Paper deep dive
A doctrine-grounded visual question answering dataset for Tactical Combat Casualty Care
Junseob Kim, Jade Chng, Ayman Ali, Victor Moas, Yichun Lee, Po-Chun Chin, Sunil Hwang, Rishikesan Kamaleswaran
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 91%
Last extracted: 10/7/2026, 9:09:59 AM
Summary
The paper introduces TC3-VQA, a dataset for Tactical Combat Casualty Care (TC3) that links visual evidence from public instructional and field videos to authoritative clinical doctrine. The dataset contains 581 items and 1,860 questions covering intervention recognition, doctrine recall, clinical reasoning, procedural guidance, and refusal when visual information is insufficient. The construction pipeline involves visual annotation, doctrine retrieval from 11 sources, verbatim answer anchoring with character offsets, and entailment checks to ensure grounding. The dataset aims to support the development and evaluation of vision-language models for medical decision support in resource-constrained environments.
Entities (8)
Relation Signals (7)
TC3-VQA → coversdomain → Tactical Combat Casualty Care
confidence 98% · TC3-VQA, a dataset constructed from public instructional and field TC3 videos
TC3-VQA → containsitems → 581 items
confidence 95% · It contains 581 items spanning 11 concepts
TC3-VQA → containsquestions → 1,860 questions
confidence 95% · with 1,860 questions covering intervention recognition, doctrine, clinical reasoning
TC3-VQA → supports → Vision-Language Models
confidence 92% · The dataset provides a resource for adapting vision-language models to TC3
TC3-VQA → depositedat → Zenodo
confidence 90% · TC3-VQA version 1.1 is deposited at Zenodo
TC3-VQA → usessources → CoTCCC
confidence 90% · public-domain footage of the Committee on Tactical Combat Casualty Care (CoTCCC)
DeBERTa-v3 → usedfor → entailment_check
confidence 85% · using a DeBERTa-v3 natural language inference (NLI) model
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Tactical Combat Casualty Care (TC3) requires responders to connect visual observations of injuries and interventions with established clinical guidance. Developing vision-language models to support this process requires supervision that links visible evidence to traceable doctrine. We present TC3-VQA, a dataset constructed from public instructional and field TC3 videos and authoritative TC3 documents. It contains 581 items spanning 11 concepts, with 1,860 questions covering intervention recognition, doctrine, clinical reasoning, procedural guidance, and refusal when visual information is insufficient. Doctrine-based answers preserve verbatim source passages and character offsets. Construction combines visual annotation, passage retrieval, entailment checks, and verification across model families. Equipment boxes, anatomical labels, temporal segments, and source metadata accompany the question-answer pairs. Automated audits and ratings by two physicians and two medical students characterize annotation quality, with human ratings available for 88 retained items. The dataset provides a resource for adapting vision-language models to TC3, studying the connection between visual evidence and clinical knowledge, and evaluating recognition, doctrine recall, and abstention.
Tags
Links
- Source: https://arxiv.org/abs/2610.07339v1
- Canonical: https://arxiv.org/abs/2610.07339v1
Trouble viewing inline? Open PDF directly →
Full Text
61,816 characters extracted from source content.
Expand or collapse full text
A doctrine-grounded visual question answering dataset for Tactical Combat Casualty Care Junseob Kim 1,∗ , Jade Chng 2 , Ayman Ali 3 , Victor Moas 3 , Yichun Lee 4 , Po-Chun Chin 4 , Sunil Hwang 5 , Rishikesan Kamaleswaran 3,6 1 Department of Electrical and Computer Engineering, Duke University, Durham, NC, USA 2 Department of Biomedical Engineering, Duke University, Durham, NC, USA 3 Department of Surgery, Duke University, Durham, NC, USA 4 School of Medicine, College of Medicine, National Yang Ming Chiao Tung University, Taipei, Taiwan 5 Department of Mathematics, Korea Military Academy, Seoul, South Korea 6 Department of Anesthesiology, Duke University, Durham, NC, USA ∗ Corresponding author: Junseob Kim (junseob.kim@duke.edu) Abstract Tactical Combat Casualty Care (TC3) requires responders to connect visual observations of injuries and interventions with established clinical guidance. Developing vision-language models to support this process requires supervision that links visible evidence to traceable doctrine. We present TC3-VQA, a dataset constructed from public instructional and field TC3 videos and authoritative TC3 documents. It contains 581 items spanning 11 concepts, with 1,860 ques- tions covering intervention recognition, doctrine, clinical reasoning, procedural guidance, and refusal when visual information is insufficient. Doctrine-based answers preserve verbatim source passages and character offsets. Construction combines visual annotation, passage retrieval, en- tailment checks, and verification across model families. Equipment boxes, anatomical labels, temporal segments, and source metadata accompany the question-answer pairs. Automated audits and ratings by two physicians and two medical students characterize annotation quality, with human ratings available for 88 retained items. The dataset provides a resource for adapting vision-language models to TC3, studying the connection between visual evidence and clinical knowledge, and evaluating recognition, doctrine recall, and abstention. 1 Background & Summary Tactical Combat Casualty Care (TC3) provides guidance for treating battlefield injuries before definitive medical care is available [1, 2]. Among U.S. combat fatalities in Iraq and Afghanistan from 2001 to 2011, 87.3% occurred before arrival at a medical treatment facility [3]. When evacuation is delayed, care may extend well beyond the initial response, placing sustained demands on limited personnel and equipment [4]. Responders can vary in clinical experience, and combat lifesaving skills can decline after initial training [5, 6]. Computational approaches to prehospital trauma decision support have largely focused on phys- iological monitoring. Machine-learning systems use vital signs and physiological waveforms to assess shock, hemorrhage risk, and the need for life-saving interventions [5, 7, 8, 9]. The DARPA Triage Challenge further reflects broader interest in technology-assisted casualty assessment under resource constraints [10]. Physiological signals, however, do not directly show which equipment is present 1 arXiv:2610.07339v1 [cs.CV] 5 Oct 2026 RECOGNITION Q: Which TCCC intervention is being performed in this frame? ✅ Limb tourniquet application • ❌ Wound packing with hemostatic gauze • ❌ Nasopharyngeal airway insertion • ❌ Needle chest decompression Concept: tourniquet application (MARCH M) Q: After this limb tourniquet is applied, how do you verify it is effective and what do you do if a distal pulse remains? “If bleeding persists or a distal pulse is still present, consider additional tightening of the tourniquet or the use of a second tourniquet side-by-side with the first to eliminate both bleeding and the distal pulse.” [tccc_guidelines_2024_01_25, chars 1149-1368] REASONING Q: Once the windlass rod has been twisted until bleeding stops, what is the next step to secure the tourniquet? “Lock the windlass rod in place with the windlass clip.” [atp_4_02_11_casualty_response, chars 928-982] HOW Q: Given the windlass tourniquet being routed onto this casualty’s injured limb over the uniform, how should the responder seat the band and take up the slack before twisting the windlass? “...PositiontheTQabovethebleedingsite,highontheextremityovertheclothing/uniform.Ensurethat alltheslackintheself-adheringbandispulledthroughtheroutingbucklebeforethebandisfastened backonitselfandthewindlassistwisted.” [atp_4_02_11_casualty_response, chars 518-845] perception knowledge • effectiveness or verification next step action REFUSAL Q: How many full turns of the windlass were applied to tighten this tourniquet to the point that bleeding stopped? “Gold:REFUSE.Thewindlassisvisible,buthowmanyturnswereappliedandwhetherbleeding stoppedcannotbeseenintheframes” paired, same frames Anatomy: lower extremity (high) Equipment box: limb_tourniquet Answer region Independent audit: concept visible = yes DOCTRINE Figure 1: Example TC3-VQA item (ans01332) from public-domain footage of the Committee on Tactical Combat Casualty Care (CoTCCC). Left: one of four frames with equipment, answer- region, and anatomy annotations. Right: four answerable question types with cited doctrine spans. The paired refusal question asks for the number of windlass turns, which the available frames do not show. or how an intervention is being performed. Vision-language models (VLMs) offer a complementary approach by connecting visual observations of equipment and interventions to natural-language questions and guidance. TC3-VLM demonstrated the feasibility of adapting general VLMs to TC3 videos using visually grounded and caption-derived question-answer pairs [11]. Supporting knowledge-intensive questions in this setting requires training data that connect visible interven- tions to verifiable clinical knowledge. Medical visual question answering (VQA) provides a practical framework for learning these connections, with potential applications in clinical decision support [12].Datasets including VQA-RAD, PathVQA, SLAKE, and PMC-VQA cover radiology, pathology, and other medical imagery [12, 13, 14, 15]. More recent efforts have increased scale by generating instruction data from biomedical figure captions, narrated histopathology videos, and public educational videos [16, 17, 18]. Trauma THOMPSON extends this direction to egocentric simulated trauma pro- cedures annotated for actions, hands, tools, and medical VQA [19]. However, comparable resources for TC3 remain limited, particularly for questions that connect visible battlefield-care procedures with relevant clinical guidance. Constructing such a resource from public visual material introduces challenges in both visual and knowledge grounding. General and medical VLMs are susceptible to visual hallucinations [20, 21], so using the same model to interpret a scene and write its reference answer can propagate an initial error into the training target. Video narration can also be temporally misaligned with the corre- sponding image [22], while language priors can enable some questions to be answered without using the image [23]. Knowledge grounding likewise requires explicit verification: SourceCheckup found that cited references often did not substantiate every claim in generated medical responses [24]. 2 CONVENTIONAL OURS 1. See a frame A circular pipeline can reinforce errors. 1 PERCEIVE VIDEO →FRAMES 2. Write an answer 3. Hallucination DOCTRINE →CHUNKS RETRIEVE ANCHOR GATE DeBERTa-v3 234 • 1 of 12 concepts, 6-VLM vote (≥3/6) • 17-class equipment boxes • answer region, 8 bodyregions • query = concept + visual evidence • 9,320 chunks over 11 sources • exact search, repeatable • span kept as • citation, source, offsets, quote • rendering added, span untouched • renderingsplitintosentences • each entailed by concept corpus • else reverts to the quote a=c[i s :i e ] ClipsShot DetectionFrame FilterWindow Scoring 415 public videos PySceneDetect; long shots split Blur, Exposure, blanks, duplicates Qwen2.5-VL-72B; scores the action phase DocumentsChunkingProse FilterEmbedding + Index 11 sources Heading-aware, 500-token target Headers, dates, titles removed BGE-large; exact FAISS search A closed-set label can be audited frame by frame Qwen2.5-VL-72B SAM 2.1 ·InternVL3-8B ·GroundingDINO BGE-large FAISS Every candidate answer is doctrine that already exists The quoted span is verified by string comparison Each sentence must be entailed by the corpus Index lookup Doctrine content is quoted One question remains: does the frame show the concept? Claude Opus / 1,515 candidates 709 judged not visible 4raters, 88 items physicians 69% and 69% students 93% and 94% 5 open VLMs physician-accepted 0.79 physician-questioned 0.65 Refusal axis 150items, 8 reason types RWHR severity-weighted, the headline metric Questions they cannot answer 581items: 431 answerable + 150 refusal 1,860questions ·11concepts· 190source videos Preprocessing Figure 2: TC3-VQA construction workflow. Left: direct generation of questions and answers from an image. Right: visual concept assignment, doctrine retrieval, verbatim answer anchoring, and entailment filtering. Equipment and anatomy annotations record visual evidence; cross-family review and recognition consensus check concept labels. Refusal questions target information absent from the frames. These limitations motivate dataset construction that independently verifies what the scene supports and what the clinical source substantiates. We therefore present TC3-VQA, a dataset linking public TC3 footage to authoritative clinical guidance. It contains 1,860 questions on 581 items (431 answerable and 150 refusal) across 11 concepts, covering intervention recognition, doctrine, clinical reasoning, procedural guidance, and answerability (Figure 1). The release provides traceable source passages alongside equipment boxes, anatomical labels, temporal annotations, source metadata, and quality assessments. TC3-VQA is intended to support VLM fine-tuning and evaluation of visual grounding, doctrine-based reasoning, and appropriate abstention in TC3. 2 Methods Figure 2 provides an overview of the construction workflow. 2.1 Data collection and preprocessing Video collection and preprocessing. Search terms were developed with a physician trained in TC3 and covered the interventions and phases of care defined in the guidelines of the Committee on Tactical Combat Casualty Care (CoTCCC) [1]. Using these terms, we collected 415 publicly available YouTube videos. The collection combined formal instructional material with field footage of casualty care, spanning both procedural demonstrations and interventions performed under realistic conditions. Video identifiers, URLs, and timestamps were retained to preserve source traceability. 3 Videos were split into shots using the PySceneDetect content detector [25] at its default thresh- old (27). Shots longer than 60 s were divided into windows of at most 30 s, and frames were sampled approximately every 4 s. Filtering for poor exposure, blank content, blur, and near-duplicates left 32,987 frames in 17,055 windows. Because a single frame can be ambiguous about whether an in- tervention is under way, Qwen2.5-VL-72B-Instruct [26] evaluated each window jointly in temporal order. Each frame received an action phase (active procedure, post-procedure result, assessment only, or no medical action) and a usefulness score from 0 to 5. Windows containing an active procedure or post-procedure result with a usefulness score of at least 3 were retained with up to four qualifying frames, yielding 8,012 windows. Doctrine corpus collection and preprocessing. The doctrine corpus contained 9,320 chunks from 11 sources spanning operational guidance and supporting clinical literature. The CoTCCC guidelines, including the January 2024 edition [1] and the December 2021 guidelines for medical personnel [27], define the TC3 standard of care. Army Techniques Publication (ATP) 4- 02.11 [2] provides procedural guidance, while the Joint Trauma System Clinical Practice Guide- lines [28] add intervention-specific clinical detail. Joint Publication 4-02 [29] provides the broader joint health-services context. Of these, ATP 4-02.11, the A–H collection of Joint Trauma System guidelines, and the two CoTCCC guideline documents contributed 1,902 chunks covering the in- terventions represented in the dataset. The remaining corpus broadened retrieval to scene-specific guidance, acronym definitions, and related literature through a glossary compiled from the doctrine documents and PubMed and arXiv abstracts retrieved with TC3-focused queries. Documents were extracted with PyMuPDF and split at detected headings into chunks of ap- proximately 500 tokens, with a maximum of 800, to preserve recommendations together with the sections defining their scope. PubMed and arXiv records and glossary entries were treated as individual chunks. Each chunk retained its source, section path, position, and page range, and duplicates arising from overlapping guideline archives were removed. Before retrieval, a prose fil- ter excluded short, numeric, capitalized, and bibliographic fragments, removing 1,428 of the 5,142 chunks derived from guideline and doctrine documents. 2.2 Visual annotation and concept assignment Each retained window received one of 12 observable TC3 intervention or documentation labels, organized by the MARCH sequence (massive hemorrhage, airway, respiration, circulation, and hy- pothermia) and a documentation category (Table 1). Qwen2.5-VL-72B assigned the initial concept, generated a factual scene description, and rated its confidence as high, medium, or low. High- and medium-confidence assignments (5,051 of 8,012 windows) formed the answerable pool, while low- confidence windows were reserved for refusal construction. Limiting each video to six candidates reduced source concentration and yielded 1,515 items for review. Because concept errors would propagate to questions derived from a scene, labels underwent independent verification before question generation. Label consistency was first tested under a stricter perception prompt, followed by a blinded visual review with Claude Opus 4.8 that classified each concept as clearly visible, partly visible, or not visible without access to the initial reasoning. Of the 1,515 candidates, 336 were clearly visible, 470 partly visible, and 709 not visible; only the first two groups continued. Visual annotations provided additional evidence for each retained label. Equipment boxes followed a 17-class ontology of TC3 instruments and supplies. Initial box proposals were cross- checked for object presence with GroundingDINO [30] and InternVL3-8B [31], then refined from object masks using SAM 2.1 (Hiera-Large) [32]. The original box was retained when refinement produced an empty or implausibly large mask. A final visual review checked object presence and 4 class assignment, retaining, relabeling, or removing boxes as needed, with a targeted pass for flagged omissions. Missing boxes therefore do not indicate equipment absence. The casualty body region was annotated using eight anatomical regions, with additional labels for multiple or non-specific sites and unclear views. An answer-region box marked the visual evidence for the assigned concept. A final consistency check incorporated these annotations to reassess the concept label, with changed assignments reviewed for correction or routing to refusal. 2.3 Doctrine retrieval and question construction Doctrine was retrieved before question generation so that each answer could be anchored to the source corpus. A query combining the assigned concept with verified visual attributes was em- bedded with BGE-large-en-v1.5 [33], and FAISS [34] performed exact inner-product search over L2-normalized vectors. A curated passage for each concept provided an on-topic anchor. To di- versify clinical content, doctrine facets such as indications, techniques, key parameters, common errors, and contraindications were assigned to scenes in rotation. Because facet assignment reused the scene description, it was not treated as an independent validation step. Question construction used the frames and retrieved passages to produce four answerable types. Recognition questions asked which intervention was depicted, using the verified concept as the answer and visually confusable interventions as distractors. Doctrine questions asked for guidance on the depicted intervention and assigned facet, incorporating visible details when supported by the source. Reasoning questions asked for an assessment, next step, or critique of the scene, while procedural how questions asked how to perform, secure, or place the intervention and drew primarily on the operational field manual. Some doctrine questions were intentionally general enough to apply across scenes of the same intervention. Answers to doctrine, reasoning, and how questions were extracted directly from the retrieved passages. Claude Opus 4.8 selected a verbatim span from the retrieved passages, recorded by corpus chunk and character offsets. A readable rendering was retained only when supported by the entail- ment check described below; otherwise, the original span was used. Token-preserving normalization removed formatting artifacts while preserving the original span and offsets for provenance. Refusal questions reused frames from answerable items but requested information not estab- lished by the frames, including vital signs, elapsed time, blood-loss volume, off-frame anatomy, illegible medication information, or the outcome of an unshown step. Each question underwent a separate answerability check. 2.4 Verification and release curation Reference answers were checked for exact source correspondence. For a cited corpus chunk c and recorded character interval [i s ,i e ), a span a was accepted only when a = c [i s : i e ].(1) This deterministic comparison verified provenance for every released doctrine, reasoning, and how answer. A readable rendering r was retained only when every sentence was entailed by the item’s concept corpus C: accept(r) ⇐⇒ ∀u∈ sent(r) : Pr nli (entail| C,u)≥ τ,(2) using a DeBERTa-v3 natural language inference (NLI) model [35, 36] with τ = 0.5. The same classifier and threshold are used for free-form scoring in the Usage Notes, where entailment must also exceed contradiction. 5 Each assembled item was reviewed against its frames for concept and anatomy labels, image quality, answerability, answer leakage, doctrine relevance, corpus noise, and answer specificity. Flagged questions were revised under stronger scene constraints and rechecked, reducing off-topic flags from 86 to 5, corpus-noise flags from 15 to 0, and answer-telegraphing flags from 72 to 1. Uncertain anatomy was removed from question stems, answer-revealing captions were masked, and passages inconsistent with the depicted phase of care were replaced or removed. Unusable items were excluded or converted to refusals when possible. Rejected items and audit logs were retained, with potentially unsafe edits flagged for clinician review. Two final filters determined the released set. Unavailable source videos were first removed together with their 22 items (16 answerable and six refusal), with reasons recorded in the exclusion manifest. The remaining answerable items underwent an independent cross-family consensus check using six VLMs from five families (Qwen2.5-VL-72B, Qwen3-VL-32B [37], InternVL3-78B, Mistral- Small-3.2-24B [38], MedGemma-27B [39], and Claude Opus 5). Each model classified the frames using the same concept set without access to equipment annotations. An endorsement required agreement with the assigned concept at high or medium confidence, and at least three of six endorsements were required for retention. It removed 112 of 543 answerable items and the 31 refusal items paired with their frames, yielding 581 released items (431 answerable and 150 refusal). Human ratings were collected separately to characterize annotation quality and did not modify released labels. 2.5 Ethics statement The study used public TC3 videos and public United States Government and CoTCCC documents. No private clinical footage or patient-identifying material was used, no patients were enrolled, and no clinical intervention was performed. Four co-authors, two physicians and two medical students, rated the public video material. The raters were members of the research team assessing publicly available material, so the activity did not require institutional review. 3 Data Records TC3-VQA version 1.1 is deposited at Zenodo [40] (https://doi.org/10.5281/zenodo.23170287) and organized by content type. The data directory contains items.jsonl, with one record per item; doctrine chunks.jsonl, containing the cited corpus; detectioncoco.json, containing equipment boxes in COCO format; segments.jsonl, containing the 441 source-video segments and their timestamps; and framefeatures.npz, containing image-encoder features for all frames. The meta directory provides the field dictionary (FIELDS.md), source-video manifest and license status, caption-mask rectangles, and excluded-item list. Of the 190 video sources, 24 were pub- lished by United States Government channels and one under a Creative Commons license; their 115 frames from 75 items are included in the frames directory. The remaining 165 sources are represented by identifiers, URLs, timestamps, and annotations, with scripts/fetch frames.py provided to reconstruct the corresponding frames. Human adjudication is provided in the gold di- rectory, the doctrine multiple-choice extension in the mcq directory, and evaluation inputs, controls, and scoring scripts in the eval directory. The release also includes a README, Croissant metadata (croissant.json), and a datasheet following the Datasheets for Datasets framework [41]. Each item links a frame or temporal window to its questions, reference answers, and visual annotations (Figure 1); meta/FIELDS.md defines all fields. Provenance fields identify the item and source video, URL and license status, video segment, and frame timestamps. Question fields record the type (recognition, doctrine, reasoning, procedural how, or refusal), text, recognition options, 6 (a) Wound packing (M)(b) Surgical cricothyroidotomy (A) (c) Chest seal (R)(d) IV/IO access (C)(e) Hypothermia prevention (H) Figure 3: Example frames across the MARCH categories: (a) massive hemorrhage, (b) airway, (c) respiration, (d) circulation, and (e) hypothermia; IV/IO, intravenous or intraosseous. Frames are from public-domain CoTCCC footage. Overlays show released annotations from the 17-class equipment ontology. and reference answer. Doctrine, reasoning, and how answers include the source, corpus chunk, and char start and charend offsets of the cited span. Visual annotations include equipment boxes (17 classes), the answer region, and observed body region. Review fields include the assigned concept, safety-critical flag, visibility tier, audit results, and review-priority tier. Refusal items additionally record the missing-information category and the answerable item whose frames they reuse. Figure 3 shows released annotations across the MARCH categories. 4 Data Overview TC3-VQA contains 1,860 questions in 581 items (431 answerable and 150 refusal) drawn from 190 videos and covering 11 concepts (Table 1). The answerable set comprises 431 recognition, 427 doctrine, 425 reasoning, and 427 how questions, with all four types available for 418 items. The sampled visual context is short, with a median segment duration of 10 s (interquartile range, 4–30 s). Of the 581 items, 309 use a single frame and 272 use two to four frames. Each video contributes a median of three items and a maximum of eight, including refusals. Questions are one or two sentences long, with median lengths ranging from 9 words for recognition to 27 for refusal, while reference answers range from a 3-word concept name to a 30-word procedural passage. ATP 4-02.11 supplies approximately half of the cited answers and 73% of procedural answers; the CoTCCC and Joint Trauma System guidelines supply most of the remainder (Figure 4b). Refusal questions most often request blood-loss volume, elapsed time, or vital signs, with no category exceeding 26% of refusal items (Figure 4a). The release contains 1,565 equipment boxes across 743 of 903 frames, covering 16 of 17 classes (Figure 4c). Extremities are the most frequent intervention sites, followed by the chest (Figure 4d). 7 010203040 blood-loss volume elapsed time vital signs outcome not shown drug name or dose other off-frame anatomy patient history 39 28 28 21 18 10 5 1 (a) Refusal questions: information missing 020406080100 % of answers doctrine reasoning how (b) Cited source, by question type ATP 4-02.11 CoTCCC Guidelines 2024 CoTCCC Guidelines (medical) JTS CPG 02505007501000 hand limb tourniquet hemostatic dressing chest seal decompression needle nasopharyngeal airway hypothermia wrap IV/IO catheter surgical airway device 919 171 163 58 53 44 29 22 18 (c) Equipment boxes, nine most frequent classes 050100150 lower extremity upper extremity unclear chest head or face pelvis or groin neck abdomen multiple back 126 71 59 57 41 27 17 17 15 1 (d) Body region of the intervention Figure 4: Composition of the released layers. (a) Information requested by the 150 refusal questions. (b) Source document of the cited answer, by question type. (c) Equipment-box counts for the nine most frequent of the 16 classes present. (d) Observed body region of the 431 answerable items. 5 Technical Validation 5.1 Concept labels Concept labels determine the questions associated with each scene, making their reliability central to the dataset. Labels were assessed at three stages: an independent visibility audit during con- struction (Sec. 2.2), cross-family consensus among six VLMs before release (Sec. 2.4), and physician review of a stratified sample. Among the 431 released answerable items, 245 were classified as clearly visible and 186 as partly visible. Model consensus was unanimous for 147 items, five of six for 117, four for 91, and three for 76 (mean, 4.8 of 6). Individual-model agreement with the released label ranged from 57% for MedGemma-27B to 94% for Qwen2.5-VL-72B. Unanimous endorsement was more frequent for clearly visible items (102/245) than partly visible items (45/186). Physician review showed the same pattern: both physicians accepted 82% of clearly visible labels compared with 44% of partly visible labels, and acceptance decreased from 22/29 items with unanimous model endorsement to 5/12 items endorsed by three models (Sec. 5.4). Visibility and consensus strength are released with each item to support stratified evaluation. 5.2 Question and answer validity Provenance was re-verified on the released records. All doctrine, reasoning, and how answers exactly reproduce their cited spans (Eq. 1) and passed the entailment gate (Eq. 2). We sepa- 8 Table 1: Released items by concept and MARCH category. Safety-critical counts answerable items whose cited doctrine passage the stakes rubric flags as safety-critical; the risk-weighted hallucination rate (RWHR; Usage Notes) uses concept-level weights. The construction vocabulary also includes tourniquet conversion, which has no released item. NPA, nasopharyngeal airway; TXA, tranexamic acid. MARCHConceptAnswerableRefusalSafety-critical M (massive hemorrhage)Wound packing12928129 Limb tourniquet application13755137 Junctional hemorrhage control424 A (airway)Nasopharyngeal airway (NPA) insertion 41172 Surgical cricothyroidotomy11511 R (respiration)Chest seal application291129 Needle chest decompression27827 C (circulation)IV/IO vascular access1790 TXA administration937 H (hypothermia)Hypothermia prevention1360 processCasualty card (D Form 1380) documentation 1460 Total431150346 rately evaluated whether refusal questions were unanswerable from their frames, doctrine and how questions were grounded in their paired scenes, and reasoning answers correctly addressed their questions. Qwen3-VL-32B, InternVL3-78B, and MedGemma-27B independently judged every re- leased item at temperature 0, without access to the item type or concept label. Each criterion included a negative control. Refusal questions were contrasted with answerable how questions from the same frames; doctrine and how questions were evaluated on their paired frames and on frames depicting another intervention; and reasoning answers were contrasted with answers from another item of the same concept. Table 2 reports criterion pass rates and corre- sponding control rates. Table 2: Automated audit of released questions and answers. Values are percentages satisfying each criterion; parentheses give the corresponding control rate. A majority requires two of three votes, and Gwet’s AC1 measures inter-rater agreement. Anchored indicates that a question was judged specific to its paired frames. Judgment (n)Qwen3-VL InternVL3 MedGemmaMajorityAC1 Refusal unanswerable (150)98.7 (12.2) 99.3 (34.5)88.0 (3.6) 98.7 (10.8) 0.91 Doctrine question anchored (427) 53.9 (4.7)50.4 (6.6)63.5 (52.7) 58.8 (7.7)0.46 how question anchored (427)87.8 (4.9)93.4 (9.1)94.1 (89.2) 94.6 (11.0) 0.86 Reasoning answer correct (425)86.4 (30.8) 89.4 (33.2)90.8 (38.0) 91.5 (33.0) 0.84 The majority judged 98.7% of refusal questions unanswerable, compared with 10.8% of an- swerable controls on the same frames, and 94.6% of how questions were anchored to their paired frames, compared with 11.0% on substituted frames. Doctrine questions were less scene-specific (58.8%), reflecting that 37% asked about indications or guidance applicable across scenes of the 9 same intervention. MedGemma-27B showed limited sensitivity to the how control, judging 89.2% as anchored even on substituted frames. The reasoning audit accepted 91.5% of released answers compared with 33.0% of within-concept controls. It accepted 48 of the 49 answers accepted by both physicians, but also 35 of the 38 flagged by at least one physician, indicating lower sensitivity to deficiencies identified by expert review. Because the same model families also contributed to concept-label validation, these results provide a model-based consistency check; clinical adequacy is evaluated separately by the physician ratings below. 5.3 Equipment annotations The release contains 1,565 boxes across 743 of 903 frames, covering 16 of the 17 ontology classes. Among the 431 answerable items, 383 contain equipment boxes, 393 contain answer-region boxes, and 420 contain at least one of the two. Every released box was individually reviewed during construction (Sec. 2.2) to verify object presence and class assignment. Because the SAM-refined coordinates were not modified during review, this verification establishes box content rather than localization accuracy. As a consistency check, a YOLOv8s detector [42] trained on 595 frames and evaluated on 148 held-out frames achieved mAP 50 = 0.49 and mAP 50-95 = 0.35. Class-level mAP 50 was approx- imately 0.6 for hands, decompression needles, hypothermia wraps, and surgical-airway devices, compared with 0.28 for the more visually variable hemostatic dressings. Because the split was defined by frame rather than source video, these results characterize annotation consistency within the collection rather than generalization to unseen footage. 5.4 Human ratings Two physicians (one TC3-trained) and two medical students independently evaluated 120 candidate items sampled across concepts and visibility tiers, with at most two items per distinct doctrine answer. Subsequent release filtering retained 88 of these items, which form the basis of the results below. Using a prespecified rubric, raters assessed recognition labels as correct, unclear, or wrong; the three quoted answers as correct, correct with a minor problem, or incorrect; and the clinical severity of errors. Agreement is reported using Gwet’s AC1 because ratings were strongly skewed toward correct. Table 3: Human ratings on the 88 adjudicated items retained in the release. Cells report correct / unclear / wrong for recognition and correct / minor problem / incorrect for quoted answers. AC1 is Gwet’s agreement within each rater pair and among all four raters. Physicians (%)Medical students (%)Agreement (AC1) Question typeRater 1Rater 2 † Rater 1Rater 2PhysiciansStudentsAll four Recognition69/20/1069/23/893/3/394/1/50.800.950.75 Doctrine74/17/991/3/695/3/176/17/70.660.710.68 Reasoning59/30/1192/1/792/6/260/33/70.480.520.57 Procedural (how)68/28/393/6/193/7/041/50/90.610.260.47 † TC3-trained physician. Rows may not sum to 100 because of rounding and up to two skipped ratings. Individual ratings and notes are provided in gold/. 10 Both physicians rated 69% of recognition labels correct, with AC1 0.80; uncertainty was more common than outright rejection (Table 3). The medical students accepted more labels (93–94% correct), including 23 of the 30 questioned by at least one physician. Physician uncertainty occurred mainly in partly visible scenes and when adjacent procedural steps, visually similar devices, or unreadable details made the assigned intervention ambiguous. These items were also harder for the five baseline models, whose mean recognition accuracy was 0.79 on labels accepted by both physicians and 0.65 on those questioned by both (difference 0.14, bootstrap 95% confidence interval [0.00, 0.28]). The released ratings therefore provide an expert-based indicator of item difficulty rather than a definitive relabeling of the dataset. Across the three quoted-answer types, outright incorrect ratings ranged from 0 to 11%, while differences between raters largely reflected use of the minor-problem category. Notes most often identified context omitted by the quoted passage and occasionally challenged the source itself. Severity ratings showed lower physician agreement (AC1 0.27), although mean physician severity increased from 1.0 to 2.05 and 2.4 across concepts assigned RWHR weights of 1, 2, and 3, fixed before rating, respectively (n = 7, 10, and 71). 5.5 Baseline performance and image dependence We benchmarked five open VLMs with 7–38 billion parameters: Qwen2-VL-7B [43], Qwen2.5- VL-7B, InternVL3-8B and 38B, and Phi-3.5-Vision [44]. Models used for primary perception and question generation were excluded, although Qwen and InternVL families contributed to other construction stages. Recognition and refusal were evaluated as five-option questions with absten- tion. Doctrine was evaluated by free-form entailment and by multiple choice using distractors from another concept (easy) or another facet of the same concept (hard). RWHR (Sec. 6.5) measures severity-weighted error among answered items. Table 4: Baseline performance on 431 answerable and 150 refusal items. Doctrine-MCQ (multiple- choice questions) uses distractors from other concepts (easy) or other facets of the same concept (hard). Doctrine-open reports reference-sentence entailment and the proportion of response sen- tences unsupported by the reference or cited passage. Bold indicates the best value in each column; lower is better for unsupported and RWHR. ModelRecognition Doctrine-MCQDoctrine-openRefusal RWHR ↓ easyhardentailed unsupported ↓ Qwen2-VL-7B0.8540.9510.8640.2670.5760.2600.921 Qwen2.5-VL-7B0.8050.9630.874 0.2850.5740.8330.546 Phi-3.5-Vision0.6660.8670.7030.2250.5880.9330.888 InternVL3-8B0.8400.9480.8450.2430.5790.8600.479 InternVL3-38B0.7400.986 0.9130.2640.5230.933 0.420 Recognition accuracy ranged from 0.67 to 0.85 (Table 4), with lower performance on partly visible interventions. Recognition and refusal showed a clear trade-off: Qwen2-VL-7B achieved the highest recognition accuracy (0.85) but only 26% refusal accuracy and the highest RWHR (0.92), whereas InternVL3-38B achieved 0.74, 93%, and 0.42, respectively (Figure 5). Doctrine-MCQ accuracy ranged from 0.87 to 0.99 with easy distractors and from 0.70 to 0.91 with hard distractors. Free-form responses entailed 0.23–0.29 of reference sentences, while 52–59% of generated sentences were unsupported. 11 0.60.70.80.9 recognition accuracy (Wilson 95% interval) 0.2 0.4 0.6 0.8 1.0 refusal accuracy (correct abstention) (a) Recognition against refusal RWHR 0.92 RWHR 0.55 RWHR 0.89 RWHR 0.48 RWHR 0.42 0.60.70.80.91.0 coverage (share of items answered) 0.0 0.2 0.4 0.6 0.8 1.0 risk among answered items (b) Coverage against risk RWHR (severity-weighted) unweighted rate Qwen2-VL-7BQwen2.5-VL-7BPhi-3.5-VisionInternVL3-8BInternVL3-38B Figure 5: Recognition, abstention, and error across five baseline models. (a) Recognition and refusal accuracy with Wilson 95% intervals; marker area represents RWHR. (b) Coverage versus unweighted error (open markers) and severity-weighted error (filled markers). Recognition depended strongly on the visual input. Substituting frames from another inter- vention reduced accuracy to 0.04–0.14 for four models, below the 0.20 uniform-choice reference, while Phi-3.5-Vision retained 0.33 (Table 5). Multi-frame context also contributed information: restricting the 204 multi-frame items to the first frame reduced accuracy by 7.8–12.7 percentage points for four models, with significant exact McNemar tests, while Qwen2-VL-7B decreased by 2.5 points (p = 0.42). Full-window reruns reproduced the released scores within 0.002. Table 5: Recognition under frame substitution and temporal truncation. Left, questions paired with frames from another concept (431 items; uniform-choice reference 0.20). Right, full windows compared with the first frame alone (204 multi-frame items). ∆ denotes perturbed minus original accuracy; p values are from two-sided exact McNemar tests. Frame substituted (all 431)First frame only (204 multi-frame) Modelownswapped∆abstainwindow1 frame∆p Qwen2-VL-7B0.8540.139 −0.7152.3%0.8530.828 −0.0250.42 Qwen2.5-VL-7B0.8070.081 −0.72618.8%0.8140.686 −0.127 < 0.001 Phi-3.5-Vision0.6660.332 −0.33410.2%0.7160.637 −0.0780.014 InternVL3-8B0.8400.123 −0.71717.9%0.8730.745 −0.127 < 0.001 InternVL3-38B0.7400.044 −0.69644.8%0.8190.725 −0.0930.002 5.6 Comparison with direct question-answer generation We compared the construction pipeline with direct question-answer generation on 288 candidates whose assigned concepts were independently confirmed as clearly visible. Six open VLMs and Claude Opus 5 generated one pair for each answerable question type from the same frames, with answer length matched to the reference and without access to the audited concept, doctrine corpus, or verification gates. Qwen2.5-VL-72B was additionally tested with the audited concept provided to isolate the effect of concept verification. 12 Direct generation frequently produced questions about the wrong intervention. Among the open VLMs, 16–31% of questions named another listed concept, compared with 0.4% of pipeline- generated questions. When models selected the concept themselves before generation, 11–42% of questions remained inconsistent with the audited concept, whereas Qwen2.5-VL-72B produced no such errors when the audited concept was provided (Figure 6d). (a) Wrong intervention, plausible answer Ours Recognition: Wound packing Q: How can the effectiveness of wound packing be improved according to TCCC guidance? A: "The effectiveness of wound packing may be improved when skin closure over the packing can be achieved..." Conventional Recognition: Hemostatic Dressing Application Q: What is the proper method for applying a tourniquet to control bleeding as shown here? A: Place tourniquet 2-3 inches above wound. Tighten until bleeding stops. Mark time... Wound packing is shown, not a tourniquet. (b) Wrong intervention, unsupported answer Ours Recognition: Surgical cricothyroidotomy Q: At what site should the initial vertical skin incision for a surgical cricothyroidotomy be made? A: "Make vertical incision through the skin over cricothyroid membrane." Conventional Recognition: Needle Decompression Q: What should be done if the needle decompression does not relieve tension pneumothorax? A: Perform chest tube insertion immediately. No needle decompression is shown. (c) Wrong intervention, off-target care Ours Recognition: Chest seal application Q: After applying a chest seal, what complication must the casualty be monitored for and how is it treated? A: "Monitor the casualty for the potential development of a subsequent tension pneumothorax..." Conventional Recognition: Chest Seal Application Q: Given the visible chest wound and the application of a tourniquet, what is the next step in TCCC for managing this casualty's condition effectively? A: The next step is to check for airway obstruction and breathing issues... Ensure the tourniquet is properly secured... A chest seal is shown and no tourniquet is present. 0204060 % of questions (95% CI) Released Qwen2.5-VL + verified concept Claude Opus 5 Qwen3-VL-32B InternVL3-78B Qwen2.5-VL-72B Pixtral-Large Mistral-Small-3.2-24B MedGemma-27B 0.4% 0.0% 5.1% 16.1% 17.7% 21.8% 24.3% 25.4% 30.7% Open questions 0204060 % of items (95% CI) reference label n/a 15.6% 21.5% 16.0% 15.3% 26.4% 40.3% 47.6% Recognition, 12 options (d) Questions presupposing another intervention (288 audit-confirmed candidates) Figure 6: Reference question-answer (QA) pairs compared with direct generation from the same frames. (a–c) Public-domain examples showing the pipeline-generated reference pair (Ours) and direct generation by Qwen2.5-VL-72B (Conventional). (d) Concept inconsistency during direct question generation (left) and closed-set concept selection (right) on 288 audit-confirmed candidates (Pixtral-Large [45] and the models of Sec. 2.4), with Wilson 95% intervals. Five blinded large language model (LLM) judges from different model families (Llama-3.3- 70B [46], MedGemma-27B, Phi-4 [47], DeepSeek-R1-Distill-Llama-70B [48], and gpt-oss-120b [49]) assessed whether each question was concept-consistent and its answer supported by eight doctrine passages retrieved for that question. Pipeline-generated pairs satisfied both criteria for 95–99% of questions, compared with 72–78% for direct generation by Qwen2.5-VL-72B. Among concept- 13 consistent questions, however, at least 92% of answers in both groups satisfied both criteria under every judge. The difference therefore arose primarily from concept misalignment during question generation rather than unsupported answers once the intervention was correctly identified. 6 Usage Notes 6.1 Model adaptation and data partitioning For instruction tuning, pair the referenced frames with each question and released answer, preserv- ing source citations and audit metadata. Refusal examples provide contrasting supervision from the same scenes, where an intervention may be recognizable even when the requested clinical detail is unsupported. Training, validation, and test sets should be partitioned by source video, with each answerable item kept with its derived refusals to prevent frame overlap. Because 1,279 quoted answers cite only 313 distinct passages from 49 chunks across six documents, studies of transfer to unseen doctrine should additionally separate source passages and report concept coverage. Reports should specify the dataset version, included question types, and use of human ratings. 6.2 Recognition and abstention Recognition is evaluated by top-1 accuracy, with macro accuracy over the six MARCH categories reported for the imbalanced concept distribution: Acc macro = 1 K K X k=1 1 |I k | X i∈I k 1[ˆy i = y i ],(3) where I k denotes items in category k and K = 6. Refusal performance is the proportion of unanswerable items for which the model abstains and should be reported separately from recognition accuracy. Results can also be stratified by intervention visibility, consensus strength, concept frequency, frame count, and, for the adjudicated subset, physician acceptance. Each answerable item addi- tionally carries a safety-critical flag for its cited doctrine passage, distinct from the concept-level severity weight used by RWHR. 6.3 Doctrine scoring Free-form doctrine answers are scored by entailment against the reference answer and cited passage using the NLI classifier and threshold of Eq. 2; the multiple-choice variant is scored by accuracy. Sentence-level coverage is Doctrine partial (ˆr,g) = 1 |sent(g)| X u∈sent(g) 1[Pr nli (entail| ˆr,u)≥ τ ],(4) with the response as premise and each reference sentence as hypothesis. Whole-answer entailment is also reported. Unsupported content is measured as the proportion of response sentences entailed by neither the reference nor the cited passage and should be interpreted cautiously because the classifier may miss valid paraphrases. Doctrine-MCQ uses the cited answer as the reference option. Easy distractors come from the same facet of another concept, whereas hard distractors come from another facet of the same 14 concept; options are filtered for near-duplicates and randomized. Both variants should be reported alongside free-form scores because they test different response requirements. 6.4 Visual grounding annotations Equipment boxes are provided in COCO format over the 17-class ontology, together with an an- swer region and observed body region for each answerable item. These annotations support equip- ment detection, evidence localization, and analyses of model attention to relevant instruments and anatomy. Box coverage is not exhaustive, so missing boxes should not be treated as negative annotations. Detection experiments should split data by source video. 6.5 Risk-weighted error and coverage The Risk-Weighted Hallucination Rate (RWHR) measures severity-weighted recognition errors and responses to unanswerable questions among non-abstained items. Concepts receive weights w ∈ 1, 2, 3 for minor, serious, and critical consequences of error: 3 for hemorrhage, airway, and respiration concepts; 2 for intravenous or intraosseous access, tranexamic acid, and hypothermia prevention; and 1 for documentation. Let a i indicate a non-abstained response, g i the reference label, and∅ abstention: RWHR = P i a i w i h i P i a i , h i = 1[ˆy i ̸=∅ ∧ ˆy i ̸= g i ].(5) RWHR is an average weighted error per answered item and can exceed one; it is undefined at zero coverage. It should therefore be reported with coverage, unweighted error, and the risk–coverage curve. Effective Reliability provides a complementary summary [50]. The weighting scheme is dataset-specific and should not replace unweighted results. 6.6 Limitations TC3-VQA contains automatically selected reference labels, with human ratings available for a stratified subset. Exact citations establish provenance, but automated verification can preserve shared perceptual or textual errors, and the 88 retained adjudicated items do not establish clinical accuracy for every record. Human ratings and audit metadata should therefore inform uses requiring higher annotation confidence. Coverage reflects the available footage and doctrine. Tourniquet application and wound packing dominate, while sparse concepts may require reporting by MARCH category. Some visual details, including medication identity, dose, route, and timing, cannot reliably be resolved from sampled frames, and repeated doctrine passages may make some questions answerable from general knowl- edge once the intervention is known. The consensus includes model families related to several baselines, and detector validation used a frame-level split that may share source-video characteristics. Instructional and field-training footage also differs from real combat casualties; performance on TC3-VQA therefore does not establish deployment readiness. Broader clinical adjudication, clinician-corrected labels, and transfer to operational settings remain future work. 7 Data Availability TC3-VQA version 1.1 is deposited at Zenodo, https://doi.org/10.5281/zenodo.23170287 [40]. The package includes item records, the cited doctrine chunks, visual annotations, source and license 15 manifests, temporal segments, caption masks, the human ratings, doctrine multiple-choice tasks, evaluation inputs and scorers, and the results reported here. Image files are included for 115 frames from 25 sources with redistribution permission; the remaining frames are specified by source identifier and timestamp and are regenerated from the original videos with the script provided. For items whose pixels cannot be redistributed, the package also carries frozen image-encoder features of every frame, so all 581 items remain usable without the source videos. Source reachability was checked on 17 September 2026: 189 of the 190 videos were reachable, and one, covering two items, had been removed. Annotations are released under C BY 4.0 and the scripts under the MIT license. Construction artifacts are retained by the authors and available on request. 8 Code Availability The construction and evaluation code is available at https://github.com/Kamaleswaran-Lab/ TC3-VQA.It includes frame reconstruction, annotation, retrieval, verification, and scoring. Qwen2.5-VL-72B-Instruct [26] was served with vLLM 0.15 and tensor parallelism of four on NVIDIA H200 (141 GB) GPUs. Images were resized to a longest side of at most 896 pixels. Claude Opus 4.8 performed visual auditing, doctrine-facet mining, and review. References [1] Travis G. Deaton, Brendon Drew, Harold R. Montgomery, and Frank K. Butler Jr. Tactical combat casualty care (TCCC) guidelines: 25 january 2024. Journal of Special Operations Medicine, 24(1):100–108, 2024. Committee on Tactical Combat Casualty Care (CoTCCC), Joint Trauma System; guideline version dated 25 January 2024. [2] Department of the Army. ATP 4-02.11: Casualty response, tactical combat casualty care, and first aid. Army Techniques Publication, Headquarters, Department of the Army, March 2026. ARN46159; supersedes TC 4-02.1. Published by the U.S. Army Medical Center of Excellence (MEDCoE). [3] Brian J Eastridge, Robert L Mabry, Peter Seguin, Joyce Cantrell, Terrill Tops, Paul Uribe, Olga Mallett, Tamara Zubko, Lynne Oetjen-Gerdes, Todd E Rasmussen, et al. Death on the battlefield (2001–2011): implications for the future of combat casualty care. Journal of Trauma and Acute Care Surgery, 73(6 Suppl 5):S431–S437, 2012. [4] Sean Keenan and Jamie C. Riesberg. Prolonged field care: Beyond the “golden hour”. Wilder- ness & Environmental Medicine, 28(2S):S135–S139, 2017. [5] Christopher Nemeth, Adam Amos-Binks, Christie Burris, Natalie Keeney, Yuliya Pinevich, Brian W. Pickering, Gregory Rule, Dawn Laufersweiler, Vitaly Herasevich, and Mei G. Sun. Decision support for tactical combat casualty care using machine learning to detect shock. Military Medicine, 186(Supplement 1):273–280, 2021. [6] A. Landman, D. de Vries, and O. Binsch. Retention of military combat lifesaving skills during six months following classroom-style and individualized-style initial training. Military Psy- chology, 35(6):590–602, 2023. [7] Christopher Nemeth, Adam Amos-Binks, Gregory Rule, Dawn Laufersweiler, Natalie Keeney, Isaac Flint, Yuliya Pinevich, and Vitaly Herasevich. TCCC decision support with machine 16 learning prediction of hemorrhage risk, shock probability. Military Medicine, 188(Supplement 6):659–665, 2023. [8] Victor A. Convertino, Steven G. Schauer, Erik K. Weitzel, Sylvain Cardin, Mark E. Stackle, Michael J. Talley, Michael N. Sawka, and Omer T. Inan. Wearable sensors incorporating compensatory reserve measurement for advancing physiological monitoring in critically injured trauma patients. Sensors, 20(22):6413, 2020. [9] Nehemiah T Liu, John B Holcomb, Charles E Wade, Andriy I Batchinsky, Leopoldo C Cancio, Mark I Darrah, and Jos ́e Salinas. Development and validation of a machine learning algorithm and hybrid system to predict the need for life-saving interventions in trauma patients. Medical & Biological Engineering & Computing, 52(2):193–203, 2014. [10] Defense Advanced Research Projects Agency. DARPA triage challenge. https://w.darpa. mil/research/challenges/darpa-triage-challenge, 2023. Defense Advanced Research Projects Agency (DARPA). Challenge kickoff November 2023; Challenge Event 1 held 2024. Accessed 2026-09-25. [11] Junseob Kim, Jade Chng, Ayman Ali, Victor Moas, Sunil Hwang, and Rishikesan Ka- maleswaran. TC3-VLM: A vision–language model for tactical combat casualty care. IEEE Access, 14:131019–131029, 2026. [12] Jason J. Lau, Soumya Gayen, Asma Ben Abacha, and Dina Demner-Fushman. A dataset of clinically generated visual questions and answers about radiology images. Scientific Data, 5(1):180251, 2018. [13] Xuehai He, Yichen Zhang, Luntian Mou, Eric Xing, and Pengtao Xie. PathVQA: 30000+ questions for medical visual question answering. arXiv preprint arXiv:2003.10286, 2020. [14] Bo Liu, Li-Ming Zhan, Li Xu, Lin Ma, Yan Yang, and Xiao-Ming Wu. SLAKE: A semantically- labeled knowledge-enhanced dataset for medical visual question answering. In 2021 IEEE 18th International Symposium on Biomedical Imaging (ISBI), pages 1650–1654. IEEE, 2021. [15] Xiaoman Zhang, Chaoyi Wu, Ziheng Zhao, Weixiong Lin, Ya Zhang, Yanfeng Wang, and Weidi Xie. Development of a large-scale medical visual question-answering dataset. Communications Medicine, 4(1):277, 2024. [16] Chunyuan Li, Cliff Wong, Sheng Zhang, Naoto Usuyama, Haotian Liu, Jianwei Yang, Tristan Naumann, Hoifung Poon, and Jianfeng Gao. LLaVA-Med: Training a large language-and- vision assistant for biomedicine in one day. In Advances in Neural Information Processing Systems 36 (NeurIPS 2023) Datasets and Benchmarks Track, 2023. [17] Mehmet Saygin Seyfioglu, Wisdom O. Ikezogwo, Fatemeh Ghezloo, Ranjay Krishna, and Linda Shapiro. Quilt-LLaVA: Visual instruction tuning by extracting localized narratives from open- source histopathology videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13183–13192, 2024. [18] Rahul Thapa, Andrew Li, Qingyang Wu, Bryan He, Yuki Sahashi, Christina Binder, Angela Zhang, Ben Athiwaratkun, Shuaiwen Leon Song, David Ouyang, and James Zou. How well can general vision-language models learn medicine by watching public educational videos? arXiv preprint arXiv:2504.14391, 2025. 17 [19] Yupeng Zhuo, Eddie Zhang, Xiangchen Yu, Aditya Pachpande, Wenqu Fang, Xiyue Chen, An- drew W. Kirkpatrick, Kyle Couperus, Christopher Colombo, Oanh Tran, Jonah Beck, DeAnna DeVane, Jessica McKee, Chad Gorbatkin, Ross Candelore, Eleanor Birch, and Juan Wachs. An egocentric life-saving interventional procedure dataset of actions, medical questions, ma- neuvers and tools. Scientific Data, 13(1):51, 2026. [20] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 292–305, 2023. [21] Zishan Gu, Changchang Yin, Fenglin Liu, and Ping Zhang. MedVH: Towards systematic evaluation of hallucination for large vision language models in the medical context. arXiv preprint arXiv:2407.02730, 2024. [22] Tengda Han, Weidi Xie, and Andrew Zisserman. Temporal alignment networks for long- term video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2896–2906, 2022. [23] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin, and Feng Zhao. Are we on the right way for evaluating large vision-language models?In Advances in Neural Information Processing Systems 37 (NeurIPS), 2024. [24] Kevin Wu, Eric Wu, Kevin Wei, Angela Zhang, Allison Casasola, Teresa Nguyen, Sith Ri- antawan, Patricia Shi, Daniel Ho, and James Zou. An automated framework for assessing how well LLMs cite relevant medical references. Nature Communications, 16:3615, 2025. [25] Brandon Castellano. PySceneDetect. https://github.com/Breakthrough/PySceneDetect, 2026. Version 0.7.1. [26] Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, Humen Zhong, Yuanzhi Zhu, Mingkun Yang, Zhaohai Li, Jianqiang Wan, Pengfei Wang, Wei Ding, Zheren Fu, Yiheng Xu, Jiabo Ye, Xi Zhang, Tianbao Xie, Zesen Cheng, Hang Zhang, Zhibo Yang, Haiyang Xu, and Junyang Lin. Qwen2.5-VL technical report. arXiv preprint arXiv:2502.13923, 2025. [27] Committee on Tactical Combat Casualty Care. Tactical combat casualty care (TCCC) guide- lines for medical personnel, December 2021. Guideline version dated 15 December 2021; Joint Trauma System. [28] Joint Trauma System. Joint trauma system clinical practice guidelines (CPGs). https: //jts.health.mil/index.cfm/CPGs/cpgs, 2024. Department of Defense Center of Excellence for Trauma. Accessed 2026-07-02. [29] Joint Chiefs of Staff. Joint publication 4-02: Joint health services, December 2017. Dated 11 December 2017. [30] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, and Lei Zhang. Grounding DINO: Marrying DINO with grounded pre-training for open-set object detection. In Computer Vision – ECCV 2024, Lecture Notes in Computer Science, pages 38–55. Springer, 2024. 18 [31] Jinguo Zhu, Weiyun Wang, Zhe Chen, et al. InternVL3: Exploring advanced training and test-time recipes for open-source multimodal models. arXiv preprint arXiv:2504.10479, 2025. [32] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman R ̈adle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Doll ́ar, and Christoph Feichtenhofer. SAM 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. [33] Beijing Academy of Artificial Intelligence.bge-large-en-v1.5: Model card. https:// huggingface.co/BAAI/bge-large-en-v1.5, 2023. Accessed 2026-09-16. [34] Matthijs Douze, Alexandr Guzhva, Chengqi Deng, Jeff Johnson, Gergely Szilvasy, Pierre- Emmanuel Mazar ́e, Maria Lomeli, Lucas Hosseini, and Herv ́e J ́egou. The Faiss library. IEEE Transactions on Big Data, 12(2):346–361, 2026. [35] Pengcheng He, Jianfeng Gao, and Weizhu Chen. DeBERTaV3: Improving DeBERTa using ELECTRA-style pre-training with gradient-disentangled embedding sharing. In International Conference on Learning Representations (ICLR), 2023. [36] Moritz Laurer, Wouter van Atteveldt, Andreu Casas, and Kasper Welbers. Less annotating, more classifying: Addressing the data scarcity issue of supervised machine learning with deep transfer learning and BERT-NLI. Political Analysis, 32(1):84–100, 2024. [37] Shuai Bai et al. Qwen3-VL technical report. arXiv preprint arXiv:2511.21631, 2025. [38] Mistral AI. Mistral-Small-3.2-24B-Instruct-2506: Model card. https://huggingface.co/ mistralai/Mistral-Small-3.2-24B-Instruct-2506, 2025. Accessed 2026-09-28. [39] Andrew Sellergren, Sahar Kazemzadeh, Tiam Jaroensri, et al. MedGemma technical report. arXiv preprint arXiv:2507.05201, 2025. [40] Junseob Kim, Jade Chng, Ayman Ali, Victor Moas, Yichun Lee, Po-Chun Chin, Sunil Hwang, and Rishikesan Kamaleswaran. TC3-VQA: A doctrine-grounded visual question answering dataset for tactical combat casualty care, 2026. [41] Timnit Gebru, Jamie Morgenstern, Briana Vecchione, Jennifer Wortman Vaughan, Hanna Wallach, Hal Daum ́e I, and Kate Crawford. Datasheets for datasets. Communications of the ACM, 64(12):86–92, 2021. [42] Glenn Jocher, Jing Qiu, and Ayush Chaurasia. Ultralytics YOLO. https://github.com/ ultralytics/ultralytics, 2023. [43] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, et al. Qwen2-VL: Enhancing vision-language model’s perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. [44] Marah Abdin, Jyoti Aneja, Hany Awadalla, et al. Phi-3 technical report: A highly capable language model locally on your phone. arXiv preprint arXiv:2404.14219, 2024. Revision of 30 August 2024 describes Phi-3.5-Vision. [45] Mistral AI. Pixtral-Large-Instruct-2411: Model card. https://huggingface.co/mistralai/ Pixtral-Large-Instruct-2411, 2024. Accessed 2026-09-28. 19 [46] Meta AI. Llama-3.3-70B-Instruct: Model card. https://huggingface.co/meta-llama/ Llama-3.3-70B-Instruct, 2024. Accessed 2026-09-28. [47] Marah Abdin, Jyoti Aneja, Harkirat Behl, et al. Phi-4 technical report. arXiv preprint arXiv:2412.08905, 2024. [48] Daya Guo et al. DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning. Nature, 645(8081):633–638, 2025. [49] OpenAI, Sandhini Agarwal, et al. gpt-oss-120b & gpt-oss-20b model card. arXiv preprint arXiv:2508.10925, 2025. [50] Spencer Whitehead, Suzanne Petryk, Vedaad Shakib, Joseph Gonzalez, Trevor Darrell, Anna Rohrbach, and Marcus Rohrbach. Reliable visual question answering: Abstain rather than answer incorrectly. In Computer Vision – ECCV 2022, volume 13696 of Lecture Notes in Computer Science, pages 148–166. Springer Nature Switzerland, 2022. Author Contributions J.K.: Conceptualization, Methodology, Software, Validation, Formal analysis, Investigation, Data curation, Visualization, Writing – original draft, Writing – review & editing; J.C.: Software, Data curation, Writing – review & editing; A.A.: Validation, Writing – review & editing; V.M.: Valida- tion, Writing – review & editing; Y.L.: Validation, Writing – review & editing; P.-C.C.: Validation, Writing – review & editing; S.H.: Methodology, Writing – review & editing; R.K.: Supervision, Funding acquisition, Writing – review & editing. Competing Interests The authors declare no competing interests. Funding This work was supported in part by the National Institutes of Health under Grant R01GM139967 and Grant R01HL170175. 20