{"id":"e1a68b6b-9591-4f48-9611-5b496ee7d6e0","arxiv_id":"2608.08648","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"AdaptivePath trains a reinforcement-learning navigator on pathologist-reviewed abnormality labels to acquire sparse multi-scale observations, then uses language models to interpret, deliberate, and adjudicate answers for whole-slide pathology questions.","lead":"AdaptivePath is a new AI system that actively chooses where to zoom in on gigapixel pathology slides, then answers questions by describing and checking the tissue evidence it saw. It beat existing pathology AI on several benchmarks and, in a blinded test, its selected image sequences helped pathologists reach 82.9% accuracy.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Question-agnostic abnormality ranking (Eq. 7-8) may discard evidence for questions whose answer requires morphology not labeled 'abnormal', leaving the central generalization claim unestablished.","rationale":"Agree with the reader's weakest assumption and conditional verdict. The concern is not that Eq. 7-8 is internally inconsistent; it is that the learned 'abnormality' construct may not cover the evidence classes that arbitrary clinical questions require. The paper's own low HER2/molecular accuracy, together with the conclusion that IHC/molecular findings are outside the system, makes the limitation concrete. A question-conditioned ranking test would decide whether the bottleneck is the Navigator or the frozen LM modules. Error bars and the missing CPathAgent comparison are also worth addressing, but they affect the strength of the SOTA claim rather than the architecture's core risk. Since the reader already conditioned acceptance on released artifacts and clearer evidence, my stress-test does not move the verdict.","tokens_in":11989,"tokens_out":7654,"duration_ms":88541,"concrete_test":"Run a controlled intervention on the existing frozen pipeline: replace Eq. 8's ranking with a question-conditioned score S(R) = AS(R) + lambda * cos_sim(CONCH(question), CONCH(region)) for each candidate region, with lambda tuned on the validation split of SlideBench-VQA and all other modules unchanged. Report per-category accuracy on the test set. If HER2 Expression and Molecular Subtype do not improve by more than about 5 points while tumor-type accuracy is stable, the abnormality-only ranking is not the bottleneck and the concern does not land. If they improve materially, the central claim would require a question-conditioned Navigator, supporting a conditional verdict.","verdict_should_be":"UNCHANGED","load_bearing_attack":"AdaptivePath's central claim is that learned abnormality-driven navigation plus explicit deliberation generalizes to arbitrary clinical questions. The Navigator is trained on binary normal/abnormal labels (Section 4.1) and at inference ranks candidate regions solely by AS(Rm) = z_T[abnormal] - z_T[normal] (Eq. 7), retaining only Top-K (Eq. 8). No question information enters until the Interpreter/Deliberator stage, after regions have already been pruned. If a question's decisive evidence is not 'abnormal' in the generic training sense (e.g., specific HER2/molecular-subtype morphology in H&E, or a negative/absence finding), the correct region can be removed at the 2.5x/5x stage and never observed. Table 1 is consistent with this failure mode: HER2 Expression (27.92) and Molecular Subtype (32.68) are far below Tumor Type (89.58), exactly the categories where generic abnormality is least likely to align with question-relevant morphology. The ROI-selection ablations (Table 3) show the Navigator beats random/PLIP/CONCH selection on aggregate accuracy, but aggregate improvement does not test sufficiency for arbitrary questions; it only shows the learned scorer is better than generic retrieval on benchmarks where relevant evidence tends to be abnormal.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper presents AdaptivePath, an agentic system for whole-slide pathology image reasoning. A Navigator trained on pathologist-reviewed abnormal-region labels uses PPO with alternating representation learning and consistency fine-tuning to decide where and at what magnification to acquire observations. At inference, candidate regions are hierarchically pruned using an abnormality score, and a frozen Patho-R1-7B Interpreter converts selected foci into question-conditioned morphology descriptions; a Qwen3-8B Deliberator and Arbiter audit evidence per option and integrate across magnifications. The paper reports state-of-the-art zero-shot results on SlideBench-VQA (60.81%), WSI-VQA (54.92%), and PathVQA (60.73%), 80.14% accuracy on six TCGA subtype classification cohorts, and a reader study in which pathologists achieve 82.9% accuracy using AdaptivePath-selected sequences. The central claim is that question-agnostic abnormality-driven active perception with explicit evidence deliberation provides an effective and traceable basis for WSI reasoning.","tokens_in":12233,"tokens_out":6246,"duration_ms":65736,"significance":"If correct, AdaptivePath is a meaningful step toward agentic WSI reasoning: it replaces heuristic region selection with a policy trained on pathologist-reviewed abnormality labels, supports multi-scale evidence acquisition under a fixed ROI budget, and makes the evidence chain explicit through interpretation, deliberation, and adjudication. The paper includes extensive comparisons, ablations of ROI selection, magnification, and budget, a two-configuration module study, and a blinded diagnostic utility test, and the code is released. The main strength is the combination of learned active perception with explicit evidence auditing, evaluated across both VQA and classification. However, the significance of the headline numbers depends on closing three gaps: confidence intervals around the benchmark differences, a demonstration that abnormality-only pruning retains question-relevant evidence for categories not aligned with generic abnormality (e.g., HER2 expression and molecular subtype), and clarity about how much of the navigation signal comes from the Patho-R1-7B-generated labels versus independent pathologist annotation.","major_comments":[{"comment":"Question-agnostic abnormality ranking can discard question-relevant evidence before the Interpreter sees it. The Navigator is trained only on binary normal/abnormal labels (Section 4.1), and candidate retention is based solely on AS(Rm); no question information enters until after top-K pruning. For questions whose decisive morphology is not 'abnormal' in the training sense (e.g., HER2 expression or molecular subtype morphology in H&E, or negative/absence findings), the correct region may be removed at the 2.5x/5x stage. Table 1 is consistent with this failure mode: HER2 Expression and Molecular Subtype accuracies are 27.92 and 32.68, far below Tumor Type at 89.58. The paper also lists question-aware reranking as future work in Section 5, which concedes the limitation. I ask the authors to report evidence recall in retained regions per question category, or to compare against a question-aware reranker; otherwise the generalization claim for arbitrary clinical questions is not established.","section":"Section 3.3.1, Eqs. (7)-(8)"},{"comment":"No error bars, confidence intervals, or numbers of independent runs are reported for the main benchmark results. Several state-of-the-art claims rest on small margins (e.g., 1.39 points over PathNavigate on SlideBench-VQA and 2.71 points on WSI-VQA), and Table 3 ablations also lack variance. Without bootstrapped confidence intervals or multi-seed evaluation, these differences cannot be distinguished from noise. Since the paper already computes paired t-tests for Figure 4, the same uncertainty reporting should be applied to the headline tables.","section":"Tables 1 and 2"},{"comment":"The abnormal-region labels are initially generated by Patho-R1-7B and later reviewed and corrected by pathologists, but the paper reports no label correction rate, inter-reviewer agreement, or proportion of labels changed. Because Patho-R1-7B is also the default Interpreter, the Navigator may be optimized toward the same model's notion of 'abnormal', and the end-to-end gain may partly reflect model self-consistency rather than independent pathology grounding. Please quantify the review process and include an ablation with labels from an independent source (e.g., manual annotation or a different model), or at least report results on a held-out set of fully manual labels.","section":"Section 4.1"}],"minor_comments":[{"comment":"The 'Overall' column needs its weighting formula; the unweighted average of the six AdaptivePath cohort accuracies is 78.95%, not 80.14%, so either the cohorts are weighted by size or the overall value is computed differently.","section":"Table 2"},{"comment":"The reader study's sample size, number of pathologists, and selection criteria for the SlideBench-VQA subset are not stated in the main text; please add them or point to a specific supplementary section.","section":"Section 4.6.1, Figure 3"},{"comment":"Please define how the aggregate heatmap H is normalized before computing D_JS and state whether H_geo is mapped back to the original coordinate system before aggregation.","section":"Section 3.2.3, Eq. (6)"},{"comment":"The Patch baseline samples 30 patches at 20x randomly; without fixing seeds or averaging over repeated samples, the reported patch-model scores are stochastic. Please report the sampling procedure in more detail.","section":"Section 4.4"},{"comment":"There are several typographical issues: 'Y ongbing Zhang' in the author block, irregular spacing in 'Quilt-LLaV A' and 'LLaV A', and inconsistent use of 'Patho-R1-7B' versus 'Patho-R1-7B' in a few places. A careful proofread is recommended.","section":"Throughout"},{"comment":"PathMMU results are mentioned only as being in the Supplementary Materials; a one-sentence summary in the main text would help readers who do not consult the supplement.","section":"Section 4.3"}],"recommendation":"major_revision","confidential_remarks":"The paper is in scope and the approach is interesting. The main risk is overclaiming generalization from question-agnostic abnormality scoring; a revision that quantifies evidence retention and adds uncertainty estimates would make the claims solid. I would not reject on the current evidence, but the missing confidence intervals and the label-source circularity need to be addressed before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The headline: this is a solid, well-tested systems paper. The genuinely new piece is a Navigator trained on pathologist-reviewed abnormality labels, then deployed at inference as a question-agnostic hierarchical region selector for WSI VQA and subtype classification. That combination—active perception grounded in pathology, plus explicit evidence deliberation—holds up empirically. I buy that the gains over PathNavigate (2.7 and 1.4 points) and especially over SlideChat on subtype classification (80.14 vs 65.29) are real; the ablations, the two model configurations, and the pathologist reader study all point the same way. The code being released is a plus.\n\nSoft spots, in proportion. First, the headline comparisons in Tables 1 and 2 have no error bars or repeated-run variance, so the smaller differences (e.g., 60.81 vs 59.42) might be noise. Second, CPathAgent, the closest method with learned navigation, is discussed in Related Work but never compared. That's a gap. Third, the stress-test concern looks legitimate: the Navigator ranks regions purely by abnormality score (Eq. 7-8), with no question information, so for questions whose decisive morphology isn't 'abnormal' in the generic sense—HER2 expression, molecular subtype—the right region could be pruned early. Table 1 shows exactly those categories low. The authors acknowledge question-aware reranking as future work, so it's a known limit rather than a hidden one. It means the title's 'agentic visual reasoning' is fair, but the claim of generalization to arbitrary clinical questions is over-broad.\n\nFourth, the 'zero-shot' framing is mostly fine (no downstream training) but slightly misleading because the Navigator was trained on a purpose-built abnormal-region dataset with labels initially produced by Patho-R1-7B—the same model family used as the Interpreter. The pathologist review reduces the circularity, but I'd want the label corrections documented.\n\nThis paper deserves a serious referee. The core mechanism is novel enough, the evaluation is unusually thorough, and the failure modes are identifiable. I'd accept it into review, with a request for statistical reporting, a CPathAgent comparison, and a sensitivity analysis of the top-K threshold under question-specific evidence.","headline":"A well-evaluated active-perception WSI system with a genuinely pathology-grounded Navigator; the main open question is how much the question-agnostic abnormality ranking limits generalization to evidence types it wasn't trained to flag.","tokens_in":69,"tokens_out":1645,"would_cite":true,"duration_ms":44854,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that a learned, abnormality-driven active-perception Navigator makes zero-shot reasoning over gigapixel pathology slides state-of-the-art and traceable.","keywords":["whole-slide pathology","active perception","visual question answering","reinforcement learning","multi-magnification navigation","cancer subtype classification","evidence deliberation","zero-shot reasoning"],"falsifier":"Take a slide-level question whose correct answer depends on regions a pathologist would label normal rather than abnormal, such as a question about the absence of tumor in a specific compartment, and run AdaptivePath with its default budget of five regions per magnification and four focus observations per region. If the retained high-magnification regions contain none of the question-critical normal regions and accuracy collapses toward random guessing, while the same question answered from the full-slide context succeeds, then the claim that abnormality ranking suffices for evidence acquisition is refuted.","tokens_in":11772,"feed_emoji":"🔬","tokens_out":8471,"duration_ms":79709,"temperature":0.7,"pith_summary":"This paper argues that the right way to reason over a gigapixel pathology slide is neither to compress the whole image into one representation nor to wander heuristically, but to actively decide where to look, at what zoom, and in what order. It proposes AdaptivePath, whose Navigator is trained on pathologist-reviewed normal/abnormal labels to choose observation locations and spatial extents, and whose Interpreter, Deliberator, and Arbiter convert those observations into question-conditioned evidence, option-by-option audits, and a final answer. The paper reports state-of-the-art zero-shot accuracy on slide-level and region-level pathology visual question answering, 80.14 percent accuracy for cancer subtype classification across six cohorts, and 82.9 percent accuracy when pathologists diagnose from AdaptivePath-selected image sequences. If these results hold, agentic visual reasoning over whole slides becomes both practical and traceable without costly question-specific trajectory annotations.","feed_headline":"Where-to-look AI tops slide pathology question benchmarks","feed_subtitle":"A learned navigator zooms from low to high magnification and feeds a reasoner five evidence-rich regions per level.","key_machinery":"The load-bearing object is the Navigator, a two-part policy consisting of FocusNet, which encodes a global glance and the sequence of focus regions into a visual state, and ActionNet, which predicts abnormality logits and the next focus action (location plus relative spatial extent). It is trained in two phases: first alternating representation learning with proximal policy optimization (a standard clipped reinforcement-learning update), using as reward the reduction in abnormality-classification loss from each observation; second, consistency-regularized fine-tuning that enforces the same focus trajectories under geometric and appearance-preserving transformations. The remaining components, the question-conditioned Morphology Interpreter, the option-wise Deliberator, and the cross-scale Arbiter, turn the acquired focus regions into explicit evidence audits. The mechanism works because the abnormality score acts as a cheap, question-agnostic proxy for where diagnostic evidence lives, and the hierarchical magnification schedule lets a small budget of regions cover sparse, scale-varying morphology.","core_discovery":"AdaptivePath's central claim is that a pathology-grounded, question-agnostic navigation policy can replace exhaustive patch compression and heuristic region retrieval for whole-slide reasoning. The Navigator is trained to recognize abnormal regions and to output the location and relative spatial extent of each next observation, using pathologist-reviewed binary normal/abnormal labels rather than question-specific evidence annotations. At inference it works hierarchically from low to high magnification, retaining the top-ranked regions by an abnormality score at each level, so a limited observation budget is spent where diagnostic morphology actually lives. Each observed region is converted into a question-conditioned morphological description; a Deliberator then audits how the directly observed morphology supports or conflicts with every answer option and rates evidence sufficiency; and an Arbiter integrates these audits across magnifications into the final answer. The paper reports that this pipeline outperforms existing agentic systems on WSI and region pathology VQA benchmarks and raises zero-shot cancer subtype classification accuracy to 80.14 percent on the same cohorts.","pith_inferences":["Not tested in the paper: because the Navigator is question-agnostic, the same trained policy could be reused for other gigapixel-image tasks, such as finding morphologically similar cases or highlighting regions relevant to survival analysis, as long as the target evidence is visible as abnormality.","A testable gap implied by the design: if a question's correct answer depends on tissue a pathologist would call normal, such as the absence of invasive cells in a particular compartment, the abnormality-based ranking may spend the budget elsewhere; a small amount of question-aware reranking data could close that gap.","The consistency-regularization phase hints at a general recipe for stabilizing reinforcement-learned visual observation policies under stain and rotation variation, which could transfer to other gigapixel imaging domains such as satellite or materials imagery.","The 82.9 percent pathologist accuracy on AdaptivePath-selected sequences, compared with 89.7 percent on recorded pathologist trajectories, sets a concrete calibration target: closing that remaining gap is a measurable next milestone."],"forward_implications":["Zero-shot slide-level visual question answering becomes practical: a frozen vision-language reasoner can answer questions over a whole slide accurately when it is fed a small set of actively chosen, multi-scale regions instead of a downsampled thumbnail or random patches.","Cancer subtype classification can be performed from H&E slides alone, with no subtype labels used in training, indicating that the learned navigation transfers to tasks beyond the abnormality-recognition objective it was trained on.","Each final answer is retraceable: the system records which regions were observed, what morphology was directly seen, and which answer options were supported, contradicted, or left neutral.","The observation budget behaves predictably: retaining five regions per magnification outperforms both smaller and larger budgets, and no single magnification matches the full low-to-high sequence's diagnostic utility in the paper's pathologist study.","Because the system currently reads only H&E morphology, questions whose evidence requires immunohistochemistry or molecular assays remain outside its directly observable evidence, as the paper states in its conclusion."],"supporting_citations":[{"why":"Supplies the pathology reasoner that generated the abnormal-region labels for the training set and serves as the frozen Morphology Interpreter during inference.","marker":"[Zhang et al. 2026]"},{"why":"Provides the clipped proximal-policy-optimization objective used to train the Navigator's observation policy.","marker":"[Schulman et al. 2017]"},{"why":"Supplies the alternating representation-learning and policy-optimization scheme and the coarse-to-fine active-perception idea that Phase I of Navigator training adapts.","marker":"[Wang et al. 2025]"},{"why":"Contributes the SlideBench-VQA benchmark used for WSI-level evaluation and the thumbnail-versus-patch protocol for comparing vision-language baselines.","marker":"[Chen et al. 2025b]"},{"why":"Contributes the WSI-VQA benchmark and the patch-aggregated baseline that AdaptivePath must outperform.","marker":"[Chen et al. 2024]"},{"why":"Provides the closest learned multi-scale navigation baseline; its reliance on model-generated rather than pathologist-reviewed trajectories motivates AdaptivePath's supervision choice.","marker":"[Sun et al. 2026]"},{"why":"Provides a training-free agentic baseline whose slide-level VQA scores AdaptivePath must beat.","marker":"[Yang et al. 2026]"},{"why":"Provides the large language model family used for the Deliberator and Arbiter components.","marker":"[Yang et al. 2025]"}],"fun_headline_variants":["Learned navigator zooms to sparse evidence in gigapixel slides","Active perception AI finds diagnostic evidence in whole-slide pathology","Question-agnostic navigator beats patch compression on slide VQA","AdaptivePath: pathologist-guided zooming for gigapixel slide reasoning","Sparse active perception tops slide pathology VQA benchmarks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything rests on the assumption that a binary normal/abnormal score, learned from pathologist-reviewed labels, is a sufficient proxy for where the evidence for any clinical question lives; if question-relevant evidence appears in tissue a pathologist would call normal, the Navigator can discard it before the reasoner ever sees it.","fun_headline_variants_meta":{"raw":{"variants":["Learned navigator zooms to sparse evidence in gigapixel slides","Active perception AI finds diagnostic evidence in whole-slide pathology","Question-agnostic navigator beats patch compression on slide VQA","AdaptivePath: pathologist-guided zooming for gigapixel slide reasoning","Sparse active perception tops slide pathology VQA benchmarks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000193,"raw_usage":{"total_tokens":1378,"prompt_tokens":998,"completion_tokens":380,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":614,"completion_tokens_details":{"reasoning_tokens":291}},"tokens_in":614,"tokens_out":380,"duration_ms":4509,"temperature":1.0,"reasoning_tokens":291,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T04:27:04.868336+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a slide-level question whose correct answer depends on regions a pathologist would label normal rather than abnormal, such as a question about the absence of tumor in a specific compartment, and run AdaptivePath with its default budget of five regions per magnification and four focus observations per region. If the retained high-magnification regions contain none of the question-critical normal regions and accuracy collapses toward random guessing, while the same question answered from the full-slide context succeeds, then the claim that abnormality ranking suffices for evidence acquisition is refuted.","supporting_citations":[],"review_version":1}