acceptodds
Withdrawn from ICLR 2027

BioTrace: A Verifiable Pathology Agent for Tracing Molecular-to-Phenotype Evidence in Whole-Slide VQA

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Whole-slide pathology visual question answering (VQA) often requires inferring clinically relevant phenotypes from evidence distributed across tissue regions and scales. Existing methods, including recent pathology agents, mainly acquire and reason over morphological evidence, leaving gene- and pathway-level biological support largely unexposed. Moreover, agent-based methods typically treat retrieved observations as additional context, without explicitly determining whether the evidence required by the question has been sufficiently resolved. To address these limitations, we introduce BioTrace, an agent system that traces question-specific evidence along a biological hierarchy with iterative evidence acquisition and verification. BioTrace organizes multiscale H&E representations into a Gene–Pathway–Phenotype evidence space, exposing concept-level predictions, concept-specific spatial responses, and supporting biological associations as queryable evidence alongside morphological observations. During inference, a pathology knowledge-guided verifier assesses evidence sufficiency, while unresolved evidence requirements guide subsequent acquisition. Experiments show that the learned evidence space performs strongly across phenotype-, pathway-, and gene-level prediction tasks. Building on this evidence space, the full BioTrace system consistently outperforms the compared pathology VQA and agent baselines on WSI-VQA and SlideBench-BCNB.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Computational Pathology, Pathology Visual Question Answering, Agentic Reasoning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_pyBcOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

BioTrace proposes an agent framework for whole-slide pathology VQA that combines a learned Gene–Pathway–Phenotype evidence space with multiscale visual acquisition and retrieval-augmented evidence verification. The evidence module is trained using phenotype annotations and paired RNA-seq, while inference relies on H&E images. The agent iteratively acquires predictions and visual observations, evaluates evidence sufficiency, and generates an answer from the accumulated evidence. The paper reports improvements on WSI-VQA and SlideBench-BCNB, which are promising.

Critical strengths

The paper addresses a relevant problem: making the evidence used by pathology VQA systems explicit and inspectable. Organizing predictions and visual observations into a structured reasoning process is a useful research direction.

The reported downstream results are promising. BioTrace achieves 60.93% overall accuracy on WSI-VQA and a 65.13% overall score on SlideBench-BCNB (Tables 1–2). The paper also includes component ablations and quantitative evaluations of phenotype, pathway, and gene prediction, providing evidence that the structured module captures useful predictive information.

Critical weaknesses
  1. Limited experimental scope. Both evaluation benchmarks are restricted to breast pathology, and the evidence module is trained on TCGA-BRCA (Sec. 4.1). These experiments do not establish whether the learned biological relations and verification policy generalize across tumor types. This is a substantial limitation for the broadly presented framework, particularly because its phenotype vocabulary and Pathway–Phenotype relations may be disease-specific.

  2. Insufficient validation of the biological evidence and verifier. I acknowledge the paired RNA-seq supervision and the authors’ distinction between predictions and measured assays. However, the inference-time molecular signals are predicted from the same WSI as the phenotype. Their agreement does not independently corroborate the biological interpretation. Aggregate prediction metrics and qualitative traces do not establish that the molecular evidence selected for individual answers is correct. Moreover, the verifier is not directly evaluated for its ability to identify erroneous, conflicting, or insufficient evidence. These gaps weaken the central claim of verifiable biological reasoning.

  3. The ablations do not establish the necessity of the complex design. Table 3 shows that phenotype-only evidence already achieves 57.71% accuracy, compared with 60.93% for the full system. Removing the verifier reduces accuracy to 59.16%. These results suggest utility, but do not isolate the contribution of the biological hierarchy from access to supervised predictions or differences in evidence acquisition. A simpler baseline using the same supervision and inference budget, together with uncertainty estimates for the component gains, is needed.

Questions for authors
  1. Generalization beyond breast cancer. Could the authors evaluate BioTrace on additional tumor types, for example using a representative subset of WSI-Bench introduced by WSI-LLaVA? Comparisons with relevant baselines would help assess whether the reported gains generalize to other cancers. Please also clarify any retraining or adaptation required. These additional results would help address my concern about the limited experimental scope.

  2. Validation of biological evidence. Could the authors provide additional validation of the gene/pathway evidence selected to support individual answers, beyond aggregate prediction metrics? In particular, can the verifier reliably detect incorrect or conflicting molecular predictions? Targeted checks against held-out molecular measurements and controlled evidence perturbations would help substantiate the verification claim.

  3. Contribution of the proposed components. Could the authors compare BioTrace with a simpler baseline using the same supervision and prediction targets under a comparable acquisition budget? This would help isolate the contribution of the biological hierarchy and verifier. Uncertainty estimates for the key ablation gains would also help assess their robustness.

Reviewer_fcDXOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces BioTrace, an agent framework for whole-slide pathology visual question answering. Its main component is a multiscale Gene–Pathway–Phenotype evidence space that exposes concept-level predictions, spatial response maps, and cross-level biological associations. Concept representations are learned from H&E features using phenotype labels and paired RNA-seq supervision, with RNA-seq required only during training. At inference, the agent combines structured predictions with observations from selected image regions. A retrieval-augmented verifier assesses evidence sufficiency and routes unresolved requirements to subsequent acquisition steps. The framework is evaluated on WSI-VQA and SlideBench-BCNB, reporting overall scores of 60.93% and 65.13%, respectively. Additional experiments evaluate phenotype, pathway, and gene prediction, component ablations, language-backbone substitutions, efficiency, and qualitative evidence traces. The results support the usefulness of structured biological predictions for pathology VQA, although the reliability and faithfulness of the verification process are less directly established.

Critical strengths

1. A useful interface between molecularly supervised representations and agent reasoning. BioTrace exposes phenotype, pathway, and gene predictions together with spatial responses and biological relations, rather than providing the agent only with unstructured image descriptions. This is a meaningful design choice: it allows acquisition to target specific concepts and makes the information available to the reasoner more inspectable. Using RNA-seq only during training also makes the inference setting practically relevant.

2. Promising results across complementary evaluations. BioTrace reports the highest overall score among the evaluated methods on both VQA benchmarks. On WSI-VQA, its 60.93% overall accuracy exceeds the strongest reported comparator’s 55.91%, with particularly strong open-ended performance. On SlideBench-BCNB, it improves over PathNavigate from 57.49% to 65.13%. Table 3 also shows a substantial reduction when the evidence space is removed, from 60.93% to 48.62%. Together with the structured prediction experiments and language-backbone substitutions, these results provide encouraging evidence that the proposed interface supplies useful information.

Critical weaknesses

1. The central claim of verifiable evidence reasoning is insufficiently validated. The experiments establish improvements in answer accuracy more directly than the reliability of the proposed evidence verification process. This leaves several questions about the paper’s central contribution.

  • (a) Correctness of sufficiency judgments. The paper does not quantitatively evaluate whether the verifier correctly distinguishes sufficient, partial, conflicting, insufficient, and unavailable evidence. Figures 5 and 10 illustrate plausible traces, but do not establish the correctness of these decisions. Is there an annotated evaluation of verifier judgments, particularly when structured predictions are incorrect or conflict with visual observations? Evidence that the verifier detects such failures would materially strengthen the central claim.
  • (b) Faithfulness of the supporting evidence. Phenotype, pathway, and gene predictions originate from the same H&E input, and upper-level representations explicitly incorporate lower-level representations through Eq. (4). Their agreement may therefore reflect shared model dependencies rather than independently corroborating evidence. Although the paper appropriately identifies molecular outputs as predictions, it does not establish whether selected regions faithfully support the cited concepts or whether answers depend on their reported evidence. What validation supports these properties, and how does the system behave when apparently consistent predictions are jointly incorrect?
  • (c) Reliability estimation and verifier specification. Equation (8) transforms aggregate validation performance into a concept/scale reliability score, which does not provide patient-specific uncertainty. How does the verifier use this score when assessing an individual case, especially under conflicting evidence? In addition, what retrieval corpus, matching function, prompts, and decision criteria implement Eqs. (10)–(11)? These details are necessary to understand and reproduce the claimed verification mechanism.

2. The experimental protocol does not sufficiently establish the source and robustness of the reported improvements. The results are promising, but baseline adaptations, incomplete controls, and ambiguities in evaluation reporting limit confidence in attributing the gains to the proposed hierarchy and verifier.

  • (a) Compatibility of adapted baselines. Table 4 reports that SlideChat receives CONCH-v1.5 features through a fixed, untrained 768-to-512 orthogonal projection. Matching dimensionality does not establish compatibility with a pretrained visual interface, so this adaptation could disadvantage the comparator. Were adapted checkpoints validated against their native feature-extraction pipelines? In particular, how does SlideChat perform with its expected features? For WSI-LLaVA, please distinguish the effect of replacing the feature encoder from its dimensional padding procedure.
  • (b) Isolation of the hierarchy and verification contributions. Removing the verifier reduces accuracy from 60.93% to 59.16%, while retaining only phenotype evidence reduces it to 57.71%. These ablations are informative, but do not separate biological structure from additional molecular supervision, or adaptive verification from additional acquisition and computation. Does “Phenotype-only Evidence” retain the same molecularly supervised training and remove only pathway/gene access during inference? How does BioTrace compare with a flat multitask evidence model trained using the same labels, and with a fixed acquisition policy under a comparable observation budget? Paired uncertainty estimates that account for questions sharing a patient or slide would help assess the robustness of these differences.
  • (c) Patient-level separation and structured prediction evaluation. Section 4.1 excludes WSI-VQA test slides from evidence-module training, but does not explicitly establish patient-level separation. Are all other slides and molecular records from those same patients also excluded? Please report patient and slide counts for the training, validation, and test splits, identify the held-out cohorts used in Tables 6–9, and clarify whether validation-based reliability scores are computed without test-set information. This clarification is necessary to assess generalization; the current description does not itself establish leakage.
  • (d) Evaluation definitions and aggregation. The definitions of open-ended accuracy and the WSI-VQA “Gold” setting require clarification. Appendix A.1 lists 390 MCQs, but 57.14% cannot arise from a single micro-accuracy over 390 binary-scored questions after rounding. Does this reflect category averaging, repeated runs, filtering, or a different evaluation subset? Please specify the aggregation rule, evaluated counts, and treatment of failed or unavailable answers for each method so that the reported comparisons can be interpreted consistently.
Questions for authors

see the weaknesses.

Reviewer_ySLEOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

BioTrace addresses pathology vqa by combining an evidence interface with evidence acquisition and verification. The model analyzes images at three scales and connects what it sees with multiple features. During inference, an agent reviews the results and decides whether it has enough evidence.

Critical strengths
  1. The paper proposes a useful system and a practically relevant interface.
  2. The results are not bad.
Critical weaknesses
  1. Paper claims existing agents as accumulating context without determining whether evidence has been satisfied. This is a central motivation, but it does not describe rightly. PathAgent evaluates evidence sufficiency and chooses it. BEACON’s evidence controller returns a sufficiency state. BioTrace does not establish that sufficiency-dependent acquisition itself is new.
  2. The hierarchy does not ensure that predictions rely on the advertised biological chain.
  3. Figure 5 labels TCGA-LL-A7T0, whereas the middle case in Figure 10 labels TCGA-AR-A0TP. Their images differ, yet both show exactly the same four values.
  4. The experimental design does not isolate the claimed mechanisms or establish a sufficiently reliable comparison.
Questions for authors

See weaknesses.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.