MedProb: Probing Internal Representations of Vision-Language Models for Medical Question Answering

1 Introduction

• Generation-based evaluation combines internal features with prompt wording, answer-option order, and answer expression, so a model may represent useful image information yet choose incorrectly.Generated explanations may also introduce unsupported reasoning.Source: §1 · pp. 1–2
• MedProb freezes the VLM and trains a linear classifier on hidden representations to predict multiple-choice answers, separating decodable answer signal from free-text generation.The framework is presented as an empirical diagnostic, not a replacement for clinical reasoning systems or a fully supervised Med-VQA solution.Source: §1 · p. 2
• Across PATH-VQA, SLAKE, and VQA-RAD, MedProb recovers substantially more answer-relevant signal than prompting and performs competitively with generation-based systems.The compared systems include medically fine-tuned VLMs and multi-agent methods.Source: §1 · p. 2
• The small-versus-large model performance gap is much smaller under probing than prompting, suggesting generation-based evaluation can underestimate smaller models' answer-relevant signal.This result motivates separating what is encoded from what is elicited through generation.Source: §1 · p. 2
• Across 14 matched general-purpose and medical VLM pairs, medical adaptation does not consistently improve linear decodability, while free-text generation shows answer-position bias of up to 10 percentage points.MedProb is also more stable under option reordering than free-text generation.Source: §1 · p. 2

2 Related Work

• Medical VLM development has pursued domain adaptation of general-purpose models and purpose-built systems, with approaches including medical instruction tuning and domain-adaptive pretraining.Examples include Med-Flamingo (Moor et al., 2023b), LLaVA-Med (Li et al., 2023), and HuatuoGPT-Vision (Chen et al., 2024a).Source: §2 · p. 2
• Unlike work that builds or extensively trains medical VLMs, MedProb probes frozen models and finds medical adaptation does not consistently improve linear decodability of answer-relevant signal.The related-work discussion also notes that reasoning-augmented variants can require extensive training resources (Huang et al., 2026; Mullappilly et al., 2026).Source: §2 · pp. 2–3
• Prompt formatting and answer-option ordering can substantially change evaluation results, including accuracy swings up to 76 points across semantically equivalent formats (Sclar et al., 2024).Prior work also reports option-order effects in multiple-choice settings (Pezeshkpour and Hruschka, 2024).Source: §2 · p. 3
• Prior probing work reports that linear probes can outperform generation-based judgment across model families and datasets (Maiya et al., 2025), and recover clinical-QA performance lost to safety alignment (Berkowitz et al., 2025a).MedProb extends this paradigm to multimodal Med-VQA and compares 14 matched model pairs across three benchmarks.Source: §2 · p. 3

3 Methodology

• Each Med-VQA example consists of a medical image, question, answer options, and gold label, with binary questions represented as two choices and other questions as four.The task is framed as multimodal multiple-choice classification.Source: §3 · p. 3
• Prompting combines the image, question, and answer options using a template, then extracts a predicted option from the VLM's free-text response.This is the generation-based comparison for probing.Source: §3 · p. 3
• MedProb freezes the VLM and tests whether correct answers are readable from hidden states rather than changing the model or adding external knowledge.A linear probe's ability to predict an answer indicates that answer-relevant information is linearly accessible at that layer.Source: §3 · p. 3
• At each transformer layer, MedProb uses the last input token's hidden state as a summary of multimodal context and trains a separate multinomial logistic-regression probe.The probe predicts a distribution over answer choices from that representation.Source: §3 · p. 3
• The VLM parameters remain fixed; regularization is selected by five-fold cross-validation, and the best layer is chosen on a held-out validation split separate from the test set.Features are standardized using training-set statistics.Source: §3 · p. 4

4 Experiments

• The model set includes publicly available general-purpose and medical VLMs, with matched comparisons restricted to models sharing a backbone and vision encoder.Unmatched model cases are excluded to isolate domain adaptation.Source: §4 · p. 4
• Evaluation uses VQA-RAD, PathVQA, and SLAKE, restricting questions to closed-ended items and removing samples with missing images or training-set leakage.PathVQA is subsampled to 500 examples; VQA-RAD and SLAKE use their full test sets.Source: §4 · p. 4
• The framework compares agentic systems, directly prompted instruction-tuned medical VLMs, and benchmark-trained medical VLMs as upper-bound references for supervised specialization.The baseline medical VLMs do not share backbone architectures with the probed models.Source: §4 · p. 4
• MedProb fits a linear probe with up to 1,000 labeled examples while prompting baselines are zero-shot, and neither setting updates VLM parameters.The experiments report macro-averaged F1 and accuracy.Source: §4 · p. 4
• The evaluation framework contrasts direct prompting, frozen-layer representation decoding, and fine-tuning on medical data followed by prompting.The compared VLMs are grouped by backbone architecture.Source: Figure 2 · p. 5
• The probe uses last-token hidden states from every transformer layer, standardizes features, tunes logistic-regression regularization by cross-validation, and selects a layer on held-out validation data.The validation split is separate from the test set.Source: §4 · p. 5

Prompting versus probing

• On PATH-VQA, prompting accuracy is 57.2%–63.6%, versus MedProb’s 82.8%–87.0% across the same four Gemma-family models.On VQA-RAD, probing ranges from 72.2% vs. 52.5% for Gemma-4B to 76.0% vs. 69.2% for Gemma-27B.Source: Figure 3 · p. 5, §5 · p. 6

Main results

• 79.32% average accuracy lets InternVL3-1B surpass most agentic and several medical-VLM baselines, while MedProb consistently outperforms both categories across three benchmarks.The comparison covers PATH-VQA, SLAKE, and VQA-RAD.Source: §5 · p. 5, Table 1 · p. 6

Binary and multiclass evaluation

• On OmniMedVQA’s true 4-way evaluation, probing outperforms prompting overall; binary gains remain large on PATH-VQA, SLAKE, and VQA-RAD.SLAKE and VQA-RAD non-yes/no items are still effectively 2-way, so they are not treated as multiclass.Source: §5 · p. 6

Fine-tuning

• MedProb still beats prompting after fine-tuning: average accuracy is 84.7% vs. 79.5% on SLAKE and 70.3% vs. 59.5% on VQA-RAD.The probe uses 1,000 examples, while the fine-tuned VLM uses the full training set; PATH-VQA is the exception, where prompting matches probing.Source: §5 · pp. 6–7, Figure 4 · p. 7

Layer-wise analysis

• Probe accuracy rises with layer depth: Qwen2.5-VL-7B peaks at 80.4% at layer 28, while MedVLThinker-7B peaks at 79.5% at layer 22.MedVLThinker-7B leads at layer 8, but Qwen2.5-VL-7B overtakes it from layer 9 onward.Source: Figure 5 · p. 7, §5 · p. 7

Evaluation confounds

• Medical adaptation shows no consistent decodability gain across 14 matched pairs, while prompting is more sensitive to answer position than probing.Prompting’s first-versus-last gap reaches 10.49 pp for InternVL3-1B and 7.33 pp for LLaVA-V0-7B, compared with probing gaps of 2.10 pp and 3.58 pp.Source: §5 · pp. 7–8, Figure 6 · p. 7, Table 2 · p. 8

Image controls and error analysis

• Shuffling images lowers probe accuracy by 6.2–18.3 percentage points, and qualitative analysis finds probes can avoid generation-driven false clinical conclusions.The identified boundary is compositional clinical reasoning or grounded semantic labels that a linear probe cannot decode from a single layer.Source: §5 · pp. 8–9, Table 3 · p. 8

Sample efficiency and scaling

• ≈12 percentage points from 50 to 1,000 examples; Qwen3-VL average accuracy rises from 80.30% at 2B to 84.64% at 30B.All three models exceed 71% accuracy with 50 examples, and gains diminish above 500 examples.Source: §5 · pp. 5, 9

• MedProb decodes clinical answers from frozen general-purpose VLM representations with a small labeled set, outperforming medical VLMs and multi-agent pipelines across PATHVQA, SLAKE, and VQA-RAD.Source: §6 · p. 9
• Across 14 matched architecture pairs, domain-adaptive pretraining does not reliably enrich internal clinical representations; general models outperform medical counterparts in 6 of 14 cases, with RL-based MediX-R1 a notable exception.Source: §6 · p. 9
• MedProb is more robust to answer-option ordering than prompting, whose positional biases vary across models.Source: §6 · p. 9
• Appendix H shows rejection-sampling reranking of open-ended generations as an initial extension, while future work should explore this setting further.Source: §6 · p. 9

• MedProb requires labeled training examples, so it is not purely zero-shot, although it uses a lightweight linear classifier without VLM parameter updates.Source: Limitations · p. 10
• Evaluation focuses primarily on closed-ended benchmarks; most non-yes/no SLAKE and VQA-RAD questions remain effectively 2-way, with 4-way OmniMedVQA reported separately.Source: Limitations · p. 10
• Linear probes measure only what is decodable from a chosen representation; missed signal may be nonlinear, distributed across layers, or require richer aggregation.Source: Limitations · p. 10
• High probing accuracy does not establish that decoded directions cause model predictions or represent clinically meaningful knowledge.Source: Limitations · p. 10
• MedProb itself provides neither faithful clinical explanations nor evidence grounding; deployment-facing systems would also need calibrated uncertainty and clinical validation.Source: Limitations · p. 10

A Dataset Details

• Closed-ended questions were converted to multiple choice using alternatives in the question or trivial binary options, without external distractors or cross-question negatives.Source: A Dataset Details · p. 14
• Ambiguous option sets or questions whose gold answer was absent were excluded; SLAKE also used parsing fallback for residual comparison questions.Source: A Dataset Details · p. 14
• Datasets were capped at 500 test and 1,000 training examples, with stratified sampling; PATH-VQA test sampling preserved the full test set's yes/no balance.Source: A Dataset Details · p. 14
• Test subsets contain 500 PATH-VQA, 415 SLAKE, and 263 VQA-RAD examples; probe training uses 1,000, 1,000, and 820 examples, respectively.Source: A Dataset Details · p. 14

B Training process of Med-VLMs

• Training corpora range from large-scale PubMed Central image-caption pairs to GPT-4-generated instruction examples and reinforcement-learning signals aimed at clinical reasoning.Source: B Training process of Med-VLMs · pp. 14–15

C.1 Reproducibility of Prompting Details

• Prompting details cover system prompts, closed-ended multiple-choice formats, and generation sampling, documenting reproduction choices where official evaluation protocols are absent.Table 7 summarizes which details are fully available, partially available, or unavailable.Source: §C.1 · p. 15, Table 7 · p. 17

C.2 Default Prompt Formats

• Gemma 3 and MedGemma use greedy decoding with a 200-token cap and no system prompt; their prompts place the question and options before a single image.MedGemma’s format is based on its fine-tuning notebook because no dedicated closed-ended evaluation protocol is publicly released (Sellergren et al., 2026); Gemma 3’s format follows that notebook (Team et al., 2025).Source: §C.1 · pp. 15–16, §C.2 · p. 20

C.1 Reproducibility of Prompting Details

• LLaVA-V0 and LLaVA-Med use VICUNA-V0 conversational templates and deterministic single-beam decoding, while undocumented evaluation choices and inferred answer formatting limit reproducibility.LLaVA-Med prompt selection searches formats and uses a model-tailored system prompt; OpenFlamingo and Med-Flamingo follow Jeong et al. (2024) generation settings, with Med-Flamingo’s closed-ended formatting only partially specified.Source: §C.1 · pp. 16–17
• BioMed-Llama, BioMed-Qwen, BioMed-InternVL, and InternVL3 use no system prompt, official multiple-choice formats, and temperature zero (Cheng et al., 2025; Zhu et al., 2025).Llama-3.2-11B-Vision instead uses a partially available format inferred from its image-token constraint and leaves sampling hyperparameters at defaults (Meta AI, 2024; Grattafiori et al., 2024).Source: §C.1 · pp. 17–18
• Qwen3-VL uses its documented multiple-choice format with temperature 0.7 and top_p 0.8, while Qwen2.5-VL and Qwen2-VL reuse those settings and format despite lacking dedicated protocols (Bai et al., 2025a; Bai et al., 2025b).These model families have no documented system prompt; Qwen2.5-VL and Qwen2-VL entries are marked not applicable for format or sampling in Table 7.Source: §C.1 · pp. 18–19
• MedVLThinker-RL uses documented system and multiple-choice prompts with temperature 0.6 and top_p 0.95, whereas MediX-R1’s structured reasoning prompt combines modality identification and answer reasoning in one user turn (Huang et al., 2026; Mullappilly et al., 2026).MediX-R1 uses temperature zero and top_p one, with the latter inactive under greedy decoding.Source: §C.1 · p. 19

C.2 Default Prompt Formats

• Prompt image-token placement differs by family, and the documented layouts follow each model’s recommended ordering because incorrect ordering can affect attention to visual input.Gemma 3 places the image after the question and options; LLaVA-V0 instead injects image patches at inference time rather than using a literal <image> placeholder.Source: §C.2 · pp. 19–20
• LLaVA-Med retains LLaVA-V0’s flat dialogue rendering and VICUNA-V0 structure, changing the system prompt to a medically adapted preamble.Image injection, option placement, and the assistant generation trigger follow LLaVA-V0 conventions.Source: §C.2 · p. 21
• OpenFlamingo and Med-Flamingo put the image before the question and use lettered options followed by “Answer:”; Med-Flamingo replaces the generic instruction with a medical-assistant preamble.The Med-Flamingo preamble tells the model to follow provided examples before answering.Source: §C.2 · p. 21
• BioMed-Llama-3.2-11B-Vision, BioMed-Qwen2-VL-2B, and BioMed-InternVL3-1B share an image-first format with “Question:” and “The choices are:” prefixes and Python-style option lists.BioMed-Qwen2-VL-2B’s prompt structure is identical to BioMed-Llama-3.2-11B-Vision’s; BioMed-InternVL3-1B follows that same format.Source: §C.2 · pp. 21–22
• Llama-3.2-11B-Vision places the image before the question because images attend only to subsequent text, then lists options separately and requests the answer letter only.Its prompt uses the model’s native special tokens and triggers generation with an assistant header.Source: §C.2 · p. 22
• Qwen3-VL, Qwen2.5-VL, Qwen2-VL, and InternVL3-1B use image-first prompts with lettered options and a closing instruction to select the correct answer.The Qwen models use an “Options:” block, while InternVL3-1B uses its own options formatting.Source: §C.2 · pp. 22–23
• MedVLThinker-RL (3B, 7B, and 32B) uses a reasoning-oriented prompt requiring <think> and <answer> tags, with the image before the question and lettered options.A dedicated system prompt instructs the model to reason before answering.Source: §C.2 · pp. 23–24
• MediX-R1 (2B and 30B) requires a modality tag, reasoning and answer tags, and embeds its system instructions in the user turn with the image after “Question:”.Its prescribed output has three steps: identify the image modality, provide reasoning in <thinking> tags, and give the answer in <answer> tags.Source: §C.2 · p. 24

D Few-Shot Probing Results

• Zero-shot probing outperforms 3-shot on average for 7 of 9 models; 3-shot helps Qwen3-VL-30B and Qwen2.5-VL-32B on VQA-RAD and PATH-VQA.The main-paper probing results use zero-shot prompts without target-benchmark examples.Source: D Few-Shot Probing Results · p. 25, Table 8 · p. 25
• For BioMed-Qwen2-VL-2B, 3-shot probing drops average performance to 30.43% from 61.61%, supporting zero-shot as the simpler, generally more effective setting.The authors suggest cross-domain in-context examples can disrupt the hidden-state structure used by the probe.Source: D Few-Shot Probing Results · p. 25

E Option-Likelihood Prompting and Statistical Significance

• 4.6 points: option-likelihood scoring improves average accuracy over free-text prompting (59.8% → 64.4%) across nine model–dataset pairs.Option-likelihood scoring selects the candidate answer with the highest conditional log-probability, avoiding generation artifacts.Source: E Option-Likelihood Prompting and… · p. 25, Table 9 · p. 26
• MedProb exceeds option-likelihood scoring by 14.2 points (64.4% → 78.6%), a gap too large to attribute to generation artifacts.Table 9 compares free-text prompting, option-likelihood scoring, and MedProb for three representative VLMs.Source: E Option-Likelihood Prompting and… · pp. 25–26, Table 9 · p. 26
• Eight of nine paired probing-versus-free-text comparisons are significant at p < 0.05; the exception is Qwen3-VL-8B on VQA-RAD (∆= +5.1, CI [−0.4, 11.0], p = 0.098).The exception uses a 263-example test set, which the authors note limits statistical power.Source: E Option-Likelihood Prompting and… · p. 26, Table 10 · p. 26

F Cross-Dataset Generalization

• PATH-VQA-trained probes show severe cross-domain collapse: InternVL3-1B reaches 57.35% on SLAKE and 57.41% on VQA-RAD, below the OmniMedVQA baseline’s 72.29% and 61.60%.The authors report this domain-locked pattern across all three models.Source: F Cross-Dataset Generalization · p. 26
• VQA-RAD-trained probes transfer to SLAKE (75.18% for InternVL3-1B; 71.33% for Qwen3-VL-2B), while OmniMedVQA training never matches in-domain performance.The authors associate the transfer with shared radiology visual concepts between VQA-RAD and SLAKE.Source: F Cross-Dataset Generalization · p. 26

G Binary vs. Multiclass Breakdown

• Probing outperforms prompting on binary yes/no items across PATH-VQA, SLAKE, and VQA-RAD, and on true 4-way OmniMedVQA items.The OmniMedVQA multiclass evaluation includes 400 examples, with 25% chance accuracy.Source: G Binary vs. Multiclass Breakdown · p. 27, Table 12 · p. 27, Table 13 · p. 27

H Rejection-Sampling with the Probe on Open-Ended Generations

• For true multiclass OmniMedVQA, the probe scores question-image-candidate inputs as correct or incorrect, then reranks multiple VLM-generated answers by score.The experiment uses 500 training and 100 test examples.Source: H Rejection-Sampling with the Probe on… · p. 27
• The rejection-sampling probe improves accuracy over single-sample no-probe generation for both models and narrows the gap to the sampled-candidate oracle.The oracle selects the correct candidate whenever one exists among the samples.Source: H Rejection-Sampling with the Probe on… · p. 27, Table 14 · p. 27
• This is an initial demonstration, not a complete open-ended Med-VQA solution; evaluation is restricted to true multiclass OmniMedVQA.The paper’s main results target explicit multiple-choice or multiclass Med-VQA.Source: H Rejection-Sampling with the Probe on… · p. 27

I Additional Layer-wise Probing Results

• All 10 final layers outperform zero-shot prompting in every reported model/dataset combination, while late-layer probing accuracy is stable.Table 15 reports mean, standard deviation, and range across the last 10 layers for two representative models.Source: I Additional Layer-wise Probing Results · pp. 27–28, Table 15 · p. 28

I.1 Gemma-4B vs. MedGemma-4B; I.2 Gemma-27B vs. MedGemma-27B

• MedGemma-4B gains a clearer deep-layer lead, while Gemma-27B and MedGemma-27B converge before MedGemma edges ahead late.The larger pair is more balanced, with Gemma-27B leading more often early.Source: §I.1 · p. 28, Figure 9 · p. 28, §I.2 · p. 28

I.3 Open-Flamingo-9B vs. Med-Flamingo-9B; I.4 InternVL3-1B vs. BioMed-InternVL3-1B

• Med-Flamingo-9B leads Open-Flamingo-9B across nearly all layers, while BioMed-InternVL3-1B leads early and middle before InternVL3-1B finishes stronger.The Flamingo advantage remains fairly stable across depth; the InternVL3 pattern reverses in the final third.Source: §I.3 · p. 28, Figure 11 · p. 28, §I.4 · p. 28

I.5 LLaVA-7B vs. LLaVA-Med-7B; I.6 LLaMA3.2-11B-Vision vs. BioMed-LLaMA3.2-11B

• LLaVA-7B and LLaVA-Med-7B nearly overlap, whereas BioMed-LLaMA3.2-11B holds a modest, persistent lead over its general-purpose counterpart.The LLaVA curves cross several times and peak at nearly the same depth and accuracy.Source: §I.5 · p. 28, Figure 13 · p. 29, §I.6 · p. 29

I.7 Qwen2-VL-2B vs. BioMed-Qwen2-VL-2B; I.8 Qwen2.5-VL-3B vs. MedVLThinker-RL-3B

• Qwen2-VL-2B leads early and middle before BioMed-Qwen2-VL-2B narrows the gap, while Qwen2.5-VL-3B leads MedVLThinker-RL-3B more clearly late.Qwen2-VL-2B retains the higher late-layer peak; Qwen2.5-VL-3B also attains the stronger peak.Source: §I.7 · p. 29, Figure 15 · p. 29, §I.8 · p. 29

I.9 Qwen2.5-VL-32B vs. MedVLThinker-RL-32B; I.10 Qwen3-VL-2B vs. MediX-R1-2B

• Qwen2.5-VL-32B leads MedVLThinker-RL-32B through most middle and late layers, and Qwen3-VL-2B opens a large, growing lead over MediX-R1-2B.The Qwen3-VL-2B pair has the largest gap among the appendix pairs.Source: §I.9 · p. 29, Figure 17 · p. 30, §I.10 · p. 30

I.11 Qwen3-VL-4B vs. MedMo-4B; I.12 Qwen3-VL-8B vs. MedMo-8B

• Qwen3-VL-4B leads for much of the network before MedMo-4B reaches a higher peak; at 8B, MedMo narrows but does not overturn Qwen3-VL’s broader lead.MedMo-8B nearly matches the general model at its best-performing depth.Source: §I.11 · p. 30, Figure 19 · p. 30, §I.12 · p. 30

I.13 Qwen3-VL-30B vs. MediX-R1-30B

• Qwen3-VL-30B remains ahead of MediX-R1-30B across almost the entire network, with the gap widening again in the deepest layers.This is a strong and stable advantage across the full depth.Source: §I.13 · p. 30

J.1 Where MedProb Improves Over Prompting; Example E-1; Example E-2

• In semantic-hallucination and rule-application cases, probes recover correct answers despite prompted reasoning that turns visual cues into false conclusions or misapplies a correct rule.The examples use layer-16 and layer-22 probes, respectively.Source: Example E-1 · p. 31, Example E-2 · pp. 31–32, §J.1 · p. 31

J.2 Where MedProb Fails to Improve; Example E-3; Example E-4

• On fine-grained ring-enhancement and bowel-pathology questions, prompting and probing can fail together, indicating perceptual limits rather than an elicitation-only problem.Two independently trained models fail on the ring-enhancement question; both methods also fail on the bowel-pathology example.Source: §J.2 · p. 32, Example E-3 · pp. 32–33, Example E-4 · p. 33

J.3 Where MedProb Underperforms Prompting; Example E-5; Example E-6

• Prompting can outperform probes on coarse tissue labels and compositional diagnosis because a single-layer linear decoder may not capture semantic categories or multi-step clinical inference.The example chain connects wall thickening to fibrosis, dilated chambers, and CHF.Source: §J.3 · p. 33, Example E-5 · p. 33, Figure 26 · p. 33

J.4 Summary

• Across the six examples, probing helps when generation adds spurious reasoning, prompting helps with compositional knowledge or grounded labels, and neither fixes perceptual limitations.Source: §J.4 · p. 34, Table 19 · p. 34

H Probing vs. Prompting Across Medical Imaging Modalities

• Across eight OmniMedVQA modalities, probing usually beats prompting; Qwen3-VL-30B averages 6.0% prompting versus 88.0% probing, an 82-point gap, with MedMo-8B and MediX-R1-30B reversing the trend.MedMo-8B has a 99.0% prompting average, while MediX-R1-30B has 76.0% prompting versus 65.0% probing.Source: H Probing vs. Prompting Across Medical… · pp. 34–35, Table 20 · p. 35

K Baseline Details; K.1 Multi-Agent Frameworks; K.2 Medical Vision-Language Models; K.3 Fine-Tuned Models

• Baselines include three multi-agent frameworks and four specialized medical VLMs; separate full fine-tuning evaluates two general-purpose models in three single-domain and one multi-domain settings.Source: §K.1 · p. 36, §K.2 · p. 36, §K.3 · p. 36

L.1 Nonlinear Probe (MLP); L.2 PATH-VQA Sampling Sensitivity

• An MLP probe does not consistently improve on logistic regression, while three resampled PATH-VQA subsets yield small accuracy variation and stable model rankings.For Qwen3-VL-8B, average accuracy changes from 81.4% to 81.6%; for Llama-3.2-11B-Vision, it changes from 81.3% to 80.9%.Source: §L.1 · p. 37, Table 21 · p. 37, §L.2 · p. 37

M Probing vs. Prompting Across All Models and Benchmarks

• On PATH-VQA, Med-Flamingo-9B scores 85.4% with probing versus 41.8% with prompting, a 43.6-point margin.BioMed-Llama-3.2-11B has the highest probing accuracy at 88.4%, while MedMO-4B-Next leads prompting at 69.4%.Source: M Probing vs. Prompting Across All… · p. 37
• On SLAKE, MedMO-4B-Next reaches 90.12% probing accuracy and 81.69% prompting accuracy, while Med-Flamingo-9B scores 79.52% and 36.87%, respectively.MedMO-8B-Next also reaches 90.12% probing accuracy; prompting scores are generally higher than on PATH-VQA.Source: M Probing vs. Prompting Across All… · p. 37
• On VQA-RAD, Qwen2.5-VL-32B scores 88.59% with probing and 82.51% with prompting, while MedMO-4B-Next is the sole reversal at 73.76% versus 76.05%.The probing-to-prompting gap narrows most for top-performing models on this benchmark.Source: M Probing vs. Prompting Across All… · pp. 37–38
• Across PATH-VQA, SLAKE, and VQA-RAD, probing consistently and substantially outperforms prompting across the evaluated models.The figures cover 24 of 28 evaluated models per benchmark, sorted by probing score.Source: M Probing vs. Prompting Across All… · p. 37, Figure 28 · p. 38, Figure 29 · p. 38

N MedGemma-27B Results (Omitted from Table 1 for Space)

• MedGemma-27B is omitted from Table 1 for space but included in Figure 3 as Gemma-27B’s matched medical counterpart, with full results in Table 23.The text says its pattern follows the other models in Table 1.Source: N MedGemma-27B Results (Omitted from… · p. 38, Table 23 · p. 38

O Use of AI Assistants

• Figure 30 compares probing and prompting accuracy on VQA-RAD for 24 of 28 evaluated models, sorted by probing score.This figure caption identifies the comparison and ordering, but does not specify individual model values.Source: Figure 30 · p. 39