The work, published in the International Journal of Computer Assisted Radiology and Surgery, introduces a technique called question-aligned semantic nearest neighbor entropy, or QA-SNNE. It is a preclinical, methodological advance rather than a clinically validated system, but its implications for the future of human-in-the-loop surgery — where an AI assists a human surgeon rather than replacing them — are substantial. The core insight is deceptively simple: knowing when to trust an AI’s answer requires not just checking whether the answer is consistent with the model’s other guesses, but whether it actually responds to what was asked.

Surgical visual question answering, or VQA, is a branch of medical AI in which a model receives an image from a surgical procedure — say, a frame from a laparoscopic video — along with a natural-language question such as “Which instrument is being used near the tissue?” or “What is the state of the dissection plane?” The model must then generate a text answer. In minimally invasive and image-guided procedures, where surgeons must rapidly interpret instruments, tissue, and anatomy through a limited field of view under time pressure, such query-conditioned information could become a powerful decision-support and training tool.

Yet accuracy alone, the researchers argue, is not enough. In a future operating room, when an AI system is uncertain, it should default to safety: abstain from answering, or route the query to a human expert. The problem is that most existing surgical VQA systems have been optimized purely for utility — benchmark scores, answer accuracy — with little built-in machinery for recognizing uncertainty, abstaining, or escalating to a human. Worse, evaluations typically rely on “in-template” questions that closely mirror the phrasing used in training data, which encourages the model to learn text-matching shortcuts. A model can appear highly competent while remaining brittle to the natural variation, negation, and subtle rewordings that characterize real clinical language.

The new research addresses this on two fronts. The first is a diagnostic insight into how current uncertainty-estimation methods fail. Over the past few years, techniques such as semantic entropy and its more recent cousin, semantic nearest neighbor entropy (SNNE), have emerged as practical ways to detect unreliable outputs from large language models and large vision-language models. The idea is elegant: instead of measuring uncertainty at the level of individual tokens, the model samples many answers to the same question at high temperature and checks whether those answers agree with each other semantically. If twenty independent samples all converge on “grasper,” confidence is presumably warranted; if they scatter across “grasper,” “hook,” and “I cannot tell,” uncertainty is high. Because these methods operate purely on the generated answers, they can be applied after the fact to black-box models — a critical practical advantage, since many clinical AI systems cannot be opened up and inspected internally.

But there is a blind spot, and it is the heart of the new paper. SNNE measures whether answers are mutually consistent. It does not check whether any of those answers actually address the question. A model sampling twenty answers to “What is the state of the tissue?” might consistently produce variations of a description of a nearby instrument — coherent, mutually reinforcing, and entirely off-target. Under standard semantic entropy, this answer set looks trustworthy: high agreement, low uncertainty. In a clinical setting, that is precisely the kind of failure that could mislead a time-pressed surgeon. The researchers describe this distinction as the difference between answer consistency and question validity — and existing methods capture only the former.

QA-SNNE closes that gap through an elegant mathematical modification. The method begins with the standard SNNE machinery: for each image-question pair, the system draws twenty high-temperature answer samples and builds a pairwise similarity matrix measuring how semantically similar each answer is to every other answer (approximated in the implementation using ROUGE-L overlap). The entropy score is then derived from this similarity structure — high entropy signals disagreement, and thus potential failure.

The innovation is a “bilateral gating” mechanism that filters this similarity structure through question relevance. First, each sampled answer receives an alignment score, measuring how well it addresses the question. The researchers propose three ways of computing this score. An embedding-based variant encodes question and answer as vectors and measures their cosine similarity. An entailment-based variant uses a natural language inference model to check bidirectional semantic compatibility, rewarding answers that the question entails and penalizing contradictions, in both directions. A cross-encoder variant runs each question-answer pair through a model that jointly attends to both texts, producing a direct relevance judgment. These alignment scores are then converted into normalized relevance weights via a softmax transformation, and the pairwise similarity matrix is rescaled: the similarity between any two answers is weighted down whenever either of them is poorly aligned with the question.

The result is a question-conditioned reshaping of the entropy calculation. Answers that drift off-topic are not simply deleted from consideration — the authors are careful to note that if a gated similarity drops to near zero, its contribution to the log-sum-exp becomes approximately one, so the method performs a gentle rescaling rather than a hard filter. The effect is that uncertainty is computed primarily from disagreement among the answers that actually engage with the question, rather than from variability caused by generic, evasive, or misdirected samples. When all sampled answers receive similar alignment scores, the gating becomes non-discriminative — a limitation the authors acknowledge — but in the common case where a mix of on-topic and off-topic answers appears, the separation between reliable and unreliable predictions sharpens considerably.

To test robustness under realistic language variation, the team constructed something the field has lacked: an out-of-template version of a standard benchmark. Starting from EndoVis18-VQA, a surgical VQA dataset derived from MICCAI EndoVis 2018 nephrectomy videos, they rephrased all thirty-five question templates — covering tools, locations, actions, and organs — while keeping the images, ground-truth answers, and dataset splits completely untouched. Each paraphrase was checked to preserve clinical intent and answer type. The result is a controlled stress test: any performance drop can be attributed purely to the change in question wording, isolating the model’s reliance on surface-form matching from genuine visual and clinical understanding. The validation split contains 2,754 image-question pairs, and the team also performed external validation on the open-ended PitVQA dataset of pituitary surgery, comprising 4,766 diverse question-answer pairs.

The evaluation was broad. Five different VQA models were tested across the two datasets, in both zero-shot and parameter-efficient fine-tuned configurations, with QA-SNNE benchmarked against three established failure-detection methods: discrete semantic entropy, standard SNNE, and VL-Uncertainty, a perturbation-based approach that probes response stability under visual and textual perturbations. The headline metric was AUROC — the area under the receiver operating characteristic curve — measuring how well each uncertainty score separates correct answers from failures.

The results are striking, if nuanced. On the in-template EndoVis18-VQA benchmark, QA-SNNE improved failure-detection AUROC for two of three zero-shot models, with relative gains of fifteen percent for Llama3.2 and twenty-one percent for Qwen2.5. Under out-of-template rephrasing — the more clinically realistic scenario — the method achieved up to eight percent relative AUROC improvement among zero-shot models, suggesting that question alignment genuinely helps when language drifts away from training templates. On external validation with PitVQA, results were mixed, a candid acknowledgment that generalization across surgical domains remains an open challenge.

The researchers were also deliberate about what their detector does and does not measure. They deliberately avoid the word “hallucination” as the primary endpoint, noting that the term conflates factual error, visual non-grounding, fabricated content, and off-question responses. Instead, a sample is labeled an operational failure when its ROUGE-L similarity to the reference annotation falls below 0.5 — a reproducible proxy for answer mismatch, not an expert-adjudicated judgment of clinical wrongness. Uncertainty, in this framework, serves purely as a detector signal, never as a failure label itself.

The authors stress that this is preclinical work: no claim of clinical validation or routine deployment is made. The intended context of use is the evaluation of automatic failure detection for future human-in-the-loop surgical VQA systems, where high-uncertainty answers would be abstained from or referred to an expert. But as large vision-language models edge closer to the operating room, the study makes a compelling case that the safety layer must be as carefully engineered as the accuracy layer — and that a truly safe system must know not only when its answers disagree with each other, but when they fail to answer the question at all. In surgery, where the cost of a confidently wrong answer can be measured in patient harm, that distinction may prove to be one of the most important lines of code in the entire pipeline.