 0.200 (0.135โ0.258) 0.149 (0.106โ0.187) 0.350 (0.328โ0.374) 0.619 (0.601โ0.638) 0.065 (0.051โ0.080) Ours 0.155 (0.120โ0.192) 0.898 (0.854โ0.937) 0.917 (0.879โ0.951) 0.772 (0.714โ0.820) 0.844 (0.642โ0.880) 0.756 (0.530โ0.826) 0.788 (0.560โ0.843) 0.765 (0.732โ0.795) 0.879 (0.860โ0.896) 0.528 (0.476โ0.577) Test GPT-5.5 0.373 (0.331โ0.417) 0.708 (0.664โ0.752) 0.679 (0.633โ0.723) 0.304 (0.260โ0.348) 0.214 (0.149โ0.299) 0.159 (0.136โ0.186) 0.140 (0.113โ0.171) 0.385 (0.374โ0.397) 0.652 (0.643โ0.661) 0.057 (0.051โ0.064) Claude Sonnet 5 0.391 (0.344โ0.441) 0.723 (0.684โ0.769) 0.657 (0.608โ0.703) 0.358 (0.311โ0.401) 0.224 (0.147โ0.302) 0.131 (0.116โ0.147) 0.118 (0.096โ0.139) 0.417 (0.401โ0.432) 0.677 (0.666โ0.688) 0.080 (0.070โ0.092) Gemini 3.5 Flash 0.309 (0.276โ0.345) 0.781 (0.740โ0.820) 0.742 (0.698โ0.783) 0.445 (0.399โ0.494) 0.380 (0.280โ0.476) 0.264 (0.232โ0.300) 0.262 (0.222โ0.304) 0.453 (0.438โ0.467) 0.689 (0.680โ0.699) 0.092 (0.082โ0.103) Qwen3.6 Plus 0.375 (0.344โ0.407) 0.781 (0.742โ0.820) 0.698 (0.657โ0.742) 0.324 (0.282โ0.367) 0.189 (0.158โ0.223) 0.165 (0.144โ0.188) 0.143 (0.122โ0.164) 0.386 (0.375โ0.398) 0.655 (0.647โ0.664) 0.059 (0.053โ0.066) Frozen VLM 0.534 (0.488โ0.582) 0.564 (0.516โ0.611) 0.479 (0.431โ0.528) 0.197 (0.161โ0.234) 0.215 (0.167โ0.256) 0.200 (0.163โ0.234) 0.135 (0.107โ0.162) 0.357 (0.341โ0.373) 0.625 (0.612โ0.638) 0.069 (0.058โ0.081) Ours 0.200 (0.168โ0.234) 0.898 (0.866โ0.927) 0.878 (0.847โ0.905) 0.698 (0.650โ0.740) 0.791 (0.626โ0.835) 0.716 (0.555โ0.769) 0.744 (0.577โ0.786) 0.730 (0.704โ0.753) 0.858 (0.843โ0.871) 0.491 (0.452โ0.523) Table 2 Comparison with retrieval-based and single-task predictors on validation and test sets. Lower MAE is better for diameter; higher ACC is better for Paris type. Bold and underlined values denote the best and second-best distinct point estimates, respectively. A dash indicates that a single-task predictor does not produce the other taskโs output. MethodOutput scope Val MAEโ Val ACCโ Test MAEโ Test ACCโ EndoFinder retrieval baselineTask-specific0.3210.6550.3110.638 Supervised ViT Paris classifierSingle-taskโ0.670โ0.657 Supervised ViT diameter regressor Single-task0.279โ0.306โ OursUnified report0.1550.7720.2000.698 main method comparisons, 95% confidence intervals are estimated from 1,000 case-level bootstrap resamples within each split. 4.2. Implementation details All local experiments used Qwen3.5-9B [16] with bfloat16 computation on one NVIDIA A100 80GB GPU. Images were supplied at their native resolution and processed by the modelโs variable-resolution visual tokenizer [42], which uses 16 ร 16 patches and a spatial merge factor of 2. The retrieval bank contains 1,024-dimensional L2-normalized EndoFinder CLS representations from the training images. Validation and test images were used only as retrieval queries. The Prompt tuning baseline inserts 128 trainable contin- uous tokens into the input prompt without retrieved cases, corresponding to 0.524 M trainable parameters. The contin- uous tokens were optimized for 20 epochs with AdamW, a peak learning rate of 10 โ4 , weight decay of 0.1, an effective batch size of 16, OneCycle scheduling, and gradient clipping at 1.0. Prompt positions were masked from the autoregres- sive loss. Unless otherwise stated, inference uses the five highest-ranked training cases and greedy decoding with at most 128 new tokens. The directly prompted models were GPT-5.5[43] , Claude Sonnet 5 [44], Gemini 3.5 Flash [45], and Qwen3.6 Plus [46]. Each model received one query image, identical field definitions, Paris categories, and output schema, without retrieved cases or specialist examples. All adaptation methods use the same Qwen3.5-9B back- bone and evaluation pipeline. Finetune updates the VLMโs 40.119 M-parameter visualโlanguage merger while keep- ing the remaining backbone frozen. LoRA inserts rank-8 attention adapters and optimizes 9.056 M parameters. Direct inference and M-RAG require no optimization, whereas Prompt tuning and Ours optimize only the continuous spe- cialist tokens. The exact trainable locations and optimization settings for the weight-adaptation baselines are provided in Supplementary Section 3 and Supplementary Fig. S1. Task- specific comparisons use the EndoFinder retrieval estimate, a supervised ViT Paris classifier, and a supervised ViT diameter regressor. Yang et al.: Preprint submitted to ElsevierPage 7 of 16 Context fusion for endoscopic polyp reporting Table 3 Main comparison on validation and test sets with bootstrap 95% confidence intervals. Bold and underlined values denote the best and second-best distinct point estimates within each split and metric, respectively. Section MethodDiameterParis typeReport generation MAEโACC>0.5โACC>1.0โACCโPrecisionโRecallโF1โBLEU-4โROUGE-LโCIDErโ Val Frozen VLM 0.569 (0.509โ0.628) 0.534 (0.471โ0.602) 0.408 (0.345โ0.476) 0.214 (0.160โ0.267) 0.240 (0.175โ0.301) 0.200 (0.135โ0.258) 0.149 (0.106โ0.187) 0.350 (0.328โ0.374) 0.619 (0.601โ0.638) 0.065 (0.051โ0.080) Finetune 0.208 (0.169โ0.250) 0.908 (0.869โ0.942) 0.908 (0.864โ0.947) 0.718 (0.655โ0.777) 0.751 (0.534โ0.818) 0.753 (0.529โ0.817) 0.740 (0.514โ0.797) 0.720 (0.687โ0.752) 0.856 (0.839โ0.875) 0.453 (0.409โ0.501) LoRA 0.195 (0.162โ0.234) 0.879 (0.830โ0.922) 0.893 (0.850โ0.932) 0.699 (0.631โ0.762) 0.758 (0.535โ0.829) 0.725 (0.499โ0.792) 0.727 (0.503โ0.787) 0.689 (0.658โ0.721) 0.837 (0.819โ0.855) 0.404 (0.361โ0.448) M-RAG 0.216 (0.172โ0.266) 0.859 (0.811โ0.903) 0.888 (0.845โ0.927) 0.641 (0.573โ0.704) 0.689 (0.570โ0.747) 0.678 (0.537โ0.752) 0.662 (0.525โ0.718) 0.694 (0.659โ0.729) 0.840 (0.821โ0.860) 0.434 (0.383โ0.484) Prompt tuning 0.244 (0.203โ0.294) 0.864 (0.820โ0.908) 0.869 (0.820โ0.913) 0.641 (0.568โ0.704) 0.505 (0.396โ0.609) 0.463 (0.391โ0.541) 0.476 (0.391โ0.551) 0.668 (0.633โ0.700) 0.826 (0.807โ0.844) 0.387 (0.342โ0.432) Ours 0.155 (0.120โ0.192) 0.898 (0.854โ0.937) 0.917 (0.879โ0.951) 0.772 (0.714โ0.820) 0.844 (0.642โ0.880) 0.756 (0.530โ0.826) 0.788 (0.560โ0.843) 0.765 (0.732โ0.795) 0.879 (0.860โ0.896) 0.528 (0.476โ0.577) Test Frozen VLM 0.534 (0.488โ0.582) 0.564 (0.516โ0.611) 0.479 (0.431โ0.528) 0.197 (0.161โ0.234) 0.215 (0.167โ0.256) 0.200 (0.163โ0.234) 0.135 (0.107โ0.162) 0.357 (0.341โ0.373) 0.625 (0.612โ0.638) 0.069 (0.058โ0.081) Finetune 0.203 (0.174โ0.236) 0.903 (0.873โ0.930) 0.883 (0.852โ0.912) 0.664 (0.618โ0.708) 0.717 (0.531โ0.767) 0.629 (0.495โ0.723) 0.659 (0.503โ0.725) 0.697 (0.673โ0.720) 0.843 (0.830โ0.856) 0.434 (0.400โ0.468) LoRA 0.226 (0.195โ0.261) 0.883 (0.852โ0.912) 0.883 (0.852โ0.912) 0.650 (0.603โ0.691) 0.676 (0.552โ0.760) 0.678 (0.513โ0.731) 0.663 (0.517โ0.724) 0.673 (0.650โ0.697) 0.826 (0.813โ0.840) 0.396 (0.365โ0.431) M-RAG 0.235 (0.198โ0.277) 0.878 (0.847โ0.908) 0.869 (0.835โ0.900) 0.635 (0.589โ0.679) 0.700 (0.638โ0.745) 0.682 (0.611โ0.734) 0.663 (0.591โ0.707) 0.683 (0.659โ0.707) 0.830 (0.816โ0.844) 0.423 (0.388โ0.459) Prompt tuning 0.281 (0.241โ0.323) 0.869 (0.837โ0.900) 0.844 (0.808โ0.876) 0.625 (0.579โ0.669) 0.524 (0.467โ0.575) 0.511 (0.454โ0.563) 0.513 (0.458โ0.557) 0.661 (0.637โ0.683) 0.822 (0.809โ0.835) 0.379 (0.347โ0.409) Ours 0.200 (0.168โ0.234) 0.898 (0.866โ0.927) 0.878 (0.847โ0.905) 0.698 (0.650โ0.740) 0.791 (0.626โ0.835) 0.716 (0.555โ0.769) 0.744 (0.577โ0.786) 0.730 (0.704โ0.753) 0.858 (0.843โ0.871) 0.491 (0.452โ0.523) Table 4 Comparison of context configurations with five multimodal context cases. Random M-RAG denotes M-RAG supplied with randomly sampled training imageโreport pairs instead of EndoFinder-retrieved cases; the number and modalities of the context cases remain unchanged. Bold and underlined values denote the best and second-best distinct point estimates within each split and metric, respectively. Section MethodDiameterParis typeReport generation MAEโ ACC>0.5โ ACC>1.0โ ACCโ Precisionโ Recallโ F1โ BLEU-4โ ROUGE-Lโ CIDErโ ValMM-RAG with random cases0.365 0.8060.786 0.369 0.415 0.272 0.216 0.5020.7300.143 Prompt tuning + random M-RAG 0.320 0.8400.811 0.597 0.576 0.384 0.411 0.6080.7860.272 M-RAG0.2160.8590.8880.6410.6890.6780.6620.6940.8400.434 Prompt tuning0.244 0.8640.869 0.6410.505 0.463 0.476 0.6680.8260.387 Ours0.155 0.8980.917 0.772 0.844 0.756 0.788 0.7650.879 0.528 Test M-RAG with random cases0.400 0.8250.771 0.377 0.351 0.255 0.204 0.5120.7350.176 Prompt tuning + random M-RAG 0.337 0.8390.810 0.523 0.356 0.316 0.321 0.5850.7750.262 M-RAG0.2350.8780.8690.6350.7000.6820.6630.6830.8300.423 Prompt tuning0.281 0.8690.844 0.625 0.524 0.511 0.513 0.6610.8220.379 Ours0.200 0.8980.878 0.698 0.791 0.716 0.744 0.7300.858 0.491 For the context-quality study, random cases are sampled without reference to the query, EndoFinder cases follow co- sine similarity in the fixed CLS space, and a label oracle pro- vides an analytical upper bound. The oracle score combines Paris-type agreement, diameter proximity, and morphology- term overlap with weights 0.40, 0.35, and 0.25, respectively. It is used only to measure the headroom available from improved evidence selection. 4.3. Benchmark of general-purpose VLMs General-purpose VLMs followed the reporting instruc- tion, but their diameter and Paris-type predictions remained unreliable (Table 1). Gemini 3.5 Flash was the strongest directly prompted API model on the test set, with a diameter MAE of 0.309 and Paris-type accuracy of 0.445, but its Paris macro-F1 was only 0.262. Direct frozen-VLM inference was weaker, with a diameter MAE of 0.534 and Paris-type accuracy of 0.197. The validation results showed the same overall ordering. The metric profile separates fluent reporting from accu- rate specialist interpretation. The strongest direct API result reached 0.689 ROUGE-L on the test set, although fewer than half of its Paris labels were exactly correct. These models of- ten reproduced the expected sentence structure and common morphology terms while making errors in quantitative or categorical fields. A single language-generation score would therefore obscure clinically important failures. Yang et al.: Preprint submitted to ElsevierPage 8 of 16 Context fusion for endoscopic polyp reporting Figure 4: Effect of explicit transduction context and implicit instruction context. Direct denotes frozen-VLM inference without either context; the remaining settings use implicit instruction context, explicit transduction context, or both. The three rows evaluate diameter estimation, Paris-type classification, and report generation after converting structured JSON outputs into the fixed clinical sentence template. For display in panel A, MAE is transformed to 1โ(1 + MAE), so all bars use a higher-is-better orientation. The combined setting yielded the highest overall values across the three metric groups. The proposed context-fusion framework reduced test di- ameter MAE to 0.200, increased Paris accuracy and macro- F1 to 0.698 and 0.744, and raised ROUGE-L to 0.858 (Table 1). Together, these results show that context fusion improves both clinically verifiable fields and overall report quality. 4.4. Comparison with specialist adaptation baselines The unified framework also exceeded the task-specific predictors (Table 2). On the test set, its diameter MAE was 0.200, compared with 0.306 for the supervised ViT regressor, while its Paris accuracy was 0.698, compared with 0.657 for the supervised ViT classifier. Unlike these single- task models, the framework generates diameter, Paris type, and morphology in one response. Weight adaptation was competitive but did not dominate across the complete report (Table 3). Finetune achieved the highest test ACC>0.5 and tied with LoRA for the high- est ACC>1.0, indicating strong separation around the two diameter thresholds. Ours achieved a slightly lower con- tinuous MAE than Finetune (0.200 versus 0.203), higher Paris accuracy (0.698 versus 0.664), and higher ROUGE- L (0.858 versus 0.843). LoRA reached 0.226 MAE, 0.650 Paris accuracy, and 0.826 ROUGE-L. The context-based framework therefore provided the strongest overall balance across numerical, categorical, and generative outputs. This balance required substantially fewer trainable pa- rameters. As shown in Fig. 5 and Table 3, Ours optimized 0.524 M parameters, whereas LoRA and Finetune required 9.056 M and 40.119 M, respectively (17.3 and 76.6 times as many). We further examined Paris-type cases that Finetune clas- sified incorrectly. Among the 61 such cases for which the top-1 retrieved report contained the target Paris type, Ours corrected 43 (70.5%). In the corresponding stratum of 201 Yang et al.: Preprint submitted to ElsevierPage 9 of 16 Context fusion for endoscopic polyp reporting Table 5 Comparison of report-generation performance across context sizes and evidence sources using internal-context multimodal RAG. At each ํพ, random, EndoFinder, and oracle evidence are evaluated with the same number of retrieved cases. Bold and underlined values denote the best and second-best distinct point estimates within each split, ํพ, and metric, respectively. Section ํพ Evidence sourceDiameterParis typeReport generationEvidence coverage MAEโ ACC>0.5โ ACC>1.0โ ACCโ Precisionโ Recallโ F1โ BLEU-4โ ROUGE-Lโ CIDErโ Paris hit@ํพโ Size hit@ํพโ Val 1 Random cases 0.311 0.8160.8010.553 0.571 0.329 0.346 0.5860.7770.2330.2520.558 EndoFinder0.2060.8500.879 0.7230.8690.7240.7690.7170.8500.4410.6550.825 Oracle0.168 0.9080.879 0.816 0.934 0.798 0.846 0.7820.881 0.5511.0000.990 3 Random cases 0.318 0.8200.825 0.597 0.580 0.372 0.398 0.6010.7840.2580.5780.859 EndoFinder0.163 0.8740.9080.7380.8170.7280.7580.7510.8700.5130.8450.932 Oracle0.117 0.9370.927 0.903 0.954 0.922 0.935 0.8680.933 0.7011.0001.000 5 Random cases 0.320 0.8400.811 0.597 0.576 0.384 0.411 0.6080.7860.2720.7570.951 EndoFinder0.155 0.8980.9170.7720.8440.7560.7880.7650.8790.5280.9080.961 Oracle0.114 0.9320.937 0.903 0.956 0.898 0.923 0.8720.936 0.7091.0001.000 7 Random cases 0.306 0.8110.816 0.617 0.600 0.400 0.429 0.6180.7960.2860.8350.976 EndoFinder0.1640.8930.8930.7520.8440.7240.7590.7490.8710.5010.9560.966 Oracle0.110 0.9320.927 0.917 0.958 0.911 0.931 0.8690.935 0.7021.0001.000 Test 1 Random cases 0.344 0.8470.800 0.479 0.343 0.258 0.250 0.5660.7620.2290.2850.613 EndoFinder0.2310.9030.8710.6470.8120.6520.6990.6930.8370.4280.6370.839 Oracle0.182 0.9120.888 0.723 0.869 0.737 0.780 0.7380.858 0.5021.0000.985 3 Random cases 0.358 0.8490.791 0.504 0.347 0.294 0.298 0.5740.7690.2420.5960.878 EndoFinder0.2040.8950.8780.7030.8220.7210.7580.7320.8590.4890.8200.939 Oracle0.130 0.9420.920 0.859 0.928 0.882 0.901 0.8390.916 0.6691.0000.998 5 Random cases 0.337 0.8390.810 0.523 0.356 0.316 0.321 0.5850.7750.2620.7620.944 EndoFinder0.2000.8980.8780.6980.7910.7160.7440.7300.8580.4910.8910.964 Oracle0.122 0.9340.925 0.881 0.944 0.893 0.914 0.8480.922 0.6791.0000.998 7 Random cases 0.346 0.8220.800 0.538 0.393 0.335 0.342 0.5880.7780.2630.8440.964 EndoFinder0.1910.8830.8780.7030.8070.7230.7530.7310.8590.4930.9320.981 Oracle0.142 0.9320.915 0.856 0.924 0.882 0.900 0.8330.914 0.6571.0001.000 cases that Finetune classified correctly, Ours changed 14 predictions to errors (7.0%). 00.529.0640.12 Trainable parameters (M) 0.2 0.3 0.4 0.5 0.6 0.7 Paris type ACC Frozen VLM Finetune LoRA M-RAG Prompt tuning Ours Figure 5: Parameter efficiency on the test set. Paris-type accuracy is plotted against the number of trainable param- eters. Direct inference and M-RAG leave the backbone unchanged and use no trainable parameters; Prompt tuning and Ours optimize only 0.524 M specialist-token parame- ters. The proposed combination attains the highest Paris- type accuracy with substantially fewer trainable parameters than LoRA and visualโlanguage merger fine-tuning. 4.5. Ablation studies The two context sources provide complementary gains. Figure 4 summarizes the component ablation across the three reporting tasks, with detailed values provided in Ta- ble 3. In panel A, implicit instruction context and explicit transduction context both improved diameter estimation over Direct, reducing MAE from 0.534 to 0.281 and 0.235, re- spectively. Combining both contexts further reduced MAE to 0.200 while retaining high agreement at both clinical size thresholds. Panel B shows a similar but task-dependent pattern for Paris classification. Implicit instruction context increased macro-F1 from 0.135 to 0.513, whereas explicit transduction context reached 0.663. Their combination further increased precision, recall, and macro-F1 to 0.791, 0.716, and 0.744, respectively. The difference between the two individual con- ditions is consistent with implicit instruction context supply- ing reusable specialist guidance, whereas explicit transduc- tion context additionally provides query-specific examples. Panel C shows that both individual contexts also im- proved report generation. Explicit transduction context pro- duced the stronger single-context result, reaching 0.683 BLEU-4, 0.830 ROUGE-L, and 0.423 CIDEr. Combining both contexts increased these scores to 0.730, 0.858, and 0.491. Together, the two contexts retained the gains of each individual source and produced the best overall result across all three tasks. Retrieval relevance, depth, and quality jointly determine the value of external context. To isolate the effect of retrieval relevance, we replaced the five EndoFinder cases with five randomly sampled train- ing cases while preserving the number and modalities of the Yang et al.: Preprint submitted to ElsevierPage 10 of 16 Context fusion for endoscopic polyp reporting Table 6 Single-task and complete-report generation with the proposed method. Check marks indicate the fields requested in each output schema. A dash denotes a metric that is not applicable because the corresponding field was not generated. For ํ and ํทํํ , language metrics are computed from the same isolated morphology-description clause; diameter and Paris-type text is excluded from both predictions and references. Bold values indicate the better result between the single-task and complete-report settings for each split and metric. Split ํท ํ ํ DiameterParis typeMorphology description MAEโ ACC>0.5โ ACC>1.0โ ACCโ Precisionโ Recallโ F1โ BLEU-4โ ROUGE-Lโ CIDErโ Val โ0.199 0.8740.888โ โโ0.723 0.808 0.735 0.757โ โ โ0.4460.6770.380 โ 0.155 0.8980.917 0.772 0.844 0.756 0.788 0.6720.826 0.623 Test โ0.225 0.9030.888โ โโ0.696 0.770 0.724 0.740โ โ โ0.4250.6730.347 โ 0.200 0.8980.878 0.698 0.791 0.716 0.744 0.6290.798 0.576 additional inputs. On the test set, this intervention increased M-RAG diameter MAE from 0.235 to 0.400, reduced Paris accuracy from 0.635 to 0.377, and reduced CIDEr from 0.423 to 0.176 (Table 4). Adding Prompt tuning improved the random condition to 0.337 MAE, 0.523 Paris accuracy, and 0.262 CIDEr, but remained below the corresponding relevant-evidence setting. The same ordering was observed on the validation set. We then varied retrieval depth and evidence quality. Table 5 compares random, EndoFinder, and oracle evidence at matched values of ํพ. Paris hit@ํพ is the proportion of queries with at least one retrieved report containing the target Paris type. Size hit@ํพ analogously measures agree- ment with the targetโs binary diameter class at the 0.5 cm threshold. These retrieval metrics quantify whether compat- ible evidence is available; the downstream metrics measure whether the VLM uses it correctly. EndoFinder outperformed random retrieval at every matched depth. On the test set, its Paris hit@ํพ increased from 0.637 at ํพ = 1 to 0.932 at ํพ = 7, whereas Paris accuracy plateaued near 0.70 from ํพ = 3 onward. The largest categorical and language gains occurred between one and three EndoFinder cases: macro-F1 increased from 0.699 to 0.758 and CIDEr from 0.428 to 0.489. From ํพ = 3 to ํพ = 7, MAE decreased from 0.204 to 0.191, while macro- F1 and CIDEr changed only marginally. Thus, additional cases mainly improved evidence coverage after the first few examples had supplied most of the useful categorical and textual context. Coverage alone was insufficient. Random retrieval reached a Paris hit rate of 0.844 at ํพ = 7, but its Paris F1 and CIDEr were only 0.342 and 0.263. Oracle retrieval remained above EndoFinder at every depth; at ํพ = 3, it achieved a Paris macro-F1 of 0.901 and diameter MAE of 0.130, compared with 0.758 and 0.204 for EndoFinder. This gap identifies remaining headroom in evidence ranking and conflict resolution. The ํพ = 5 setting achieved the best overall validation performance across diameter estimation, Paris classification, and report generation and was therefore selected as the principal setting. Joint multi-task reporting improves morphology genera- tion. Table 6 compares the complete diameterโParisโreport (ํทํํ ) schema with the three single-task schemas ํท, ํ , and ํ . Definitions, output rules, and JSON keys for absent fields are removed, while the query image, frozen VLM, top- 5 retrieved cases, and decoding settings remain fixed. Metrics are computed only for requested fields. To com- pare morphology fairly, the language metrics for ํ and ํทํํ are calculated from the Description clause alone; diameter, Paris type, and their fixed template words are excluded from both predictions and references. The compar- ison therefore measures how the requested schema changes the shared morphology output rather than rewarding the longer ํทํํ sentence. Requesting ํท and ํ together with ํ improved all three morphology metrics on both splits. On the test set, BLEU- 4, ROUGE-L, and CIDEr increased from 0.425, 0.673, and 0.347 under ํ alone to 0.629, 0.798, and 0.576 under ํทํํ . Because diameter and Paris text is excluded from this scoring, the improvement cannot be attributed to the additional fixed clauses in the complete sentence. Within the same trained model, producing the quantitative and categor- ical fields therefore provides useful context for morphology generation. The model was trained with complete ํทํํ re- ports; the reduced schemas are inference-time interventions on that model rather than independently trained single-task baselines. 4.6. Case studies Figure 6 links the predictions to the query and retrieved images in three representative cases. The left case contains a pedunculated polyp. Direct inference predicted a compos- ite sessile type, whereas the retrieved set contained four Yang et al.: Preprint submitted to ElsevierPage 11 of 16 Context fusion for endoscopic polyp reporting Figure 6: Representative context-attribution cases. Each case column contains the query, five retrieved imageโreport pairs, and predictions from Direct, Prompt tuning, M-RAG, and Ours. Token-conditioned overlays visualize the attribution of the generated Diameter(cm), ParisType, and Description values to the query and evidence images; the six image maps within a case share one normalization scale. In the reports, blue denotes content that agrees with the reference, red denotes incorrect content, and orange denotes spurious content not supported by the reference report. Yang et al.: Preprint submitted to ElsevierPage 12 of 16 Context fusion for endoscopic polyp reporting 0-Ip cases and several images with visible stalks. M- RAG recovered 0-Ip but followed the retrieved thin-stalk descriptions and overestimated diameter. Prompt tuning re- covered the target thick-stalk attribute with a smaller size error. Ours combined the correct Paris evidence with the internal context, restoring the target diameter and conges- tion, although the thin-stalk attribute remained in the final report. The attribution is concentrated on the query image and Evidence #1, whose polyp shapes are visually similar. Evidence #1 is annotated with a thin stalk, and this at- tribute appears in both the M-RAG and combined outputs. The visualization therefore captures both the benefit of a compatible categorical example and the propagation of one conflicting morphology attribute. In the middle case, the target is a 0.6 cm sessile polyp. The first four retrieved reports describe 0-IIa lesions, whereas only Evidence #5 carries the target 0-Is label. Direct infer- ence predicted 0-Ip, and M-RAG recovered the sessile type but overestimated diameter to 1.0 cm. Prompt tuning recovered the target diameter and Paris type; Ours retained both corrections despite the conflicting retrieved Paris la- bels. Visual attribution appears mainly in Evidence #3โ 5, with little response in the first two retrieved images, showing that retrieval rank alone did not determine the visual evidence emphasized by the model. In Evidence #3 and #4, attribution is concentrated around the polyp margins; their reports contain smoothness and congestion, which are also present in the target morphology. Evidence #5 provides the target 0-Is label and broad-base morphology, and the latter is retained in the final report. These patterns identify the later retrieved cases as the more relevant visual and textual support for this prediction. The right case provides a complementary supported- retrieval pattern. All five retrieved reports describe sessile polyps, and their diameters and morphology closely match the target. Direct inference instead predicted a larger com- posite 0-Is+0-IIa lesion with a depressed area. M-RAG corrected all three fields, and Ours retained the correction. Attribution is strongest on Evidence #1, whose image and re- port closely match the target in size, Paris type, and morphol- ogy. These examples do not establish a causal explanation of the modelโs decision, but they show how compatible and conflicting evidence coexist with the observed corrections. Figure 7 compares the final reports produced by general- purpose VLMs, direct frozen-VLM inference, and the pro- posed framework. All structured outputs are verbalized by the same fixed rule, so differences arise from predicted content rather than a separate paraphrasing model. Direct models often preserve grammatical structure and common morphology terms while making errors in diameter, Paris type, or a specific descriptive attribute. The combined con- text setting corrects these fields while retaining coherent report language. The examples also show why ROUGE-L should be read together with diameter and Paris-type met- rics: a report may share most words with the reference and still contain a clinically important numerical or categorical error. 5. Discussion This study approaches specialist adaptation as a problem of context organization rather than VLM weight modifica- tion. Explicit transduction context supplies query-specific clinical precedents, while implicit instruction context pro- vides a compact condition learned across cases. Because the retrieved cases are supplied as explicit evidence rather than absorbed into model weights, their relevance and report content can be examined for each prediction. This offers a lightweight alternative to specialist systems that train dedi- cated visual or language components [47, 48, 8, 10]. The task-composition experiment revealed an interaction that is not captured by the main aggregate comparison. Jointly requesting diameter and Paris type improved mor- phology generation even though their text was excluded from report scoring (Table 6). The quantitative and categorical predictions may therefore act as intermediate clinical con- straints: estimating lesion scale and gross morphology first can narrow the description that follows. This interpretation is limited to inference-time interventions on a model trained with complete reports, but it suggests that multi-task report- ing can contribute more than output convenience. The hard-case analysis further clarifies when retrieved context is useful. When Finetune failed and the top-ranked retrieved report contained the target Paris type, Ours cor- rected 70.5% of those errors while changing 7.0% of the cor- responding Finetune-correct predictions to errors. This im- balance suggests that a relevant precedent can supply miss- ing case-level evidence for difficult categorical decisions without broadly destabilizing predictions that are already correct. The attribution examples also show the boundary of this benefit: compatible evidence can correct size or Paris type, whereas a repeated conflicting attribute can still enter the generated morphology (Fig. 6). Several limitations define the scope of the present re- sults. Diameter estimation from a monocular image remains uncertain without a calibrated instrument or another scale reference [6]. The structured subset contains 2,056 images, and rare Paris types are under-represented. All adaptation experiments use a single frozen VLM, so the findings do not yet establish cross-backbone generality. The current schema covers diameter, Paris type, and morphology but omits other report fields such as lesion location. Retrieval also relies on one fixed EndoFinder encoder. Future work should evaluate the framework on larger structured cohorts with calibrated size references, broader Paris-type coverage, and external clinical validation. Apply- ing the same context construction to other frozen VLMs would test whether the observed gains extend across back- bones. Field-aware and conflict-aware retrieval could se- lect evidence separately for size, Paris type, morphology, and additional report fields. The oracle results indicate that improved evidence selection could provide further gains without changing the reporting backbone. Yang et al.: Preprint submitted to ElsevierPage 13 of 16 Context fusion for endoscopic polyp reporting Figure 7: Representative case-study comparison on the test set. Each column shows one query case, and the rows compare target reports with direct outputs from general VLMs and the local frozen VLM, followed by the proposed context-fusion setting. Structured JSON predictions are converted into fixed clinical sentences for visual comparison. In model-generated reports, blue denotes content that agrees with the reference, red denotes incorrect content, and orange denotes spurious content not supported by the reference report. In these examples, fluent text can contain clinically important field errors, whereas the proposed setting corrects size or Paris-type content while retaining morphology descriptions. 6. Conclusion We presented a context-fusion framework that combines query-specific imageโreport pairs with learned specialist guidance to adapt a frozen VLM for endoscopic polyp re- porting. It jointly estimated lesion diameter, classified Paris type, and generated morphology descriptions, achieving the strongest overall balance among the evaluated methods with trainable parameters equal to only 0.006% of the frozen VLM. Within the evaluated backbone and dataset, these results support context fusion as a lightweight approach to specialist VLM adaptation. Yang et al.: Preprint submitted to ElsevierPage 14 of 16 Context fusion for endoscopic polyp reporting Data availability The annotations used in this study are available through the EndoVL dataset [37] on Figshare at https://doi.org/ 10.6084/m9.figshare.32320998.v1. This study uses the 2,056- image polyp-reporting subset defined in Section 4.1. References [1] Ann G. Zauber, Sidney J. Winawer, Michael J. OโBrien, Iris Lansdorp-Vogelaar, Marjolein van Ballegooijen, Benjamin F. Hankey, Weiji Shi, John H. Bond, Marc Schapiro, Jerome F. Panish, Edward T. Stewart, and Jerome D. Waye. Colonoscopic polypectomy and long- term prevention of colorectal-cancer deaths. New England Journal of Medicine, 366(8):687โ696, 2012. doi: 10.1056/NEJMoa1100370. [2] Susan G. Coe, Chakri Panjala, Michael G. Heckman, Mihir Patel, Bashar J. Qumseya, Yize R. Wang, Benjamin Dalton, Philip Tran, William Palmer, Nancy Diehl, Michael B. Wallace, and Massimo Raimondo. Quality in colonoscopy reporting: An assessment of compliance and performance improvement. Digestive and Liver Disease, 44(8):660โ664, 2012. doi: 10.1016/j.dld.2012.03.022. [3] Samir Gupta, David Lieberman, Joseph C. Anderson, et al. Rec- ommendations for follow-up after colonoscopy and polypectomy: A consensus update by the US Multi-Society Task Force on colorectal cancer. Gastroenterology, 158(4):1131โ1153.e5, 2020. [4] Michal F. Kaminski, Siwan Thomas-Gibson, Michal Bugajski, et al. Performance measures for lower gastrointestinal endoscopy: a eu- ropean society of gastrointestinal endoscopy quality improvement initiative. Endoscopy, 49:378โ397, 2017. [5] The Paris Endoscopic Classification Group. The paris endoscopic classification of superficial neoplastic lesions: esophagus, stomach, and colon. Gastrointestinal Endoscopy, 58(6 Suppl):S3โS43, 2003. doi: 10.1016/S0016-5107(03)02159-X. [6] Giulio Antonelli, Federico Desideri, Sara Schiavone, et al. Clinical implications of computer-aided real-time size estimation of colorectal polyps during colonoscopy: a prospective study. Endoscopy, 58(3): 290โ294, 2026. doi: 10.1055/a-2695-1978. [7] Davide Massimi, Luca Carlini, Yuichi Mori, et al. Large language model for interpreting the paris classification of colorectal polyps. Endoscopy International Open, 13:a27030209, 2025. doi: 10.1055/ a-2703-0209. [8] Jun-yan Qu, Zhen Li, Jing-ran Su, Ming-jun Ma, Chang-qin Xu, Ai- jun Zhang, Cheng-xia Liu, Hai-peng Yuan, Yan-liu Chu, Cui-cui Lang, Liu-ye Huang, Lin Lu, Yan-qing Li, and Xiu-li Zuo. Devel- opment and validation of an automatic image-recognition endoscopic report generation system: A multicenter study. Clinical and Trans- lational Gastroenterology, 12(1):e00282, 2021. doi: 10.14309/ctg. 0000000000000282. [9] Roger Fonolla, Quirine E. W. van der Zander, Ramon M. Schreuder, Sharmila Subramaniam, Pradeep Bhandari, Ad A. M. Masclee, Erik J. Schoon, Fons van der Sommen, and Peter H. N. de With. Automatic image and text-based description for colorectal polyps using BASIC classification. Artificial Intelligence in Medicine, 121:102178, 2021. doi: 10.1016/j.artmed.2021.102178. [10] Shasha Zhang, Yuang Cai, Yijun Chen, Xiang Cai, and Peng Li. SAM- ColonPolypGen: Enhancing automated colon polyp report generation via reinforcement learning and prompt chaining. Biomedical Signal Processing and Control, 120:110084, 2026. doi: 10.1016/j.bspc. 2026.110084. [11] Yasuhide Miura, Yuhao Zhang, Emily Tsai, Curtis Langlotz, and Dan Jurafsky. Improving factual completeness and consistency of image-to-text radiology report generation. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5288โ5304, 2021. doi: 10.18653/v1/2021.naacl-main.416. [12] Michael Moor, Oishi Banerjee, Zaid S. H. Abad, et al. Foundation models for generalist medical artificial intelligence. Nature, 616:259โ 265, 2023. [13] Kai Zhang et al. A generalist vision-language foundation model for diverse biomedical tasks. Nature Medicine, 2024. [14] Zishan Gu, Jiayuan Chen, Fenglin Liu, Changchang Yin, and Ping Zhang. MedVH: Toward systematic evaluation of hallucination for large vision language models in the medical context. Advanced Intel- ligent Systems, page 2500255, 2025. doi: 10.1002/aisy.202500255. [15] Wei Lou, Yue Wu, Pusheng Xu, Weiyi Zhang, Xiaolan Chen, Jiancheng Yang, Mingguang He, and Danli Shi. Key concept learning for medical vision language model with reasoning capabilities. npj Digital Medicine, 9:512, 2026. doi: 10.1038/s41746-026-02676-5. [16] Qwen Team. Qwen3.5: Towards native multimodal agents. https: //qwen.ai/blog?id=qwen3.5, February 2026. [17] Ruiqing Jiang, Boru Chen, Zehua Dong, Xiaoquan Zeng, Hang You, Yanxia Li, Yunchao Deng, Ganggang Mu, Jing Wang, Li Huang, Jia Li, Du Cheng, Wei Zhou, et al. Domain specific multimodal large language model for automated endoscopy reporting with multicenter prospective validation. npj Digital Medicine, 9:394, 2026. doi: 10.1038/s41746-026-02569-7. [18] Chunyuan Li, Cliff Wong, Sheng Zhang, et al. LLaVA-Med: Training a large language-and-vision assistant for biomedicine in one day. In Advances in Neural Information Processing Systems, 2023. [19] Songtao Jiang, Yuan Wang, Sibo Song, Tianxiang Hu, Chenyi Zhou, Bin Pu, Yan Zhang, Zhibo Yang, Yang Feng, Joey Tianyi Zhou, Jin Hao, Zijian Chen, Ruijia Wu, Tao Tang, Junhui Lv, Hongxia Xu, Hongwei Wang, Jun Xiao, Bin Feng, Fudong Zhu, Kenli Li, Weidi Xie, Jimeng Sun, Jian Wu, and Zuozhu Liu. Hulu-Med: A transparent generalist model towards holistic medical vision-language understanding, 2025. [20] Ryutaro Tanno, David G. T. Barrett, Andrew Sellergren, Sumedh Ghaisas, Sumanth Dathathri, Abigail See, Johannes Welbl, et al. Collaboration between clinicians and visionโlanguage models in ra- diology report generation. Nature Medicine, 31:599โ608, 2025. doi: 10.1038/s41591-024-03302-1. [21] Gang Liu, Xiaotian Tang, Jinlong He, Pengfei Li, Zhaolin Chen, and Shenjun Zhong. PeFoMed: Parameter efficient fine-tuning of multimodal large language models for medical CXR. Scientific Reports, 16:18729, 2026. doi: 10.1038/s41598-026-47871-2. [22] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691, 2021. [23] Menglin Jia, Luming Tang, Bor-Chun Chen, et al. Visual prompt tuning. In European Conference on Computer Vision, 2022. [24] Yan Zhu, Te Luo, Pei-Yao Fu, Zhen Zhang, Zi-Long Wang, Yi-Fan Qu, Zi-Han Geng, Jia-Qi Xu, Lu Yao, Li-Yun Ma, Wei Su, Wei- Feng Chen, Quan-Lin Li, Shuo Wang, and Ping-Hong Zhou. GI- Bench: A panoramic benchmark revealing the knowledge-experience dissociation of multimodal large language models in gastrointestinal endoscopy against clinical standards, 2026. URL https://arxiv.org/ abs/2601.08183. [25] Lingrui Mei, Jiayu Yao, Yuyao Ge, Yiwei Wang, Baolong Bi, Yujun Cai, Jiazhi Liu, Mingyu Li, Zhong-Zhi Li, Duzhen Zhang, Chenlin Zhou, Jiayi Mao, Tianze Xia, Jiafeng Guo, and Shenghua Liu. A survey of context engineering for large language models, 2025. [26] Yongchao Zhou, Andrei Ioan Muresanu, Ziwen Han, Keiran Paster, Silviu Pitis, Harris Chan, and Jimmy Ba. Large language models are human-level prompt engineers. In International Conference on Learning Representations, 2023. [27] Chengrun Yang, Xuezhi Wang, Yifeng Lu, Hanxiao Liu, Quoc V. Le, Denny Zhou, and Xinyun Chen. Large language models as optimizers. In International Conference on Learning Representations, 2024. [28] Liang Wang, Nan Yang, and Furu Wei. Learning to retrieve in- context examples for large language models. In Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics, pages 1752โ1767, 2024. doi: 10.18653/ v1/2024.eacl-long.105. [29] Yucheng Zhou, Xiang Li, Qianning Wang, and Jianbing Shen. Visual in-context learning for large vision-language models. In Findings of the Association for Computational Linguistics: ACL 2024, pages Yang et al.: Preprint submitted to ElsevierPage 15 of 16 Context fusion for endoscopic polyp reporting 15890โ15902, 2024. doi: 10.18653/v1/2024.findings-acl.940. [30] Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Kuttler, Mike Lewis, Wen-tau Yih, Tim Rocktรคschel, Sebastian Riedel, and Douwe Kiela. Retrieval-augmented generation for knowledge-intensive nlp tasks. In Advances in Neural Information Processing Systems, 2020. [31] Darren Edge, Ha Trinh, Newman Cheng, Joshua Bradley, Alex Chao, Apurva Mody, Steven Truitt, and Jonathan Larson. From local to global: A graph RAG approach to query-focused summarization, 2024. [32] Madhavi Kumari, Rohit Chauhan, Rashi Jain, and Prabha Garg. A novel context-aware retrieval framework for biomedical knowledge integration with large language models. Information Fusion, 127: 103902, 2026. doi: 10.1016/j.inffus.2025.103902. [33] Mark Endo, Rayan Krishnan, Viswesh Krishna, Andrew Y. Ng, and Pranav Rajpurkar. CXR-RePaiR: Retrieval-based chest x-ray report generation using a pre-trained contrastive language-image model. In Proceedings of the Machine Learning for Health NeurIPS Workshop, 2021. [34] Peng Xia et al. MMed-RAG: Versatile multimodal retrieval- augmented generation for medical vision-language models, 2025. [35] Liwen Sun, James Jialun Zhao, Wenjing Han, and Chenyan Xiong. Fact-aware multimodal retrieval augmentation for accurate medical radiology report generation. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Compu- tational Linguistics: Human Language Technologies, pages 643โ655, 2025. doi: 10.18653/v1/2025.naacl-long.28. [36] Ruijie Yang, Yan Zhu, Peiyao Fu, Yizhe Zhang, Zhihua Wang, Quan- lin Li, Pinghong Zhou, Xian Yang, and Shuo Wang. EndoFinder: Online image retrieval for explainable colorectal polyp diagnosis. In Medical Image Computing and Computer Assisted Intervention โ MICCAI 2024, volume 15010 of Lecture Notes in Computer Science, pages 251โ262. Springer Nature Switzerland, 2024. doi: 10.1007/ 978-3-031-72117-5_24. [37] Peiyao Fu, Siyuan Li, Yan Zhu, Shuo Wang, Xiushi Li, Xinyang Liu, Mengjiang He, Quanlin Li, and Ping-Hong Zhou. EndoVL: An expert annotated dataset of public gastrointestinal endoscopic images with uncertainty labels, 2026. URL https://doi.org/10.6084/m9.figshare. 32320998.v1. [38] Franklin C. Tsai and Williamson B. Strum. Prevalence of advanced adenomas in small and diminutive colon polyps using direct mea- surement of size. Digestive Diseases and Sciences, 56(8):2384โ2388, 2011. doi: 10.1007/s10620-011-1598-x. [39] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. BLEU: A method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, 2002. [40] Chin-Yew Lin. ROUGE: A package for automatic evaluation of summaries. In Text Summarization Branches Out, 2004. [41] Ramakrishna Vedantam, C. Lawrence Zitnick, and Devi Parikh. CIDEr: Consensus-based image description evaluation. In Pro- ceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2015. [42] Qwen Team. Qwen2-VL: Enhancing vision-language modelโs per- ception of the world at any resolution, 2024. [43] OpenAI. OpenAI model documentation. https://platform.openai. com/docs/models, 2026. [44] Anthropic. Claude model documentation. https://docs.anthropic. com/, 2026. [45] Google DeepMind. Gemini model documentation. https://ai. google.dev/gemini-api/docs/models, 2026. [46] Alibaba Cloud. Qwen model documentation. https://help.aliyun. com/zh/model-studio/, 2026. [47] Qikui Zhu, Yeyu Lin, Weitao Fu, Wenhui Tang, Jiaxing Li, Ying Zhang, Bo Li, Xiaolong Guo, Fang Wang, Haiyan Qi, Chongke Sun, Xiangkui Zhu, Zhonghao Liu, Li He, Zhiguo Zheng, Bo Du, Jie Yang, Zhuan Bian, and Liuyan Meng. Towards clinical-level interpretation of dental panoramic radiography using an instance-guided vision- language model. Nature Biomedical Engineering, 2026. doi: 10.1038/ s41551-026-01713-8. [48] Anabia Sohail, Iyyakutti Iyappan Ganapathi, Basit Alawode, Sajid Javed, Mohammed Bennamoun, and Arif Mahmood. ConVLM: Context-guided vision-language model for fine-grained histopathol- ogy image classification. Information Fusion, 127:103737, 2026. doi: 10.1016/j.inffus.2025.103737. Yang et al.: Preprint submitted to ElsevierPage 16 of 16