Publication Plan

Here is my plan for publishing research papers from November 2025 to September 2026. I will target a mix of computer vision, medical informatics, and machine learning venues to reach different audiences.

Planned Papers

TitleVenueStatusDeadline
Cheap Robustness for Medical VLMs: Projector-only LoRA cuts paraphrase flips on MIMIC-CXRCVPR 2026PlannedNov 13, 2025
PEFT over Frozen Backbones for Radiology VQA: Robustness and calibration on MIMIC-CXRMIDL 2026PlannedDec 5, 2025
From LoRA to lower risk: calibrating VLMs for chest X-ray triageAMIA 2026 Informatics SummitPlannedDec 2, 2025
Measuring paraphrase flip rate in radiology VQAISBI 2026 (1-page abstract)PlannedJan 13, 2026
Paraphrase-Consistent Fine-Tuning: A simple loss that stabilizes VLMs without full retrainingICML 2026PlannedLate Jan/Feb 2026
Human-in-the-loop robustness: single-radiologist audit for VLM safety on MIMIC-CXRMICCAI 2026PlannedLate Feb/Mar 2026
Robust VLMs via projector-centric PEFT: analysis across paraphrase familiesECCV 2026PlannedMarch 2026
ROBORAD: A reproducible benchmark for paraphrase robustness in radiology VLMsNeurIPS 2026PlannedMid-May 2026
Clinically aware robustness for chest X-ray VQA with single-expert auditsCHIL 2026PlannedSpring 2026

Publication Strategy

Early Papers (Nov 2025 - Jan 2026)

The first three papers focus on establishing the problem and showing initial solutions:

  • CVPR 2026: Focuses on computer vision community, shows that projector-only LoRA can reduce paraphrase sensitivity
  • MIDL 2026: Medical imaging venue, covers robustness and calibration aspects
  • AMIA 2026: Medical informatics audience, emphasizes clinical triage applications

Core Methodology Papers (Feb - May 2026)

Mid-timeline papers dive deeper into methods:

  • ICML 2026: Machine learning community, presents the paraphrase-consistent loss function
  • MICCAI 2026: Medical imaging, includes human expert validation
  • ECCV 2026: Computer vision, analyzes different paraphrase families

Benchmark and Evaluation Papers (Spring - Fall 2026)

Later papers focus on reproducibility and comprehensive evaluation:

  • NeurIPS 2026: Premier ML venue for the benchmark dataset release
  • CHIL 2026: Healthcare informatics, clinical validation aspects

Target Impact

This publication plan aims to:

  1. Establish the problem early in top-tier venues
  2. Present solutions across different communities
  3. Enable reproducibility through benchmark and code releases
  4. Validate clinical relevance through expert studies

The mix of venues ensures the work reaches computer vision researchers, medical imaging specialists, machine learning practitioners, and clinical informaticists - all stakeholder communities for medical AI safety.

Timeline Considerations

The deadlines are aggressive but achievable given the overlapping research phases. Early papers will be based on initial experiments and proof-of-concept results. Later papers will have more comprehensive evaluations and human studies.

All planned papers will contribute to the dissertation while serving different audiences and highlighting different aspects of the robustness problem and solutions.