Medical VLMs frequently hallucinate clinically unsupported findings. We identify a stability gap: visually grounded tokens maintain relatively stable logits under semantics-preserving image perturbations, whereas visually misgrounded tokens exhibit substantially larger shifts. We introduce LENS (Logit-ENhanced Stability), which exploits this signal at both training and inference. During training, Semantics-Preserving Preference Optimization (SPPO) scores preference pairs using clinical faithfulness and cross-view stability; at inference, stability-aware contrastive decoding penalizes tokens whose logits diverge across original and counterfactual views. Constrained counterfactuals use attribution-guided masking and feature-space proximity guards to reduce disruption of lesion- and device-level evidence. Across radiology VQA, chest X-ray reporting, and gastrointestinal endoscopy, LENS consistently outperforms strong decoding-time and preference-optimization baselines. On MIMIC-CXR, entity-level RadGraph recall improves from 21.0 to 22.8 over PCCD, while expert evaluation shows a hallucination-rate reduction from 17.0% to 8.5%.

We present the CanSA system for the MedEx-ACT@ACL 2026 shared task, which requires extracting and classifying clinical decisions from ICU discharge summaries into nine DIC-TUM categories. We have developed three approaches: (1) a training-free system which consists of a preprocessing module that normalizes text and an inference engine combining zero shot LLMs with a RAG ensemble, (2) a supervised fine-tuning method which required training, and (3) a training-free retrieval-augmented pipeline employing TF–IDF-based lexical retrieval to surface in-context exemplars from the development corpus, combined with section aware chunking and structured extraction calls to a large language model. Our team’s best submission achieved a Final Score of 0.41, ranking 34th out of 37 on the official test leaderboard.

Sample-efficient pretraining under a developmental data budget, such as the ten million words and ten epochs of the BabyLM Strict-Small track, is usually approached by adding training signal: a new objective, optimizer, curriculum, or teacher. The budget itself receives less scrutiny. We show that much of it is spent after learning has stopped. A 33.0M-parameter GPT-BERT hybrid trained under the track rules saturates on downstream evaluations by roughly 70M words seen, about 70% of its permitted exposure, while training loss continues to fall; the final epochs improve fit, not competence. From this diagnosis we derive three design principles, timing, preservation, and alignment, and instantiate them as Minimal-Pair Preference Optimization (MPO): a short post-training phase, begun at the saturation checkpoint, that teaches the model to prefer real corpus sentences over length-preserving corruptions of them under a frozen-reference penalty. A controlled study of seven in-training interventions supports the principles: none improves on the baseline, and the unanchored ones damage the same cluster of long-distance dependencies first. MPO improves five of the seven official task averages, concedes under a point on the other two, is alone in repairing that cluster, and matches the fully trained baseline with 7.2 rather than 10 epochs.