PathoTool: A Tool-Using Agent for Reliability-Aware Pathology Diagnosis

ACL ARR 2026 May Submission17151 Authors

Keywords: tool-using, pathology diagnosis, VLM agent

Abstract: Recent advances in VLMs have moved toward agentic systems capable of invoking external tools to refine reasoning process. Pathology diagnosis naturally fits this paradigm, as morphological patterns in HE images may require additional immunophenotypic evidence for reliable decision-making. Recent virtual staining methods enable the generation of virtual IHC images from HE, creating a new opportunity for tool-using agents in diagnosis. However, such generated evidence may contain unreliable or conflicting signals, which can interfere with reasoning when directly incorporated.To address this, we propose PathoTool, a tool-using agent for reliability-aware pathology diagnosis. PathoTool performs morphology-based diagnosis from HE images while estimating the confidence, and invokes a virtual staining tool to generate IHC evidence. The HE confidence is further updated through a Confidence Re-evaluation tool conditioned on the generated IHC, which determines whether virtual staining influences the final decision. In addition, an Immunophenotype Conflict Filter is introduced to suppress inconsistent or contradictory IHC signals, ensuring coherent reasoning. Experiments on NSCLC and BRCA classification demonstrate that PathoTool achieves a more balanced and reliable performance improvement.

Paper Type: Long

Research Area: Clinical and Biomedical Applications

Research Area Keywords: clinical decision support, clinical and biomedical language models

Contribution Types: NLP engineering experiment

Languages Studied: English

EMNLP 2026 AI Reviewing Experiment: yes

Reassignment Request Area Chair: This is not a resubmission

Reassignment Request Reviewers: This is not a resubmission

Visa Needs: yes

Country Of Origin: CN

A1 Limitations Section: This paper has a limitations section.

A2 Potential Risks: Yes

A2 Elaboration: In the Section of Ethical Consideration

B Use Or Create Scientific Artifacts: Yes

B1 Cite Creators Of Artifacts: Yes

B1 Elaboration: In the Section 4 and Section of Reference

B2 Discuss The License For Artifacts: Yes

B2 Elaboration: In the Section 4

B3 Artifact Use Consistent With Intended Use: Yes

B3 Elaboration: In the Section 4 - Implementation Detail

B4 Data Contains Personally Identifying Info Or Offensive Content: Yes

B4 Elaboration: In the Section of Ethical Consideration

B5 Documentation Of Artifacts: N/A

B6 Statistics For Data: Yes

B6 Elaboration: In the Section 4 - Datasets

C Computational Experiments: No

C1 Model Size And Budget: No

C1 Elaboration: Our VLM backbone is from previous work which is finetuned on the well-known Qwen2.5-VL-7B-Instruct

C2 Experimental Setup And Hyperparameters: Yes

C2 Elaboration: In the Section 4 - Implementation Detail

C3 Descriptive Statistics: Yes

C3 Elaboration: To ensure the reliability and certainty for pathological diagnosis, we set temperature=0 to guarantee consistent results.

C4 Parameters For Packages: N/A

D Human Subjects Including Annotators: Yes

D1 Instructions Given To Participants: No

D1 Elaboration: No risks for annotators, and the descriptions related to annotations are in Section 4 - Datasets

D2 Recruitment And Payment: N/A

D3 Data Consent: Yes

D3 Elaboration: In the Section of Ethical Consideration

D4 Ethics Review Board Approval: Yes

D4 Elaboration: The data is part of a program with papers already published. And it is a retrospective study

E Ai Assistants In Research Or Writing: Yes

E1 Information About Use Of Ai Assistants: No

E1 Elaboration: We only use AI for writing polish with manual modification, all the thinking and framework building process are done by human

Author Submission Checklist: yes

Submission Number: 17151