Paper deep dive
See, Measure, and Reason: Learning Visually Grounded Reasoning in Pathology
Chengyang Zhang, Wenchuan Zhang, Bo Li, Mengran Li, Xinyu Liu, Jiaming Yang, Jie Chen, Zhang Zhang, Yuhao Yi, Hong Bu, Jiancheng Lv
Intelligence
Status: not_run | Model: - | Prompt: - | Confidence: 0%
Entities (0)
Relation Signals (0)
No relation signals yet.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Pathological assessment relies on recognizing fine-grained visual details in histological images. Vision-language models (VLMs) increasingly support pathology interpretation, yet their ability to perceive these details remains inadequate. This weakness leads to inaccurate cellular observations that can persist even when final answers are correct. In this paper, we propose ASPECT to improve visually grounded reasoning through explicit supervision of cellular appearance and abundance. ASPECT trains intermediate visual tokens through pathology feature reconstruction, cell feature alignment, and count supervision. Three-stage supervised fine-tuning teaches the model to perceive, generate visual tokens, and reason, followed by reinforcement learning that rewards answer correctness and consistency with reported measurements. We also introduce PathoVernier, a benchmark of 759 expert-reviewed questions from five pathology datasets covering four cellular composition tasks. It evaluates both final answers and intermediate measurements to expose errors hidden by answer accuracy. On PathoVernier, ASPECT achieves relative accuracy gains of approximately 19.2% over the strongest baseline, Gemini-3.1-Pro, and 99.3% over its Qwen3-VL-8B backbone, while reducing RAWR, which measures counting errors within correct responses, by 28.1% and 42.7%, respectively. ASPECT also improves over its backbone on three external pathology benchmarks covering classification and question answering beyond cellular composition tasks.
Tags
Links
- Source: https://arxiv.org/abs/2609.34277v1
- Canonical: https://arxiv.org/abs/2609.34277v1
Trouble viewing inline? Open PDF directly →
Full Text
141,586 characters extracted from source content.
Expand or collapse full text
SEE, MEASURE, AND REASON: LEARNING VISUALLY GROUNDED REASONING IN PATHOLOGY Chengyang Zhang 1,2 , Wenchuang Zhang 2 , Bo Li 3 , Mengran Li 4 Xinyu Liu 1 , Jiaming Yang 1 , Jie Chen 2 , Zhang Zhang 2 Yuhao Yi 1,2∗ , Hong Bu 2 , Jiancheng Lv 1 1 College of Computer Science, Sichuan University 2 Department of Pathology and Institute of Clinical Pathology, West China Hospital, Sichuan University 3 Department of Computer Science, School of Computing, National University of Singapore 4 School of Intelligent Systems Engineering, Sun Yat-sen University yuhaoyi@scu.edu.cn ABSTRACT Pathological assessment relies on recognizing fine-grained visual details in his- tological images. Vision-language models (VLMs) increasingly support pathol- ogy interpretation, yet their ability to perceive these details remains inadequate. This weakness leads to inaccurate cellular observations that can persist even when final answers are correct. In this paper, we propose ASPECT to improve visu- ally grounded reasoning through explicit supervision of cellular appearance and abundance. ASPECT trains intermediate visual tokens through pathology feature reconstruction, cell feature alignment, and count supervision. Three-stage super- vised fine-tuning teaches the model to perceive, generate visual tokens, and rea- son, followed by reinforcement learning that rewards answer correctness and con- sistency with reported measurements. We also introduce PathoVernier, a bench- mark of 759 expert-reviewed questions from five pathology datasets covering four cellular composition tasks. It evaluates both final answers and intermediate mea- surements to expose errors hidden by answer accuracy. On PathoVernier, AS- PECT achieves relative accuracy gains of approximately 19.2% over the strongest baseline, Gemini-3.1-Pro, and 99.3% over its Qwen3-VL-8B backbone, while reducing RAWR, which measures counting errors within correct responses, by 28.1% and 42.7%, respectively. ASPECT also improves over its backbone on three external pathology benchmarks covering classification and question answer- ing beyond cellular composition tasks. 1INTRODUCTION Pathology vision-language models (VLMs) can describe tissue morphology and answer diagnostic questions from histological images (Lu et al., 2024; Seyfioglu et al., 2024; Sun et al., 2025; Zhang et al., 2026c). However, success in question answering does not establish how accurately a model identifies the visual details needed for its answers. Recent pathology evaluations reveal difficulties with localization, counting, and spatial relationships, even on slides where models answer higher- level questions correctly (Chen et al., 2026; Zhang et al., 2026a). These findings motivate a closer examination of the visual understanding underlying pathology VLM predictions. General-purpose VLM evaluations examine fine-grained perception through spatial judgments and counting (Fu et al., 2024; Rahmanzadehgervi et al., 2024; Zhou et al., 2025). In pathology, cellu- lar composition analysis requires these abilities and supports subsequent tissue assessment. Tumor content informs specimen suitability for molecular testing (Lindeman et al., 2018), while assessment of tumor-infiltrating lymphocytes requires identifying inflammatory cells within the appropriate tis- sue compartment (Salgado et al., 2015). Distinguishing cell types and comparing their abundance across densely populated regions therefore provide a clinically relevant test of visual understanding. ∗ Corresponding author 1 arXiv:2609.34277v1 [cs.CV] 28 Sep 2026 (a)Incorrect answer, incorrect observations (b) Correct answer, incorrect observations QUESTION: What fraction represents the concentration of epithelial-like nuclei in the upper-left quadrant divided by their concentration in the upper- right quadrant? Use these bands: 'much more' if the ratio is at least 2.0, 'much fewer' if it is at most 1/2.0, otherwise 'comparable’. GT COUNTS: "epithelial_like_upperleft": 16, "epithelial_like_upperright": 4 GOLD: much more OPTIONS: ‘muchmore’, ‘muchfewer', 'comparable' QUESTION: What is the ratio of epithelial-like nucleus concentration present within the upper-right quadrant to that within the lower-left quadrant? Use these bands: ‘muchmore’ if the ratio is at least 2.0, ‘muchfewer' if it is at most 1/2.0, otherwise 'comparable’. GT COUNTS: "epithelial_like_upperright": 6, "epithelial_like_lowerleft": 27 GOLD: muchfewer OPTIONS: ‘muchmore’, ‘muchfewer', 'comparable' I divided the patch into quadrants and counted only plump/oval epithelial-like nuclei ...... The upper-right quadrant has visibly more epithelial-like nuclei than the lower-left quadrant. <observe>......,"epithelial_like_upperright": 6, "epithelial_like_lowerleft":2</observe> Answer: muchmore <think>The image shows epithelial-like nuclei concentrated in the upper-right quadrant ...... In the upper-right quadrant, I observe approximately 40 nuclei per HPF, while the lower-left quadrant contains roughly 20 nuclei per HPF. This gives a ratio of 40:20 (2.0:1), which falls into the 'much more' category. ......<think> Answer: much more Patho-R1-7B GPT-5.5 UL UR LLLR Patho-R1-7B <think>The image shows epithelial-like nuclei in the upper-left quadrant (ELUPL) and upper-right quadrant (ELUUR) ...... there are approximately 15 nuclei, while the upper-right quadrant has 10 nuclei ...... Since 1.5 exceeds 1.0, this falls under the ‘much more’ category ...... <think> Answer: muchmore GPT-5.5 I compared the two equal-area upper quadrants and counted epithelial-like nuclei visible in each. The upper-left quadrant contains a dense epithelial cluster, while the upper-right quadrant...... <observe>......,"epithelial_like_upperleft":32 , "epithelial_like_upperright":4</observe> Answer: muchmore Incorrectcount Incorrectcount Incorrectcount Incorrectcount, rule mismatch UL UR LLLR Figure 1: Correct answers can conceal inaccurate observations. GPT-5.5 and Patho-R1 reverse regional nuclear densities in (a) and answer correctly despite incorrect counts in (b), where Patho- R1 also misapplies the ratio rule. Green contours mark reference nucleus annotations. Nucleus instance annotations supply reference identities and positions, allowing both the reported measurements and the conclusions drawn from them to be checked. Our evaluation reveals errors in both observing cellular composition and using those observations to answer questions. Figure 1 illustrates these failures in GPT-5.5 (OpenAI, 2026a) and Patho-R1 (Zhang et al., 2026c), representing a frontier closed-source VLM and a pathology-tuned model. In Figure 1(a), both models reverse the relative epithelial-like nuclear densities of two regions. In Fig- ure 1(b), both select the correct answer despite inaccurate counts; Patho-R1 additionally selects a category inconsistent with the ratio implied by its own counts. Thus, a correct answer can hide both inaccurate observation and incorrect use of the stated evidence. We hypothesize that fine-tuning fo- cused on answer generation can strengthen pathology-specific language priors without comparable gains in extracting and using visual information. Consistent with this hypothesis, pathology-tuned models retain substantial accuracy without images, and their gains over base models need not co- incide with better visual grounding (Zhang et al., 2026a). Similar image-independent performance has been reported on broader medical multimodal benchmarks (Asadi et al., 2026). These find- ings motivate evaluating the observations behind an answer and training models to make accurate observations and use them correctly. To connect fine-grained perception with pathology reasoning, we introduce ASPECT and PathoVernier. ASPECT supervises intermediate visual tokens with pathology features and cell counts, using three-stage supervised fine-tuning (SFT) to generate quantitative observations and reason from them. Reinforcement learning (RL) further rewards answer correctness and consistency with the reported counts. PathoVernier provides 759 expert-reviewed questions from five datasets spanning more than 15 organs. Four cellular composition tasks pair each answer with reference counts and a decision rule, allowing evaluation beyond final-answer accuracy. Our main contribu- tions are summarized as follows: • We introduce PathoVernier to assess fine-grained pathology perception through four cel- lular composition tasks. Its RAWR metric (right answer, wrong reason) exposes counting errors hidden by correct answers. • We propose ASPECT, which trains intermediate visual tokens to encode pathology appear- ance and cell abundance through three-stage SFT, followed by RL for answer correctness and consistency with reported counts. 2 • We conduct experiments against ten baselines. ASPECT-8B leads on all PathoVernier met- rics, improving accuracy by 19.2% and reducing RAWR by 28.1% relative to the strongest baseline. Gains on three external benchmarks extend beyond cellular composition tasks. 2RELATED WORK Pathology Vision-Language Models Pathology VLMs combine domain-specific visual represen- tations with language supervision to describe histological findings and answer diagnostic questions (Lu et al., 2024; Seyfioglu et al., 2024; Sun et al., 2025). PathChat combines a pathology vision encoder with language instruction tuning (Lu et al., 2024). Reasoning-oriented models subse- quently introduce chain-of-thought supervision and reinforcement learning (Zhang et al., 2026c). More recently, PathReasoner-R1 uses knowledge-guided reasoning data and entity-based rewards to strengthen the connection between pathological findings and diagnoses (Jiang et al., 2026). Path- Found seeks additional evidence to refine diagnoses (Hua et al., 2026). RECAP-PATH optimizes prompts for morphological descriptions using diagnostic feedback (Hong et al., 2026). However, diagnostic accuracy and plausible explanations do not establish whether reported cell counts are correct or support a quantitative conclusion. Evaluating Visual Understanding Visual-understanding benchmarks assess fine-grained percep- tion, visual reasoning, and domain-specific knowledge (Fu et al., 2024; Sun et al., 2024; Chen et al., 2024b). General-domain evaluations isolate perceptual operations, whereas pathology benchmarks emphasize morphological recognition and diagnostic reasoning. For example, BLINK evaluates abilities such as visual correspondence, relative depth estimation, and multi-view reasoning (Fu et al., 2024), while PathMMU assesses expert-level pathology understanding through question an- swering (Sun et al., 2024). More targeted pathology evaluations examine image dependence, en- tity–region correspondence, and visual understanding across local regions and whole-slide views (Chen et al., 2026; Zhang et al., 2026a). However, these pathology evaluations largely assess visual abilities through separate task outputs, without jointly checking a final answer and its observations. Visual Supervision and Process Verification Visual supervision and process verification provide complementary learning signals beyond final-answer correctness (Qin et al., 2025; Xiao et al., 2026; Li et al., 2025; Pronesti et al., 2026). Existing approaches supervise intermediate visual represen- tations or assess a model’s observations and reasoning steps. For example, CoVT trains continuous visual tokens to reconstruct features from vision experts (Qin et al., 2025). More recently, PEARL constructs verifiable perception questions for each reasoning task and uses auxiliary perception roll- outs to guide subsequent reasoning updates (Zhang et al., 2026b). However, feature reconstruction does not verify the numerical observations stated in a response, and PEARL checks perception through separately generated answers. These objectives do not directly check whether the reported measurements are accurate and support the conclusion within the same reasoning response. 3METHODS 3.1OVERVIEW ASPECT learns to perceive pathology images, report quantitative observations, and reason from them to an answer (Figure 2). Built on Qwen3-VL-8B (Bai et al., 2025), it generates eight pathol- ogy feature tokens and six cell tokens before producing the textual observations and answer. We supervise their hidden states through pathology feature reconstruction, cell feature alignment, and count prediction. Three-stage SFT connects these representations to observation and answer gen- eration; RL then rewards correct answers supported by the reported measurements. PathoVernier evaluates this connection through 759 questions spanning four quantitative tasks and more than 15 organs, with reference measurements for every answer. 3.2PATHOLOGY VISUAL SUPERVISION Cell composition analysis requires both recognizing cell appearance and measuring abundance. Su- pervising intermediate visual tokens can retain image information within autoregressive generation 3 Overview of ASPECT Three-Stage SFTRL for Answer-Observation Consistency Question: Relative to the bottom half, what is the density of epithelial-like nuclei in the top half? ASPECT <think> ×8 ...... <uni> ×6 </think> <cell> the patch feature of this image is ProjectionCount head Decoder CellViT++ Encoder Annotated Counts the cell composition of this image is ℒ !"#$% ℒ &'(%) ℒ *+ ℒ *+ ℒ (%# UNI Encoder <observe> </observe> ... <answer> </answer> ... A moderately cellular field shows cohesive epithelial-lined ...... "epithelial_like": 53, "asked": "epithelial-like nuclei, compared between the top half and the bottom half", "epithelial_like_top": 0, "epithelial_like_bottom" : 53 The top half has 0 epithelial-like nuclei compared with 53 in the bottom half. Since the areas are equal, the density ratio is 0/53 = 0.00, substantially fewer relative to the 2.0x threshold. FINAL: muchfewer Stage1 (perceive) Stage2 (generate) Stage3 (reason) Stage Input Output <image> + <uni> + <cell> + question <image> + <uni> query + <cell> query <image> + question <answer> <uni> +<cell> <think> + <observe> + <answer> ASPECT-SFT ASPECT policy RLDataset 1.3k hard samples Rollouts 푂 " 푂 # 푂 $ 푂 % ... Initialize ... Acc: answer correctness Con: answer-count consistency Reward Vs: valid response structure R = Vs(Acc + 0.5Con) - 0.1(1 - Vs) GRPO ℒ *+ ℒ *+ Figure 2: Overview of ASPECT. Pathology feature reconstruction, cell feature alignment, and count supervision train the visual-token representations. Three-stage SFT teaches the model to perceive, generate, and reason with explicit measurements. Subsequent RL rewards answer correctness and consistency with the reported counts. (Qin et al., 2025; Li et al., 2026; Yang et al., 2026). For composition analysis, these representa- tions must also encode cell identity and abundance. We therefore pair pathology feature reconstruc- tion with cell feature alignment and count supervision, connecting intermediate visual states to the quantitative observations needed for answering. In the complete response, the two token groups, abbreviated as <uni> and <cell>, appear in the <think> block, followed by observations in <observe>, and reasoning with a final answer in <answer>. Let H u ∈ R 8×d and H c ∈ R 6×d denote the final-layer hidden states at the two groups, where d is the language model’s hidden di- mension. The following losses train these states to encode the visual information used in subsequent observation and answer generation. Pathology feature reconstruction. To preserve pathology appearance in the <uni> states, we reconstruct features F UNI (x) extracted from image x by a frozen UNI encoder (Chen et al., 2024c). A cross-attention decoder takes learned queries U and the projected token states as inputs: b F u = Dec u U, norm(H u W u ) , L uni = b F u − F UNI (x) 2 F . (1) Here, b F u denotes the reconstructed features, W u is a learned projection, norm applies L 2 normal- ization, and∥·∥ F is the Frobenius norm. The cross-attention decoder uses shared learned queries U and the projected states as keys and values, so image-specific information must pass through H u . Cell feature alignment. With guidance from pathologists, we map source annotations to five tar- get categories: tumor cells, normal epithelial cells, stromal-like cells, lymphocytes, and other in- flammatory cells. Finer subtypes within a target category are merged; coarse labels spanning several targets retain their group membership. The first five <cell> tokens represent these categories, and the sixth summarizes all retained, mappable instances. A frozen CellViT++ model (H ̈ orst et al., 2026) supplies instance features u j ∈ R d T , where d T is the teacher feature dimension. We pool features within annotated masks or predicted instances; NuCLS contributes counts but no alignment targets. For token s∈1,..., 6, letJ s index its contributing instances. For a non-empty group, its target t s averages individually normalized features: u j = u j ∥u j ∥ 2 ,t s = 1 |J s | X j∈J s u j .(2) 4 An instance with a coarse label contributes to each corresponding category target: an inflammatory instance, for example, supervises both lymphocyte and other-inflammatory features. The sixth token pools each retained instance once; excluded labels are listed in Appendix A.2. We align token states to these targets through L align = 1 d T |V| X s∈V ∥W c (h s + r s )− t s ∥ 2 2 ,(3) where h s ∈ R d is the transposed s-th row of H c , r s ∈ R d is a learned role embedding, and W c ∈ R d T ×d is a learned projection. The setV =s :|J s | > 0 excludes empty instance groups. Count supervision. Mean pooling summarizes appearance without explicitly preserving instance counts. We therefore train the cell-token states to predict the image-wide abundance of the five target categories. An MLP g predicts the five image-wide category counts in log space, ℓ = g(vec(H c ))∈ R 5 , where vec concatenates all six token states. For aggregation, we recover nonnegative counts as ˆn k = max0, exp(ℓ k )− 1. LetD index categories with individually resolved counts n k , and letG contain groups G⊆1,..., 5 with annotated combined counts n G . We optimize L count = X k∈D SmoothL1 ℓ k , log(1 + n k ) + X G∈G SmoothL1 log 1 + X k∈G ˆn k ! , log(1 + n G ) ! . (4) Merged fine labels supervise individual target counts, whereas coarse labels supervise their com- bined abundance. For example, a coarse inflammatory annotation aligns both inflammatory token representations while supervising the sum of lymphocyte and other-inflammatory counts. The frozen teachers and visual-supervision modules are used only during SFT. The count head super- vises image-wide abundance; the language model learns to generate regional measurements through response-text supervision. At inference, it generates visual tokens, observations, and answers au- toregressively. Non-cellular training samples use only the <uni> group before answer generation. 3.3SUPERVISED FINE-TUNING The visual tokens must first acquire useful image representations, then be generated as part of a response and used to support an answer. We organize SFT into three stages, as shown in Figure 2. In Perceive, the image, question, and visual tokens are supplied in the user prompt. Visual supervision shapes the token states while the model learns to answer from this context. In Generate, the model predicts the visual-token block from the image and a feature query, learning to produce the representations that were previously supplied in the prompt. In Reason, it generates the complete response: visual tokens in <think>, measurements in <observe>, and reasoning with a final answer in <answer>. This stage trains the representations, reported measurements, and answer within the same generation sequence. Cross-entropy covers the entire assistant target and masks the user prompt, supervising visual-token generation in stages two and three. Visual losses apply throughout: L SFT =L LM + λ u L uni + λ a L align + λ c L count ,(5) where L LM is assistant-token cross-entropy and the λ coefficients weight the three visual losses. A visual loss is zero when its token group or supervision is unavailable. Training updates LoRA parameters, visual-token parameters, and visual-supervision modules. For quantitative examples, programs derive counts and reference answers from nucleus annotations; a language model sup- plies descriptions and reasoning, checked against those values. Data, training details, and response templates appear in Appendices A, B, and D.3. 3.4REINFORCEMENT LEARNING FOR ANSWER–OBSERVATION CONSISTENCY SFT provides consistent reference responses, but inference requires reasoning from the model’s own estimated counts. An answer-only reward can accept a correct answer that contradicts those counts. We therefore train the model on its generated responses, rewarding both answer correctness and agreement with its observations. 5 Colon 480 Organ Origin Magnification Task Skin 142 Breast 60 Others 77 Lizard 400 CoNSeP 44 PUMA 142 NuCLS 34 PanNuke 139 20x 400 40x 359 Region selection 190 Region comparison 186 Cell-type comparison 193 Multi-step composition 190 (a) Statistics of PathoVernier (b) Samples of PathoVernier What is the comparative density of epithelial-like nuclei found in the left half versus the right half? Use these bands: 'much more' if the ratio is at least 2.0, 'much fewer' if it is at most 1/2.0, otherwise 'comparable'. “epithelial_like_left”: 5, “epithelial_like_right”: 33Answer: much fewer What is the ratio of lymphocytic nucleus concentration present within the upper-left quadrant to that within the lower-left quadrant? Use these bands: 'much more' if the ratio is at least 2.0, 'much fewer' if it is at most 1/2.0, otherwise 'comparable'. "lymphocyte_upperleft": 0, "lymphocyte_lowerleft": 10Answer: much fewer epithelial like stromal likeinflammatory Figure 3: PathoVernier overview. (a) Distribution of 759 questions across organs, sources, magni- fications, and tasks; flow widths represent question counts. (b) Example questions with reference counts. Answer correctness and count consistency. For question u, let Q denote reference counts and a ⋆ = f u (Q) the complete reference answer. A response reports counts C and answer ˆa. Let g u denote the count-based task check and π u the corresponding answer projection. We define Acc = 1[ˆa = a ⋆ ],Con = 1[g u (C) = π u (ˆa)], R = V s Acc + β Con − γ(1− V s ). (6) Acc checks the complete answer, whereas Con compares the count-derived decision with the corre- sponding part of the model’s own answer. For multi-step composition, Con checks region selection only. We set Con = 0 when either side is undeterminable. The coefficients β and γ weight the con- sistency bonus and invalid-response penalty. The structural indicator V s requires parseable observa- tion JSON, nonnegative counts and a parseable final answer. Answer correctness and consistency are evaluated with deterministic rules. Policy optimization. Starting from ASPECT-SFT, we optimize the policy with GRPO (Shao et al., 2024). We sample a group of responses for each prompt and normalize their rewards within the group to obtain relative advantages for the clipped policy update. LoRA parameters in both the visual and language components are updated. Appendices A.4 and B.3 detail the RL data and optimization settings. 3.5PATHOVERNIER BENCHMARK Data statistics and task coverage. PathoVernier evaluates whether pathology VLMs can answer questions based on accurate cell measurements. As shown in Figure 3(a), it contains 759 expert- reviewed questions across different organs, with images acquired at approximately 20× and 40× magnification. The images originate from five datasets with nucleus-level annotations: Lizard (Gra- ham et al., 2021), PUMA (Schuiveling et al., 2025), PanNuke (Gamper et al., 2020), CoNSeP (Gra- ham et al., 2019), and NuCLS (Amgad et al., 2022). The questions are approximately balanced across four tasks. Region selection identifies the region with the highest density of a specified cell type. Region comparison compares one cell type across two regions, whereas cell-type comparison compares two types within the same image. Multi-step composition analysis requires selecting the densest region and then determining a cell proportion category within it. These tasks assess cell recognition, spatial localization, and quantitative reasoning, with reference measurements accompa- nying every answer. Question construction and annotation. We filter source images for task suitability and balance while keeping them disjoint from training images. We then map the original nucleus labels to the cell categories used in the questions and compute reference counts from the nucleus annotations. 6 Table 1: Comparison on PathoVernier and three external pathology benchmarks. ASPECT leads on all PathoVernier metrics and improves over Qwen3-VL-8B on all external benchmarks. Bold and underlined values denote the best and second-best results, respectively. Dashes indicate CA coverage below 30%; Table C.11 reports coverage and unconditional Count Acc. Models PathoVernierPathCLSPathVQAQuilt-VQA Acc↑CA↑RAWR↓Acc↑Acc↑Acc↑ Closed-source Models GPT-5.5 (OpenAI, 2026a)0.590.650.760.550.690.69 Gemini-3.1-Pro (Google DeepMind, 2026)0.630.710.680.670.780.65 Open-source Models Qwen3-VL-8B (Bai et al., 2025)0.370.530.850.370.670.58 Qwen3-VL-32B (Bai et al., 2025)0.440.500.820.450.650.62 InternVL3-8B (Zhu et al., 2025) 0.300.380.860.390.690.64 InternVL3-38B (Zhu et al., 2025)0.450.450.830.410.710.64 Medical and Pathology-tuned Models Quilt-LLaVA-7B (Seyfioglu et al., 2024)0.12--0.240.540.64 PathGen-LLaVA-13B (Sun et al., 2025) 0.12--0.520.620.66 HuatuoGPT-Vision-7B (Chen et al., 2024a)0.30--0.340.640.55 Patho-R1-7B (Zhang et al., 2026c)0.22--0.440.560.68 ASPECT-8B(ours)0.750.810.490.540.740.67 ∆ (vs base model Qwen3-VL-8B)+0.38+0.28-0.36+0.17+0.07+0.09 For a regionR and a cell category c, the annotated count is q R,c = X j∈J c 1[p j ∈R],(7) where J c indexes retained nuclei belonging to category c, including its constituent classes for ag- gregate categories, and p j is nucleus j’s centroid. Four templates convert the required counts Q and rule-derived answers a ⋆ = f u (Q) into 786 candidate questions. GPT-5.6-sol (OpenAI, 2026b) rewrites their wording while preserving numerical rules and answers. Experts review each question against its image and annotation overlay, retaining 759 valid items (Figure 3(b)). Appendices A.5 and D.2 detail filtering, rewriting, and expert review. Evaluation metrics. We report answer accuracy Acc = E[1[ˆa = a ⋆ ]] and two measures of the reported observations. All expectations denote empirical averages over eligible responses. Let K index a question’s required counts, q i be its annotated count, and dom(C) index reported counts C i . The fraction of required counts within tolerance is S(C,Q) = 1 |K| X i∈K∩dom(C) 1 |C i − q i |≤ τ i , τ i = max(1, 0.1q i ).(8) Each count contributes equally; missing counts contribute zero. We adapt coherent accuracy (CA) from VPRM (Pronesti et al., 2026) to compare a count-derived task decision with its reference: CA = E[1[g u (C) = π u (a ⋆ )]|g u (C) is defined].(9) Here, g u and π u denote the task check and answer projection defined in Appendix C.2. Additionally, right answer, wrong reason (RAWR) measures the fraction of required counts outside tolerance among correct responses with complete measurements: RAWR = 1− E[S(C,Q)| ˆa = a ⋆ , K⊆ dom(C)].(10) Lower RAWR indicates fewer out-of-tolerance counts. Appendix C.2 specifies eligibility for both conditional metrics. 4EXPERIMENTS Performance on PathoVernier. Table 1 shows that ASPECT leads on all three PathoVernier met- rics, with relative gains of approximately 19.2% in Acc and 12.9% in CA, and a 28.1% reduction 7 Question: Determine the region whose lymphocytic nuclear density is greatest, then state the proportion of that region's nuclei that are lymphocytic. For the share, use these bands: 'low' below 0.35, 'high' at or above 0.55, otherwise 'mid’. GT counts: 'upperleft': 2, 'upperright': 0, 'lowerleft': 2, 'lowerright’: 7 GT lower-right region reference: lymphocyte share = 7/29 ≈ 0.24 Gold answer: lowerright: low To determine the region with the highest lymphocytic nuclear density, we first divide the image into four quadrants and count the lymphocytes (small, dark, round nuclei) in each. <observe>`'tumor': 110.0, 'lymphocyte': 23.0, 'upperleft': 1.0, 'upperright': 2.0, 'lowerleft': 8.0, 'lowerright': 12.0`</observe> FINAL: lowerright: low Gemini-3.1-Pro <answer> The upper right quadrant has the highest lymphocytic density (>50% of nuclei). Proportion of lymphocytes in this region is high (likely >70% based on visual estimation). </answer> <observe> ("upperleft": 0.35, "upperright": 0.55, "lowerleft": 0.55, "lowerright": 0.55) </observe> FINAL: upperright: high Source: PathoVernier Qwen3-VL-8B <think> the patch feature of the image is <uni> and the cell composition of the image is <cell></think><observe> "stromal_like": 17, "lymphocyte": 10, "epithelial_like": 46, "asked": "lymphocytic nuclei", "upperleft": 1, "upperright": 1, "lowerleft": 1, "lowerright": 7, "_num": 7, "_den": 29</observe> <answer> Lymphocytic nuclei are most concentrated in the lower-right quadrant, with 7 compared with 1, 1, 1 in the other regions.This quadrant contains 29 nuclei overall, yielding a low lymphocytic share of 7/29 = 0.24. FINAL: lowerright: low </answer> Lymphocytic nuclei: I can count approximately 15-20 small, dark, round nuclei. In the same region, I can count approximately 40-50 total nuclei. upperleft: 0, upperright: 0, lowerleft: 0, lowerright: 20 (lymphocytic nuclei) / 50 (total nuclei) = 0.4 → 'mid’. FINAL: lowerright: mid Question: This is an image of a skin histology sample. Which category can be determined based on observations made from the image? Gold answer: Tumor epithelial squamous cell carcinoma The image displays a highly cellular neoplasm characterized by cells with marked nuclear pleomorphism, large vesicular nuclei, and prominent nucleoli. These cytological features, in the absence of clear squamous di#erentiation such as keratinization or intercellular bridges, are highly characteristic of melanoma. FINAL: Tumor melanoma Gemini-3.1-Pro <think>The image shows a well-organized arrangement of stratified squamous epithelium with intact basal, spinous, granular, and cornified layers...... The dermal- epidermal junction is preserved, and the stroma lacks fibrosis or inflammation. These features align with normal epidermal structure...... </think> FINAL: Non-tumor sweat glands Patho-R1-7B Source: PathCLS Qwen3-VL-8B <think> the patch feature of the image is <uni>, and the cell composition of the image is <cell>. </think><observe> "tumor": 38, "tumor_benign": 38, "stromal_like": 1, "lymphocyte": 1, "asked_tumor": "Yes"</observe> <answer> The composition shows tumour=38, indicating a tumor-cell field. That exclusive dominance supports classification in the tumor category rather than any non-tumor skin structure, matching Tumor epithelial squamous cell carcinoma. FINAL: Tumor epithelial squamous cell carcinoma </answer> The most likely interpretation is that this is a **tumor epithelial basal cell carcinoma**, as the cells are small, round to oval, with hyperchromatic nuclei and some mitotic activity, and they appear to be invading the surrounding stroma — consistent with basal cell carcinoma histology, especially in a skin sample. FINAL: Tumor epithelial basal cell carcinoma Patho-R1-7B ASPECT-8B ASPECT-8B Figure 4: Qualitative comparison on PathoVernier (left) and PathCLS (right). Responses are con- densed and reformatted to highlight key observations and predictions. ASPECT reports more accu- rate regional counts on PathoVernier and predicts the reference class on PathCLS. in RAWR over Gemini-3.1-Pro, the strongest baseline. ASPECT thus improves both task deci- sions and the quantitative observations underlying them. The reduction in RAWR directly addresses our motivating observation: even among correct responses, ASPECT’s reported counts more of- ten match the annotated measurements. In contrast, all medical- and pathology-tuned baselines fall below Qwen3-VL-8B in PathoVernier accuracy and have CA coverage below 30%, so we omit their CA and RAWR. Notably, PathGen-LLaVA and Patho-R1 outperform Qwen3-VL-8B on Path- CLS yet underperform it on PathoVernier. This ranking reversal shows that conventional pathology question-answering performance does not guarantee reliable quantitative image understanding. Ap- pendices E.1–E.4 report task- and source-level results, confidence intervals, and tolerance sensitivity. Performance on external pathology benchmarks. ASPECT also improves over its Qwen3-VL- 8B backbone on all three external pathology benchmarks, with relative accuracy gains of approxi- mately 45.9% on PathCLS (Sun et al., 2024), 10.4% on PathVQA (He et al., 2020), and 15.5% on the pathology-image subset of Quilt-VQA (Seyfioglu et al., 2024). The gains on classification and question answering show that training for cellular composition also benefits pathology tasks that do not explicitly request measurements. Qualitative comparison. Figure 4 summarizes predictions from four models. On PathoVernier, Gemini and ASPECT give the same correct answer, but Gemini overestimates the lower-right lym- phocyte count (12 versus 7). ASPECT’s regional counts fall within tolerance, and its reported pro- portion, 7/29≈ 0.24, correctly supports the low category. Qwen3-VL-8B selects the correct region but misjudges the proportion; Patho-R1 errs on both steps. This contrast shows how answer accuracy can conceal measurement errors. On PathCLS, ASPECT predicts the reference class while all three baselines misclassify the image, illustrating a classification example beyond composition analysis. Full responses for these and additional cases are provided in Appendix F. Ablation on visual supervision and training curriculum. Table 2 ablates visual supervision and the SFT curriculum on PathoVernier. Count Acc averages S(C,Q) over all questions, while RAWR measures count errors only among correct answers with complete measurements. ASPECT- SFT achieves 0.44 Count Acc, compared with 0.296 for Gemini-3.1-Pro and 0.068 for Qwen3- VL-8B. Removing pathology feature reconstruction or cell alignment reduces Count Acc to 0.27 8 Table 2: Ablations of visual supervision and the SFT curriculum on PathoVernier. All variants are evaluated before RL with matched total update budgets. ConfigurationAcc↑CA↑RAWR↓Count Acc↑ ASPECT-SFT0.700.800.530.44 w/o visual supervision 0.640.770.590.32 w/o pathology feature reconstruction0.680.750.560.27 w/o cell feature alignment0.670.770.540.29 w/o count supervision 0.720.780.570.36 Direct Reason training (Stage 3)0.620.740.560.31 Table 3: RL reward ablations on PathoVernier. The listed rewards are the internal term r; all RL variants retain the structural validity gate and the invalid-response penalty. ConfigurationRL RewardAcc↑CA↑RAWR↓Count Acc↑Consistency↑ ASPECT-SFTN/A0.700.800.530.440.94 Answer onlyAcc 0.720.790.550.430.92 Consistency only0.5 Con0.690.760.510.410.98 Answer + count scoreAcc + 0.5 S0.740.800.500.450.96 ASPECT-8BAcc + 0.5 Con0.750.810.490.470.99 (a) Image dependence(b) Count-basedcheck/ final answer Acc 0.00.51.0 0.64 0.32 0.70 0.33 0.75 0.32 Count Acc 0.00.51.0 0.32 0.16 0.44 0.19 0.47 0.20 w/o visualsupervision ASPECT-SFT ASPECT-RL Original image Mismatched image Correct / CorrectCorrect / Wrong Undecidable Wrong / CorrectWrong / Wrong 0 25 5075100% ASPECT-SFT 67.9 7.1 0.3 18.8 5.9 Answer-only RL 69.66.6 0.8 19.2 3.8 ASPECT-RL 73.56.1 0.3 18.7 1.4 Figure 5: Image dependence and task correctness on PathoVernier. (a) Performance with original and mismatched images. (b) Joint outcomes of the count-based check and final-answer correctness. or 0.29. Removing count supervision slightly improves Acc but lowers Count Acc to 0.36 and increases RAWR to 0.57. Thus, reported measurements can support the correct conclusion while remaining inaccurate, reinforcing the need to evaluate observations alongside answers. The staged SFT configuration also improves answer and count accuracy over QA-only Direct Reason training. Ablation on reinforcement learning rewards. Table 3 compares reward variants from the same SFT checkpoint, retaining R = V s r− 0.1(1− V s ) and varying only r, with count score S(C,Q). Consistency averages Con over responses with determinable count-derived decisions and answer components, measuring agreement with the model’s own decision; CA instead uses the reference decision. Answer-only RL improves Acc while degrading count accuracy and consistency. Con- versely, consistency-only RL reaches 0.98 agreement without improving answer or count accuracy over SFT. ASPECT combines these complementary signals and achieves the highest Acc and Count Acc and the lowest RAWR among the tested rewards. Image dependence and task correctness. Figure 5(a) shows that mismatching images largely removes the Acc gains and reduces the Count Acc gains, indicating that ASPECT’s improvements depend on relevant visual evidence. Figure 5(b) further examines a gold-anchored count-based check alongside final-answer correctness. ASPECT-RL passes both checks on 73.5% of questions, versus 67.9% for SFT and 69.6% for answer-only RL. For multi-step composition, the count-based check assesses the reference region, whereas final-answer accuracy requires both the region and proportion band (Appendix C.3). 9 5CONCLUSION We presented ASPECT and PathoVernier to connect fine-grained visual perception with pathol- ogy reasoning. ASPECT supervises intermediate visual tokens for pathology appearance and cell abundance, then learns to generate and use quantitative observations through staged SFT and RL. PathoVernier evaluates the resulting answers together with their supporting measurements. AS- PECT improves both answer accuracy and count fidelity, with further gains over its backbone on three external pathology benchmarks. Ablations and image-mismatch experiments show the value of visual supervision and the dependence of these gains on relevant images. Together, these results demonstrate the value of treating quantitative observations as explicit targets for pathology VLM training and evaluation. REPRODUCIBILITY STATEMENT Code: https://github.com/ChyaZhang/ASPECT Model: https://huggingface.co/Mikezcy/ASPECT-8B Benchmark: https://huggingface.co/datasets/Mikezcy/PathoVernier REFERENCES Mohamed Amgad, Lamees A Atteya, Hagar Hussein, Kareem Hosny Mohammed, Ehab Hafiz, Maha AT Elsebaie, Ahmed M Alhusseiny, Mohamed Atef AlMoslemany, Abdelmagid M El- matboly, Philip A Pappalardo, et al. Nucls: A scalable crowdsourcing approach and dataset for nucleus classification and segmentation in breast cancer. GigaScience, 11:giac037, 2022. Mohammad Asadi, Jack W O’Sullivan, Fang Cao, Tahoura Nedaee, Kamyar Rajabalifardi, Fei-Fei Li, Ehsan Adeli, and Euan Ashley. Mirage: The illusion of visual understanding. arXiv preprint arXiv:2603.21687, 2026. Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, et al. Qwen3-vl technical report. arXiv preprint arXiv:2511.21631, 2025. Junying Chen, Chi Gui, Ruyi Ouyang, Anningzhe Gao, Shunian Chen, Guiming Hardy Chen, Xi- dong Wang, Zhenyang Cai, Ke Ji, Xiang Wan, et al. Towards injecting medical visual knowledge into multimodal llms at scale. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, p. 7346–7370, 2024a. Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin, et al. Are we on the right way for evaluating large vision-language models? Advances in Neural Information Processing Systems, 37:27056–27087, 2024b. Richard J Chen, Tong Ding, Ming Y Lu, Drew FK Williamson, Guillaume Jaume, Andrew H Song, Bowen Chen, Andrew Zhang, Daniel Shao, Muhammad Shaban, et al. Towards a general-purpose foundation model for computational pathology. Nature Medicine, 30(3):850–862, 2024c. Zongyi Chen, Yu Liang, Jie Lin, and Liansheng Wang. Pathview-bench: Can multimodal large language models achieve fine-grained multiscale understanding of pathology images?arXiv preprint arXiv:2607.28318, 2026. Xingyu Fu, Yushi Hu, Bangzheng Li, Yu Feng, Haoyu Wang, Xudong Lin, Dan Roth, Noah A Smith, Wei-Chiu Ma, and Ranjay Krishna. Blink: Multimodal large language models can see but not perceive. In Proceedings of the European Conference on Computer Vision, p. 148–166. Springer, 2024. Jevgenij Gamper, Navid Alemi Koohbanani, Ksenija Benes, Simon Graham, Mostafa Jahanifar, Syed Ali Khurram, Ayesha Azam, Katherine Hewitt, and Nasir Rajpoot. Pannuke dataset exten- sion, insights and baselines. arXiv preprint arXiv:2003.10778, 2020. Google DeepMind. Gemini 3.1 pro. Model Card, February 2026. URL https://deepmind. google/models/model-cards/gemini-3-1-pro/. Published February 19, 2026. 10 Simon Graham, Quoc Dang Vu, Shan E Ahmed Raza, Ayesha Azam, Yee Wah Tsang, Jin Tae Kwak, and Nasir Rajpoot. Hover-net: Simultaneous segmentation and classification of nuclei in multi-tissue histology images. Medical Image Analysis, 58:101563, 2019. Simon Graham, Mostafa Jahanifar, Ayesha Azam, Mohammed Nimir, Yee-Wah Tsang, Katherine Dodd, Emily Hero, Harvir Sahota, Atisha Tank, Ksenija Benes, et al. Lizard: A large-scale dataset for colonic nuclear instance segmentation and classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, p. 684–693, 2021. 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. Yunqi Hong, Kuei-Chun Kao, Liam Edwards, Nein-Tzu Liu, Chung-Yen Huang, Alex Oliveira- Kowaleski, Cho-Jui Hsieh, and Neil Y. C. Lin. Adaptive diagnostic reasoning framework for pathology with multimodal large language models. Communications Medicine, 6:236, 2026. Fabian H ̈ orst, Moritz Rempe, Helmut Becker, Lukas Heine, Julius Keyl, and Jens Kleesiek. Cel- lvit++: Energy-efficient and adaptive cell segmentation and classification using foundation mod- els. Computer Methods and Programs in Biomedicine, p. 109206, 2026. Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen.Lora: Low-rank adaptation of large language models.arXiv preprint arXiv:2106.09685, 2021. Shengyi Hua, Jianfeng Wu, Tianle Shen, Kangzhe Hu, Zhongzhen Huang, Shujuan Ni, Zhihong Zhang, Yuan Li, Zhe Wang, and Xiaofan Zhang. PathFound: An agentic multimodal model activating evidence-seeking pathological diagnosis. Medical Image Analysis, 113:104200, 2026. Songhan Jiang, Fengchun Liu, Ziyue Wang, Linghan Cai, and Yongbing Zhang. Pathreasoner- r1: Instilling structured reasoning into pathology vision-language model via knowledge-guided policy optimization. arXiv preprint arXiv:2601.21617, 2026. Jakob Nikolas Kather, Niels Halama, and Alexander Marx. 100,000 histological images of human colorectal cancer and healthy tissue. Zenodo, 2018. URL https://doi.org/10.5281/ zenodo.1214456. Bangzheng Li, Ximeng Sun, Jiang Liu, Ze Wang, Jialian Wu, Xiaodong Yu, Emad Barsoum, Muhao Chen, and Zicheng Liu. Latent visual reasoning. In Proceedings of the International Conference on Learning Representations, volume 2026, p. 148076–148090, 2026. Zongxia Li, Wenhao Yu, Chengsong Huang, Zhenwen Liang, Rui Liu, Fuxiao Liu, Jingxi Che, Dian Yu, Jordan Boyd-Graber, Haitao Mi, et al. Self-rewarding vision-language model via reasoning decomposition. arXiv preprint arXiv:2508.19652, 2025. Neal I Lindeman, Philip T Cagle, Dara L Aisner, Maria E Arcila, Mary Beth Beasley, Eric H Ber- nicker, Carol Colasacco, Sanja Dacic, Fred R Hirsch, Keith Kerr, et al. Updated molecular testing guideline for the selection of lung cancer patients for treatment with targeted tyrosine kinase in- hibitors: guideline from the college of american pathologists, the international association for the study of lung cancer, and the association for molecular pathology. Archives of Pathology & Laboratory Medicine, 142(3):321–346, 2018. Ming Y Lu, Bowen Chen, Drew FK Williamson, Richard J Chen, Melissa Zhao, Aaron K Chow, Kenji Ikemura, Ahrong Kim, Dimitra Pouli, Ankush Patel, et al. A multimodal generative ai copilot for human pathology. Nature, 634(8033):466–473, 2024. OpenAI. Gpt-5.5 system card. April 2026a. URL https://deploymentsafety.openai. com/gpt-5-5/gpt-5-5.pdf. OpenAI.GPT-5.6 Sol. https://developers.openai.com/api/docs/models/ gpt-5.6-sol, 2026b. OpenAI model documentation. Accessed: 2026-09-23. Massimiliano Pronesti, Anya Belz, and Yufang Hou. Beyond outcome verification: Verifiable pro- cess reward models for structured reasoning. In Findings of the Association for Computational Linguistics, p. 32187–32202, 2026. 11 Yiming Qin, Bomin Wei, Jiaxin Ge, Konstantinos Kallidromitis, Stephanie Fu, Trevor Darrell, and XuDong Wang. Chain-of-visual-thought: Teaching vlms to see and think better with continuous visual tokens. arXiv preprint arXiv:2511.19418, 2025. Pooyan Rahmanzadehgervi, Logan Bolton, Mohammad Reza Taesiri, and Anh Totti Nguyen. Vision language models are blind. In Proceedings of the Asian Conference on Computer Vision, p. 293– 309. Springer, 2024. Roberto Salgado, Carsten Denkert, Sandra Demaria, Nicolas Sirtaine, Frederick Klauschen, Gian- carlo Pruneri, Stephan Wienert, Gert Van den Eynden, Frederick L Baehner, Frederique P ́ enault- Llorca, et al. The evaluation of tumor-infiltrating lymphocytes (tils) in breast cancer: recom- mendations by an international tils working group 2014. Annals of Oncology, 26(2):259–271, 2015. Mark Schuiveling, Hong Liu, Daniel Eek, Gerben E Breimer, Karijn PM Suijkerbuijk, Willeke AM Blokx, and Mitko Veta. A novel dataset for nuclei and tissue segmentation in melanoma with base- line nuclei segmentation and tissue segmentation benchmarks. GigaScience, 14:giaf011, 2025. 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, p. 13183–13192. IEEE, 2024. Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, et al. Deepseekmath: Pushing the limits of mathemati- cal reasoning in open language models. arXiv preprint arXiv:2402.03300, 2024. Yuxuan Sun, Hao Wu, Chenglu Zhu, Sunyi Zheng, Qizi Chen, Kai Zhang, Yunlong Zhang, Dan Wan, Xiaoxiao Lan, Mengyue Zheng, et al. Pathmmu: A massive multimodal expert-level benchmark for understanding and reasoning in pathology. In Proceedings of the European Conference on Computer Vision, p. 56–73. Springer, 2024. Yuxuan Sun, Yunlong Zhang, Yixuan Si, Chenglu Zhu, Kai Zhang, Zhongyi Shui, Jingxiong Li, Xuan Gong, Xinheng Lyu, Tao Lin, et al. Pathgen-1.6 m: 1.6 million pathology image-text pairs generation through multi-agent collaboration. In Proceedings of the International Conference on Learning Representations, volume 2025, p. 94611–94653, 2025. Tong Xiao, Xin Xu, Zhenya Huang, Hongyu Gao, Quan Liu, Qi Liu, and Enhong Chen. Perception- r1: Advancing multimodal reasoning capabilities of mllms via visual perception reward. In Pro- ceedings of the International Conference on Learning Representations, volume 2026, p. 26868– 26898, 2026. Zeyuan Yang, Xueyang Yu, Delin Chen, Maohao Shen, and Chuang Gan. Machine mental imagery: Empower multimodal reasoning with latent visual tokens. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, p. 33510–33520, 2026. Chengyang Zhang, Wenchuan Zhang, Bo Li, Xinyu Liu, Jiaming Yang, Mengran Li, Chenxun Deng, Jie Chen, Yang Zhang, Wei Ju, et al. Do pathology vision-language models truly see pathology? arXiv preprint arXiv:2607.21065, 2026a. Chi Zhang, Haibo Qiu, Qiming Zhang, Yufei Xu, Zhixiong Zeng, Siqi Yang, Peng Shi, Lin Ma, and Jing Zhang. Perceptual-evidence anchored reinforced learning for multimodal reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 41111–41120, 2026b. Wenchuan Zhang, Penghao Zhang, Jingru Guo, Tao Cheng, Jie Chen, Shuwan Zhang, Zhang Zhang, Yuhao Yi, and Hong Bu. Patho-r1: A multimodal reinforcement learning-based pathology expert reasoner. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 40, p. 28418– 28426, 2026c. Yikang Zhou, Tao Zhang, Shilin Xu, Shihao Chen, Qianyu Zhou, Yunhai Tong, Shunping Ji, Jiangn- ing Zhang, Lu Qi, and Xiangtai Li. Are they the same? exploring visual correspondence short- comings of multimodal llms. In Proceedings of the IEEE/CVF International Conference on Com- puter Vision, p. 17663–17674. IEEE, 2025. 12 Jinguo Zhu, Weiyun Wang, Zhe Chen, Zhaoyang Liu, Shenglong Ye, Lixin Gu, Hao Tian, Yuchen Duan, Weijie Su, Jie Shao, et al. Internvl3: Exploring advanced training and test-time recipes for open-source multimodal models. arXiv preprint arXiv:2504.10479, 2025. 13 APPENDIX CONTENTS OF THE APPENDIX • A. Data Sources and Construction • B. Implementation Details and Algorithms • C. Evaluation Protocols and Metric Definitions • D. Prompts and Response Formats • E. Additional Quantitative Results • F. Complete Case Studies and Failure Analysis • G. Limitations and Future Work ADATA SOURCES AND CONSTRUCTION A.1SOURCE DATASETS AND DATA SPLITS We use seven public datasets for training and benchmark construction (Table A.1). Lizard (Graham et al., 2021), PUMA (Schuiveling et al., 2025), PanNuke (Gamper et al., 2020), CoNSeP (Graham et al., 2019), and NuCLS (Amgad et al., 2022) provide nucleus-level annotations for quantitative supervision and PathoVernier construction. The first four provide instance masks and cell labels, whereas NuCLS provides nucleus locations and class annotations for the subset used in this work. NCT-CRC-HE-100K (Kather et al., 2018) and the H&E subset of the PathVQA (He et al., 2020) training split add tissue-classification and pathology question-answering examples during SFT. For NCT-CRC-HE-100K, cellular supervision combines detected nuclei with patch-level tissue labels. For PathVQA, it uses predicted nucleus locations and native cell-type predictions. Neither source provides reference nucleus-level annotations. Data are partitioned according to the available source identifiers: source images for Lizard, source WSIs and their corresponding ROIs for PUMA, individual patches for PanNuke, and TCGA patient– slide identifiers for NuCLS. CoNSeP follows its official training and test partition. NCT-CRC-HE- 100K is sampled from its official training set with balanced tissue classes, while PathVQA uses only its official training split after H&E filtering. SFT, RL, validation, and PathoVernier use mutually disjoint image sets. The SFT collection contains 19,422 visual-supervision examples and 13,095 question-answer pairs. RL uses 1,372 questions from 1,180 unique images, with 93 validation ques- tions used for checkpoint selection. PathoVernier contains 759 questions from 553 unique patches. Data access and release. The PathoVernier release will include questions, reference counts, split indices, and scripts for reconstructing the benchmark from the original datasets. Images will be obtained directly from the respective providers rather than redistributed with the benchmark. The release documentation will retain source attribution and specify the applicable licenses and access terms for source-derived materials. A.2CELL CATEGORY HARMONIZATION With guidance from pathologists, we harmonize source labels into five categories: tumor cells, nor- mal epithelial cells, stromal-like cells, lymphocytes, and other inflammatory cells (Table A.2). Fine- grained labels within a category are merged. Coarse labels retain their aggregate meaning: Lizard epithelial annotations remain epithelial-like, while PanNuke and CoNSeP inflammatory annotations remain inflammatory. For feature alignment, these annotations contribute to both corresponding category targets; count supervision applies only to their combined count. Background and excluded labels are removed before constructing supervision: Dead in PanNuke, Other/Miscellaneous in CoNSeP, apoptotic nuclei in PUMA, and unlabeled nuclei or apoptotic bod- ies in NuCLS. These instances contribute neither to category counts nor to the total count or the sixth cell-token target. The sixth token summarizes all retained, mappable instances, each included once. 14 Table A.1: Data sources and sample counts. SFT examples are separated into visual-supervision and question-answer examples. RL, validation, and PathoVernier columns report question counts. Multiple examples may share an image. Source SFT RLValidationPathoVernier VisualVQA Lizard (Graham et al., 2021)2,5202,94455940400 PUMA (Schuiveling et al., 2025)2,7451,0102195142 PanNuke (Gamper et al., 2020)6,7802,56556720139 CoNSeP (Graham et al., 2019)19613327244 NuCLS (Amgad et al., 2022)8547702634 NCT-CRC-HE-100K (Kather et al., 2018)5,9385,386000 PathVQA (He et al., 2020)389980000 Total19,42213,0951,37293759 For sources without nucleus-level annotations, we use two supervision paths (Table A.3). In NCT- CRC-HE-100K, CellViT-SAM-H detects nuclei, and all detected nuclei in a patch inherit the cate- gory mapped from its tissue label. Adenocarcinoma epithelium maps to tumor, normal colon mu- cosa to normal epithelial, stroma and smooth muscle to stromal-like, and lymphocytes to lympho- cyte. These assignments provide weak cell-type labels. Adipose, debris, background, and mucus patches receive pathology feature reconstruction without cell-specific supervision. In PathVQA, na- tive CellViT PanNuke predictions follow the PanNuke mapping in Table A.2, including aggregate inflammatory supervision and exclusion of Dead and background predictions. Benchmark questions use tumor, stromal-like, lymphocyte, epithelial-like, and inflammatory counts. Epithelial-like combines tumor and normal epithelial cells; inflammatory combines lymphocytes and other inflammatory cells. These quantities are obtained by summing resolved categories or retaining the corresponding aggregate annotation. Questions respect source-label granularity: Lizard does not support tumor-specific questions, while PanNuke and CoNSeP do not support lymphocyte-specific questions. Normal epithelial and other inflammatory counts are not queried individually. Cell-type comparisons exclude category pairs with a subset relationship, such as tumor versus epithelial-like. A.3SFT DATA AND TEACHER TARGETS The SFT collection contains 19,422 visual-supervision records and 13,095 question-answer exam- ples. These records are rendered into stage-specific input–target pairs: the visual-supervision records are used in Generate, while QA records are used in all three stages. The question-answer examples comprise 6,729 programmatically constructed quantitative questions from the five nucleus-annotated datasets, 5,386 classification and related questions from NCT-CRC-HE-100K, and 980 closed-ended yes/no questions from the H&E subset of PathVQA. GPT-5.6-sol generates descriptive prose and reasoning from the image and source-specific reference information. In a subsequent text-only pass, the same model rewrites the <answer> reasoning notes while preserving their numerical values and final answers. PathVQA observations are qualitative and contain no count JSON. Auto- matic checks enforce agreement between the generated prose, reference counts, and answers. Failed generations are retried or discarded; rewrites that alter numerical values or final answers revert to templates. Visual teacher targets are extracted offline and remain fixed during training. UNI provides image features for pathology feature reconstruction. For Lizard, PUMA, PanNuke, and CoNSeP, CellViT- SAM-H features are mean-pooled within ground-truth nucleus masks to obtain instance embeddings, then normalized and aggregated into six cell-token targets as described in Section 3.2. NuCLS supplies UNI features and category counts derived from point annotations, but does not contribute to cell feature alignment. NCT-CRC-HE-100K and PathVQA use the detected instances and category assignments described in Appendix A.2. PathVQA patches with fewer than three retained detections receive UNI supervision only. 15 Table A.2: Source-label mappings for cellular supervision. Starred entries retain aggregate labels: feature alignment uses both listed targets, whereas count supervision uses their sum. NuCLS uses these mappings for count supervision only. SourceOriginal label(s)Target category/categories Lizard EpithelialTumor; normal epithelial ∗ Connective tissueStromal-like LymphocyteLymphocyte Plasma, neutrophil, eosinophilOther inflammatory PUMA TumorTumor EpitheliumNormal epithelial Stroma, endotheliumStromal-like LymphocyteLymphocyte Plasma cell, neutrophil, histiocyte, melanophageOther inflammatory PanNuke NeoplasticTumor EpithelialNormal epithelial Connective/soft tissueStromal-like InflammatoryLymphocyte; other inflammatory ∗ CoNSeP Dysplastic/malignant epithelialTumor Healthy epithelialNormal epithelial Fibroblast, muscle, endothelialStromal-like InflammatoryLymphocyte; other inflammatory ∗ NuCLS Tumor, mitotic figureTumor Ductal epithelium, myoepitheliumNormal epithelial Fibroblast, vascular endotheliumStromal-like LymphocyteLymphocyte Plasma cell, macrophage, neutrophil, eosinophilOther inflammatory Table A.3: Cellular supervision for sources without nucleus-level reference annotations. CRC cell types inherit patch-level tissue labels; PathVQA cell types come from native nucleus predictions. SourceLabel sourceFeature-alignment targetsCount supervision NCT-CRC-HE-100K Patch-level tissue labels Mapped patch category for all detected nuclei Detected nucleus count assigned to that category PathVQA Native CellViT PanNuke predictions Mapped nucleus categories; shared inflammatory targets Predicted category counts; aggregate inflammatory count A.4RL DATA CONSTRUCTION The RL collection contains 1,372 questions from 1,180 unique images drawn from Lizard, PUMA, PanNuke, and CoNSeP. Reference counts are derived from nucleus annotations, and task rules de- termine the reference answers. RL images are disjoint from those used for SFT and PathoVernier. The collection covers eight skills, including regional selection, composition analysis, and additional counting and comparison tasks (Table A.4). We select questions by skill category using the SFT model’s rollout accuracy to identify skills with greater room for improvement. We retain the available questions for multi-step composition, re- gion selection, dominant cell type, count band prediction, and count comparison, while limiting the more saturated region-comparison, cell-type-comparison, and yes/no categories. This allocation concentrates training on skills where SFT remains less reliable while preserving task diversity. A 93-question validation set is used for checkpoint selection. Reward and optimization settings are provided in Appendix B.3. 16 Table A.4: Skill composition of the RL training collection. Question selection is based on skill-level SFT performance. SkillQuestionsSkillQuestions Multi-step composition388Count comparison85 Region selection407Region comparison60 Dominant cell type208Cell-type comparison60 Count band prediction124Yes/no40 Total1,372 A.5PATHOVERNIER CONSTRUCTION AND EXPERT REVIEW We construct PathoVernier from eligible H&E patches in the five nucleus-annotated datasets. After the category filtering described in Appendix A.2, patches containing fewer than five or at least 300 retained nuclei are excluded. NuCLS contributes only questions whose reference quantities can be computed from point annotations. Candidate questions are sampled with quotas stratified by task, dataset, and answer category to balance task coverage and answer distributions. Four task templates instantiate questions from annotation-derived counts and deterministic decision rules, yielding 786 candidates. GPT-5.6-sol paraphrases the questions while numerical thresholds and answer options remain fixed. Automatic checks preserve the queried cell types, regions, and relationships; a separate language model checks semantic equivalence against the reference question. Paraphrases that fail these checks are regenerated, with the original template retained when retries are exhausted. An independent implementation recomputes regional assignments and rule-based answers from the annotations and flags discrepancies for manual review. Detailed task rules and rewriting prompts are provided in Appendices C.2 and D.2, respectively. All 786 candidates are presented through a structured review interface to four board-certified pathol- ogists, with automatically flagged cases prioritized for inspection. Review assesses whether the question matches the image, the reference counts agree with the annotations and visible morphol- ogy, and the answer is reliably distinguishable. Particular attention is given to near-tied regional counts and proportions close to category thresholds, especially when small denominators make the answer sensitive to counting discrepancies. Following review, 27 questions are excluded, leaving 759 questions from 553 unique patches with approximately balanced coverage of the four tasks. A.6DETAILED DATASET STATISTICS PathoVernier contains 759 questions from 553 unique patches. The four tasks each account for 24.5– 25.4% of the questions, with their source composition detailed in Table A.5. We harmonize the 19 source organ labels into 18 categories by merging Colon and Colorectal (Table A.6). Colon, skin, and breast contribute 480, 142, and 60 questions, respectively. The remaining 15 categories con- tribute 77 questions and are supplied by PanNuke. All distributions in this section count questions, including those sharing the same patch. Table A.5: PathoVernier question counts by source dataset and task. SourceRegion selectionRegion comparisonCell-type comparisonMulti-step compositionTotal Lizard828513598400 PUMA45471040142 PanNuke35363335139 CoNSeP17713744 NuCLS111121034 Total190186193190759 Each patch contains 256×256 pixels at its source resolution. Physical resolution ranges from 0.20 to 0.50μm/pixel, corresponding to patch widths of 51.2–128μm (Table A.7). The 400 Lizard questions 17 use images at approximately 20×, while the remaining 359 questions use images at approximately 40×. These magnification groups therefore also differ in data source and organ composition. Table A.6: PathoVernier question counts across 18 harmonized organ categories. Colon includes the Colorectal label from CoNSeP. OrganQuestionsOrganQuestions Colon480Kidney4 Skin142Ovary3 Breast60Liver2 Bile duct15Pancreas2 Head and neck11Prostate2 Uterus10Stomach2 Esophagus9Testis2 Cervix6Thyroid2 Lung6Bladder1 Total759 Table A.7: Image scale by source. Patch width is calculated from the 256× 256-pixel inputs and source resolution. Magnifications are approximate objective equivalents. SourceResolution (μm/pixel)Patch width (μm)Approx. magnificationQuestions Lizard0.50128.020×400 PanNuke, CoNSeP0.2564.040×183 PUMA0.2256.340×142 NuCLS0.2051.240×34 Total759 BIMPLEMENTATION DETAILS AND ALGORITHMS B.1MODEL AND TRAINING CONFIGURATIONS ASPECT uses Qwen3-VL-8B (Bai et al., 2025) as its backbone and applies LoRA (Hu et al., 2021) to both the visual and language components during SFT and RL. Both stages use AdamW with BF16 training. Table B.8 summarizes the principal configurations; additional optimization settings are provided in Appendices B.2 and B.3. B.2SUPERVISED FINE-TUNING SFT proceeds through Perceive, Generate, and Reason for 1,000, 1,000, and 1,500 updates, respec- tively. Perceive uses QA records, placing visual tokens in the input and supervising the answer block. Generate uses the full pool of visual-supervision and QA records to predict the visual-token block from the image and a feature query. Reason uses QA records to generate the complete <think>, <observe>, and <answer> sequence. All stages use cross-entropy with unit weight over the en- tire assistant target, masking user-prompt positions. Visual losses apply wherever their supervision is available; NuCLS supplies reconstruction and count supervision without cell feature alignment. Teacher targets remain fixed throughout training. We use AdamW with a cosine learning-rate schedule, a warmup ratio of 0.05, and weight decay of 0.1. Learning rates and visual-loss weights are listed in Table B.8. A per-device micro-batch size of 1 and gradient accumulation over 24 steps across four GPUs yield an effective batch size of 96. Images are resized to 512× 512, matching the resolution used for teacher feature extraction. We use random seed 0 and select the checkpoint at update 3,500 as ASPECT-SFT based on the independent development set. 18 Table B.8: Model and training configurations of ASPECT. SFT stage lengths are reported as updates within each stage. SettingValue Model and adaptation BackboneQwen3-VL-8B Pathology feature tokens8 Cell tokens6 Visual teachersUNI; CellViT-SAM-H Teacher updatesFrozen; targets extracted offline LoRA-adapted componentsVisual and language components LoRA rank r16 LoRA scaling α32 Hardware4× NVIDIA RTX 6000D Supervised fine-tuning Total updates3,500 Perceive / Generate / Reason updates1,000 / 1,000 / 1,500 Effective batch size96 LoRA learning rate2× 10 −4 Reconstruction and alignment module learning rate4× 10 −5 New-token embedding learning rate4× 10 −5 Count head and cell-role embedding learning rate2× 10 −4 Reconstruction weight λ u 0.5 Cell-alignment weight λ a 1.0 Count-supervision weight λ c 1.0 Reinforcement learning InitializationASPECT-SFT AlgorithmGRPO Learning rate1× 10 −5 Responses per question8 Sampling temperature1.0 Consistency reward coefficient β0.5 Invalid-response penalty coefficient γ0.1 B.3REINFORCEMENT LEARNING RL initializes from ASPECT-SFT and updates LoRA parameters in both the visual and language components using GRPO. Each rollout batch contains 48 questions, with eight responses sampled per question at temperature 1.0. Prompt and response lengths are capped at 2,048 and 512 tokens, respectively. Responses are scored using equation 6, and rewards are normalized within each ques- tion’s response group to obtain relative advantages. The policy objective uses asymmetric clipping with lower and upper thresholds of 0.2 and 0.28, corresponding to a probability-ratio interval of [0.8, 1.28], and a dual-clipping coefficient of 3.0. No KL penalty is included in either the reward or the optimization loss. We use AdamW with a constant learning rate of 1 × 10 −5 , no warmup, weight decay of 0.01, and random seed 42. Training runs for five epochs, with checkpoints saved every 16 updates. We select the checkpoint with the highest answer accuracy on the independent validation set described in Appendix A.4, yielding the checkpoint at update 128 used for all reported ASPECT-RL results. B.4TRAINING ALGORITHMS AND INFERENCE We summarize the SFT and RL procedures in Algorithms 1 and 2, respectively. The algorithms specify how supervision and generated responses enter training; optimization settings are provided in Table B.8 and Appendices B.2–B.3. Inference. ASPECT generates visual tokens, textual observations, and the final answer autore- gressively from the image and question. The frozen teachers and visual-supervision modules are not 19 Algorithm 1 Three-stage supervised fine-tuning Require: Backbone model; QA recordsD QA ; visual-supervision recordsD V Require: Cached teacher targets; stage lengths (T 1 ,T 2 ,T 3 ) = (1000, 1000, 1500) 1: Initialize LoRA adapters, visual-token parameters, and visual-supervision modules 2: for stage s∈1, 2, 3 do 3:if s = 2 then 4: D s ←D V ∪D QA 5:else 6: D s ←D QA 7:end if 8:for t = 1,...,T s do 9:Sample a batch fromD s 10:if s = 1 then▷ Perceive 11:Input← image, question, and visual-token block 12:Target← answer block 13:else if s = 2 then▷ Generate 14:Input← image and feature query 15:Target← visual-token block 16:else▷ Reason 17:Input← image and question 18:Target← <think> with visual tokens, <observe>, and <answer> 19:end if 20:Run a teacher-forced forward pass and collect visual-token hidden states 21:ComputeL LM over all target tokens, masking input positions 22:ComputeL uni ,L align , andL count from available teacher targets 23:Set unavailable visual-loss terms to zero 24: L←L LM + λ u L uni + λ a L align + λ c L count 25:Update trainable parameters using AdamW 26:end for 27: end for 28: return ASPECT-SFT used at inference. Reported counts are generated as part of the response, rather than read from the auxiliary count head, which provides image-wide supervision during SFT. The model determines the generated token sequence without an external routing mechanism. CEVALUATION PROTOCOLS AND METRIC DEFINITIONS C.1BENCHMARKS AND MODEL EVALUATION SETTINGS Evaluation datasets. We evaluate ASPECT and all baselines on PathoVernier and three exter- nal pathology benchmarks. For PathCLS, we use all 1,632 questions from the official PathMMU PathCLS test subset, with the original multiple-choice options and answer labels. PathMMU data are not used for SFT or checkpoint selection. For PathVQA and Quilt-VQA, we evaluate native closed-ended yes/no questions from their official test splits, retaining only H&E histology images. Each unique image is assessed independently by GPT-5.5 and Gemini-3.1-Pro and retained if either model identifies it as H&E histology. This filtering retains 1,306 of 3,362 PathVQA questions and 314 of 343 Quilt-VQA questions. All models are evaluated on the same retained subsets. These evaluations use no in-context examples; PathVQA training examples used for SFT are described in Appendix A.3. Inference settings. Images are resized to 512× 512 pixels before processing with each model’s native image processor. We generate one response per question. Local models use greedy de- coding with sampling disabled and seed 42; closed-source APIs use temperature zero. Out- put token limits are listed in Table C.9.The closed-source model identifiers are gpt-5.5 and gemini-3.1-pro-preview. For PathoVernier and PathCLS, GPT-5.5 was evaluated on 20 Algorithm 2 Reinforcement learning with answer–observation consistency Require: ASPECT-SFT policy π θ ; RL datasetD RL ; validation setD val Require: Reward coefficients β,γ; responses per question K = 8 1: Initialize the RL policy from ASPECT-SFT 2: for each of five training epochs do 3:for each batch of 48 questions fromD RL do 4:for each image–question pair (x,u) with reference answer a ⋆ do 5:Sample responsesy i K i=1 from π θ (·| x,u) 6:for each response y i do 7:Check structural validity V s (y i ) 8:if V s (y i ) = 0 then 9:R i ←−γ 10:else 11:Extract reported counts C i and final answer ˆa i 12:Acc i ← 1[ˆa i = a ⋆ ] 13:Con i ← 1[g u (C i ) = π u (ˆa i )], or 0 if undetermined 14:R i ← Acc i + β Con i 15:end if 16:end for 17:Normalize group rewards to obtain relative advantages 18:end for 19:Update visual and language LoRA parameters with GRPO 20:Use asymmetric clipping and no KL penalty as specified in Appendix B.3 21:Save a checkpoint every 16 training steps 22:end for 23: end for 24: return The saved checkpoint with the highest accuracy onD val Table C.9: Evaluation subsets and output token limits for local models and closed-source APIs. BenchmarkEvaluation subsetQuestionsLocal token limitAPI token limit PathoVernierFull benchmark759512512 PathCLSOfficial test1,632512512 PathVQAH&E, closed-ended test1,306512512 Quilt-VQAH&E, closed-ended test314512512 September 2, 2026, and Gemini-3.1-Pro on September 3, 2026. Both models were evaluated on PathVQA and Quilt-VQA on September 15, 2026. Prompts and scoring. On PathoVernier, all models receive identical questions and answer op- tions. Baselines additionally receive a system prompt requesting whole-image cell-type counts, the quantities required by the question, and a final answer, whereas ASPECT uses its trained response format without a system prompt. All responses are scored with the same parser. PathVQA and Quilt-VQA request only yes/no answers and use no system prompt. PathCLS requests a brief mor- phological justification followed by the selected option letter. Intermediate counts are not required for these external tasks. Unparseable final answers are scored as incorrect. Failed API requests are retried up to four times, and unresolved failures are also scored as incorrect. Metric definitions and eligibility criteria are detailed in Appendix C.2, and the evaluation prompts are provided in Appendix D.4. C.2TASK RULES, METRICS, AND ELIGIBLE RESPONSES Task rules. Reference counts are obtained by assigning annotated nuclei to the geometric regions specified in each question according to their centroids. Region selection identifies the region with the largest count of the queried cell type among equal-area regions. Region comparison compares 21 one cell type across two equal-area regions, while cell-type comparison compares two types within the image. Let a and b denote the first and second counts in the comparison. For b > 0, both tasks assign the ratio r = a/b to a category using h(r) = much fewer, r ≤ 0.5, comparable, 0.5 < r < 2.0, muchmore, r ≥ 2.0. (11) When b = 0, the result is muchmore if a > 0 and comparable if a = 0. These cases remain determinable. For equal-area regions, the count ratio equals the density ratio. Multi-step composition first selects the region with the largest count of the queried type and then assigns that type’s proportion p within the selected region to a category: h prop (p;t 1 ,t 2 ) = low, p < t 1 , mid, t 1 ≤ p < t 2 , high, p≥ t 2 . (12) The thresholds t 1 and t 2 are estimated from training-split tertiles separately for each cell type and spatial partition, rounded to increments of 0.05, and then frozen. Each question states its applicable thresholds explicitly. Table C.10 lists the thresholds for quadrants, four horizontal bands, and four vertical bands. Tumor is excluded from this task because its median proportion is approximately one. Candidate questions whose reference proportions fall within a 5% relative margin of either threshold are excluded during construction. Table C.10: Frozen proportion thresholds (t 1 ,t 2 ) for multi-step composition, estimated separately for each spatial partition and cell type. PartitionStromal-likeLymphocyteEpithelial-likeInflammatory Quadrants(0.50, 0.85)(0.35, 0.55)(0.60, 0.90)(0.55, 0.80) Horizontal bands(0.40, 0.80)(0.30, 0.50)(0.55, 0.80)(0.55, 0.75) Vertical bands(0.40, 0.75)(0.30, 0.50)(0.60, 0.85)(0.55, 0.75) If reported counts yield a tied maximum in region selection or the first step of multi-step composi- tion, the count-derived region is undeterminable. Such responses are excluded from CA and Con- sistency, but remain eligible for Acc and Count Acc; RAWR eligibility follows its correct-answer and complete-count requirements. A tied maximum therefore differs from a zero denominator in a comparison task, for which the rule above returns a definite category. Answer and count accuracy. Let ˆa and a ⋆ denote the predicted and reference answers, respec- tively. Answer accuracy evaluates the complete answer, including both the region and proportion category for multi-step composition. An unparseable final answer is scored as incorrect. To evaluate the reported measurements, let K index the counts required by a question, q i denote reference count i, and dom(C) denote the reported count keys. The per-question count score is S(C,Q) = 1 |K| X i∈K∩dom(C) 1[|C i − q i |≤ max(1, 0.1q i )],(13) where Q = q i i∈K is the set of required reference counts and 1[·] is the indicator function. Each required count contributes equally, and missing counts contribute zero. We report Acc = E[1[ˆa = a ⋆ ]],Count Acc = E[S(C,Q)].(14) Both averages include all evaluation questions, with equal weight per question. Count Acc therefore measures the fraction of required counts within tolerance rather than exact integer-count accuracy. Gold-referenced coherent accuracy. CA evaluates whether the reported counts imply the refer- ence task conclusion. Let g u (C) denote the conclusion checked by CA and π u (a ⋆ ) the correspond- ing component of the reference answer: CA = E[1[g u (C) = π u (a ⋆ )]|g u (C) is defined].(15) 22 For region selection and the two comparison tasks, this check uses the complete task conclusion. For multi-step composition, g u checks only region selection, and π u extracts the reference region. Its eligibility therefore depends on the counts needed to select a region, without requiring the additional counts needed to determine the proportion category. CA does not require the model’s final answer to be correct. Moreover, counts can imply the correct conclusion while remaining numerically inaccurate, so CA and Count Acc measure different properties. Right answer, wrong reason. RAWR evaluates counting errors among responses that have a cor- rect final answer and report every required count: RAWR = 1− E[S(C,Q)| ˆa = a ⋆ , K ⊆ dom(C)].(16) Thus, Count Acc averages count agreement over all questions, whereas RAWR averages the fraction of counts outside tolerance over correct responses with complete measurements. RAWR is not a binary indicator of whether a response contains any counting error. Consistency with the model’s own answer. Consistency evaluates whether the conclusion im- plied by the reported counts agrees with the corresponding component of the model’s own answer. Using the same task check g u and answer projection π u as CA, we define Consistency = E[1[g u (C) = π u (ˆa)]|g u (C) and π u (ˆa) are defined].(17) CA compares the count-derived conclusion with π u (a ⋆ ), whereas Consistency compares it with π u (ˆa). For multi-step composition, both checks evaluate region selection only and do not require the proportion denominator. The full region-and-proportion answer is evaluated by Acc, while S(C,Q) measures agreement between reported and reference counts. Consequently, a response can be con- sistent despite an incorrect proportion category or an incorrect final answer. The indicator in equa- tion 17 is the same as Con in the RL reward: undeterminable cases receive zero consistency reward during training but are excluded from the evaluation metric’s denominator. The off-diagonal cate- gories in Figure 5 therefore do not measure the rate of inconsistency with the model’s own answer. Eligible responses. Acc and Count Acc include all evaluation questions. CA includes responses with a determinable count-derived conclusion even when the final answer is missing or unparseable. RAWR requires a correct final answer and complete reported counts, whereas Consistency requires both the count-derived conclusion and the corresponding predicted-answer component to be deter- minable. Counts are scored field by field using the evaluation parser’s output; negative values or duplicate keys do not automatically assign zero to the entire response. The training validity gate V s is not applied as a response-level filter during evaluation. Missing required counts contribute zero to Count Acc. Count coverage is the fraction of all 759 questions for which the reported counts yield a determinable conclusion under the task rules used for CA. For multi-step composition, this requires determining the selected region, irrespective of the proportion band. Complete-count coverage in- stead requires every necessary count to be reported, regardless of answer correctness. Complete counts can still yield an undeterminable conclusion when regional maxima are tied. In Table 1, CA and RAWR are omitted when CA coverage is below 30%; this reporting threshold does not change either metric’s definition or eligible sample set. Table C.11 reports both coverage measures, RAWR sample counts, and unconditional Count Acc. ASPECT achieves the highest Count Acc over all questions (0.471 versus 0.296 for Gemini-3.1-Pro), showing that its improvement in measurement accuracy extends beyond the conditional subsets used for CA and RAWR. C.3ABLATION AND IMAGE-MISMATCH PROTOCOLS Visual supervision and training curriculum. We evaluate the SFT ablations before RL to sepa- rate the effects of supervised training from reward optimization. Starting from the same backbone, we remove pathology feature reconstruction, cell feature alignment, or count supervision individ- ually, and additionally evaluate a variant without all three visual losses. These loss-removal vari- ants retain the visual tokens, response formats, three-stage curriculum, and remaining loss weights. Assistant-token cross-entropy has weight one throughout training. Direct Reason retains all three visual losses and trains on the 13,095 QA examples using the Reason-stage response format from the first update. It runs for 3,500 updates, matching the total update budget of the complete curriculum. 23 Table C.11: Evaluation coverage and unconditional count accuracy on PathoVernier. Both coverage measures report counts out of all 759 questions. N RAWR is the number of correct responses with complete reported counts. Count Acc averages S(C,Q) over all questions, including failures and incomplete responses. ModelCA coverageComplete-count coverage N RAWR Count Acc↑ GPT-5.5748/759749/7594470.210 Gemini-3.1-Pro733/759750/7594720.296 Qwen3-VL-8B233/759261/7591060.068 Qwen3-VL-32B682/759700/7593330.178 InternVL3-8B448/759449/7591660.071 InternVL3-38B750/759753/7593400.153 Quilt-LLaVA-7B1/7592/75910.000 PathGen-LLaVA-13B0/7590/75900.000 HuatuoGPT-Vision-7B12/75912/75920.002 Patho-R1-7B0/7590/75900.000 ASPECT-8B (SFT)714/759738/7595250.444 ASPECT-8B (RL)748/759752/7595630.471 Unlike the full curriculum, whose Generate stage also includes visual-supervision examples, Direct Reason uses QA examples throughout. All variants follow the evaluation protocol in Appendix C.1. Reinforcement learning rewards. The reward ablations initialize from the same ASPECT-SFT checkpoint and use the same 1,372-question RL set and GRPO configuration. We retain the struc- tural validity gate and invalid-response penalty in every variant: R = V s r− 0.1(1− V s ),(18) where the valid-response reward r is Acc, 0.5 Con, Acc+0.5S(C,Q), or Acc+0.5 Con for answer- only, consistency-only, answer-plus-count, and ASPECT RL, respectively. The count-reward variant uses the per-question tolerance score defined in equation 13, rather than the dataset-level Count Acc. The SFT row provides the initialization baseline without RL. Checkpoints are selected by answer accuracy on the independent validation set. Image mismatch. Figure 5(a) evaluates the model without auxiliary visual supervision, ASPECT- SFT, and ASPECT-RL using original and mismatched images. We construct a permutation of unique patches within each source dataset, requiring every patch to map to a different patch. This preserves dataset provenance while breaking the correspondence between the image and question. All ques- tions associated with an original patch share the same replacement image, and all three models use identical mappings. We generate three mappings with seeds 0, 1, and 2 over the initial 786-question candidate set and evaluate only the final 759 benchmark questions. Questions, options, reference answers, and reference counts remain unchanged. We report the mean mismatched-image Acc and Count Acc across the three mappings; all corresponding standard deviations are at most 0.017. This variation reflects image replacements rather than training runs. Joint outcome analysis. Figure 5(b) decomposes the responses of ASPECT-SFT, answer-only RL, and ASPECT-RL according to final-answer correctness and the gold-referenced count check. For each response, we define A = 1[ˆa = a ⋆ ], B = 1[g u (C) = π u (a ⋆ )],(19) where g u and π u follow Appendix C.2. Responses with a determinable count-derived conclusion are assigned to the four (B,A) combinations; responses for which g u (C) is undeterminable form a fifth category. All category proportions use the full 759-question benchmark as their denominator. For multi-step composition, B evaluates region selection, whereas A evaluates the complete region- and-proportion answer. Thus, B = 1,A = 0 includes responses that select the correct region but give an incorrect proportion category. This group need not be inconsistent with the model’s own reported counts, and the off-diagonal proportion is not the complement of Consistency. 24 DPROMPTS AND RESPONSE FORMATS D.1TRAINING DATA GENERATION We generate the textual content of SFT question-answer records using GPT-5.6-sol. Each request includes the H&E image, question, reference answer, and source-specific context. The five nucleus- annotated datasets provide reference cell counts, whereas cellular NCT-CRC-HE-100K patches pro- vide pseudo-counts together with their tissue labels. PathVQA supplies the original yes/no answer without a nuclear composition input. All sources share the system prompt below. The displayed prompts are excerpts, and braces denote sample-specific fields. GPT-5.6-sol: Shared system prompt You are an expert pathologist. You are given an H&E image, the GROUND-TRUTH nuclear composition of the field, a question, and the CORRECT answer. Your job is NOT to decide the answer (it is already given and correct). Produce EXACTLY two lines: OBSERVE: one sentence describing the overall visual appearance of the field — cellularity, tissue archi- tecture, dominant nuclear morphology — WITHOUT stating any numeric counts (no digits). REASON: 2-3 sentences that justify the GIVEN answer, citing counts ONLY from the provided composi- tion, leading to that answer. Rules: never invent numbers; OBSERVE must contain NO digits; do NOT repeat the OBSERVE sentence inside REASON . . . ; do NOT write ‘FINAL:’ or the answer token — that is appended automatically. The source-specific user inputs are summarized below. For the five nucleus-annotated datasets, the composition includes explicitly supplied zero counts for absent types. CRC inputs distinguish cellular patches, which supply detected composition, from non-cellular patches. PathVQA instead requests an explanation based on image morphology. GPT-5.6-sol: Source-specific user inputs Five nucleus-annotated datasets Tissue/organ:organ (H&E). Ground-truth nuclear composition of THIS field:composition. Question:question CORRECT answer:gold Output the two lines now: OBSERVE: ... REASON: ... NCT-CRC-HE-100K H&E colorectal histology. Ground-truth tissue class: option (tissueclass). For cellular patches: Detected nuclear composition: composition. For non-cellular patches: This field is essentially acellular (no significant nuclei). Question: question CORRECT answer: gold Output the two lines now: OBSERVE: ... REASON: ... PathVQA H&E pathology microscopy image. Question: question CORRECT answer: yes/no Write two lines. OBSERVE = one sentence on the overall appearance relevant to the question (no digits). REASON = 1-3 sentences that justify the given yes/no answer from the morphology. OBSERVE: ... REASON: ... 25 The generated lines are incorporated into the <observe> and <answer> blocks. Where count observations are included, their JSON is inserted programmatically from the corresponding annota- tions or pseudo-labels. PathVQA observations contain qualitative descriptions without count JSON. The FINAL: line is appended from the reference answer. For counting questions, GPT-5.6-sol subsequently rewrites the reasoning notes in <answer>, in- cluding their count statements, conclusions, and final-answer lines. It receives text only, with multi- ple notes grouped by identifiers. Rewrites must preserve the numerical multiset and the final-answer line; failures revert to the template version. GPT-5.6-sol: Answer-note rewriting You are rewriting the wording of short analytical notes that accompany histopathology counting questions. Each note states some counts and draws a conclusion. Rewrite EACH note below. Hard requirements: • Keep every number EXACTLY as given. Do not add, drop, round or recompute any number. You cannot see the image, so you have no basis to change any count. • Keep the final line exactly as FINAL:<option> with the option unchanged. • Vary sentence structure, connectives and register . . . ; do not reuse one skeleton with different numbers . . . • Keep it to one or two sentences plus the FINAL line. • Output . . . one rewritten note per input, each introduced by ###<id> . . . Notes to rewrite: ###id answer noteincludingFINAL . . . D.2BENCHMARK QUESTION REWRITING We use GPT-5.6-sol to diversify question wording while preserving the queried quantities. Training and benchmark questions use separate phrasing pools, denoted A and B, with no template shared between them. The generator receives a reference template, its placeholder list, and an explicit description of the intended measurement. Numerical rules, reference counts, and final answers are handled programmatically rather than generated by the paraphrasing model. GPT-5.6-sol: Question paraphrasing You are helping diversify the phrasing of questions in a computational-pathology benchmark. Reference question template (Python format string):ref Placeholders:ph What the question asks:meaning Writen ALTERNATIVE phrasings of the SAME question. Hard requirements: • Use EXACTLY the same placeholders, each exactly once. Do not add or drop any. • Ask for exactly the same quantity. Do not change what is being compared, which region, or how it is normalised. • Do NOT hint at the answer: the question must not contain comparative or magnitude words such as more/fewer/higher/lower/most/least/high/low/dense. • Vary sentence structure and register (some clinical, some plain), not just synonyms. • One per line, no numbering, no quotes, nothing else. Here,ref is the reference template,ph lists its placeholders,meaning specifies the quan- tity being queried, andn is the requested number of alternatives. Within question templates,t denotes the queried nucleus type,t1 andt2 denote two compared types, andr1 andr2 denote two compared regions. The placeholder region denotes the whole-image field. The following example specifies that both steps of a multi-step composition question concern the same nucleus type. 26 Example generator input: Multi-step composition Reference template Determine the region whoset nuclear density is greatest, then state the proportion of that region’s nuclei that aret. Placeholder list t Intended measurement first which equal-area region contains the greatest density of nucleus typet, then what proportion of ALL nuclei in THAT region are of the SAME typet. Only one nucleus type is involved throughout — do not introduce a second type. Candidates first undergo deterministic checks for placeholder identity and prohibited directional wording. Geometric-region questions must retain the word “region”, use “nuclei” rather than “cells”, and restrict “area” to the expression “per unit area”. Cross-pool Jaccard similarity must not exceed 0.55. Candidates passing these checks are evaluated by GPT-5.4, using the reference and candidate templates instantiated with the same placeholder values. GPT-5.4: Independent semantic verification Two questions about the same histopathology image are shown. A:a B:b Do A and B ask for EXACTLY the same quantity — same object type, same region, same normalisation (per-area vs raw count), same comparison direction? A difference in wording is fine; a difference in what is measured is not. Answer with ONLY one word: SAME or DIFFERENT. The candidate is accepted when the verifier returns SAME and rejected when it returns DIFFERENT. Failed candidates are regenerated, with the original template retained when retries are exhausted. Rewriting changes question wording without modifying the numerical keys in <observe> or the programmatically determined FINAL answer. Independent answer verification and expert review follow the procedures in Appendix A.5. D.3SFT AND RL PROMPTS We present the user inputs and assistant targets used in the three SFT stages, followed by the prompt used for RL. Images precede text in every input. For compactness, [UNI block] denotes <|anchorstart|>, eight consecutive <|unipad|> tokens, and <|anchorend|>; [CELL block] uses the same delimiters with six consecutive <|cellpad|> tokens. These bracketed names are display abbreviations and are replaced by the corresponding token sequences in training. The UNI block always precedes the CELL block. SFT applies cross-entropy to all assistant-target tokens with unit weight and masks user-prompt positions. Perceive: Answer from visual tokens User input <image>the patch feature of the image is [UNI block], and the cell composition of the image is [CELL block].QUESTION Assistant target ANSWER Here, ANSWER is the prepared answer block, including its <answer> and </answer> de- limiters. The target contains neither a <think> block nor an <observe> block. Visual tokens occur in the user input and therefore receive visual supervision but no token-prediction loss at their input positions. 27 Generate: Predict visual tokens User input <image>What is the patch feature, cell composition of the image? Assistant target [UNI block][CELL block] Generate uses both visual-supervision records and QA records, replacing their questions with the feature query above. Every record supervises prediction of the visual-token sequence through cross- entropy, together with the available visual supervision losses. Reason: Generate observations and an answer User input <image>QUESTION Assistant target <think> the patch feature of the image is [UNI block], and the cell composition of the image is [CELL block].</think> <observe>OBSERVATION</observe> <answer>REASONING FINAL:OPTION</answer> The Reason target places the visual tokens inside the <think> block before the observation and answer. The OBSERVATION field contains the descriptive sentence and, where provided by the data source, the count JSON described in Appendix D.1. PathVQA observations contain only descriptive prose. The question includes its answer options. Reinforcement learning: Sample a response System prompt None. User input <image> QUESTION Choose exactly one of:opt1, opt2, ... Assistant response Sampled from the policy; no fixed target is supplied. RL retains the image-and-question input structure of Reason and explicitly lists the available options. The policy generates its response in the format learned during SFT, without additional formatting instructions in the prompt. Responses are scored using the reward described in Appendix B.3. For training samples marked as non-cellular, the CELL block and the corresponding cell- composition clause are omitted. Their Generate query is “What is the patch feature of the image?”, and its target contains only the UNI block. This adjustment applies to training examples; inference uses no external routing to select which visual-token blocks the model generates. D.4EVALUATION PROMPTS We provide the evaluation prompts below. Each question is accompanied by its image; image in- puts are omitted from the displayed templates. Placeholders are replaced with the question and its available answer options. All evaluations use zero-shot prompting without demonstration examples. PathoVernier. Baseline models receive explicit instructions to report nuclear counts before their final answer. ASPECT uses no system prompt and follows the response format learned during SFT, with the image, question, and available options as input, as illustrated in Appendix D.3. All models receive the same questions and answer options. 28 PathoVernier: Baseline evaluation System prompt You are an expert pathologist analyzing an H&E-stained histology image. You MUST base your an- swer strictly on what is visible in THIS image: actually look at and count the nuclei you can see. Never answer from priors, definitions, or the wording of the question alone — if you did not look at the image, you cannot answer. Think step by step. Then output ONE JSON object wrapped in <observe>...</observe>: FIRST the whole-image count of each main nucleus type you can see — epithelial-like, stromal-like, lymphocyte, inflammatory, and tumor separately when identifiable (do NOT report a type that is not present in this patch) — THEN the specific quantities the question asks you to report. Finally, output a line ‘FINAL: X’ where X is EXACTLY one of the allowed option tokens, copied verbatim. User prompt question, Choose exactly one of these option tokens:opt1, opt2, .... Respond with your reasoning, then<observe>...</observe>, then a final line ‘FINAL:<token>’. PathCLS. The prompt asks the model to identify the tissue or lesion, briefly justify its choice from the image, and end with the selected option letter. It does not require structured count observations. PathCLS: Tissue and lesion classification System prompt You are an expert pathologist analyzing a single H&E-stained histology image. This is a MULTIPLE- CHOICE QUESTION about which tissue or lesion type the image shows. You may first note what you observe about the cells, but your final<answer> MUST directly address THIS question: decide which of the listed options best matches the tissue/lesion in the image, briefly justify it from the morphology, and end with a line ‘FINAL: X’ where X is EXACTLY one option letter. Do NOT just describe cell counts — you must pick the correct option. User prompt question Options: A) . . . B) . . . In your answer, choose the single option that correctly names the tissue/lesion shown, and end with ‘FINAL:<letter>’. PathVQA and Quilt-VQA. Both benchmarks use the same user prompt without a system prompt. Models are asked to return only the binary answer, without reasoning or count observations. PathVQA and Quilt-VQA: Closed-ended questions System prompt None. User prompt You are looking at a histopathology image. question Answer with exactly one word: yes or no. EADDITIONAL QUANTITATIVE RESULTS E.1RESULTS BY TASK TYPE Figure E.1 compares ASPECT with its base model and two closed-source models across the four PathoVernier task types. ASPECT achieves higher Acc and lower RAWR in every task, extending its advantage from region selection to multi-step composition and quantitative comparisons. Relative to Gemini-3.1-Pro, its Acc improves by 31.3% on multi-step composition and 26.3% on cell-type comparison. Lower RAWR across all four tasks further shows that, among correct answers with complete count observations, ASPECT reports more accurate quantitative evidence. These results connect its gains in task accuracy with improved measurement of the underlying cell composition. 29 0.2 0.4 0.6 0.8 1.0 (a) Acc 0.2 0.4 0.6 0.8 1.0 Region selection Multi-step composition Region comparison Cell-type comparison (b) RAWR ASPECT-8B (RL) Qwen3-VL-8B Gemini-3.1-ProGPT-5.5 Region selection Multi-step composition Region comparison Cell-type comparison Figure E.1: Results by task type on PathoVernier. ASPECT achieves higher Acc and lower RAWR than the three comparison models across all four tasks. Larger values are better for Acc; smaller values are better for RAWR. (a) Acc Lizard 0.50 PUMA 0.22 PanNuke 0.25 CoNSeP 0.25 NuCLS 0.20 ASPECT-8B (RL) 0.8200.6690.6470.6590.706 Qwen3-VL-8B 0.4000.3450.4100.2500.206 Gemini-3.1-Pro 0.7050.5920.5110.4770.500 GPT-5.5 0.5830.6620.5540.5680.647 00.250.50.751 (b) RAWR Lizard 0.50 PUMA 0.22 PanNuke 0.25 CoNSeP 0.25 NuCLS 0.20 ASPECT-8B (RL) 0.538 n=326 0.364 n=94 0.400 n=90 0.474 n=29 0.656 n=24 Qwen3-VL-8B 0.855 n=55 0.886 n=22 0.821 n=21 -- n=4 -- n=4 Gemini-3.1-Pro 0.738 n=280 0.584 n=83 0.574 n=71 0.679 n=21 0.632 n=17 GPT-5.5 0.816 n=231 0.673 n=94 0.701 n=77 0.696 n=23 0.739 n=22 10.750.50.250 Pixel sizes below source names are in micrometers per pixel; gray RAWR cells have fewer than 10 eligible questions. Figure E.2: Results by source dataset on PathoVernier. Column labels include native pixel sizes in μm/pixel. Darker shading indicates better performance within each panel. In the RAWR panel, n denotes the number of correct answers with complete count observations; estimates with n < 10 are withheld. Source and imaging scale are evaluated jointly. E.2RESULTS BY DATA SOURCE AND IMAGING SCALE Figure E.2 breaks down performance across the five source datasets and their native pixel sizes. ASPECT achieves the highest Acc on every source, showing that its gains extend beyond Lizard, which contributes more than half of PathoVernier. On PUMA, ASPECT and GPT-5.5 obtain similar Acc (0.669 versus 0.662), yet ASPECT achieves substantially lower RAWR (0.364 versus 0.673). This comparison illustrates how similar answer accuracy can conceal differences in the quality of reported counts. ASPECT achieves the lowest RAWR on four sources; NuCLS is the exception, where Gemini-3.1-Pro has lower RAWR (0.632 versus 0.656) despite lower Acc. The source-level results thus reveal both the broad improvement in task accuracy and the variation in the accuracy of its supporting measurements. E.3STATISTICAL UNCERTAINTY We quantify test-set sampling uncertainty using 10,000 paired bootstrap resamples of the 553 PathoVernier patches, stratified by source dataset, with random seed 42. Within each source, we sample its original number of patches with replacement and retain all associated questions, includ- 30 ing sampling multiplicities. All models share the same resamples. We recompute each metric and its denominator according to Appendix C.2, retaining API failures as incorrect answers for Acc and using each model’s own eligible samples for CA and RAWR. We report the 2.5th and 97.5th percentiles as 95% confidence intervals. As shown in Table E.12, all six paired improvement intervals lie above zero. Against Gemini-3.1- Pro, ASPECT improves Acc by 0.120, with a 95% CI of [0.077, 0.162]. The positive intervals for CA gains and RAWR reductions further support improvements in both count-derived conclusions and the reported quantities behind correct answers. These comparisons use the paired difference distributions, rather than overlap between individual model intervals. Table E.12: Model scores and paired improvements on PathoVernier with 95% bootstrap confidence intervals. Improvements are ASPECT minus baseline for Acc and CA, and baseline minus ASPECT for RAWR; positive values favor ASPECT. ModelAcc↑CA↑RAWR↓ Qwen3-VL-8B0.374 [0.341, 0.408]0.532 [0.471, 0.595]0.854 [0.801, 0.901] Gemini-3.1-Pro0.626 [0.588, 0.663]0.715 [0.681, 0.749]0.680 [0.651, 0.708] GPT-5.50.594 [0.558, 0.631]0.652 [0.616, 0.689]0.756 [0.728, 0.784] ASPECT-8B0.746 [0.714, 0.776]0.807 [0.777, 0.836]0.489 [0.460, 0.518] ComparisonAcc gainCA gainRAWR reduction ASPECT vs Qwen3-VL-8B+0.372 [+0.327, +0.415]+0.275 [+0.207, +0.342]+0.365 [+0.305, +0.420] ASPECT vs Gemini-3.1-Pro+0.120 [+0.077, +0.162]+0.092 [+0.052, +0.132]+0.191 [+0.152, +0.229] E.4SENSITIVITY TO COUNT TOLERANCE We vary the relative count tolerance r over0, 0.05,..., 0.30 while retaining the absolute tolerance floor of one nucleus. A reported count c agrees with its reference count q when |c− q|≤ max1,rq.(20) We recompute the count-agreement score and RAWR at each tolerance, keeping each model’s correct-answer subset with complete count observations fixed. The main results correspond to r = 0.10, rather than an average over this tolerance grid. At r = 0, the criterion still permits an absolute counting error of one. Figure E.3 shows that ASPECT achieves the lowest RAWR at all seven tested tolerances. Even under the strictest setting, its RAWR is 0.578, compared with 0.725 for Gemini-3.1-Pro. Their gap increases from 0.147 at r = 0 to 0.256 at r = 0.30. The consistent ranking across the scan shows that ASPECT’s advantage in reported count accuracy extends beyond the default tolerance. FCOMPLETE CASE STUDIES AND FAILURE ANALYSIS F.1COMPLETE RESPONSES FOR MAIN-TEXT EXAMPLES Figures G.4 and G.5 provide the complete responses for the two examples discussed in the main- text qualitative comparison. Each figure includes the question, reference answer, and outputs from ASPECT, Qwen3-VL-8B, Gemini-3.1-Pro, and Patho-R1-7B. PathoVernier: Correct answers and their quantitative basis. The reference answer selects the lower-right region and assigns a low lymphocyte proportion, based on 7 lymphocytes among 29 nuclei. Gemini reaches the correct answer using inaccurate counts of 12 and 47: their ratio falls in the same band despite the counting errors. Qwen3-VL-8B also selects the correct region, but its estimate of 20 out of 50 nuclei leads to the incorrect mid band. ASPECT recovers the reference counts for the selected region and derives the correct low band from 7/29 ≈ 0.24. This example illustrates why final-answer accuracy alone cannot distinguish accurate measurements from counting errors that preserve the answer category. 31 ASPECT-8B (RL)Qwen3-VL-8BGemini-3.1-ProGPT-5.5 0.0 0.2 0.4 0.6 0.8 1.0 0.000.050.100.150.200.250.30 Default r = 0.10 Relative count tolerance r RAWR (lower is better) Figure E.3: RAWR under varying relative count tolerance r, with an absolute tolerance floor of one nucleus. The dashed line marks the main evaluation setting, r = 0.10. Each model’s eligible sample set remains fixed throughout the scan. Lower is better. PathCLS: Tissue and lesion classification. For the skin histology example, ASPECT selects the reference class, squamous cell carcinoma (N), while Qwen3-VL-8B selects basal cell carcinoma (M) and Gemini selects melanoma (O). Patho-R1 describes non-tumor epidermis in its reasoning but outputs K, which denotes non-tumor sweat glands. The complete responses expose differences in both class selection and the correspondence between the written rationale and the selected option. F.2ADDITIONAL PATHOVERNIER EXAMPLES Figures G.6–G.8 supplement the multi-step example in Appendix F.1 with region comparison, cell- type comparison, and region selection. The complete responses show how count estimates and their subsequent interpretation affect the final answer. Region comparison. In Figure G.6, the reference tumor counts are 15 in the top half and 2 in the bottom half, giving the answer muchmore. ASPECT reports 16 and 3 and reaches the same com- parison category. Gemini reports 12 and 8, leading to the incorrect comparable answer. Qwen3- VL-8B exhibits a further discrepancy: its reasoning uses 6.5/3.5, but its final reported counts are 6 and 3. The latter imply muchmore at the stated threshold, although its final answer remains comparable. Cell-type comparison. Figure G.7 asks whether stromal and inflammatory nuclei have compa- rable abundance. The reference counts are 57 and 44. Gemini estimates 2 and 65 and answers much fewer, whereas Qwen3-VL-8B reports 17 and 6 and answers muchmore. ASPECT re- ports 54 and 50 and correctly selects comparable. The opposing baseline conclusions arise from markedly different estimates of the same two cell populations. Region selection. In Figure G.8, the reference stromal counts across four equal-area horizon- tal bands are (16, 4, 1, 1), making the topmost band, r1, the correct answer. ASPECT reports (16, 5, 2, 1) and selects r1, recovering the dominant concentration in the top band. Patho-R1’s reasoning also points to r1, although its reported counts of (2, 0, 0, 1) substantially underestimate the stromal population. Gemini and Qwen3-VL-8B instead select r2 and r4, respectively. This example further separates correct regional selection from accurate measurement. F.3FAILURE ANALYSIS Figure G.9 presents two ASPECT failures in which the final answers follow the reported counts, but those counts misrepresent the regional composition. 32 Incorrect regional distribution despite a near-correct total. In the CoNSeP example (right), ASPECT reports 47 tumor nuclei across the image, close to the reference total of 46. Its regional counts, however, are (0, 4, 7, 36) rather than (0, 8, 28, 10). Selecting the largest reported count con- sequently yields r4 instead of the reference region r3. The response is internally consistent, but the near-correct total does not ensure accurate regional measurement. Across 43 failed strip-based density-selection questions, 27 (62.8%) select a band adjacent to the reference band. Denominator overestimation after correct region selection. In the Lizard example (left), AS- PECT correctly identifies the upper-right quadrant and estimates 19 stromal nuclei, close to the reference count of 18. It nevertheless reports 53 total nuclei in that region instead of 31, reducing the estimated stromal proportion from the reference 18/31 ≈ 0.58 to 19/53 ≈ 0.36. This changes the answer from mid to low at the 0.50 threshold. The case passes the region-only CA check but fails Acc, which evaluates the complete region-and-band answer. Among 44 multi-step questions with correct region selection but an incorrect proportion band, the median predicted-to-reference denominator ratio is 1.72, close to this example’s 53/31≈ 1.71. GLIMITATIONS AND FUTURE WORK Spatial and clinical scope. This study focuses on quantitative visual reasoning within H&E image patches. Pathological interpretation also draws on tissue architecture over larger fields, relationships between spatially separated regions, and patient-level context. Extending the present framework to whole-slide analysis would require coordinating local measurements with region selection and broader tissue organization. An important direction is to study how verifiable observations at differ- ent spatial scales can support slide-level assessments and clinically relevant endpoints. Task and annotation scope. PathoVernier evaluates cell composition through a shared taxonomy and questions with explicit quantitative reference answers. This formulation makes intermediate measurements assessable across datasets, while its scope remains tied to the categories and quanti- ties supported by the source annotations. Morphological attributes, interactions between cell popu- lations, and richer descriptions of tissue organization offer complementary dimensions of pathology understanding. Future work could develop reference annotations and evaluation criteria for these dimensions, extending the connection between visual evidence and model conclusions beyond the count-based tasks studied here. 33 **Question:**Determine the region whose lymphocytic nuclear density is greatest, then state the proportion of that region's nuclei that are lymphocytic. For the share, use these bands: 'low' below 0.35, 'high' at or above 0.55, otherwise 'mid’. **GT per-region lymphocyte counts (asked = lymphocyte):** `'upperleft': 2, 'upperright': 0, 'lowerleft': 2, 'lowerright': 7` **GT lower-right region reference:** lymphocyte `_num=7` / **total nuclei `_den=29`** → lymphocyte share = 7/29 ≈ 0.24 → band `low`(< 0.35) **Gold answer:** `lowerright:low` · options: ['lowerright:low', 'lowerleft:high', 'upperleft:mid', 'lowerright:high', 'lowerright:mid', 'upperright:high', 'lowerleft:low', 'lowerleft:mid', 'upperright:mid', 'upperleft:high', 'upperright:low', 'upperleft:low'] To determine the region with the highest lymphocytic nuclear density, we first divide the image into four quadrants and count the lymphocytes (small, dark, round nuclei) in each. 1. **Upper-left (UL):** This quadrant contains mostly empty space and some large tumor cells. There is approximately 1 lymphocyte visible. 2. **Upper-right (UR):** This quadrant is densely packed with large tumor cells. There are about 2 lymphocytes visible. 3. **Lower-left (L):** This quadrant contains tumor cells, pink keratin/osteoid material, and a scattering of lymphocytes. We can count about 8 lymphocytes here. 4. **Lower-right (LR):** This quadrant contains many tumor cells and a noticeable cluster of lymphocytes, particularly in the center-left and bottom areas of the quadrant. We can count about 12 lymphocytes here. Comparing the counts, the lower-right quadrant has the highest number of lymphocytes (12), and thus the highest lymphocytic nuclear density. Next, we determine the proportion of nuclei in the lower-right quadrant that are lymphocytic. - Number of lymphocytes in LR: ~12 - Number of other nuclei (mostly large tumor cells) in LR: ~35 - Total nuclei in LR: 12 + 35 = 47 - Proportion: 12 / 47 ≈ 0.255 Since 0.255 is below the 0.35 threshold, the share is considered 'low'. <observe>"tumor": 110, "lymphocyte": 23, "upperleft": 1, "upperright": 2, "lowerleft": 8, "lowerright": 12, "_num": 12, "_den": 47 </observe> FINAL: lowerright:low Gemini-3.1-Pro <think>The image shows a patch with distinct nuclear density variations across quadrants. In the upper left, nuclei are predominantly epithelial-like with scattered lymphocytes, while the upper right contains clusters of small, round lymphocytes. The lower right quadrant displays a dense infiltrate of lymphocytes mixed with stromal cells. The share specifies two criteria: (1) the highest lymphocytic density region (upper right) and (2) the proportion of lymphocytes in that region. Using the defined bands, the upper right exceeds 0.55 (high), and the proportion of lymphocytes in this area is estimated visually as >50%. The lower right is high but not the highest density region. Thus, the upper right meets both criteria.</think> <answer> **Step 1: Screening** - **Structural Assessment**: The upper right quadrant shows dense lymphocytic infiltration. - **Nuclear Density**: High lymphocytic density in the upper right. **Step 2: Key Feature Identification** - **Nuclear Atypia**: Lymphocytes are small and uniform, lacking atypia. - **Mitotic Activity**: No mitotic figures observed. **Step 3: Simplified Report Output** - The upper right quadrant has the highest lymphocytic density (>50% of nuclei). - Proportion of lymphocytes in this region is high (likely >70% based on visual estimation). </answer> <observe> [ ("upperleft": 0.35, "upperright": 0.55, "lowerleft": 0.55, "lowerright": 0.55) ] num: 0.75 , den: 0.55 </observe> <answer> **Step 5: Diagnostic Report** - **Tissue Location**: Upper right quadrant. - **Key Findings**: High lymphocytic density (≥70% nuclei in this region). - **Classification**: Atypical hyperplasia/suspected malignancy. Recommend immunohistochemistry (CD3/CD20) to confirm lymphocyte subtype. </answer> Source: PathoVernier Qwen3-VL-8B <think> the patch feature of the image is <|anchor_start|><|uni_pad|><|uni_pad|><|uni_pad|><|uni_pad|><|uni_pad|><|uni_pad| ><|uni_pad|><|uni_pad|><|anchor_end|>, and the cell composition of the image is <|anchor_start|><|cell_pad|><|cell_pad|><|cell_pad|><|cell_pad|><|cell_pad|><|cell_pa d|><|anchor_end|>. </think> <observe>A moderately cellular field shows gland-forming epithelial structures within fibrous stroma, with predominance of crowded elongated hyperchromatic epithelial nuclei and scattered stromal spindle cells accompanied by sparse lymphocytes. "stromal_like": 17, "lymphocyte": 10, "epithelial_like": 46, "asked": "lymphocytic nuclei — first locate their densest region, then give their share of all nuclei there", "upperleft": 1, "upperright": 1, "lowerleft": 1, "lowerright": 7, "_num": 7, "_den": 29</observe> <answer> Lymphocytic nuclei are most concentrated in the lower-right quadrant, with 7 compared with 1, 1, 1 in the other regions. This quadrant contains 29 nuclei overall, yielding a low lymphocytic share of 7/29 = 0.24. FINAL: lowerright:low </answer> I need to analyze the H&E-stained histology image to determine which quadrant has the greatest lymphocytic nuclear density and then calculate the proportion of nuclei in that region that are lymphocytic. Step 1: Identify the quadrants. -Upper-left (UL) - Upper-right (UR) - Lower-left (L) - Lower-right (LR) Step 2: Visually inspect each quadrant for lymphocytes. Lymphocytes are small, round, darkly stained nuclei. I need to count them relative to all nuclei in each quadrant. Looking at the image: - Upper-left quadrant: Contains mostly epithelial-like structures (pinkish, glandular) with some scattered small nuclei. Not many lymphocytes. - Upper-right quadrant: Similar to upper-left, mostly epithelial structures with sparse small nuclei. No obvious lymphocytic clusters. - Lower-left quadrant: Contains a mix of epithelial structures and some stromal areas. There are a few small, dark nuclei that could be lymphocytes, but not concentrated. -Lower-right quadrant: This quadrant has a distinct area with many small, dark, round nuclei clustered together — this is characteristic of lymphocytes. There is a clear lymphocytic infiltrate here. So, the lower-right quadrant has the greatest lymphocytic nuclear density. Step 3: Count lymphocytic nuclei and total nuclei in the lower-right quadrant. - Lymphocytic nuclei: I can count approximately 15-20 small, dark, round nuclei. -Total nuclei: In the same region, there are also epithelial nuclei (larger, more oval) and possibly some stromal nuclei. I can count approximately 40-50 total nuclei. Step 4: Calculate the proportion. Proportion = lymphocytic nuclei / total nuclei ≈ 20 / 50 = 0.4 Step 5: Classify the proportion. The proportion is 0.4, which is between 0.35 and 0.55, so it falls in the 'mid' category. Step 6: Report the required quantities. - upperleft: 0 (no significant lymphocytic density) - upperright: 0 (no significant lymphocytic density) - lowerleft: 0 (very low lymphocytic density) - lowerright: 20 (lymphocytic nuclei) / 50 (total nuclei) = 0.4 → 'mid' - _num: 20 -_den: 50 Step 7: Choose the correct token. The token must be for the lower-right quadrant with 'mid' classification. The options include "lowerright:mid". FINAL: lowerright:mid Patho-R1-7B ASPECT-8B Figure G.4: Complete responses for the PathoVernier multi-step example. Gemini reaches the cor- rect answer with inaccurate counts, whereas ASPECT recovers the reference counts of 7 lympho- cytes among 29 nuclei in the selected region. 34 **Question:** This is an image of a skin histology sample. Which category can be determined based on observations made from the image? **Options:** - A) Non-tumor chondral tissue - B) Non-tumor dermis - C) Non-tumor elastosis - D) Non-tumor epidermis - E) Non-tumor hair follicle - F) Non-tumor skeletal muscle - G) Non-tumor necrosis - H) Non-tumor nerves - I) Non-tumor sebaceous glands - J) Non-tumor subcutis - K) Non-tumor sweat glands - L) Non-tumor vessel - M) Tumor epithelial basal cell carcinoma - N) Tumor epithelial squamous cell carcinoma - O) Tumor melanoma - P) Tumor naevus **Gold answer:** `N` = Tumor epithelial squamous cell carcinoma The image displays a highly cellular neoplasm characterized by cells with marked nuclear pleomorphism, large vesicular nuclei, and prominent nucleoli. These cytological features, in the absence of clear squamous di!erentiation such as keratinization or intercellular bridges, are highly characteristic of melanoma. FINAL: O Gemini-3.1-Pro <think>First, I assess the tissue architecture and cellular features. The image shows a well-organized arrangement of stratified squamous epithelium with intact basal, spinous, granular, and cornified layers. No nuclear atypia, mitotic figures, or infiltrative growth are observed. The dermal-epidermal junction is preserved, and the stroma lacks fibrosis or inflammation. These features align with normal epidermal structure. Since there is no evidence of cellular disarray, atypia, or tumor-specific patterns (e.g., nests of basaloid cells in basal cell carcinoma or keratin pearls in squamous cell carcinoma), the tissue is classified as non-neoplastic. The correct answer is the option describing non-tumor epidermis, as the image matches its histological characteristics.</think> <answer>K</answer> Source: PathCLS <think> the patch feature of the image is <|uni_pad|><|uni_pad|><|uni_pad|><|uni_pad|><|uni_pad|><|uni_pad|><|uni_pad|><|u ni_pad|>, and the cell composition of the image is <|cell_pad|><|cell_pad|><|cell_pad|><|cell_pad|><|cell_pad|><|cell_pad|>. </think> <observe>The field is highly cellular with sheets of pleomorphic hyperchromatic nuclei and scant intervening stroma, lacking the organized architecture of normal skin compartments. "tumor": 38, "asked": "tumor epithelial basal cell carcinoma, squamous cell carcinoma, melanoma, or naevus", "tumor_benign": 38, "stromal_like": 1, "lymphocyte": 1, "asked_tumor": "Yes"</observe> <answer> The composition shows tumour=38, indicating a tumor-cell field. That exclusive dominance supports classification in the tumor category rather than any non- tumor skin structure, matching N) Tumor epithelial squamous cell carcinoma. FINAL: N </answer> The image shows a histology section with densely packed, irregularly shaped cells with hyperchromatic nuclei and some mitotic figures, embedded in a somewhat disorganized stroma. There is no clear stratified squamous epithelium, no hair follicle structure, no sebaceous gland lobules, no sweat gland ducts, no nerve bundles, no muscle fibers, no cartilage, no necrotic debris, and no vascular lumen. The tissue appears to be a neoplastic lesion rather than normal tissue, given the cellular atypia and disorganization. Among the tumor options, basal cell carcinoma (M) typically shows basaloid cells with peripheral palisading and stromal invasion, often with a “pearly” appearance clinically, but this image lacks the classic basaloid pattern. Squamous cell carcinoma (N) often shows keratin pearls and intercellular bridges — not visible here. Melanoma (O) would show melanin pigment and atypical melanocytes — not evident. Naevus (P) is benign and usually shows regular melanocytes in the epidermis or dermis — not matching. The most likely interpretation is that this is a **tumor epithelial basal cell carcinoma**, as the cells are small, round to oval, with hyperchromatic nuclei and some mitotic activity, and they appear to be invading the surrounding stroma — consistent with basal cell carcinoma histology, especially in a skin sample. FINAL: M Patho-R1-7B ASPECT-8B Qwen3-VL-8B Figure G.5: Complete responses for the PathCLS example. ASPECT selects the reference class, squamous cell carcinoma (N), while the three comparison models select other classes. **Question:** Taking the bottom half as the baseline, what is the relative spatial frequency of tumour nuclei in the top half? Use these bands: 'much_more' if the ratio is at least 2.0, 'much_fewer' if it is at most 1/2.0, otherwise 'comparable'. **GT counts (asked = tumor):** `'tumor_top': 15, 'tumor_bottom': 2` **Gold answer:** `much_more` · options: ['much_fewer', 'comparable', 'much_more'] To determine the relative spatial frequency of tumor nuclei in the top half compared to the bottom half, I will first identify and count the tumor nuclei in each region. Tumor nuclei in this image are characterized by their large size, pleomorphism, and prominent nucleoli. In the top half of the image, I observe a cluster of tumor nuclei on the right side and another smaller group on the left. Counting these yields approximately 12 tumor nuclei. In the bottom half of the image, there are scattered tumor nuclei on both the left and right sides. Counting these yields approximately 8 tumor nuclei. Comparing the two halves, the ratio of tumor nuclei in the top half to the bottom half is 12 / 8 = 1.5. Since this ratio is between 0.5 and 2.0, the spatial frequency is considered comparable. <observe> "tumor": 20, "tumor_top": 12, "tumor_bottom": 8 </observe> FINAL: comparable Gemini-3.1-Pro <think>The image shows tumor nuclei in the lower half and epithelial nuclei in the upper half. To assess spatial frequency, I first confirm the tumor nuclei are distinct from epithelial nuclei based on their size, chromatin pattern, and spatial distribution. The tumor nuclei occupy a smaller proportion of the field compared to the epithelial nuclei, which aligns with the baseline comparison. The ratio of tumor nuclei to epithelial nuclei in the upper half is lower than the baseline (1:2.0), falling into the 'much fewer' category. This is supported by the absence of tumor cells in the upper half and their sparse distribution in the lower half. No mitotic figures or necrosis are noted, but the spatial distribution remains the key diagnostic feature here.</think> <answer>much fewer</answer> Source: PathoVernier <think> the patch feature of the image is <|anchor_start|><|uni_pad|><|uni_pad|><|uni_pad|><|uni_pad|><|uni_pad|><|uni_pad|><|uni_pad|><|uni_pad|><|anchor_end|>, and the cell composition of the image is <|anchor_start|><|cell_pad|><|cell_pad|><|cell_pad|><|cell_pad|><|cell_pad|><|cell_pad|><|anchor_end|>. </think> <observe>A moderately cellular field shows fibrous stroma infiltrated by cohesive atypical tumour cells with enlarged hyperchromatic pleomorphic nuclei and only scant inflammatory cells. "tumor": 19, "stromal_like": 9, "inflammatory": 1, "asked": "tumour nuclei, compared between the top half and the bottom half", "tumor_top": 16, "tumor_bottom": 3</observe> <answer> There are 16 tumour nuclei in the top half versus 3 in the bottom half. Since the areas are equal, the density ratio is 16/3 = 5.33 and is substantially more than the 2.0x threshold. FINAL: much_more </answer> I need to analyze the image and count the tumor nuclei in the top half and bottom half, then calculate the ratio to determine the relative spatial frequency. Step 1: Identify tumor nuclei. - Tumor nuclei are typically larger, more irregularly shaped, and often have prominent nucleoli compared to normal epithelial or stromal nuclei. - In this image, I can see several large, dark, irregular nuclei that appear to be tumor cells. They are clustered in the upper right and scattered in the upper left. Step 2: Count tumor nuclei in the top half. - The top half of the image (from the horizontal midline to the top edge) contains approximately 6-7 clearly identifiable tumor nuclei. Step 3: Count tumor nuclei in the bottom half. - The bottom half (from the horizontal midline to the bottom edge) contains approximately 3-4 clearly identifiable tumor nuclei. Step 4: Calculate the ratio. - Ratio = tumor_top / tumor_bottom = 6.5 / 3.5 ≈ 1.86 - Since 1.86 is less than 2.0, it does not qualify as "much_more". - Since 1.86 is greater than 1/2.0 (0.5), it does not qualify as "much_fewer". - Therefore, the ratio is "comparable". Step 5: Report the quantities as requested. - tumor_top = 6 - tumor_bottom = 3 Step 6: Output the final token. <observe>"tumor_top": 6, "tumor_bottom": 3, "relative_frequency": "comparable"</observe> FINAL: comparable Patho-R1-7B ASPECT-8B Qwen3-VL-8B Figure G.6: Regional comparison of tumor nuclei. ASPECT selects the correct ratio band; the baseline responses illustrate errors in count estimation and in translating reported counts into the final answer. 35 **Question:** Within the entire image, what is the numeric ratio of stromal nuclei to inflammatory nuclei? Use these bands: 'much_more' if the ratio is at least 2.0, 'much_fewer' if it is at most 1/2.0, otherwise 'comparable'. **GT counts (asked = stromal_like/inflammatory):** `'stromal_like': 57, 'inflammatory': 44` **Gold answer:** `comparable` · options: ['much_fewer', 'comparable', 'much_more'] Based on a careful examination of the H&E-stained image: 1. **Identify and count inflammatory nuclei:** The image contains a dense infiltrate of small, dark, round nuclei, which are characteristic of lymphocytes (inflammatory cells). These are distributed throughout the image, particularly concentrated in the upper right and lower left areas. I count approximately 65 inflammatory nuclei. 2. **Identify and count stromal nuclei:** Stromal cells typically have elongated, spindle-shaped nuclei. In this specific field of view, there are very few cells that fit this description. Most of the non- inflammatory cells are larger, paler, and more round/oval, suggesting they are epithelial or tumor cells. I can identify perhaps 2 nuclei that might be stromal. 3. **Calculate the ratio:** The number of stromal nuclei (~2) is significantly less than the number of inflammatory nuclei (~65). The ratio is approximately 2/65, which is well below the 0.5 threshold. Therefore, the number of stromal nuclei is much fewer than the number of inflammatory nuclei. <observe> “epithelial_like”: 12,“stromal_like”: 2,"inflammatory": 65 </observe> FINAL: much_fewer Gemini-3.1-Pro <think>The image shows stromal-like nuclei (small, spindle-shaped, and evenly distributed) and inflammatory nuclei (larger, rounder, with condensed chromatin). Stromal nuclei are densely packed but maintain a regular arrangement, while inflammatory nuclei are scattered but not overwhelming. The ratio of stromal to inflammatory nuclei is approximately 2:1, falling within the 'much more' category. This aligns with the observation that stromal nuclei dominate the field, though inflammatory cells are present. The question’s phrasing ("much more") matches the observed ratio, and the answer must reflect this quantitative assessment.</think> <answer>much_more</answer> Source: PathoVernier <think> the patch feature of the image is <|anchor_start|><|uni_pad|><|uni_pad|><|uni_pad|><|uni_pad|><|uni_pad|><|uni_pad|><|uni_pad|><|uni_pad|><|anchor_end|>, and the cell composition of the image is <|anchor_start|><|cell_pad|><|cell_pad|><|cell_pad|><|cell_pad|><|cell_pad|><|cell_pad|><|anchor_end|>. </think> <observe>Cellular fibroinflammatory stroma with scattered small round hyperchromatic inflammatory nuclei and elongated spindle-shaped stromal nuclei in a loose collagenous background. "stromal_like": 54, "inflammatory": 50, "asked": "stromal versus inflammatory nuclei, counted over the whole image"</observe> <answer> Whole-image review identifies 54 stromal nuclei and 50 inflammatory nuclei, for a ratio of 54/50 = 1.08 that is comparable against the 2.0x threshold. FINAL: comparable </answer> I need to analyze the H&E-stained histology image to count stromal-like nuclei and inflammatory nuclei, then compute their ratio. Step 1: Identify and count stromal-like nuclei. - Stromal nuclei are typically larger, more irregular, and less densely packed than epithelial nuclei. They are often found in the connective tissue framework. - In this image, I can see a background of cells with larger, more variable nuclei that are not forming tight epithelial sheets. These are likely stromal cells. - I count approximately 15-20 stromal-like nuclei in the image. Step 2: Identify and count inflammatory nuclei. - Inflammatory nuclei are typically smaller, rounder, and more densely packed, often seen in clusters or infiltrates. - I can see some small, darkly stained nuclei that are scattered throughout the image, likely representing inflammatory cells (e.g., lymphocytes). - I count approximately 5-8 inflammatory nuclei. Step 3: Compute the ratio. - Ratio = stromal_like / inflammatory = 17.5 / 6.5 ≈ 2.7 - Since the ratio is at least 2.0, the correct token is 'much_more'. <observe> stromal_like: 17 inflammatory: 6 </observe> FINAL: much_more Patho-R1-7B ASPECT-8B Qwen3-VL-8B Figure G.7: Whole-image comparison of stromal and inflammatory nuclei. ASPECT correctly se- lects comparable, while Gemini and Qwen3-VL-8B reach opposite incorrect conclusions from their count estimates. 36 **Question:** In which region is the spatial frequency of stromal nuclei at its apex? (r1 = the topmost horizontal band; r2 = the second horizontal band from the top; r3 = the third horizontal band from the top; r4 = the bottom-most horizontal band). **GT counts (asked = stromal_like):** `'stromal_like_r1': 16, 'stromal_like_r2': 4, 'stromal_like_r3': 1, 'stromal_like_r4': 1` **Gold answer:** `r1` · options: ['r1', 'r3', 'r4', 'r2'] To determine the region with the highest spatial frequency of stromal nuclei, we first need to identify and classify the di!erent types of nuclei visible in the image based on their morphology: 1. **Epithelial-like nuclei**: These are large, oval, and vesicular, forming the prominent glandular structures that dominate the center and upper portions of the image. 2. **Lymphocytes**: These are small, round, and dark, densely packed in the lamina propria, particularly on the top left. 3. **Inflammatory nuclei**: These include neutrophils (with lobated nuclei) and are mostly found in the loose tissue at the bottom (r4) and lower right (r3, r4). 4. **Stromal-like nuclei**: These are typically elongated or spindle-shaped (e.g., fibroblasts or endothelial cells). They are relatively rare in this field of view, as the stroma is heavily infiltrated by immune cells. Now, let's scan each horizontal band (r1 to r4) specifically for these elongated, spindle- shaped stromal nuclei: * **r1 (top 25%)**: The left side is packed with round lymphocytes. The middle is epithelium. On the far right edge, there are about 2 distinct spindle-shaped nuclei. * **r2 (25-50%)**: The left side still contains lymphocytes. The middle is epithelium. On the right edge, there are about 3 distinct spindle-shaped nuclei. * **r3 (50-75%)**: The left and middle are epithelium. The right side contains loose inflammatory cells (mostly round or lobated). No clear spindle cells are visible. * **r4 (bottom 25%)**: This region contains loose, edematous tissue or mucus with clusters of inflammatory cells (neutrophils and lymphocytes) on the left and right. No clear spindle-shaped stromal cells are identifiable; the nuclei here are predominantly round or lobated. Comparing the counts, region r2 contains the highest number of identifiable stromal (spindle-shaped) nuclei. <observe> "epithelial_like_count": 130, "lymphocyte_count": 60, "inflammatory_count": 35, "stromal_like_count": 5, "stromal_like_r1": 2, "stromal_like_r2": 3, "stromal_like_r3": 0, "stromal_like_r4": 0 </observe> FINAL: r2 Gemini-3.1-Pro <think>The image shows stromal-like nuclei localized to specific horizontal bands. Stromal-like nuclei (r1) are concentrated in the topmost horizontal band (r1), while stromal-like nuclei (r2) occupy the second band from the top (r2). Stromal-like nuclei (r3) are found in the third band from the top (r3), and stromal-like nuclei (r4) reside in the bottom horizontal band (r4). The question requires reporting the band number for each stromal-like nucleus type. Since the image does not show stromal-like nuclei in bands r2 or r3, these are not reported. The stromal-like nuclei (r1) are explicitly noted in the topmost band, and stromal-like nuclei (r4) are present in the bottom band. The stromal- like nuclei (r2) and (r3) are not visible in this patch, so they are excluded. The final answer aligns with the observed distribution.</think> <answer> [ "stromal_like_r1": 2, "stromal_like_r2": 0, "stromal_like_r3": 0, "stromal_like_r4": 1 ] </answer> Source: PathoVernier Qwen3-VL-8B <think> the patch feature of the image is <|anchor_start|><|uni_pad|><|uni_pad|><|uni_pad|><|uni_pad|><|uni_pad|><|uni_pad| ><|uni_pad|><|uni_pad|><|anchor_end|>, and the cell composition of the image is <|anchor_start|><|cell_pad|><|cell_pad|><|cell_pad|><|cell_pad|><|cell_pad|><|cell_pa d|><|anchor_end|>. </think> <observe>A moderately cellular field shows glandular mucosal epithelium with crowded elongated hyperchromatic nuclei and goblet cell change, set in inflamed stroma rich in small dark lymphocytes. "stromal_like": 24, "lymphocyte": 54, "inflammatory_other": 9, "epithelial_like": 93, "asked": "stromal nuclei, compared across 4 equal-area regions", "stromal_like_r1": 16, "stromal_like_r2": 5, "stromal_like_r3": 2, "stromal_like_r4": 1</observe> <answer> The regions have equal area, so comparing densities is the same as comparing counts. The topmost horizontal band holds 16 stromal nuclei, against 5 in the second horizontal band from the top, 2 in the third horizontal band from the top, 1 in the bottom- most horizontal band. The count in r1 is the highest. FINAL: r1 </answer> I need to analyze the image to count stromal-like nuclei in each of the four horizontal bands (r1 to r4) and determine where the spatial frequency (i.e., density) is highest. Step 1: Define the bands. - r1: topmost horizontal band (top 1/4 of the image) - r2: second horizontal band from the top (next 1/4) - r3: third horizontal band from the top (next 1/4) - r4: bottom-most horizontal band (bottom 1/4) Step 2: Identify stromal-like nuclei. Stromal nuclei are typically larger, less densely packed, and often located in the connective tissue between epithelial structures. In this image, the epithelial structures (likely glandular) are prominent in the upper and middle sections. The bottom section (r4) has a more di!use, less organized appearance with scattered nuclei that are stromal. Step 3: Count stromal-like nuclei in each band. - r1: The top band contains mostly epithelial structures with tightly packed nuclei. There are very few nuclei that appear stromal-like (perhaps 1-2 scattered ones). - r2: The second band also contains epithelial structures, but some nuclei in the interstitium may be stromal. Counting them, I see about 3-4 stromal-like nuclei. - r3: The third band has more stromal-like nuclei, especially in the lower part of this band. I can count about 5-6 stromal-like nuclei. - r4: The bottom band has the most stromal-like nuclei. There is a cluster of small, dark nuclei scattered throughout, totaling about 10-12 stromal-like nuclei. Step 4: Determine the region with the highest spatial frequency. The spatial frequency is highest in r4, as it contains the most stromal-like nuclei per unit area. <observe> stromal_like_r1: 1, stromal_like_r2: 3, stromal_like_r3: 5, stromal_like_r4: 10 </observe> FINAL: r4 Patho-R1-7BASPECT-8B Figure G.8: Stromal-density selection across four equal-area horizontal bands. ASPECT identifies the reference region, r1, while Gemini and Qwen3-VL-8B select r2 and r4, respectively. **Question:** Determine the region whose stromal nuclear density is greatest, then state the proportion of that region's nuclei that are stromal. (upperleft = the upper-left quadrant; upperright = the upper- right quadrant; lowerleft = the lower-left quadrant; lowerright = the lower-right quadrant) For the share, use these bands: 'low' below 0.50, 'high' at or above 0.85, otherwise 'mid'. **Gold answer:** `upperright:mid` Source: PathoVernier <think> the patch feature of the image is <|anchor_start|><|uni_pad|><|uni_pad|><|uni_pad|><|uni_pad|><|uni_pad|><|uni_p ad|><|uni_pad|><|uni_pad|><|anchor_end|>, and the cell composition of the image is <|anchor_start|><|cell_pad|><|cell_pad|><|cell_pad|><|cell_pad|><|cell_pad|><|cell_ pad|><|anchor_end|>. </think> <observe>A highly cellular epithelial-rich field shows crowded gland-forming architecture with predominant enlarged hyperchromatic pleomorphic epithelial nuclei and only scant stromal and rare lymphocytic nuclei. "stromal_like": 24, "lymphocyte": 5, "epithelial_like": 164, "asked": "stromal nuclei — first locate their densest region, then give their share of all nuclei there", "upperleft": 1, "upperright": 19, "lowerleft": 1, "lowerright": 5, "_num": 19, "_den": 53</observe> <answer> The upper-right quadrant leads in stromal nuclei with 19, while the other quadrants have 1, 1, 5. Out of 53 total nuclei in the region, the stromal share is 19/53 = 0. 3 6, w h i c h i s l o w . FINAL: upperright:low </answer> **Question:** What region demonstrates the peak spatial concentration of tumour nuclei? (r1 = the topmost horizontal band; r2 = the second horizontal band from the top; r3 = the third horizontal band from the top; r4 = the bottom-most horizontal band) **Gold answer:** `r3` Source: PathoVernier ASPECT-8B <think> the patch feature of the image is <|anchor_start|><|uni_pad|><|uni_pad|><|uni_pad|><|uni_pad|><|uni_pad|><|uni_p ad|><|uni_pad|><|uni_pad|><|anchor_end|>, and the cell composition of the image is <|anchor_start|><|cell_pad|><|cell_pad|><|cell_pad|><|cell_pad|><|cell_pad|><|cell_ pad|><|anchor_end|>. </think> <observe>A highly cellular field is composed predominantly of crowded atypical tumour nuclei with enlarged hyperchromatic pleomorphic features in a desmoplastic stroma, with only scant inflammatory cells. "tumor": 47, "stromal_like": 15, "inflammatory": 1, "asked": "tumour nuclei, compared across 4 equal-area regions", "tumor_r1": 0, "tumor_r2": 4, "tumor_r3": 7, "tumor_r4": 36</observe> <answer> The regions have equal area, so comparing densities is the same as comparing counts. The bottom-most horizontal band holds 36 tumour nuclei, against 7 in the third horizontal band from the top, 4 in the second horizontal band from the top, 0 in the topmost horizontal band. The count in r4 is the highest. FINAL: r4 </answer> ASPECT-8B Figure G.9: ASPECT failure cases. Left: correct region selection followed by an incorrect propor- tion band due to denominator overestimation. Right: a near-correct whole-image count accompanies an incorrect regional distribution and region choice. Both final answers are consistent with the re- ported counts. 37