Selected research projects, peer-reviewed publications, and patents.
We audit the striking zero-shot claims of pathology vision–language models on whole-slide image VQA and find them compromised by data leakage at two levels: patient-level (slides from one case split across train and test) and institutional-level (different cases sharing staining-batch and scanner signatures through a common Tissue Source Site). Tracing slide, case, and TSS identifiers across public resources, we document 92.3–100% case-level train–test overlap on TCGA-derived benchmarks, show both leakage levels are linearly decodable from foundation-model features, and find reported accuracies concentrate on the most contaminated benchmarks — so current evaluation cannot separate genuine multimodal reasoning from nearest-neighbor retrieval over memorized artifacts. We close with concrete recommendations for contamination-free evaluation.
Accuracy alone does not measure evidence use: it can conflate dataset contamination, prior knowledge, and image evidence. In a lymph-node metastasis study, most public pathology VLMs barely changed when given the whole slide, an annotated lesion, or no image at all. We introduce CleanSlide, a TCGA-based VQA benchmark of 149K audited multiple-choice questions over 9,985 slides with patient- and tissue-source-disjoint splits, where every question is audited for option shortcuts, stem leakage, cross-split duplication, and blind solvability. We further propose Pair-DPO, a preference loss over counterfactual pairs of real slides with opposite, verified findings for the same question and source, which cancels the question-attributable signal and leaves image evidence as the source of preference. Our model gains 15.29% from image evidence on CleanSlide, versus 2.81% for the best published model, and reaches 57.6% and 59.0% on the external CPTAC and BCNB cohorts, outperforming all evaluated models by 9.7% and 3.4%.
A whole-slide image viewer that pathologists and AI agents read under the same rules. Today's slide-reading agents pan and zoom inside their own harnesses, where pixels are free, coordinates are private, and only the final answer is scored. SlideTwin gives people and agents one environment: microscope-style rule profiles, physical units in micrometres, a virtual clock that charges every action, a replayable log of every pixel served, and protocol checks on every report (cited views must exist, positive findings must be seen at 20×, measurements must fit the reported size, and negative calls must be backed by tissue coverage). A pathologist in the browser and an agent over REST, MCP, or a gym-style Python Env share the same session, so every reading can be replayed exactly and checked against what was actually seen. In one replay, an open VLM reported a micro-metastasis after viewing only 0.4% of the tissue at 10× or higher.
Self-evolving loops improve LLM agents by iteratively rewriting their harnesses, and many rely on a sample of benchmark units to accept a candidate, select one to deploy, and report its gain, yet never score the units the sample leaves out. Replaying published sampling designs on AgentDojo, Agent-SafetyBench, and Terminal-Bench 2, where every candidate was run on every unit, we find errors in all three decisions: accept tests misjudge 13–37% of clearly separated pairs, selection deploys a worse candidate in up to 54% of runs, and reported gains are off by 5 points or more in up to 78% of cases. Samples drawn from the incumbent's failures are the least reliable: an unchanged copy of the harness appears to gain 16–55 points and is often accepted. Because real loops advance in steps smaller than this sampling error, we recommend using a random half of the benchmark to reject candidates but a paired test on all units to accept one.
A VLM-specific synthesis of computational pathology organized along two axes — a training-paradigm axis (contrastive dual-encoder, generative instruction-tuned, reasoning/RL-enhanced, agent-based, and VLM-augmented multiple-instance learning) and a granularity axis (patch, region, whole-slide). Its centerpiece is an evidence-graded paradigm–task suitability matrix, calibrated with controlled probe studies and head-to-head comparisons against pure-vision foundation models rather than leaderboard numbers.
A three-level LLM-based image-description framework over 15 public pathology datasets yielding the disentangled Balance9k dataset; fine-tuning DINOv3 on 9k curated samples rivals million-scale fine-tuning. Cell-Guided Attention and organ-agnostic learning improve cross-organ generalization by ~30%, and Text-Guiding Pooling mitigates LLM hallucination effects.