# Research Papers — 2026-09 · bundle page 167 of 266 > Full-text markdown of papers #8301–8350 of 13298 published in 2026-09, newest first. AI-readable mirrors of the full 13298-paper 2026-09 archive (50 papers per file). Canonical page: https://aipod.dev/papers/archive/2026-09/page/167 [← Previous bundle](https://aipod.dev/papers/archive/2026-09/page/166.md) [Next bundle →](https://aipod.dev/papers/archive/2026-09/page/168.md) ### Beyond Accuracy: Robustness, Cost, and Governance Trade-offs for Vision-Language Models in Templated Document Extraction Authors: Kushal Patel, Pushkal Shrivastava, Mackenzie Lees, Qirui Lu, Bhargobjyoti Saikia, Liying Li, Junlin Jiang Published: 2026-09-14 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.15706v2 Vision-language models (VLMs) are increasingly used to extract structured fields from business documents, yet most evaluations report accuracy on clean benchmarks and offer little guidance to practitioners choosing an approach for a given task complexity. We address this gap with a measurement-grounded study and an open-source release. Across eleven systems (three commercial, two reasoning, five open-source VLMs in pretrained and fine-tuned form, and a non-LLM OCR->regex floor) scored on a 750-document held-out pool of synthetic checks, fine-tuning on 3K samples lifts the best open-source VLMs above F1 0.98-above every zero-shot commercial system on this task-while GPT-5 leads the commercial pool on F1 and Claude Sonnet 4.5 collapses on Date. To turn these measurements into actionable choices, we introduce a practitioner-oriented selection framework that maps a task profile (quality, latency, governance, volume) to a recommended approach via filtering and total-cost minimization, illustrated on a hypothetical mid-volume document-extraction scenario. --- ### Backward SDEs-based Diffusion for Physics-Constrained Generation Authors: Zihao Wang Published: 2026-09-14 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.15702v1 Pretrained score-based diffusion models provide strong unconditional priors, yet enforcing measurement or physics consistency in inverse problems is often handled by heuristic guidance, intermittent projections, or task-specific conditional training, with limited guarantees of feasibility at the end of inference. We propose terminal-conditioned inversion for score-based SDE priors. Given a frozen Score-SDE prior and a task-defined terminal feasibility specification, we construct an associated backward stochastic differential equation whose adapted solution defines a principled inverse map from the terminal requirement to a prior state at a chosen noise level. Under standard regularity conditions, we establish existence and uniqueness of the adapted solution and obtain terminal consistency by construction. We further develop a practical neural BSDE solver that composes arbitrary pretrained diffusion priors with domain constraints without modifying the score-defined coefficients, producing an anchored prior state that enables neighborhood sampling for uncertainty characterization. Experiments on toy datasets validate stable terminal-conditioned inversion and distributionally consistent neighborhood sampling. As a real-world case study, we apply the framework to sparse-view CT reconstruction and achieve improved reconstruction quality over representative training-free baselines while satisfying strict measurement feasibility under the prescribed terminal specification. Project is available in: \href{https://laplacelab.github.io/BSDEDiffusion/}{https://laplace.center/icmlbsdeI/} --- ### More Than Just Access: Generative AI as Communication Intermediary for Blind and Low-Vision Users Authors: Protik Dey, Mohd Saifuzzaman, Taslima Akter Published: 2026-09-14 Categories: cs.HC, cs.AI, cs.ET arXiv: https://arxiv.org/abs/2609.15696v1 Generative AI (GenAI) tools are increasingly woven into how blind and low-vision (BLV) people communicate, not only with digital information, but with the physical world and with other people. Tools such as ChatGPT, Google Gemini, Be My AI, and Seeing AI translate visual and textual content into accessible form, and are beginning to substitute for interpersonal requests for help, such as asking a family member to read a label or describe a scene. Drawing on semi-structured interviews with 19 BLV participants, we examine GenAI as a communication intermediary and how it succeeds and fails as an alternative for reading, describing, and even asking another person for help. We also investigated what BLV users gain and risk when these tools take over that role. We conclude with design and policy implications for GenAI systems that communicate uncertainty honestly, protect information, and support BLV users' independence rather than substitute for it unsafely. --- ### NoteVQA: Benchmarking VLMs on Real-Life Questions from Human Communities Authors: Haonan Jiang, Guojian Zhan, Jiancong Xie, Shijun Wan, Dongiia Zhao, Cheng Chen, Yahui Liu, Chuan Mu Published: 2026-09-14 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.15695v2 Vision-language models (VLMs) increasingly power consumer-facing AI search, yet evaluating them on the diversity of everyday visual questions remains challenging. Existing benchmarks often target predefined capabilities, such as multi-hop retrieval or long-form synthesis, whereas users ask photo-grounded questions spanning a long tail of everyday scenarios. Despite advances in VLMs, users on Xiaohongshu, a mainstream Chinese image-sharing platform, continue to turn to other people for help with everyday visual questions. Motivated by this behaviour, we curate NoteVQA from these questions, yielding 252 items across 12 topical categories and 7 user intents. Each item includes a concise reference distilled from expert community responses and a human-audited interleaved reference answer that combines textual explanations with supporting visual evidence. We evaluate both short-answer correctness and interleaved-answer quality. To support the latter, we introduce AgenticInterleave, a single-agent ReAct framework for retrieval-supported answer generation, together with IVR-12, a 12-dimensional rubric for assessing the content, presentation, and image quality of interleaved references and model outputs. Across 9 frontier VLMs, the highest short-answer accuracy is 52.8\%, while adding agentic search to Qwen3.5-397B-A17B improves accuracy by only 2.0\%. For interleaved answers, the same model running AgenticInterleave scores 3.52 under IVR-12, compared with 4.65 for the human-audited references, with the largest gap in content quality. These results highlight the challenges that everyday visual questions pose for current VLMs in both answer accuracy and the quality of visually grounded explanations. --- ### Complete Suffix Prediction for Recommendation via Latent Retrieval over Process Graphs Authors: Sarra Madad, Myriam Maumy, Fr{é}d{é}ric Bertrand, Yoann Valero Published: 2026-09-14 Categories: cs.NE arXiv: https://arxiv.org/abs/2609.15692v1 Complete suffix prediction is challenging in sequential decision settings, where the same prefix can remain compatible with several plausible suffixes. We propose a graphbased metric-learning framework that reformulates complete suffix prediction as latent retrieval over process graphs. Prefixes and suffixes are represented as directed attributed graphs and encoded by edge-conditioned graph neural networks, allowing event-level activities and transition-level durations to be modelled jointly. Prefix representations are projected into the latent suffix space through a predictor trained with a joint reconstruction and contrastive objective strengthened using process-aware hard negatives. To stabilise the learned retrieval geometry, spectral normalisation, and retrieval robustness, spectral normalisation is applied to enforce a Lipschitz constraint on both encoders and predictor. Experiments on two real-life process datasets demonstrate that the proposed framework achieves the best overall results across nearly all evaluated criteria. It improves semantic suffix accuracy measured by normalized Damerau-Levenshtein distance, yields strong retrieval quality through Recall@1, Recall@5, and MRR@5, and maintains temporal plausibility according to Mean Absolute Error. These results show that graph-based latent retrieval is an effective alternative to sequential suffix prediction for recommendation-oriented process monitoring under structural and KPI-related constraints. --- ### EEG-Xplain: Decoding Neural Black-Boxes of EEG Foundation Models Authors: Hansong Ma, Junxiao Wang Published: 2026-09-14 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.15687v1 EEG foundation models such as BIOT, LaBraM, and EEGMamba have achieved remarkable performance in neural signal decoding, but their black-box nature limits clinical trust and neuroscientific validation. We propose a unified attribution framework for interpreting EEG foundation models across heterogeneous architectures. The framework integrates gradient-, perturbation-, and activation-based explanation methods to analyze model behavior in spatial, temporal, and frequency dimensions. Spatially, it identifies critical EEG channels and visualizes their distributions using topographic maps. Temporally, it highlights decision-relevant signal segments through attribution heatmaps. In the frequency domain, it quantifies the contributions of canonical EEG rhythms via spectral perturbation analysis. To assess explanation reliability, we introduce a population-level evaluation combining Area Over the Perturbation Curve (AOPC) and cross-method consistency analysis. The framework further leverages Large Language Models (LLMs) to transform structured attribution outputs into natural-language reports, bridging low-level neural representations and high-level semantic reasoning. Experiments on benchmark datasets, including Mumtaz2016 and TUAB, demonstrate that the generated explanations are consistent with established neurophysiological markers, validating meaningful neural representations while exposing potential dependencies on artifacts and spurious patterns. The proposed framework provides a standardized approach for evaluating the interpretability, reliability, and physiological plausibility of EEG foundation models. --- ### Can a Neural Encoding Model Replicate an fMRI Visualization Study? Authors: Erfan Nasirzadeh Orang, Zack While Published: 2026-09-14 Categories: cs.HC, cs.CV arXiv: https://arxiv.org/abs/2609.15685v1 Most knowledge of graphical perception comes from behavioral studies. Understanding from a neural perspective is much more limited due in part to neuroimaging studies' expensiveness and difficulty to conduct. In this paper, we evaluate whether Meta's Tribe V2 neural encoding model can recover neural contrasts from a visualization fMRI study. Specifically, we evaluate Tribe V2 through a conceptual replication of the visualization-viewing component of a prior comparison of Bubble charts and three-dimensional Surface charts in color and grayscale. We generate TRIBE-predicted cortical responses for the original stimuli and compare the resulting contrasts with those reported in the human study. The model reproduced the direction of 11 of 14 reported cortical effects, with agreement concentrated in visual-processing regions. This agreement characterizes the model's alignment with the prior human-generated fMRI results rather than independently confirming them. We discuss the limitations encountered when working with this model for in-silico replication and hope to encourage future work exploring this new avenue for neuroimaging studies in visualization. Supplemental materials are available at https://osf.io/8a96x/. --- ### RESKILL: Explicit Failure Attribution and Structured Repair for Interactive Language Agents Authors: Mengyi Deng, Xin Li, Duyi Pan, Zilin Wang, Zhiwei Li, Zhijiang Guo, Wei Wang Published: 2026-09-14 Categories: cs.CL arXiv: https://arxiv.org/abs/2609.15684v1 Language agents increasingly rely on reusable skills, but post-failure repair is often handled by opaque one-shot reflection: a model generates a skill patch without explicitly maintaining how failure explanations relate to candidate repairs or how unsuccessful retests should influence later edits. We introduce RESKILL, a structured repair framework that maintains an explicit repair state across repair rounds. Given a failed rollout, the framework links failure hypotheses to candidate skill patches, selects local repairs through coverage-based attribution, retests the edited skill set in the environment, and uses retest outcomes to guide subsequent repair updates. The language model supplies structured repair factors, while the repair procedure records them, compares local skill patches by how well they address active failure explanations, and carries unsuccessful retest outcomes into later repair rounds. We evaluate RESKILL on ALFWorld and TextCraft across three model sizes under fixed repair budgets. RESKILL obtains the strongest final success in all six benchmark-model settings, improving average final success by 3.7 percentage points over direct repair and 3.3 points over hypothesis-conditioned repair. These results suggest that explicit attribution alone is insufficient; durable improvement emerges when attribution is integrated with repair selection and persistent retest-conditioned update. --- ### V-ICAL Bench: Evaluating Video In-Context Learning for Multimodal Agents in Interactive Environments Authors: Ziqian Fan, Shibo Xu, Junjie Li, Xiangyu Zhao, Shengyuan Ding, Yifan Yang, Zhenjie Yang, Haodong Duan, Yue Zhou, Zhihang Zhong, Xue Yang Published: 2026-09-14 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.15683v1 While In-Context Learning (ICL) enables models to adapt from exemplars without parameter updates, multimodal ICL remains largely underexplored, particularly regarding video demonstrations in interactive environments. For multimodal agents, learning from videos presents unique challenges: they must translate in-context demonstrations into executable policies, ground these policies in novel visual states, and iteratively refine actions based on environmental feedback. We introduce V-ICAL, a novel benchmark designed to evaluate video-based ICL for multimodal agents. Comprising 342 interactive tasks across 37 environments, V-ICAL utilizes human-curated demonstration videos as task-specific behavioral exemplars, evaluating agents through sustained interaction from a target initialization. The benchmark seamlessly connects in-context knowledge induction with core agentic capabilities, including state grounding, temporal memory, planning, and adaptation in dynamic environments. Extensive evaluations across 19 state-of-the-art multimodal agents reveal significant limitations: the best-performing model, Seed-2.1-Pro, achieves a score of only 54.4/100, while other leading models (e.g., Gemini-3.1-Pro, GPT-5.6) fail to surpass 50, far below the human baseline of 83.6. Controlled comparisons further demonstrate that current agents struggle to reliably translate video exemplars into effective policies, failing to yield consistent performance gains. Ultimately, V-ICAL exposes a critical gap in the ICL capabilities of multimodal agents, underscoring an urgent need for future research. --- ### Scalability and Performance Evaluation of Federated Learning Frameworks: A Comparative Analysis Authors: Bassel Soudan, Sohail Abbas, Ahmed Kubba, Manar Wasif Abu Talib, Qassim Nasir Published: 2026-09-14 Categories: cs.DC, cs.AI arXiv: https://arxiv.org/abs/2609.15681v1 This paper presents a systematic examination and experimental comparison of the prominent Federated Learning (FL) frameworks FedML, Flower, Substra, and OpenFL. The frameworks are evaluated experimentally by implementing Federated Learning over a varying number of clients, emphasizing a thorough analysis of scalability and key performance metrics. The study assesses the impact of increasing client counts on total training time, loss and accuracy values, and CPU and RAM usage. Results indicate distinct performance characteristics among the frameworks, with Flower displaying an unusually high loss, FedML achieving a notably low accuracy range of 66% to 79%, and Substra demonstrating good resource efficiency, albeit with an exponential growth in total training time. Notably, OpenFL emerges as the most scalable platform, demonstrating consistent accuracy, loss, and training time across different client counts. OpenFL's stable CPU and RAM underscore its reliability in real-world scenarios. This comprehensive analysis provides valuable insights into the relative performance of FL frameworks, offering good understanding of their capabilities and providing guidance for their effective deployment across diverse user bases. --- ### Projection-Free Multi-level Algorithms for Stochastic Constrained Compositional Optimization Authors: Wei Jiang, Sifan Yang, Wenhao Yang, Yibo Wang, Yuanyu Wan, Zechao Li, Lijun Zhang Published: 2026-09-14 Categories: math.OC, cs.LG arXiv: https://arxiv.org/abs/2609.15679v1 This paper studies projection-free algorithms for stochastic constrained multi-level compositional optimization. In this context, the objective function is a nested composition of several smooth functions, and the decision set is closed and convex. Since projection onto the constraint set can be computationally expensive, we develop projection-free methods that rely on linear minimization oracles. For non-convex objectives, we propose variance-reduced projection-free algorithms and establish complexity guarantees under both the Frank-Wolfe gap and the gradient mapping criteria. We also develop momentum-based methods that achieve convergence guarantees under weaker smoothness assumptions. Additionally, by using a stage-wise design, we derive a parameter-free variant that preserves the same complexities for the Frank-Wolfe gap. Such a design can be further used to develop algorithms for convex and strongly convex functions whose rates match those of single-level projection-free counterparts. Finally, we consider finite-sum problems and derive complexities for non-convex, convex, and strongly convex objectives. Numerical experiments across multiple tasks demonstrate the effectiveness of the proposed methods. --- ### MambaMPD: A Mamba-Driven Segmentation Framework for Marine Pollution Detection from Remote Sensing Imagery Authors: Shuaiyu Chen, Wei Han, Peng Ren, Chunbo Luo, Zeyu Fu Published: 2026-09-14 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.15676v1 Accurate marine pollution detection (MPD) is essential for protecting coastal ecosystems and marine biodiversity. Vision Mamba models have shown promise in remote-sensing semantic segmentation by efficiently capturing long-range dependencies and global context, yet their potential for MPD remains underexplored. MPD is particularly challenging because of low signal-to-noise ratios, fragmented pollution patterns, and indistinct boundaries caused by the visual similarity between pollutants and the surrounding sea. To address these issues, we propose MambaMPD, an enhanced Mamba-based framework incorporating two complementary structural priors: Frequency-Aware Augmentation (FAA) and multi-scale Edge-Guided Attention (EGA). FAA integrates wavelet transforms into the encoder to decompose features into multi-scale frequency subbands, enabling the model to capture low-frequency contextual semantics and high-frequency structural details needed to identify small, low-contrast, and irregular pollution patterns. EGA adaptively fuses hierarchical, Laplacian-derived boundary cues with deep semantic representations, refining encoder features before decoding to sharpen boundaries and reduce ambiguity in visually confusing, spatially fragmented scenes. Together, these modules improve sensitivity to subtle pollution signals while preserving fine boundary structures. A U-Net-style decoder with squeeze-and-excitation attention and deep supervision progressively restores and refines semantic and spatial information across scales. Extensive experiments on two benchmark MPD datasets show that MambaMPD achieves higher mIoU than competing methods while requiring substantially less computation than foundation-model-based approaches. On MADOS, it improves F1 by 3.6% over OSDMamba; on M4D, it raises Oil Spill IoU by 6.82% over TransOilSeg. --- ### Don't Send What You Don't Need: Question-Guided Token Pruning as a Privacy Defense for Vision-Language Models Authors: Md Khalid Syfullah, Alvi Ataur Khalil Published: 2026-09-14 Categories: cs.CV, cs.AI, cs.CR arXiv: https://arxiv.org/abs/2609.15671v1 Visual Question Answering (VQA) with Vision-Language Models (VLMs) is increasingly used in privacy-sensitive and bandwidth-constrained settings. Federated Learning (FL), Split Learning (SL), and U-Shaped Split Learning (USL) keep raw data local, but transmitting all visual tokens across a model partition remains costly and can expose private information. We propose QPriv-VL, a question-guided, privacy-aware token-pruning framework for FL, SL, and USL that prunes visual tokens before transmission based on task utility and privacy sensitivity. Its core component is a lightweight Dynamic Threshold Predictor (DTP) that jointly estimates a sample-specific pruning ratio and a token-level retention mask in one forward pass. DTP combines question relevance, computed from cross-modal similarity between visual patches and the pooled question embedding, with a sensitivity signal derived from frozen DINOv2 features. This allows the model to suppress potentially sensitive regions while preserving patches useful for answering the question, without requiring sensitivity labels. We evaluate QPriv-VL on GQA, OK-VQA, VQAv2, SLAKE, VQA-RAD, and PathVQA against four privacy attack families: FSHA, FORA, iDLG, and attribute-inference membership inference attacks. DTP matches or outperforms fixed-ratio pruning while using substantially fewer transmitted tokens. On VQA-RAD, it reduces membership-inference attack success from 0.99 to 0.76-0.79, lowers FSHA and FORA reconstruction PSNR relative to fixed-ratio pruning, and preserves competitive VQA accuracy using about 40% of the original visual-token budget. A sensitivity exclusion ratio of 1.20 +/- 0.18 indicates preferential removal of privacy-sensitive patches, while explainability analysis shows that retention adapts to question semantics rather than generic visual saliency. --- ### Benchmarking Intra-Patient 3D Deformable Multimodal Image Registration Authors: Matteo Barbieri, Giammarco La Barbera, Juan Pablo De La Plata, Sabine Sarnacki, Isabelle Bloch, Pietro Gori Published: 2026-09-14 Categories: cs.CV, cs.AI arXiv: https://arxiv.org/abs/2609.15669v1 Multimodal image registration is a key component of many clinical workflows, yet it remains challenging because corresponding anatomical structures often exhibit substantially different image intensities across modalities. In this work, we present a comprehensive benchmark of intra-patient 3D multimodal deformable registration methods across three datasets covering different anatomical regions and difficulty levels, including both synthetic deformation recovery and real clinical scenarios. We evaluate classical optimization-based approaches and modern learning-based methods, including recent deep learning and foundation models, using complementary metrics: Average Dice similarity coefficient (DSC), average 95th-percentile Hausdorff distance (HD95), and a modality-independent structural similarity measure based on the MIND self-similarity context (MIND-SSC). Results show high variability across datasets, with learning-based methods demonstrating superior performance on large synthetic benchmarks, while only limited improvements are observed in real pelvic registration. A key finding of this study is the consistent disagreement between geometric metrics (DSC, HD95) and image-based similarity metrics (MIND-SSC), highlighting that improved overlap does not necessarily imply better global multimodal correspondence. Furthermore, anatomy-guided approaches achieve the highest overlap scores but exhibit degraded performance outside of segmented regions, revealing a trade-off between label-driven alignment and global structural coherence. Overall, our results indicate that no current method achieves robust performance across anatomies and modalities. We demonstrate that intra-patient 3D multimodal registration requires multi-criteria evaluation, including deformation-based metrics, and remains an open problem. --- ### Circuit-MLLM: Topological Logic-Guided Latent-Space Visual Reasoning for Circuit Schematic Understanding Authors: Jinyuan Deng, Yuqi Jiang, Wenjing Huang, Xin Li, Qi Sun, Cheng Zhuo Published: 2026-09-14 Categories: cs.CV, cs.AI, cs.LG arXiv: https://arxiv.org/abs/2609.15668v1 Through pre-training on extensive text and image datasets, current multi-modal large language models (MLLMs) achieve strong performance on general tasks. However, circuit schematics present a unique challenge for MLLMs due to their dense component layouts and distinct topological logic, demanding fine-grained structural parsing to extract the electrical semantics. To address this, we propose Circuit-MLLM, a multimodal reasoning framework that reformulates circuit topology analysis as a process of device localization, path tracing, and sequential reasoning within the latent space. We introduce a circuit knowledge mining mechanism that deeply aligns the model's latent representations with structurally rich features derived from multi-granularity circuit vision experts, enabling the model to effectively internalize topological semantics. Building upon these internalized semantics, we devise a topology-guided sequencing strategy that decouples reasoning from the rigid raster-scan order, enforcing stepwise inference along the circuit's topological logic in latent space. Across diverse circuit analysis tasks, Circuit-MLLM consistently outperforms strong baselines, notably achieving a 25% higher average score than GPT-5.1, which demonstrates the effectiveness of our framework in circuit schematic topology analysis. Code is publicly available at https://github.com/IC-Yuan/Circuit-MLLM. --- ### CIDERS: Cloud-Edge LLM Collaborative Learning via Accelerating Personalized Bilevel Optimization Authors: Victor H. Chen, Hairui Yu, Stella K. Chung, Hong Yan Published: 2026-09-14 Categories: cs.DC, cs.AI arXiv: https://arxiv.org/abs/2609.15664v1 Amid the rapid advancement of physical-world intelligence, cloud-edge collaborative large language models (LLMs) have emerged as a promising roadmap for practical LLM deployment. However, existing cloud-edge paradigms struggle to balance global consensus with local personalization, which fails to satisfy the need for a unified knowledge foundation on the cloud and domain-specific adaptation at the edge. To address this, we introduce, for the first time, a personalized bilevel optimization framework that formalizes cloud-edge LLM collaboration as a dual structure: the upper level optimizes edge-side personalization, while the lower level governs cloud-side knowledge transfer, reaching cloud-edge evolving in coordination. We then propose CIDERS, an efficient solver that decomposes the model into a learnable backbone and a messenger. While the cloud performs knowledge transfer to the learnable backbone, the key lies in embedding global trajectories into each local personalization step via consensus-variate correction to reconcile personalization with consensus. We provide a comprehensive theoretical analysis, including a geometric characterization of the local trajectory and a full convergence guarantee, revealing an explicit trade-off structure between personalization and global convergence. Extensive experiments demonstrate that CIDERS consistently outperforms competitive baselines on the compressed edge path, with 3.1x and 1.7x gains on mathematical reasoning and code generation, respectively, and a 10\% relative gain on instruction metrics. Mechanism experiments attribute these gains to early consensus-corrected coordination and task-aware distillation. Overall, CIDERS offers a viable path toward consensus-guided continuous personalization in cloud-edge LLM systems. --- ### CiteShade: Citation Laundering in Multi-Source Retrieval-Augmented Generation and Its Counterfactual Defense Authors: Guo Fuzheng Published: 2026-09-14 Categories: cs.CR, cs.CL arXiv: https://arxiv.org/abs/2609.15660v1 Retrieval-augmented generation (RAG) grounds a language model's answers on retrieved external knowledge and returns each answer with citations that identify its sources. Those citations are the user's audit trail: they let a reader verify a claim without trusting the model. Prior security work on RAG asks whether an attacker can corrupt the answer, leaving the citation channel unexplored. We show that this channel is a new and practical attack surface. We propose CiteShade, the first citation laundering attack to RAG, in which an attacker controlling a single source induces a model to produce an attacker-chosen wrong answer and to attribute it to a trusted source that does not support it, while the evidence for the correct answer remains in context. We formulate the attack as an optimization problem, derive three necessary conditions (retrieval, generation, and citation) and construct sources satisfying them without any instruction. On multi-source multi-hop question answering the attack raises the wrong-answer rate from 0.01 to 0.68, and source deletion confirms the malicious source is the causal driver in every measured case. Vulnerability tracks a model's propensity to cite rather than its scale, reaching CLR 0.84 under explicit instruction and 0.64 with no instruction at all on the most citation-prone model tested. We then show that perplexity filtering and citation-support checking are each insufficient, and propose a counterfactual defense that verifies which source actually drove the answer. --- ### KaiNinja: Extending Native 3D Generators to the Part Level Authors: Ruihan Yu, Lian Fu, Muyao Niu, Zheng-hui Huang, Yu-Ju Tsai, Sho Kuno, Fengbo Lan, Yonghao Yu, Erwin Wu, Ming-Hsuan Yang, Kaipeng Zhang, Zhixiang Wang Published: 2026-09-14 Categories: cs.GR, cs.AI, cs.CV arXiv: https://arxiv.org/abs/2609.15659v2 Native 3D generators turn one image into a single mesh. TRELLIS.2 and its peers deliver high-fidelity non-watertight geometry with materials, but the output is one fused object, while downstream work such as editing, rigging and simulation operates on part-level assets. A naive idea is to run a 3D segmentation network on the fused mesh that TRELLIS.2 generates, but such pipelines are slow and bounded by the accuracy of the segmentation. We want a simple way to extend an existing native 3D generator to the part level. But we face a critical problem: the O-Voxel grid stores one sheet of surface per voxel, so a single volume cannot represent the interface where two parts touch, at any resolution. We introduce a dual-volume representation to solve this problem and put forward KaiNinja, a part-level extension of TRELLIS.2 built on a dual-volume form of its O-Voxel representation. KaiNinja keeps the generation speed and quality of TRELLIS.2 while extending it to the part level, with no mask or segmenter in the pipeline. Its training data come from sources of many kinds, including CAD models and assets authored by an LLM-driven agent; to our knowledge it is the first 3D generative model trained on agent-authored part data. Surprisingly, we also find that whole-object fidelity improves over the same backbone fine-tuned on the same dataset. Against part generation pipelines of different paradigms, it lowers whole-object Chamfer distance by 40% and raises strict part F-score by 16%. --- ### Predictive Likelihood Ratios for Language Model Watermark Detection Authors: Li Ma Published: 2026-09-14 Categories: stat.ML, cs.AI, cs.LG, stat.AP arXiv: https://arxiv.org/abs/2609.15657v1 Keyed watermark detection tests dependence between observed tokens and pseudorandom variables reconstructed from a secret key. Building on the pivotal framework of Li et al. (2025), we construct predictive likelihood ratios that average over uncertain probability deficits and residual-tail distributions. The aim is robust detection power across alternative specifications without requiring a single signal-strength tuning. A mixture prior combines tail shape and effective width; hierarchical extensions allow within-document variation in deficit or width. The test maximizes prior-averaged power at a fixed size, but is not generally uniformly most powerful or minimax. Under the exact conditional pivot null, normalized predictive alternatives selected before each observation yield a Bayes factor that is also a test martingale: Type I error control is unaffected by alternative misspecification and remains valid under optional stopping. This guarantee does not cover violations of the conditional null, and the interpolated implementation has no certified anytime guarantee. Gumbel marginal likelihoods are evaluated by fixed quadrature. Across the evaluated tail-shape and tail-width alternatives and three horizons, the union-tail mixture has maximum observed Type II error regret .0080, compared with .0962 for the equal-tail mixture, relative to the best tested rule. On temperature-matched outputs from two open models, it improves AUC over the equal-tail baseline in all eight non-saturated model-temperature cells, although the leading reference score generally has higher AUC. Supplementary experiments show retained power under independent null-like replacement and smaller changes from hierarchical dependence modeling. The evidence supports robustness across the evaluated alternatives, not uniform power guarantees or resistance to arbitrary text edits. --- ### Empathy Is Steerable but Multi-Axial: Mechanism Geometry and Persona Effects in LLMs Authors: JuHeon Ha, Byounghan Lee, Yunseo Choi, Kyung-Ah Sohn Published: 2026-09-14 Categories: cs.CL arXiv: https://arxiv.org/abs/2609.15654v1 Activation steering has been used to control traits such as honesty, refusal, and sycophancy, yet supportive empathy is evaluated along multiple dimensions that need not correspond to independently controllable activation directions. Using the EPITOME framework, which decomposes supportive empathy into Emotional Reactions, Interpretations, and Explorations, we study three instruction-tuned LLMs and ask whether candidate directions derived from these labels produce distinguishable intervention effects or instead share structure, and how persona prompts interact with those directions. We find that contrastive activation addition yields a stable middle-layer intervention that consistently shifts the EPITOME proxy scores across models, moving empathy analysis beyond response-level scoring. However, the recovered directions are only partially separable: steering one direction induces off-target shifts, and hand-crafted prompting shifts the empathy profile rather than isolating a single dimension. Persona prompts substantially change EPITOME scores, but a paired activation-shift decomposition shows that the recovered subspace captures only approximately 3 percent of persona-induced squared activation-shift magnitude at layer 15. Under this EPITOME-based definition, expressed empathy is steerable but multi-axial, and controlling persona-conditioned empathy requires targeting structure beyond individual mechanism directions. --- ### From Model Patterns to Abstract Semantics in Compositional Zero-Shot Learning Authors: Weize Li, Zhicheng Zhao, Fei Su Published: 2026-09-14 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.15649v1 Compositional Zero Shot Learning aims to recognize unseen compositions by recombining learned primitives. Recent methods rely on vision language models and attempt to explicitly model contextual variations of primitives through multiple representations. However, such approaches are limited by fixed variant capacity and competition between abstract and concrete semantics. In this work, we present a new perspective that views primitive variations as the context-driven activation of concrete visual cues rather than independent entities. Based on it, we propose CLEAR, a CLoze-style rEAsoning-based Re-ranking framework inspired by human perceptual processes. CLEAR extracts conditional variants from the primitive candidate set in a coarse-to-fine manner, performs cloze-style reasoning to infer high-level semantics, and re-ranks predictions to correct biases toward salient concrete primitives. Extensive experiments demonstrate that CLEAR consistently improves the Base Model and outperforms state-of-the-art methods on the challenging C-GQA and MIT-States datasets. Code is available at https://github.com/buptLwz/CLEAR. --- ### MedDiME: Efficient Latent Diffusion with Adaptive Masking for Medical Counterfactual Generation Authors: Yan Zeng, Changlu Guo, Anders Nymark Christensen, Morten Rieger Hannemose, Anders Bjorholm Dahl Published: 2026-09-14 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.15647v1 Medical counterfactual generation modifies images to change model predictions for interpretability. However, existing diffusion-based approaches are often prohibitively slow and memory-intensive, making them difficult to apply in high-resolution settings. Moreover, existing masking strategies are tightly coupled with pixel-space representations, making them incompatible with latent-space diffusion editing. To address these challenges, we propose MedDiME, a latent-space classifier-guided diffusion framework that reduces computational and memory overhead while introducing a latent-compatible, gradient-driven adaptive masking mechanism for spatially precise medical counterfactual generation. Extensive experiments demonstrate that MedDiME achieves high-quality counterfactual generation with significant efficiency gains compared to prior classifier-guided diffusion baselines, achieving up to 40 times faster inference and 13 times lower peak GPU memory usage. --- ### Principal-timestep Restricted Init via Sparse Matrix-decomposition in Flow-matching Authors: Jiayang Gu, Zheng Fang, Lichaun Xiang, Fanghui Liu, Xu Cai, Hongkai Wen Published: 2026-09-14 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.15643v2 Flow-matching diffusion models have recently emerged as a strong paradigm for high-fidelity visual generation. However, their prohibitively high fine-tuning cost limits scalability to downstream tasks. While Low-Rank Adaptation (LoRA) combined with spectral initialization has demonstrated accelerated convergence and improved performance in autoregressive language models by better aligning gradient directions, we find that it fails to deliver similar gains in diffusion fine-tuning, often yielding marginal or even negative improvements over vanilla LoRA.We attribute this discrepancy to a fundamental mismatch between LoRA's low-rank parameterization and the intrinsically high-rank gradients induced by the flow-matching objective. In particular, stochastic timestep sampling introduces directionally heterogeneous gradient signals across training steps, leading to misaligned updates under low-rank constraints.To address this issue, we propose Prism-LoRA,a Principal-timestep Restricted Init via Sparse Matrix-decomposition framework that improves gradient alignment during fine-tuning. Our method consists of two key components: (i) principal timestep selection, which restricts initialization gradients to a subset of dominant timesteps to suppress effective gradient rank, and (ii) principal channel filtering, which removes task-irrelevant channels, enabling the one-step spectral initialization gradient to better align with the long-horizon optimization trajectory. Extensive experiments demonstrate that our method consistently improves both convergence speed and final performance across multiple diffusion fine-tuning benchmarks, including subject-driven generation, controllable generation, and deblurring, achieving not only performance improvement but also earlier stages of convergence over baseline LoRA and other spectral-init methods. --- ### Human-Grounded Calibration for Long-Text Image-Text Congruence in Vision-Language Models Authors: Alessandro Gambetti, Qiwei Han Published: 2026-09-14 Categories: cs.CV, cs.CL arXiv: https://arxiv.org/abs/2609.15640v1 Long-text image--text congruence scoring is increasingly important for vision-language systems that must evaluate whether detailed textual descriptions match visual content. However, raw similarity scores from dual-encoder models are difficult to interpret as calibrated congruence measures, especially under the modality gap between image and text embeddings. This paper proposes Congruency Score (CS), a lightweight calibration layer that maps image--text similarity evidence into a bounded score. Using DOCCI and Urban1k, we evaluate four frozen vision-language backbones and show that observed reductions in post-projection centroid distance do not uniformly improve image--text retrieval performance. Human-grounded evaluations on DOCCI further reveal a trade-off: direct post-hoc calibration preserves high association with human judgments, whereas selected projection-based configurations can reduce threshold-relevant slope and intercept distortions at the cost of retrieval performance and association strength. These results establish long-text image--text congruence scoring as a calibrated score-estimation problem, where retrieval performance, human association, and threshold calibration must be evaluated as distinct objectives. CS provides a lightweight way to expose and operationalize this separation. --- ### SAM3D-Part: Interactive Part Selection and Generation from 3D Objects Authors: Jiahao Chang, Dong Du, Wanhu Sun, Yujian Zheng, Chuanyu Pan, Bowen Zhao, Chongjie Ye, Yuanming Hu, Xiaoguang Han Published: 2026-09-14 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.15639v1 Part-level control is essential for modern 3D asset creation, where objects are frequently edited, reused, animated, or fabricated through their individual components. In many such workflows, users need only several specific components rather than a complete object decomposition. However, existing 3D generation methods produce all parts regardless of user intent, while promptable 3D segmentation methods typically output partial surfaces instead of reusable complete meshes. In addition, image-conditioned part generators further struggle to preserve hidden geometry and accurate placement without directly conditioning on the source mesh. To address these problems, we present SAM3D-Part, a prompt-driven framework for selective part generation from input 3D object meshes. Given a source mesh and a part prompt, SAM3D-Part first encodes the source geometry into compact mesh features and aligns them with the rendered image, selective mask, and point-map observations via pixel-wise channel fusion. The fused representation conditions a feed-forward generative model to produce only the queried component as a completed mesh. To place the generated part back into the source coordinate frame, SAM3D-Part predicts dense per-voxel correspondences and estimates the part transformation from distributed spatial evidence rather than a single global pose code. For sequential multi-part queries, previously generated parts are stored in a part cache and reused as contextual constraints, reducing conflicts among independently requested components. Extensive experiments and ablations demonstrate that SAM3D-Part can significantly improve source alignment, reduce conditioning cost, and enable consistent selective part generation, achieving state-of-the-art. Code and weights will be available at https://github.com/Jiahao620/sam3d-part. --- ### Potential of Artificial Intelligence Algorithms for Identification of Relevant Diagnostic and Prognostic Biomarkers of Early-Stage Liver Cancer Authors: Ali Bou Nassif, Darko Castven, Manar Abu Talib, Jibran Sualeh Muhammad, Ahmed Ammar Kubba, Jens Marquardt, Abdalla Sayed Ali Published: 2026-09-14 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.15638v2 This study explores the use of deep learning and explainable artificial intelligence to diagnose hepatocellular carcinoma (HCC) and define effective biomarkers across five different stages of disease development using a transcriptomic biomarker HCC dataset constructed via semi-supervised learning from three source datasets. Several deep learning experiments were conducted with different feature extraction techniques and gene sets to identify the most effective features for training high-accuracy models with minimal loss. The best-performing model, using 15 selected genes with the SelectKBest algorithm, achieved 90.74% accuracy, while the model with the lowest recorded loss of 0.3187 was obtained using 20 selected genes. To address the issue of class imbalance in the dataset, a weighted training approach was conducted, and for model transparency and interpretability a SHAP-based XAI analysis provided insights into the model's decision-making, consistently finding DNAJB14 as the most influential gene. Functional validation in this study has provided compelling evidence that DNAJB14 plays an important role in the adverse properties of HCC and that its inhibition effectively reverses tumour cell migration, invasion, colony and sphere formation. The main limitation of this study is the dataset's class imbalance, and while weighted training helped mitigate this, further research and additional data are needed to guarantee model generalizability. Future studies should also explore the influence of genetic variations, environmental factors, and clinical differences on model performance across diverse populations. --- ### ModaLens: Measuring Image Sensitivity in Report-Conditioned Medical VLMs Authors: Sebastián Andrés Cajas Ordóñez, Maximin Lange, Quang Bui, Anqi Peter Li, Felipe Ocampo Osorio, Rafi Al Attrach, Kushul Reddy Palakala, Sahil Kapadia, Zakaria Laouabdia Sellami, Xinyue Zhang, Ashley Zhang, Leo Anthony Celi Published: 2026-09-14 Categories: cs.CV, cs.AI arXiv: https://arxiv.org/abs/2609.15635v1 A radiology report can already answer a clinical question, so it is hard to tell whether a vision-language model also uses the image. ModaLens, a paired image-swap audit, measures how report availability changes image sensitivity: MedGemma-27B on 3,199 paired MIMIC-CXR cases from 293 patients, all 14 questions per case (13 finding-specific and one composite), each image replaced by one from another study, usually of the same patient, with question and report fixed. Under an explicit answer instruction, the model's generated answer changes on 4.26 percent of trials with the report and 20.94 percent without it, a paired increase of 16.7 points (patient-clustered 95 percent CI 15.6 to 17.7), so report availability reduces image-swap sensitivity under this protocol; the original prompt with a lowercase first-token readout gives 4.70 percent against 17.07 percent, and substitutions also move continuous answer scores where the binary prediction does not change. The labels are derived from reports, which limits conclusions about visual correctness; the direction replicates in two further model lineages. Code, the exact prompts and a run record for every number are at https://github.com/criticaldata/MODALENS. --- ### FedLTLib: A Comprehensive Benchmark for Federated Long-Tail Learning Authors: Changkun Lin, Junxiao Wang Published: 2026-09-14 Categories: cs.LG, cs.AI arXiv: https://arxiv.org/abs/2609.15625v1 Driven by the escalating demand for privacy-preserving computing, Federated Learning (FL) has witnessed remarkable progress, becoming a cornerstone technology for bridging distributed data silos in mobile edge networks. However, in real-world mobile computing environments, data is generated by heterogeneous mobile devices with varying user behaviors, leading to a significant Long-Tail Distribution. Unlike idealized balanced datasets, data in the wild manifests an acute imbalance where a minority of head classes dominate the sample space while a vast number of tail classes, often representing rare but critical edge-case events, are extremely scarce. This data heterogeneity, which we formally characterize as "Double Heterogeneity", referring to the superposition of global class imbalance and local statistical skew, precipitates severe performance deterioration on tail classes, thereby spurring the vital research direction of Federated Long-Tail Learning (FL-LT). To standardize evaluation and accelerate research in this field, we introduce FedLTLib, a comprehensive benchmark tailored for FL-LT. Addressing the critical issues of inconsistent experimental configurations and unfair comparisons in prior work, FedLTLib establishes a standardized evaluation framework. The platform not only incorporates diverse benchmark datasets reflecting mobile data characteristics but also implements 13 state-of-the-art FL algorithms (4 traditional FL algorithms and 9 FL-LT algorithms). By leveraging FedLTLib, researchers can perform fair and reproducible evaluations of algorithm robustness and generalization capabilities under a unified experimental protocol, ultimately advancing the deployment of robust intelligence in mobile computing ecosystems. --- ### Beyond AI Literacy: A Structured Review and Exploratory Meta-Analysis of Measures for Competent Generative-AI Use Authors: Daniele Veri' Published: 2026-09-14 Categories: cs.HC, cs.AI arXiv: https://arxiv.org/abs/2609.15624v1 Researchers assessing competent generative-AI use at work must choose among self-reports, objective tests, and measures of oversight and reliance. We conducted a structured, seeded review of 24 focal empirical publications, starting from the 2024 COSMIN-based review and adding a targeted update through 17 August 2026. We grouped the measures into four domains: knowledge and use, epistemic oversight, reliance calibration, and operational control of tool-using agents. In an exploratory meta-analysis, we pooled three direct subjective-objective correlations from one research program (REML r = .055; Hartung-Knapp 95% CI (-.047, .156); combined reported N = 2,765). We could not resolve a discrepancy between the largest study's reported correlation and p-value, leaving its weight uncertain. Adding a synthetic mean of 12 cross-factor correlations from a fourth study gave r = .079 (95% CI (-.025, .181)). This sensitivity analysis concerns a broader comparison. From this small evidence base, we cannot establish a population correlation, validate workplace cutoffs, or justify substituting self-ratings for performance scores. We identified tests of foundation knowledge (AICOS-S and GLAT) and measures of verification, reliance, trust, and dependency. We found no validated individual-level instrument in the focal corpus that tests the full combination of agent scope, permissions, recovery, state isolation, independent review, and evidence-based closure; some cover subsets. We propose a four-layer workplace battery with non-compensatory decision rules, but have not tested its thresholds or whether it improves on other assessment approaches. --- ### Where to Compute and How to Interact: Operator-Readable Adaptation with Gauge-Aware Transport Authors: Zixuan Shen, Quanxu Wan, Bingchuan Wang, Zhi Wang, Biao Luo Published: 2026-09-14 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.15620v1 Adaptive meshes enable neural operators for partial differential equations (PDEs) to allocate spatial samples and computation according to local physical structures. Existing approaches, however, mainly address where to compute, with less attention to how information should interact after node relocation. Mesh adaptation changes local sampling scales, neighborhood structures, and geometric contexts, so representations formed at different nodes may not be directly comparable. Direct aggregation can therefore entangle physical variation with discretization-induced representation variation. Because allocation and interaction are jointly optimized through the same output objective, their individual roles are also difficult to distinguish from final errors alone. We introduce operator readability, requiring an adaptive operator to account for and test why computation is allocated to particular locations and how representations interact under the resulting nonuniform discretization. Based on this principle, we propose the Gauge-Aware Adaptive Mesh Neural Operator (GA-AMNO). Physics-informed adaptive allocation answers where to compute, while geometry-conditioned low-rank Gauge transport maps source features into target representation contexts before aggregation, answering how to interact. This makes mesh-to-solver information exchange inspectable and intervenable. We establish sufficient conditions for representation-consistent aggregation and analyze approximate transport errors and continuity under topology-preserving mesh deformations. Experiments on five PDE benchmarks demonstrate improved predictive accuracy, while controlled interventions and geometric-mismatch analyses verify the roles of allocation and interaction and show that Gauge transport improves cross-discretization representation compatibility under strong geometric mismatch. --- ### IROH: Insightful Ranking Of Humor using Multi-Stage Hybrid Retrieval with Rationale-Distilled LLM Judges for JOKER 2026 Track Task 1 English Authors: Ana-Maria Luisa Mocanu, Sebastian Mocanu, Ciprian-Octavian Truica, Elena-Simona Apostol Published: 2026-09-14 Categories: cs.IR, cs.CL arXiv: https://arxiv.org/abs/2609.15618v2 Our team, VANGUARD, presents IROH (Insightful Ranking of Humor), a three-stage retrieval system for JOKER Task 1 English at CLEF 2026, achieving first place on the leaderboard with 0.6347 MAP. Our pipeline combines hybrid sparse-dense retrieval, cross-encoder reranking, and a LoRA-adapted Large Language Model judge ensemble. We employ Gemma 4 to generate query-aware rationales under two prompt strategies, generic and typed, and produce up to four types of structured hard negatives for training data construction. Through an ablation across three cross-encoder architectures, four dense embedders, and eight judge configurations, our key findings are threefold: (1) the rationale-distilled judge is the primary driver of ranking quality, whereas appending rationales to the first-stage index contributes negligibly; (2) structured hard negatives degrade generalisation in nearly all configurations despite inflating local validation scores; and (3) across the components we ablate, the lighter, better-calibrated model is competitive with or stronger than its larger counterpart, with the generic-rationale Qwen2.5-7B judge (0.6055 MAP) outperforming every Gemma-4-31B configuration, and the advantage of generic over typed rationales is concentrated almost entirely in the smaller model. --- ### Multi-View Molecular Representation Learning with Hierarchical Graphs and Contextualized Fingerprints Authors: Gwang-Hyeon Yun, Jong-Hoon Park, Bing Hu, Helen Chen, Anita Layton, Young-Rae Cho Published: 2026-09-14 Categories: cs.LG, cs.AI arXiv: https://arxiv.org/abs/2609.15611v1 Molecular property prediction requires representations that generalize from limited labeled data to structurally novel compounds. Existing molecular pretraining methods often rely on a single view: graph-based approaches model atom-bond topology but provide limited fragment-level supervision, whereas fingerprint descriptors encode chemical patterns but are typically used as fixed auxiliary features. We propose HiFi-Mol, a multi-view framework that separately pretrains a hierarchical graph encoder and a contextualized fingerprint encoder before downstream integration. The graph branch uses fragment-aware masking with multi-resolution supervision to capture substructure-aware representations, while the fingerprint branch tokenizes active entries from seven fingerprint families and applies masked language modeling to learn contextualized embeddings. During fine-tuning, HiFi-Mol combines projected multi-resolution graph features with fingerprint embeddings for downstream prediction. Evaluated on MoleculeNet benchmarks under the scaffold split, HiFi-Mol achieves a 2.77% improvement in average ROC-AUC over the best baseline across eight classification tasks while maintaining competitive performance on three regression tasks. Further analyses reveal that fragment-aware masking improves graph representation quality, and classification results demonstrate dataset-dependent strengths of the individual graph and fingerprint variants, confirming that the two views provide complementary predictive signals. --- ### Diversified and Perceptible Counterfactual Examples Leveraging Expert Knowledge Authors: Akram Bensalem, Fahima Djelil, Marie-Jeanne Lesot, Gr{é}gory Smits Published: 2026-09-14 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.15609v1 CounterFactual Examples (CFEs) are a cornerstone of eXplainable Artificial Intelligence (XAI), offering local, post hoc, and model-agnostic explanations by identifying minimal input modifications that alter a model's prediction. Yet, in order to be intelligible, these modifications must also be semantically meaningful to the explainee. This paper proposes to integrate knowledge expressed as a fuzzy linguistic vocabulary to represent the explainee's perception and interpretation of the data. The domain induced by this fuzzy vocabulary imposes structural constraints that make the features dependent, preventing the use of gradient-based optimisation methods for CFE generation, e.g., DiCE. The paper proposes a continuous data embedding in this linguistic domain and exploits it to define a variant of DiCE that allows personalisation for the explainee, named DiCEf. As illustrated by experimental results on a real-world dataset, this extension of the DiCE method enables the generation of CFEs that are linguistically perceptible while preserving cost minimality, sparsity, and diversity. --- ### Through the Eyes of the Beholder: Biometric and Demographic Conditioning for Multimodal Sexism Detection Authors: Ana-Maria Luisa Mocanu, Sebastian Mocanu, Ciprian-Octavian Truică, Elena-Simona Apostol Published: 2026-09-14 Categories: cs.CL, cs.AI, cs.CV, cs.LG arXiv: https://arxiv.org/abs/2609.15608v1 Detecting sexism on the internet is a fundamentally subjective task; our team, VANGUARD, addresses this challenge in the EXIST 2026 Task 2 by proposing a human-centered multimodal framework that analyses and incorporates the psychological and demographic characteristics of human annotators into the detection pipeline. We fuse five input modalities through a cross-attention architecture with Feature-wise Linear Modulation conditioning. Meme text is extracted and visually described with Gemma 4, then augmented by automatic translation between English and Spanish with NLLB-200. Text and image representations are produced by LoRAadapted XLM-RoBERTa and CLIP encoders and fused with sensor features encoded by a pretrained autoencoder. To model annotator subjectivity, we frame Subtask 2.1 as a label distribution learning problem, optimizing a Kullback-Leibler divergence loss over the full annotator label distribution. At inference time, predictions are produced by soft-voting between the deep multimodal network and a complementary SVM trained on stylometric and physiological features. Our best submission ranks 29th out of 114 on Subtask 2.2 (source intention) under soft evaluation, and the normalized ICM scores remain above the baseline on Subtasks 2.1 and 2.2, indicating that annotator-centered conditioning contributes a usable signal. We release our full pipeline and analysis to support reproducible human-centered modeling. --- ### VideoScout: Learning Agentic Active Exploration with Adaptive Reasoning Pacing for Long Video Understanding Authors: Weixin Xu, Zhenyu Yang, Bing Wang, Shengsheng Qian, Changsheng Xu Published: 2026-09-14 Categories: cs.CV, cs.AI arXiv: https://arxiv.org/abs/2609.15606v1 Multimodal Large Language Models (MLLMs) have achieved remarkable progress on short video understanding yet remain limited on long videos due to the limited visual context window. Prevailing approaches rely on uniform frame sampling or recent coarse-to-fine agentic zooming, both of which struggle to localize sparse, decisive evidence in sufficiently long videos. We formulate long video understanding as a \textbf{Sequential Evidence Acquisition (SEA)} problem, in which an agent reads the video turn by turn along the temporal axis, deciding at each turn how fast to watch, what evidence to retain, when to revisit uncertain segments, and when to stop and answer. Inspired by this view, we propose \textbf{VideoScout}, a multi-turn reasoning agent that instantiates the SEA paradigm through adaptive reasoning pacing. Specifically, by dynamically controlling the viewing pace, VideoScout enables efficient traversal of long videos within a bounded visual context window, allowing the agent to access more video content while balancing content analysis depth with reading efficiency. To train VideoScout, we construct VideoScout-66K, a set of over 66K high-quality exploration turns from 10K answer-verified trajectories, and adopt a two-stage pipeline: cold-start supervised fine-tuning teaches the agent per-turn output format, while the Decoupled Clip and Dynamic sAmpling Policy Optimization (DAPO) algorithm performs trajectory-level reinforcement learning with a composite reward that jointly considers answer accuracy, output format compliance, and the temporal alignment between the agent's viewing progress and the teacher's answer timing measured by intersection-over-union (IoU). Extensive experiments on long video understanding and reasoning benchmarks demonstrate that our 7B model achieves strong performance compared with existing trained 7B agentic models. --- ### A Unified Vision-Language Model for PSMA PET/CT Report Generation, Visual Question Answering, and Lesion Segmentation Authors: Yang Xing, Jiong Wu, Savas Ozdemir, Yang Zhou, Boxiao Yu, Ying Zhang, Zheren Zhu, Chenyu You, Wei Shao, Yang Lu, Kang Wang, Tinsu Pan et al. Published: 2026-09-14 Categories: cs.CV, cs.AI arXiv: https://arxiv.org/abs/2609.15603v1 Accurate PSMA PET/CT interpretation is central to prostate cancer management, yet existing PET/CT AI models typically address isolated tasks. We propose a unified PSMA PET/CT vision-language model for report generation, visual question answering, and lesion segmentation. The framework adopts an LLaVA-style architecture, comprising a PET/CT vision encoder, an MLP-Mixer projection module, a LoRA-tuned large language model, and a 3D segmentation branch. Training followed a four-stage strategy: vision encoder pretraining, projection-layer alignment, VLM fine-tuning, and final multitask tuning. Language tasks used 5,747 PSMA PET/CT datasets with paired reports, while segmentation used the PSMA subset of AutoPET. The model outperformed PET2REP and a CT-based baseline across standard report-generation metrics, improved performance across VQA question types, and achieved higher Dice and lesion-level overlap F1 than SegAnyPET and nnUNet. These results support the feasibility of a unified framework for structured, interactive, interpretable PSMA PET/CT analysis with voxel-level grounding within a single multitask model architecture. --- ### Self-Evolving Memory for Generative Recommendation Authors: Xinyu Lin, Zhuosong Jiang, Zixiao Suo, Siqin Wang, Hanqing Zeng, Hanchao Yu, Yinglong Xia, Jiang Zhang, Aashu Singh, Fei Liu, Wenjie Wang, Fuli Feng et al. Published: 2026-09-14 Categories: cs.IR, cs.AI arXiv: https://arxiv.org/abs/2609.15598v1 Generative recommendation has emerged as a promising end-to-end paradigm for personalized recommendation. However, user preferences continuously evolve over time, making self-evolving an essential capability for generative recommender systems. Existing evolving strategies, such as continual retraining and distillation-based adaptation, directly update the shared model parameters using streaming interactions. Nevertheless, we find that directly applying such strategies to generative recommendation introduces a critical issue, termed evolution conflict. Specifically, heterogeneous preference shifts from different users are optimized within a fully shared autoregressive parameter space, causing dominant behavioral patterns to progressively dominate the model evolution process while underrepresented patterns become increasingly overlooked. To address this issue, we propose a self-evolving memory paradigm for generative recommendation, aiming to enable effective evolution across heterogeneous behavioral patterns. We further identify three key principles for effective self-evolving recommendation systems, including isolated memorization, reinforced evolution, and scalable application. Guided by these principles, we develop LION, a simple yet effective framework centered on a sparse Key-Value memory layer. Specifically, LION introduces sparse memory activation to isolate the evolution of different behavioral patterns, while a consolidation loss is designed to reinforce the learning of underrepresented preference dynamics during continual adaptation. Extensive experiments on diverse real-world datasets demonstrate the effectiveness of LION under various continual evolution settings (e.g., per-period evaluation, user/item group evaluation, and evolution convergence analysis). The codes are released at https://github.com/JazyJiang/Self-Evolving-Memory-for-Generative-Recommendation. --- ### What a gated sensing pipeline never looks at: bandwidth reduction and the misses behind it Authors: Raghu Venkat, Tricha Anjali Published: 2026-09-14 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.15579v1 An airborne sensor on a contested link cannot send video, so the appealing move is to send findings instead and report the ratio between the two. We evaluate a gated sensing pipeline that does this, combining learned object detection and image-text comparison with deterministic scheduling, gating, evidence accumulation and transmission rules. On staged footage with the semantic stage live it sends 38,736 bits over 211 s, a reduction of 41,977x, and names 1 of 4 staged events with no false report. That detection has since been superseded: a correction to how the tracker measures speed removed the measurement artefact the normality model had been learning from, and the flight no longer warms. Under the corrected code the pipeline names 1 of 7 staged events across four flights, and we report both. On a control flight where nothing was staged it reports nothing, a reduction of 155,830x: the largest number in the study and the least informative, because a reduction ratio measures the scene. The mechanisms that produce the reduction also decide which observations ever reach a decision, so the two cannot be reported apart. We give a tick-level trace of one flight (3,187 rows) that places each of three missed events at the stage where it stopped progressing: one produced no track, one failed the structural place test at 0.129, and one passed 539 structural ticks but reached only 3.192 against a boundary of 3.807. We also report one instance of a known failure mode, an online normality model absorbing the object it will later judge, measured against the threshold that object then failed. The evidence is one detection and six misses across four staged flights, beside one clean control, and we treat it as a case study. We give the reproduction protocol and generated results, identify which supporting artifacts are not distributed, and state which experiments did not run. --- ### GRIN+: Towards Fast Yet Effective Machine Unlearning for Imbalanced Medical Data Authors: Minghui Huang, Junxiao Wang Published: 2026-09-14 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.15571v1 As deep learning models become fundamental to modern healthcare, the "Right to be Forgotten" mandated by privacy regulations like GDPR and HIPAA necessitates effective machine unlearning (MU) to remove sensitive patient data from trained models. However, existing MU techniques often struggle with a fundamental "privacy-efficiency-utility" (PEU) trilemma, particularly in medical scenarios where data is frequently characterized by severe class imbalance and long-tailed distributions. In such cases, standard unlearning methods can fail to protect key clinical knowledge or mistakenly delete features essential for diagnosing rare conditions due to the gradient dominance of majority classes. To address these challenges, we propose GRIN+, a novel machine unlearning framework designed for fast and precise data erasure in imbalanced medical scenarios. GRIN+ decouples unlearning-specific knowledge from generalized representations at the parameter level by analyzing the gradient contributions of both "forget" and "retain" sets. It introduces a class-adaptive influence scoring mechanism to rectify gradient dominance and employs a direction-constrained update strategy to prevent the unintended erosion of vital clinical knowledge. Comprehensive benchmarking across multiple medical datasets, including skin cancer (ISIC), brain tumor (MRI), and breast ultrasound (BUSI), demonstrates that GRIN+ achieves an optimal balance of the PEU trilemma. Experimental results show that GRIN+ maintains high diagnostic accuracy and robust privacy while significantly enhancing runtime efficiency compared to existing baselines. We open-source the GRIN+ code and benchmarks to support further research. --- ### Big Brains and Changing Environments: Cause or Consequence? Authors: Sian Heesom-Green, Jonathan Shock, Geoff Nitschke Published: 2026-09-14 Categories: cs.NE, cs.AI, q-bio.PE arXiv: https://arxiv.org/abs/2609.15569v1 Large brains are metabolically costly, and associations with changing environments do not imply they evolved there, as the Cognitive Buffer Hypothesis (CBH) would suggest. They may instead evolve in stable conditions and later facilitate colonization of changing environments. Using neuro-evolution in an artificial seasonal foraging task, we compared agents evolving exclusively in changing environments to agents first evolved in static environments before transitioning. Results show that larger neural networks in dynamic environments arise mainly from prior static evolution, achieving superior performance under unpredictable changes. Our results challenge strict CBH predictions, provide agent-based (computational) support for a colonization-based account and highlight the role of evolutionary history in brain size evolution. --- ### PIVOT: Physics-Grounded Verification for AI-Generated Audio-Video Detection Authors: Bo Zheng, Kangran Zhao, Xiaoyu Zhang, Weinan Guan, Zhiheng Li, Yize Chen, Haizhou Li, Qingshan Liu, Siwei Lyu, Baoyuan Wu Published: 2026-09-14 Categories: cs.CV, cs.AI, cs.MM arXiv: https://arxiv.org/abs/2609.15562v1 As generative models continue to advance, AI-generated content (AIGC) is becoming increasingly realistic, weakening the artifact cues commonly exploited by existing detectors. Nevertheless, faithfully reproducing the physical behavior of real-world events remains challenging for current generators. We therefore explore detecting AIGC by assessing whether the depicted event satisfies measurable constraints derived from physical laws. We introduce PIVOT, a physics-grounded AIGC detector, instantiated here for audio-video clips, that estimates physical quantities from video and audio, selects physical laws relevant to each clip, and verifies their measurable constraints. Beyond a real/fake decision, PIVOT returns supporting evidence that records the verification outcome, relevant time window, and supporting quantities for each applicable law. Although instantiated and evaluated here on audio-video data, the framework can, in principle, extend to other AIGC modalities whenever the physical quantities required for verification can be estimated reliably. We also introduce PhysForensics-Bench, comprising paired real and generated audio-video clips from nine event-centric scene families and two recent audio-video generators. On PhysForensics-Bench, PIVOT achieves 70.30% accuracy and 64.29% F1 score on Real+Seedance, and 72.16% accuracy and 65.82% F1 on Real+VEO. In comparison, direct inspection with Gemini 3.1 Pro obtains 53.96% accuracy and 60.09% F1 on Real+Seedance, and 57.22% accuracy and 63.44% F1 on Real+Veo. These results demonstrate the practical promise of physical-consistency verification as a structured and inspectable source of evidence that complements artifact-based AIGC detection. --- ### Can We Trust the Judges? Validation of Factuality Evaluation Methods via Answer Perturbation Authors: Sarra Gharsallah, Adele Robaldo, Mariia Tokareva, Giovanni Gatti Pinheiro, Ilyana Guendouz, Raphaël Troncy, Paolo Papotti, Pietro Michiardi Published: 2026-09-14 Categories: cs.CL arXiv: https://arxiv.org/abs/2609.15561v1 Evaluating the factual correctness of large language models (LLMs) is vital for many applications. But are our evaluation tools themselves trustworthy? Despite the rise of factuality-based metrics, their sensitivity and reliability remain underexplored. This paper introduces a meta-evaluation framework that systematically tests these metrics using controlled corruptions of gold standard answers. Our method generates ranked outputs with known degrees of degradation to probe how metrics capture nuanced changes in truthfulness. Our experiments reveal that pipeline-based methods, such as the RAGAS's factual correctness metric, better track degradation than LLM-as-judge approaches. We also propose a new variant of the factual correctness metric that provides a competitive and cost-efficient. --- ### Don't Count the Edits, Judge by the Outcome Alone: Reward-Based Evaluation for Grammatical Error Correction Authors: Hayeong Ryu, Sunhee Jo, Seunguk Yu, YoungBin Kim Published: 2026-09-14 Categories: cs.CL, cs.AI arXiv: https://arxiv.org/abs/2609.15559v1 Grammatical error correction (GEC) evaluation has traditionally relied on reference or edit overlap, which can penalize valid rewrites that differ from gold corrections. Reference-free metrics reduce this dependence, but evaluating whether a fluent output is a valid correction of the source remains challenging. We propose SURE, a source-conditioned reward evaluator trained on within-source preferences spanning minimal-edit and rewrite-oriented corrections. SURE jointly learns an overall reward with criteria-level supervision for grammaticality, faithfulness, and fluency, together with span-level grounding for source-side error resolution. Experiments on SEEDA show that SURE performs competitively against strong baselines, with particular gains on rewrite-style corrections and more disentangled criteria-level diagnostics. Our code is available at https://github.com/hayeonggg/SURE. --- ### Bayesian Optimisation Using Product-of-Experts Gaussian Process Models with Uncertainty Calibration Authors: Yean Hoon Ong Published: 2026-09-14 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.15555v1 Bayesian optimisation (BO) typically relies on a single global Gaussian process (GP) model as its surrogate model. However, GP regression has cubic computational complexity in the number of training data points, limiting its applicability to large-scale optimisation problems. The product-of-experts Gaussian process model with uncertainty calibration (GP-pro-c) mitigates this limitation by combining multiple local GP experts, enabling improved uncertainty quantification, reduced computational cost, and preservation of global correlations. Despite these desirable properties, the use of GP-pro-c in BO has not been thoroughly studied. This paper introduces BO-pro-c, a Bayesian optimisation algorithm that uses GP-pro-c as its surrogate model, and evaluate its performance across a diverse range of BO settings. Experimental results suggest that BO-pro-c maintains competitive optimisation performance while achieving a 0.9% reduction in simple regret and a 39.4% reduction in computational overhead relative to a BO algorithm based on a single global GP model. --- ### Multi-Stage NeRF for Efficient 3D Coronary Artery Reconstruction from Two Narrow-Angle Angiographic Projections Authors: Deyu Meng, Mojtaba Lashgari, Yiying Wang, Abhirup Banerjee Published: 2026-09-14 Categories: cs.CG, cs.CV arXiv: https://arxiv.org/abs/2609.15550v1 X-ray coronary angiography is the clinical gold standard for coronary artery disease during real-time cardiac interventions, but provides only 2D projections of inherently 3D vessels. Existing learning-based 2D-to-3D reconstruction methods typically require wide angular coverage or multiple views, assumptions that are rarely satisfied in routine practice where only two projections with narrow angular separation are available. To address these challenges, we propose NeCA++, a multi-stage self-supervised neural radiance field (NeRF) framework tailored to clinically realistic acquisition constraints. The framework decomposes reconstruction into two stages that progressively refine spatial support and representation capacity. In the first stage, a coarse 3D representation of the vasculature is reconstructed, restricting the subsequent optimisation to regions with a higher likelihood of vessel presence, termed an active region. Afterward reconstruction is restricted to this region while higher-resolution representations are progressively activated to recover fine vascular details. This multi-stage strategy focuses learning on anatomically plausible regions, mitigates gradient dilution under extreme sparsity, and stabilises global topology before recovering fine vascular branches. Furthermore, two vessel-specific regularisations are introduced: a ray-aligned constraint to reduce projection-induced ambiguity, and a bimodal density penalty to enable early vessel-background separation. Extensive experiments across three datasets (ImageCAS, ASOCA, and Synthetic RCA) and four angular configurations demonstrate consistent superiority over state-of-the-art baselines, particularly under clinically realistic narrow-angle settings, while achieving reconstruction within 58 seconds per case. --- ### The Token Before the Value Is the Key: How Hybrid Architectures Organize Induction Circuits Authors: Ke Cheng, Xin Xu, Yixiao Chen, Lei Xin, Jianbo Zhao, Fanhu Zeng, Yue Liu, Jun Zhang, Jie Jiang Published: 2026-09-14 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.15545v2 Hybrid language models can improve capability as well as efficiency, raising the question of how architectural complementarity becomes learned computation. We examine the established induction roles of Carrying predecessor information, Matching a source by content, and Copying its value. How are these position-sensitive and content-based computations allocated across heterogeneous layers? We introduce layer-type-agnostic paired probes that track Carrying and Matching through a common block-update interface. In recurrent--global and local--global hybrids, Carrying concentrates in efficient layers and Matching in global receivers. The measured local contribution concentrates on lag one: the token immediately before the historical value. Changing predecessor support through lag-one masking, convolution removal, or early learning-rate reduction can relocate Carrying and Matching between stages. Source-key restoration and fixed-value selection trace the receiver's dependence on the prepared source. These interventions also change natural-text recall, with outcomes depending on configuration and target. Varying local windows and induction-enriched training text changes the early development of functional Carrying and Matching, connecting architectural priors and training evidence to formation timing. Together, the probes and interventions shift the explanatory focus upstream: the organization of Matching follows how Carrying is learned. The token before the value provides a concrete link between a hybrid's architecture, circuit development, and recall. Code is available in https://github.com/ckpassenger/bind-match-copy/tree/main. --- ### Specifying Reward Functions for RL Without Environment Sampling Authors: Stephane Hatgis-Kessell, W. Bradley Knox, Emma Brunskill Published: 2026-09-14 Categories: cs.LG, cs.AI arXiv: https://arxiv.org/abs/2609.15544v1 Enabling human stakeholders to specify reward functions that lead to their desired outcomes is a key challenge in deploying reinforcement learning agents. Preference-based methods such as online RLHF can reduce the burden of manual reward design, but they require repeatedly training policies, sampling trajectories from the real world, and eliciting feedback, making them impractical in settings where environment interaction is computationally expensive or unsafe. We introduce Experience-Free Autonomous Reward Specification (EARS), a method for learning reward functions from preferences without environment interaction. Our approach uses a structured LLM-mediated process to construct a small set of expressive reward features from a task description and the environment observation space, then strategically samples imagined trajectories in this feature space and learns feature weights from preferences over the imagined trajectory pairs. We evaluate on three long-horizon domains: pandemic lockdown regulation design, insulin administration for diabetes patients, and autonomous vehicle control on a highway. We compare EARS to baselines that also enable reward specification without environment interaction--namely, methods that directly prompt an LLM to generate a reward function. When learning from either ground-truth preference labels or preferences labeled by a LLM, EARS designs reward functions that are more aligned with the ground truth reward function that produced the preferences or LLM context than these baselines. These results suggest that preference-based reward specification remains effective without environment sampling, enabling practical reward design in settings where collecting real trajectories is costly or infeasible. --- ### Strong and Compact Policies for Submodular Markov Decision Processes via LP-Based Submodular Orienteering Authors: Lars Rohwedder, Rico Zenklusen Published: 2026-09-14 Categories: cs.DS, cs.LG arXiv: https://arxiv.org/abs/2609.15539v1 Finding policies for Markov Decision Processes (MDPs) is a central problem in areas such as Reinforcement Learning and Operations Research. Here, we have to repeatedly choose an action that should be performed by an agent. Depending on the action and the current state of the agent, the agent collects a reward and randomly transitions into a new state. The goal is to maximize the reward in expectation over a finite time horizon of length $H$. We consider a recently introduced variant that generalizes the traditionally additive reward function in the model to a monotone submodular one, which allows for capturing a range of interesting applications. Without the stochastic component, this problem is equivalent to the Submodular Orienteering problem, where the goal is to find an $s$-$t$ walk in a directed graph maximizing a monotone submodular function under a length constraint. We present a novel LP-based algorithm for Submodular Orienteering using ideas from the Sherali-Adams hierarchy and Round-or-Cut. Our guarantees are comparable to the known quasi-polynomial time logarithmic approximation for Submodular Orienteering, but also extend to the setting of Submodular Markov Decision Processes. In the polynomial time regime, we present an $O(n^{\varepsilon})$-approximation (and $O(H^{\varepsilon})$ for Submodular MDPs) for every $\varepsilon >0$, where $n$ is the number of vertices, which was unknown even for Submodular Orienteering. Prior to our work, the best known approximation guarantee for Submodular MDPs had an approximation ratio linear in $H$. Beyond these algorithmic results, our methods reveal a trade-off between the approximation guarantee and the number of previously visited vertices on which an agent conditions its decision. --- ### The Misery of Mechanistic Interpretability: A Formal Perspective Authors: Tobias Ladner, Matthias Althoff Published: 2026-09-14 Categories: cs.LG, cs.AI arXiv: https://arxiv.org/abs/2609.15533v1 Mechanistic interpretability has become the dominant lens for understanding frontier language models, as their inner workings are complex and inherently black boxes. To gain insights into these models, interpretable replacement networks (IRNs) are trained at all layers, exposing interpretable features through sparsely activated neurons. However, the faithfulness of an IRN is usually evaluated only empirically on clean data, and we show that even semantically minor input perturbations flip the dominant IRN features-and thus the human-understandable interpretation-across five open-weight model families (GPT-2 small, Gemma 2 2B, Gemma 3 1B, Llama 3.2 1B, R1-Distill-Qwen 1.5B). We propose the first formal verification framework for the faithfulness of an IRN, where reachability analysis certifies a sound upper bound of the faithfulness gap in adversarial scenarios. Moreover, we show that verification-aware training of IRNs substantially tightens this certified bound, restoring a feature-level interpretation that safety auditors can act on. Together, these results give, to the best of our knowledge, the first formal guarantees for mechanistic interpretability of large language models. --- ### Option-Aware Retrieval and Task-Specific VLM Adaptation for Medical VQA Authors: Tristan Kirscher, Niklas C. Koser, Soren Pirk Published: 2026-09-14 Categories: cs.AI, cs.CL, cs.CV arXiv: https://arxiv.org/abs/2609.15530v1 We describe our submission to the MedReason 2026 challenge, covering multiple-choice (MCQ) and open-ended (OE) medical visual question answering (VQA) under fully offline, containerized inference. Our first finding is that MCQ retrieval must compare answer \emph{semantics} rather than answer labels: labels are independently assigned per question, so copying a retrieved neighbor's label transfers no useful information, whereas scoring each current option's text against correct-answer text from similar training cases raises retrieval-only accuracy from 20.0\% to 57.5\% on a 200-case retrieval-excluded development holdout. Our second finding attributes the submitted system's accuracy: holding the task-specific MCQ Low-Rank Adaptation (LoRA) adapter fixed and varying the number \(k\) of in-prompt retrieved examples changes accuracy by at most one case --- 187/200 (93.5\%) at both \(k=0\) and the adapter's training-time \(k=1\), 188/200 (94.0\%) at the packaged runtime's default \(k=3\) --- and the submitted confidence-gated override adds no net accuracy on top of \(k=3\), selecting the VLM in 198/200 cases. With the final MCQ adapter fixed, retrieval changes accuracy by at most one case, and gating provides no net gain. On 20 OE cases, token-F1 and RaTEScore~\cite{zhao2024ratescore} decrease as \(k\) grows, but paired sign tests on token-F1 differences are nonsignificant (\(p \ge 0.29\)); a single-annotator comparison found 6/20 wrong-anchor errors for the final configuration and 14/20 for an earlier configuration that jointly differed in routing, adapter, and prompting. The system reaches 94.0\% MCQ accuracy on the development holdout and 93.20\% on the organizer's official pre-evaluation, versus 29.43\% for the off-the-shelf reference baseline, while both of the organizer's open-ended scores are lower than that baseline's (ground-truth agreement 1.245 versus 1.588, visual accuracy 1.995 versus 2.696, each out of 4). ---