Supervised fine-tuning (SFT) is an efficient approach for downstream task adaptation and often serves as the initialization stage for reinforcement learning (RL), but it can show weaker generalization than RL. A key limitation is its off-policy objective: SFT fits fixed demonstrations token by token, including targets poorly aligned with the model's pretrained distribution, which can lead to overfitting. A recent line of work addresses this issue by assigning larger training weights to tokens better aligned with the current model's predictive distribution, with the intuition that fitting these tokens are less distortive to the model's knowledge and representations. However, computing the token weights from the model that is currently fine-tuned entangles token weights with the optimization trajectory, inducing a self-reinforcing loop which concentrates the supervision signal to a narrow set of initially preferred tokens. To address this, we propose PriFT (Prior-support guided Fine-Tuning), which derives token weights from a frozen pretrained reference to obtain a stable reweighting signal unaffected by fine-tuning. This signal estimates prior support: the extent to which each target token is supported by the pretrained distribution. By reweighting tokens according to their prior support, PriFT limits excessive adaptation to targets requiring substantial deviation from the pretrained distribution and improves SFT generalization. We introduce two instantiations: PriFT-prob uses pretrained target-token probability, while PriFT-mass selects tokens by cumulative probability mass under the pretrained distribution. Across mathematical reasoning, code generation, and medical QA, PriFT achieves state-of-the-art results among SFT baselines, while also providing a stronger initialization for subsequent RL.
https://arxiv.org/abs/2606.09396
UK Biobank is a British clinical study which, in addition to large amounts of medical data collected, plans to include detailed Magnetic Resonance Imaging (MRI) examina- tions from 100,000 participants. The image data can be analyzed for example by auto- mated analysis with neural networks for image segmentation or classification to expand the collection of image-derived phenotypes for medical research. In this work, we pro- pose, present and evaluate a methodology that uses image-based regression to estimate numerical properties from the collected neck-to-knee water-fat MRI scans. It applies a ResNet50-style network to 2D multi-channel collages that represent targeted and study- specific anatomical sub-regions from the MRI scan. The targeted sub-region sampling is achieved by propagating manually defined subregions from a reference scan to the other >30,000 analyzed scans using image registration. The resulting collages can be compact while retaining high-resolution, for example by providing multiple slices of kidney tissue for a kidney volume regression network. This approach also enables group-wise saliency analysis that highlight image and anatomical regions used by the network for the pre- dictions. The method was applied for the regression endpoints of kidney volume, heart volume, and age. Predictive performance was found between R2 = 0.731-0.984 and the group-wise saliency maps were found to highlight medically relevant regions. We con- clude that the method can be used to standardize and restrict the information presented to a deep regression analysis pipeline, increasing information density and opening up a wide range of future medical inference studies.
https://arxiv.org/abs/2610.35339
Medical image segmentation (MIS) remains challenging due to composite degradations and multi-scale geometric variations, which jointly impair lesion-background separability, structural consistency, and optimization stability. Existing single-stream frameworks entangle heterogeneous cues within a single representation pathway, while hybrid-stream approaches often rely on generic or static fusion, limiting effective coordination across domains and failing to explicitly address optimization instability. To overcome these limitations, we propose a new optimization-aware dual-domain framework for MIS (O2D-MedNet). First, a degradation-aware dualdomain interaction (DDI) module integrates Fourier - and Wavelet-domain features to suppress degradation-sensitive responses while preserving structural and fine-detail information. Second, a cross-backbone contour-guided fusion (CGF) module enables selective interaction between CNN-based local detail modeling and visual state-space-based global structure modeling, improving boundary accuracy and representation consistency. Third, a gradient-context optimization (GCO) strategy stabilizes training by adaptively adjusting learning dynamics based on loss trends and gradient statistics, mitigating instability caused by class imbalance and heterogeneous sample difficulty. Extensive experiments on six benchmarks demonstrate consistent improvements over state-of-the-art MIS methods across multiple imaging modalities. Further analyses confirm enhanced sensitivity, structural consistency, and optimization stability, validating the effectiveness of the proposed O2D-MedNet for robust MIS under heterogeneous conditions.
https://arxiv.org/abs/2610.35220
Positron Emission Tomography (PET) provides important metabolic information for tumor detection and staging but remains significantly less accessible than Computed Tomography (CT) due to its high cost, infrastructure requirements, and reliance on injected radioactive tracers. CT-to-PET image translation using generative models aims to address this disparity, yet state-of-the-art diffusion-based methods are limited by the scarcity of paired training data. To overcome this, we propose a domain-adapted latent diffusion framework that transfers natural-image priors using a relatively small dataset of 21,930 paired CT/PET slices, incorporating a signed power transform to resolve extreme intensity sparsity in PET data. Compared to CPDM, a state-ofthe-art baseline for CT-to-PET image translation, our approach yields superior performance in pixel-wise reconstruction and localized tumor fidelity, demonstrating effective transfer of natural-image priors to the medical domain. We further provide a systematic analysis of evaluation metric reliability for sparse PET data, showing that standard full-image metrics can be misleading indicators of translated image quality, and propose background-filtered and mask-restricted metrics as more clinically meaningful alternatives.
https://arxiv.org/abs/2610.35132
GLP-1 medications are increasingly used for weight management, and many people share their experiences online. These narratives show why people take the drugs and what they experience, but they remain unstructured and disconnected from formal pharmacovigilance. We therefore present AskGLP1, a system that transforms online patient narratives into a knowledge graph (KG) and answers questions about medication experiences. Through an LLM-assisted pipeline grounded in the MedDRA terminology, AskGLP1 screens nearly 8.5 million Reddit posts for medication experiences and distills 163,135 of them, with 2,506 WebMD reviews, into a KG of 385,790 drug-side-effect and drug-purpose relations. We find that GLP-1 experiences are source- and drug-specific. Online narratives share the most common gastrointestinal symptoms with FAERS, but they foreground everyday burdens such as fatigue and food aversion, whereas FAERS reports more often include injection-site reactions and selected serious diagnoses such as pancreatitis. We further find that posts report less nausea and vomiting for tirzepatide than for semaglutide, more fatigue for Wegovy than Ozempic and for Zepbound than Mounjaro. Evaluations show that the agent answers with verifiable evidence and declines requests for medical advice. Our findings reveal gaps between what FAERS captures and what people report online, and they call for patient-centered approaches to pharmacovigilance. Code, prompts, and evaluation questions are available at this https URL.
https://arxiv.org/abs/2610.34987
Reusing pretrained medical-imaging models offers an alternative to training from scratch. A model trained for the same task may perform poorly at a new hospital because imaging protocols and preprocessing differ. Effective adaptation depends on the model architecture, domain shift, and available labels and compute. Adaptation can improve scores on a small labeled set yet reduce performance on unseen data. We present MedAdaptAgent, an automatic harness that turns a supplied model, target data, and a free-text request into a new-site adaptation workflow. Its LLM-driven Adaptation Planner uses model-site evidence to choose compatible experiments and revise subsequent adjustments. The harness requires evidence of improvement over the pretrained model, then delivers the selected model as a complete, runnable package. Evaluation spans 26 segmentation and classification transfers across architectures and compute settings, with segmentation annotation budgets ranging from zero-shot transfer to few-shot adaptation. With 20 local annotations, mean segmentation Dice rises from 0.483 to 0.774 across 10 transfers, a 60.1% relative gain, with no harmful replacements. Mean classification AUROC rises from 0.810 to 0.837 across 13 transfers, a 3.4% relative gain. Comparing adapted models with the pretrained model reduces harmful classification replacements by 30.8% versus selecting the highest-scoring model on the available labels, using identical candidates. Seven public-model deployments demonstrate adaptation and reproducible package delivery across contrast and modality shifts. A CT-to-MRI study extends the evaluation to four LLM planners, including two open-weight models. We release the code on GitHub.
https://arxiv.org/abs/2610.34884
Disease diagnosis is a key application of large language model (LLM) agents in healthcare. However, existing medical agents are primarily trained and evaluated with static question-answering data, providing limited opportunities to learn active information acquisition, examination decision-making, and dynamic clinical interaction. To address this limitation, we present ClinVerse, an interactive, real-data-grounded multi-agent environment designed for training and evaluating medical diagnostic agents. Built on ClinPatient, a scalable dataset comprising 150k diagnostic cases derived from real-world electronic health records, ClinVerse supports coordinated interactions among six specialized clinical roles. Specifically, a doctor agent starts from an initial patient complaint, progressively gathers clinical evidence through patient interviews and examination requests, and ultimately produces a diagnosis. For each interaction trajectory, a rule-based scoring framework quantifies information acquisition, examination ordering, and diagnostic accuracy. These trajectory-level scores serve as reward signals for agent training while also enabling scalable evaluation. In experiments, we evaluate various general-purpose and medical-domain models and show that nine self-evolution methods across three categories consistently improve agents' diagnostic capabilities within ClinVerse. We further analyze the interplay among model capability, environment feedback, and self-evolution methods.
https://arxiv.org/abs/2610.34878
Objective: The rapid advancement of large language models (LLMs) in healthcare has opened new possibilities for AI-assisted clinical decision-making; however, safe and effective deployment remains challenging because of hallucination and domain misalignment. This study introduces DocTalk, a doctor-assistive question-answering (QA) system designed to generate concise, chart-grounded responses from de-identified discharge summaries. Methods: We developed a high-fidelity data generation pipeline that produces 500 clinically relevant QA pairs through sentence-level decomposition, chain-of-thought-guided question formulation, and expert review. Six open-source LLMs, including three general-purpose and three biomedical models, were evaluated before and after parameter-efficient instruction tuning using Low-Rank Adaptation (LoRA). The framework integrates prompt engineering, in-context learning, and medically aligned evaluation to support factual and interpretable responses. Results: On DocTalk-QA, instruction tuning produced consistent improvements across lexical, semantic, and clinical fidelity metrics. The best-performing tuned model, Llama-3-Med42-8B, achieved ROUGE-L = 0.40, BERTScore-F1 = 0.76, and UMLS-F1 = 0.49. To examine external generalization beyond the DocTalk-QA dataset, the two best-performing models, Llama-3.1-8B-Instruct and Llama-3-Med42-8B, were evaluated on PubMedQA before and after instruction tuning. Accuracy improved from 0.931 to 0.977 for Llama-3.1-8B-Instruct and from 0.935 to 0.943 for Llama-3-Med42-8B. Conclusion: DocTalk was deployed as a secure web-based platform for interactive EHR querying. These findings suggest that clinically grounded instruction tuning can improve medical QA performance and support practical clinical decision support.
https://arxiv.org/abs/2610.34871
Progress in AI-based retinal image analysis has accelerated with the adoption of foundation models, yet evaluating their real-world reliability remains challenging. Performance reported on a single dataset does not capture how models behave under dataset shift, across clinical definitions, or for different patient subgroups. This limitation is particularly critical in medical imaging, where robustness, calibration, and fairness are essential for safe deployment. We introduce FOCUS (Foundation Ophthalmic Cross-Dataset Understanding under Shift), a cross-dataset benchmark evaluating retinal fundus models across vision-only encoder models (VM), vision-language dual-encoder models (VLM), and multimodal large language models (MLLM). FOCUS harmonizes binary diabetic retinopathy, referable diabetic retinopathy, and glaucomatous optic neuropathy tasks across ten public datasets spanning diverse geographies, acquisition conditions, and label protocols. The benchmark evaluates models through a unified analysis layer that measures ranking performance, calibration, subgroup disparities, and image-quality robustness. We present a large-scale evaluation covering 532 base configurations and 228 MLLM configurations adapted via LoRA fine-tuning. Results show no model family consistently dominates across tasks and datasets: general VM encoders achieve the strongest average ranking performance, medical MLLMs are competitive but variable, and dual-encoder VLMs benefit substantially from lightweight adaptation. Fine-tuning improves in-domain performance but exhibits heterogeneous transfer to external datasets, particularly in calibration. FOCUS provides a practical framework and public benchmark to assess generalization, reliability, and robustness beyond single-dataset leaderboards.
https://arxiv.org/abs/2609.33158
Vision-language anomaly detectors describe normal and abnormal appearance with text, and their gains are commonly credited to the semantic content of these descriptions. This attribution is plausible in the zero-shot setting, but few- and full-shot detectors interpose a trained visual adapter, and in that setting it has not been isolated. Changing a description changes both its meaning and its geometry, so accuracy comparisons cannot separate the two. We ask what a fixed pair of text embeddings, or anchors, contributes to such a detector: semantic content, effective scale, or joint anchor geometry. For a single anchor pair, the answer is effective scale alone. The score reads the anchors only through their difference, and a trained adapter can rotate to follow any replacement direction, leaving that direction, and with it the linguistic content, unidentifiable. We prove that replacement anchors with the same difference-vector length permit exactly the same probability maps and attainable losses. Across seven medical datasets, three vision backbones, both supervision regimes, and three adapter families, scale-matched random and orthogonal anchors track correct prompts; breaking the scale match opens a gap that restoring it closes without text. The same pattern holds in MVFA and when MadCLIP's learned prompts are replaced by a text-free learned scale. When one adapter serves many categories, the answer expands to their joint arrangement. Rotating all category axes together changes nothing, as predicted. For the tested CLIP prompts, these axes are nearly collinear, and meaning-free rotations that spread them apart improve performance on both industrial benchmarks. Even under frozen zero-shot transfer, a random anchor frame transfers as well as the prompt frame: coordinate consistency, not meaning. What these detectors read from their prompts is scale and arrangement, not semantic content.
https://arxiv.org/abs/2610.34610
Objective. Deformable image registration remains challenging because corresponding anatomy may exhibit different intensity relationships. Hand-crafted modality-robust descriptors such as Modality-Independent Neighborhood Descriptor with Self-Similarity Context (MIND-SSC) and features from pretrained vision encoders provide distinct representations for correspondence estimation, but how they should be combined remains unclear. We investigate whether tighter integration improves registration over a sequential, scale-separated design. Approach. We introduce FlexiMIND, a training-free and label-free two-stage framework in which frozen FlexiCT-2D encoder features establish contextual correspondence and MIND-SSC refines the deformation within ConvexAdam. Five alternative feature-coupling strategies were evaluated under a protocol across CT-CT (DIR-Lab 4DCT 4DCT, 10 cases), MR-MR (ACDC, 50 patients), and MR-CT (Learn2Reg Abdomen MR-CT, 8 pairs) registration. Performance was assessed using target registration error (TRE), Dice similarity coefficient (Dice), SDlogJ, and local folding. Main results. FlexiMIND achieved a TRE of 1.27 mm on DIR-Lab 4DCT, a DSC of 0.807 on ACDC, and an organ-averaged DSC of 0.825 on Abdomen MR-CT. Across 15 coupling-strategy dataset comparisons, only four outperformed the reference method, with a maximum absolute improvement of 0.0043 in Dice, whereas degradation reached 20.6 % in TRE and 37.9 % in Dice. Changes in registration accuracy were closely associated with changes in deformation-field regularity, and strategies that better preserved regularity generally remained closest to the reference performance. Landmark-level analysis showed that the pretrained encoder stage primarily reduced large correspondence errors rather than uniformly improving local precision. Code is available upon acceptance at this https URL.
https://arxiv.org/abs/2610.34600
Despite advances in automated liver lesion segmentation, a gap remains between image-derived predictions and clinically interpretable radiology reports. Bridging this gap requires quantitative measurements that are both reproducible and traceable. We present an evidence-grounded agentic AI framework that connects liver and lesion segmentation with deterministic quantification and structured reporting. TotalSegmentator-based nnU-Net models delineate the liver, Couinaud segments, and lesions, while computational tools extract volumes, dimensions, counts, and anatomical locations. A large language model (LLM) organizes these measurements into a predefined reporting schema, with numerical consistency checks against source measurements. On 51 LiTS cases, the segmentation stage achieved a liver Dice score of 0.947, lesion sensitivity of 75.6%, and positive predictive value of 68.9%. Exploratory screening of 25,494 MERLIN reports identified 10,208 (40.0%) with liver-related abnormalities. By connecting quantitative imaging evidence with agentic report generation, this work takes a step toward human-centered medical AI that prioritizes traceability and clinical relevance.
https://arxiv.org/abs/2610.34599
Adapting general-purpose vision-language models (VLMs) to medical visual question answering (VQA) is valuable for helping clinicians read images and make diagnoses, but annotated data are scarce and provide only short answers, without reasoning annotations. Supervised fine-tuning (SFT) fits these labels but overfits to them and amplifies their biases, eroding out-of-domain (OOD) generalization. Reinforcement learning (RL) better preserves generalization but never uses the gold answer as a training target: GRPO-style algorithms learn from the contrast between correct and incorrect answers within a group of sampled responses, so all-correct or all-wrong groups give no correctness signal. Existing SFT-RL hybrids either apply SFT and RL in separate stages or switch between them per question, or need reasoning traces from experts or stronger models, which medical VQA data lack. We propose GRASP, which combines RL and SFT in every update using only short answers: RL lets the model explore its own reasoning, SFT teaches the gold answer after that reasoning to supply the direct supervision RL lacks, and their balance adapts to how well the model answers each question. Tuning only about 1% of the parameters with LoRA, GRASP outperforms all compared training objectives both in domain and on OOD open-ended questions across Qwen3-VL-8B, InternVL3-8B and Gemma 3-4B. On an OOD benchmark spanning eight imaging modalities, GRASP-trained InternVL3-8B and Qwen3-VL-8B beat all 10 compared general and medical VLMs in open-ended accuracy, including models trained on up to 2,600 times more data or with over 250 times more trainable parameters. We release MedTerse, a dataset of 12.9K short-answer questions, and our code at this https URL.
https://arxiv.org/abs/2610.34561
Vision--language models applied to medical question answering can reach correct diagnoses without using the diagnostic image, inferring instead from the clinical vignette. We ask whether this is visible in how \emph{firmly} a model holds its answer: does a model defend a diagnosis derived from a blank placeholder as tenaciously as one derived from a real medical image? Crossing an image-substitution protocol with two levels of adversarial pressure on 60 Italian State Examination items, we find no evidence that ungrounded verdicts are more fragile. In a paired item-level analysis, fabricated consensus (L4) overturns 94.4\% of verdicts with the real image and 98.1\% without ($\Delta = +0.037$; one-sided 95\% upper bound $+0.105$); contentless doubt (L1) overturns 62.5\% and 51.8\% ($\Delta = -0.107$, 95\% CI $[-0.209, -0.001]$), the point estimate running opposite to the direction a grounding detector requires. The two levels nonetheless produce opposite \emph{dynamics}: under L4 the model flips once, always to the exact option argued for (110/110 first-turn flips), and then freezes, averaging 1.9 verdict changes over ten turns; under L1 it never settles, averaging 7.4 changes. Instability under contentless doubt tracks intrinsic item ambiguity rather than evidence availability, and so does not serve as a grounding signal. Flips are almost entirely corrupting (99/100 across both levels). The model is not unaware of the missing evidence---it flags the blank image at baseline on 35 of 60 items and abstains on 4--6---yet those abstentions do not survive contact with pressure. None of the epistemic signals a user could plausibly consult distinguishes a grounded diagnosis from a confabulated one.
https://arxiv.org/abs/2610.34520
Large language model agents are increasingly used to analyze potential software vulnerabilities, but their outputs can silently violate domain knowledge, limiting their reliability in safety-critical settings such as medical devices. Prior work either treats that output as a prediction to be scored or constrains it to walks within a single knowledge graph; neither checks whether reasoning over a binary is entailed by an independent body of domain knowledge. We present EntailLLM, a framework that includes a specialized LLM agent and a validation mechanism. The agent proposes a traversal path of the function call graph for a vulnerability, which is checked against a separate knowledge graph and a logic program by an entailment procedure under annotated logic. We evaluate EntailLLM by examining six specialized agents across three CWE classes, and eight binaries varying in size from 405 to 12,696 functions. The agents improve pooled entailment from 77% to 98%, with entailment decreasing in only 3 of 118 configurations. We found that the best-performing agent combines self-refinement and a priori domain knowledge. EntailLLM is deployed end-to-end on real medical-device binaries, reaching 98% pooled entailment without per-device tuning. Our framework inherits the formal guarantees of generalized annotated logic, providing logical verification of agent output that is both explainable and grounded in well-defined semantics.
https://arxiv.org/abs/2608.01763
Feature analysis is a crucial step in medical image analysis, particularly in breast cancer classification. Traditional image analysis methods often struggle with optimal patterns along the diagonal axis, computational efficiency and feature buffer storage. Different classical operators require high-efficiency hardware for the complex operations, such as image convolution and gradient analysis, which are used for processing complex machine learning models for training and test operations. Here, in this research, we introduce a non-interpolation virtual trellis-based (NVHT) operator that utilizes a filter kernel over the same square image without transformation to a hexagonal trellis. This proposed method is very efficient in reducing computational and storage costs. The proposed operator is used for improved feature extraction efficiency and accuracy in disease classification. This method outperforms conventional approaches by leveraging a hexagonal pixel representation while maintaining high sensitivity in different directions with less storage.
https://arxiv.org/abs/2610.34352
Cross-modality medical image translation is challenged by anatomically inconsistent synthesis that can compromise clinical diagnosis. We propose StaX-Net, a Stackelberg interleaved asymmetric network that incorporates a partially coupled encoder-discriminator structure, and a 2D feature aggregation and spatial transformation (FAST) layer that distills target-supervised priors under the learning using privileged information paradigm. The shared blocks between the generator and discriminator enhance architectural feature alignment. To stabilize the leader-follower training, a Stackelberg-inspired asymmetric optimization scheme employs explicit shared-parameter freezing and asymmetric learning rates. The FAST layer distills target-domain priors during training and applies a source-only transform at inference. This constrains the synthesis process through target-informed feature representations relative to standard generative architectures. The code is available on GitHub for public use.
https://arxiv.org/abs/2610.34214
Modern surgical AI increasingly requires multimodal understanding rather than isolated perception, yet existing datasets remain largely limited to single-task supervision such as recognition or templated question answering. Current medical VQA resources often rely on synthetic templates, lack natural descriptive supervision, or provide sparse annotations that hinder reasoning about spatial relations and procedural context. We introduce MIS-VQA, a surgical vision-language dataset generated through an annotation-driven pipeline that converts heterogeneous surgical annotations into structured textual scene representations. Using EndoVis18 and CholecT50, we transform phase, instrument, interaction, triplet, and segmentation labels into unified contexts that guide large language models to generate clinically coherent question-answer pairs. MIS-VQA supports four response types-concise and detailed descriptions, perception-level grounding, and multi-step reasoning-enabling hierarchical evaluation from atomic visual grounding to compositional reasoning. To address annotation sparsity, we incorporate confidence-filtered segmentation and cross-annotation consistency checks, enabling spatially grounded supervision at scale. MIS-VQA establishes a unified benchmark for surgical vision-language understanding with spatial supervision and multi-tier reasoning evaluation. To demonstrate its utility, we introduce LLaVA-SurgXpert, a family of surgical vision-language models with visual encoders to analyze how visual representations influence multimodal reasoning. Experiments show consistent improvements over existing surgical VQA baselines while revealing capability gaps between perceptual grounding and higher-level reasoning hidden in prior benchmarks.
https://arxiv.org/abs/2610.34210
More people are seeking out general-purpose AI models for medical advice, including diagnosis. Yet, AI models are susceptible to sycophancy, like validating users' beliefs even when they are untrue, which could lead to false positive diagnoses. Here, we evaluate how nine AI models change their diagnostic decision when a patient presents a belief that they have a specific (but incorrect) medical condition. We vary how close the user belief is to the clinical evidence presented. Models sycophantically adopt the patient's belief on average 23.2\% of the time---sometimes even when the patient proposes a fictional, made-up condition. We also varied different personas, including a Rare Disease Claimer that got models to accept their belief 43.0\% of the time. Additional analyses find that medical fine-tuning and model scale makes models less susceptible to changing their diagnoses. Sycophantic bias in medical diagnosis raises implications for safety, especially as more people are using AI for health support.
https://arxiv.org/abs/2610.34013
Query-time filters that screen each retrieved document before it reaches the generator are the dominant published defense against retrieval-augmented generation poisoning, and they report near-perfect detection on domain benchmarks. We show that this performance is predicted by how the poison set was built rather than by what the filter reads: removing a single construction cue destroys the deployed decision before it destroys the ranking. Two label-free statistics of a poison set and its clean corpus profile the cue: containment, how often the target question appears verbatim in the poison, and length overlap. PoisonedRAG's released code prepends the target question to every inserted passage. Of the 21 published defense evaluations we survey, 8 verifiably run poison built this way and 5 substitute a length cue; separately, 12 report no false-positive rate at the operating point they deploy. Deleting only that prefix from the attack authors' own passages (0.93 still retrieved for their target) takes RAGuard, a published content filter we re-implement, from AUROC 0.953 to 0.758 and its miss rate from 0.191 to 0.763. Across four corpora, four attack families and ten per-document defenses on one shared candidate set, the profile predicts where detection saturates. Padding medical poison to clean length with word-shuffled filler that carries no content takes TrustRAG's set-level stage from a miss rate of 0.000 to 0.827, and one learned scorer of ours from 0.000 to 0.986 (0.321 at a matched 5% false-positive rate) while its AUROC holds at 0.953. The ranking survives; the decision does not. A defense-aware attacker defeats our own scorer, so we claim no defense. A detection result is evidence only when it is reported with its poison set's construction profile and the false-positive rate at the threshold actually deployed; we release the script that computes the profile.
https://arxiv.org/abs/2610.33849