Selected work, newest first. Each entry carries a short note on what it does. The complete list is on Google Scholar.
ICML 2026, SPIGM workshop, Seoul · OpenReview · Code
A multilingual generative model for reading scanpaths. Joint training across languages captures shared gaze patterns, beats monolingual baselines, and generalises to languages with no eye-tracking data at all. Synthetic gaze from the model improves a multilingual question-answering benchmark — including for languages absent from gaze training.
ICLR 2026, UCRL workshop · OpenReview · Code
Concept bottleneck models gain transparency by predicting interpretable concepts before the label, but the concepts have to be annotated. LangCBM shows that a vision–language model can serve as the bottleneck itself: it generates textual descriptions of visual concepts, which a lightweight pipeline then extracts and classifies.
EMNLP 2025 · ACL Anthology
RLHF aligns language models with human preferences but is computationally expensive, and a preference label discards how the annotator arrived at it. Two ways of injecting modelled human gaze are tested: gaze-aware reward models, and gaze-based distribution of sparse rewards at token level. Gaze-informed RLHF converges faster while maintaining or slightly improving performance.
NeurIPS 2025 · OpenReview · NeurIPS
LogitGaze-Med, a multimodal transformer that predicts where a radiologist will look on a chest X-ray. It combines a domain-specific visual encoder, textual embeddings of diagnostic labels, and semantic priors read out of a medical vision–language model via the logit lens. Scanpath similarity improves 20–30% over baselines, and predicted fixations as extra training data yield over 5% gain in downstream pathology classification.
IEEE Access, 2025 · doi · Code
Automated segmentation has not reached expert accuracy, so correction remains manual and slow. MedSAM is fine-tuned to accept gaze points as prompts for interactive mask adjustment, trained on abdominal CT from the public WORD database. Studies with medical experts showed that gaze-assisted correction raised average segmentation performance by nearly 62% on difficult cases.
IEEE Access, 2025 · doi
A model that reports a cardiac measurement without saying how sure it is cannot be used clinically. On about 800 X-rays labelled by four practising radiologists, an uncertainty-aware negative log-likelihood loss trains models to predict both the value and its variance. Natural label disagreement between radiologists is measured and folded in.
EMNLP 2025, HCI+NLP workshop, Suzhou · ACL Anthology
Synthetic scanpaths are the standard answer to the scarcity of eye-tracking data, but the literature evaluates them on isolated metrics rather than on cognitive plausibility as a whole. This is a systematic evaluation framework: current models reproduce basic gaze patterns but fail on part-of-speech-dependent features and degrade across text domains.
CVPR 2025, MMFM-BIOMED workshop · arXiv
Interactive segmentation models such as SAM-2 and MedSAM-2 rely on manually provided bounding boxes and clicks. This work introduces eye gaze as a new input modality — the first application of eye tracking to 3D medical image segmentation. Against bounding boxes, gaze is more time-efficient at slightly lower segmentation quality.
NeurIPS 2025, TS4H workshop · OpenReview
Not a static gaze point but the continuous temporal gaze stream — a physiological time series — is fused with spatial image data to steer SAM-2 and MedSAM-2 zero-shot. 62 seconds per volume against 88 for manual bounding boxes, with a modest accuracy trade-off.
ICLR 2025, Reasoning and Planning for LLMs workshop · PDF
Scanpath prediction usually rests on low-level visual features and misses the semantic and contextual factors that actually guide the eye. LogitGaze extracts word-level representations from a vision–language model using the logit lens and uses them as semantic priors, improving all key scanpath metrics by roughly 15% on average.
ICLR 2025, SynthData workshop · OpenReview
Retrieval systems that let a physician find relevant radiology reports and patient descriptions are held back by the scarcity of high-quality labelled data under privacy constraints. Synthetic records generated by large language models improve retrieval both when training entirely on synthetic data and when mixing it with real data.
ICLR 2025, Building Trust in Language Models workshop · OpenReview
A fractal measure — the self-similarity exponent S — is proposed as a reference-free metric of how human-like generated text is. Human-written text sits at S = 0.57; non-instruct models score higher, while instruction-tuned models approach the human value.
IEEE Access, 2024 · doi
Cardiothoracic indices are computed by hand and cost radiologists considerable time. Four experienced radiologists each labelled 800 chest X-rays; a deep model trained on them reached performance comparable to a professional. Replacing the indices' central points with vertebra-based landmarks improved accuracy further: Moore from 85.94% to 87.34%, Lupi from 87.55% to 90.67%.
NeurIPS 2024, AIM-FM workshop · arXiv · Code
Semi-supervised segmentation in which human gaze recorded while reading abdominal images serves as the prompt for fine-tuning MedSAM. Validated on the public WORD database — 120 CT scans, 16 abdominal organs. Average Dice 90.5%, against 85.8% for nnUNetV2, 86.7% for ResUNet and 81.7% for the original MedSAM.
Preprint · arXiv
Two questions: how to train a model to estimate the uncertainty of its own predictions, and what to do with uncertain predictions once they appear. An uncertainty-aware negative log-likelihood loss is proposed for N-class classification, and augmentation is extended to test time so the model gives multiple predictions per sample.
EMNLP 2023 · ACL Anthology · Code
Concept bottleneck models assume training examples are annotated with high-level concepts — and choosing those concepts, then labelling every example with them, is the main obstacle to using them. Here the text accompanying the images during training, such as radiology reports, guides the induction of concepts, yielding interpretable, disentangled concepts and suppressing reliance on shortcut features.
Journal of Digital Imaging, 36(3), 2023 · doi
A randomised experiment on how gaze coverage of a chest X-ray changes with radiological workload and with the abnormality present. Four radiologists read 400 chest X-rays on an eye-tracker-equipped workstation. Average lung coverage ranged 55–65% per radiologist and fell by 1.3–7.6% for every 100 X-rays read.
IEEE Journal of Biomedical and Health Informatics, 2022 · doi
Around 60–80% of radiological errors are attributed to overlooked abnormalities, and the rate rises at the end of a shift. Four radiologists read 400 chest X-rays while their eye movements were recorded. Gaze-travelled distance, X-ray coverage and lung coverage deteriorated significantly (p < 0.01) with cumulative work done for three of four radiologists.
SPIE Medical Imaging 2022: Image Perception, Observer Performance, and Technology Assessment · doi
It was known that radiologists' performance deteriorates towards the end of a shift, but the quantitative gaze features predictive of fatigue had not been identified. Gaze features extracted over automatic lung segmentation were the strongest fatigue predictors, at a correlation of 0.82.
SPIE Medical Imaging 2022: Image Processing · doi
Despite large annotated public databases, computer-aided diagnosis still fails on rare abnormalities. Physician gaze was recorded over 400 chest X-rays sampled from CheXpert, RSNA and SIIM-ACR; a ResNet34 was trained to map the X-ray to both the gaze map and the pathology label. Accuracy rose to 0.714 AUC from 0.681 for the same network trained on labels alone.
