Leaderboards, eval harnesses and ablations — the contested business of deciding which model is actually better.
arXiv:2609. 16933v1 Announce Type: cross Abstract: Modern predictive systems expose multiple quantities that are commonly interpreted as measures of confidence.
The paper introduces ResLRP, an extension of Layer-wise Relevance Propagation that explicitly handles residual connections in Vision Transformers to prevent attribution explosion. It demonstrates that residual cancellation causes instability in ViT explanations, and that ResLRP improves faithfulness and localization across a wide range of ViT architectures, including Vision Language Models. The method also provides a diagnostic measure for predicting attribution degradation and successfully localizes Sparse Autoencoder features.
The paper introduces DA‑EGO, an efficient global optimization algorithm that dynamically aggregates high‑dimensional design spaces into low‑dimensional subspaces for surrogate‑based search. The algorithm updates subspace variables each iteration using variable‑interaction analyses, perturbation, and ANOVA, and adaptively adjusts search ranges based on previous results. Tests on 21 benchmark functions and real turbomachinery problems demonstrate DA‑EGO’s effectiveness, especially on separable and partially separable problems, while noting case‑dependent performance on non‑separable functions.
RiskChainBench is a new benchmark that pairs 3,600 synthetic token‑text restoration inputs with 600 human‑labeled local web environments to evaluate how well models can recover obfuscated platform messages and then investigate the associated websites. The benchmark measures both message restoration accuracy and the subsequent web‑investigation decision, using a fixed multimodal evidence judge to assess faithfulness, sufficiency, completeness, and consistency. Across ten models, performance varies widely, with entry recovery ranging from 35.2% to 95.2% and web decision accuracy from 26.3% to 62.8%, highlighting execution failures as the main bottleneck.
BenSparX introduces the first Bengali conversational speech dataset for Parkinson’s disease detection and pairs it with a robust, explainable machine learning framework. The framework uses diverse acoustic features, systematic feature selection, and advanced classifiers, achieving 95.67% accuracy, 95.62% F1, and 0.990 AUC. SHAP analysis is employed to explain feature contributions, and the model outperforms state‑of‑the‑art methods on other language datasets.
BiCLIP is a bidirectional multimodal framework that enhances medical image segmentation by allowing visual features to iteratively refine textual representations, improving semantic alignment. It incorporates an augmentation consistency objective to stabilize learning against perturbed inputs. Experiments on QaTa-COV19 and MosMedData+ show that BiCLIP outperforms state‑of‑the‑art image‑only and multimodal baselines, achieving strong performance even with only 1% labeled data and resisting common clinical artifacts such as motion blur and low‑dose CT noise.
arXiv:2609.17521v1 Announce Type: cross Abstract: Interactive control for video generation is moving from coarse prompts toward fine-grained, physically meaningful manipulation of dynamic scenes. Yet...
RAM‑H1200 is a new public benchmark comprising 1,200 hand radiographs from six medical centers, each annotated with whole‑hand bone instance segmentation, pixel‑level bone erosion masks, joint regions of interest, and joint‑level Sharp‑van der Heijde (SvdH) scores for bone erosion and joint space narrowing. The dataset enables unified multi‑level analysis of anatomical structure, localized erosive pathology, and clinically standardized RA severity, filling a gap in existing resources that lack full‑hand coverage and fine‑grained annotations. Initial benchmark results show strong performance in bone segmentation but highlight that bone erosion segmentation remains a significant challenge, underscoring the dataset’s potential to advance quantitative RA analysis.
The paper titled "Position: AI Is Not Ready for Strategic Conflicts" argues that language‑model (LM) based open‑ended strategic wargames, while useful for simulating adversaries, institutions, and crisis response, pose significant safety risks. It identifies five failure modes—decision laundering, adjudication opacity, role collapse, escalation‑through‑adjudication, and failure of strategic imagination—and contends that such wargames should not inform real‑world planning or policy without an auditable safety case. Instead, the authors suggest using these simulations primarily as stress tests to expose potential failures in decision‑influencing LM agents.
QueryFormer is a unified architecture designed for post‑click conversion rate prediction, addressing both feature interactions and sequential user behaviors. It introduces a stackable field–sequence block that generates query tokens via cross‑attention and packs sequence queries into shared‑parameter attention, improving efficiency and accuracy. The model won first place in the KDD Cup 2026 Tencent UniRec Challenge Industrial Track with an AUC of 0.83254, and scaling studies show that increasing view width slightly boosts validation AUC while maintaining low latency.
The paper introduces a coverage-aware virtual IMU augmentation framework for human activity recognition. It selects diverse and scarce data points in a learned sensor embedding space, generates virtual IMU samples as prompts, ranks them by proximity and label consistency, and incorporates them into training with reliability-based weights. Experiments on public benchmarks demonstrate consistent performance gains over existing baselines, with ablation studies confirming the framework’s effectiveness.
The paper evaluates strategies for increasing plan‑execution flexibility by converting sequential plans into partial‑order plans through deordering and reordering. It compares block deordering methods, which restructure causal dependencies, with MaxSAT‑based approaches that optimize within existing causal structures. The study finds that block deordering consistently outperforms MaxSAT in both effectiveness and efficiency, offering anytime solutions and higher flexibility gains per computation time.
The paper investigates whether natural language inference (NLI) can be performed using only interpretable, graph-based representations of evidence. It introduces a pipeline that decomposes sentences into atomic propositions, maps them to ConceptNet triples, and feeds three graphs—premise, hypothesis, and a retrieved ConceptNet subgraph—into a fine‑tuned 0.8‑billion‑parameter language model. On the SNLI dataset the graph‑only model reaches 89.7% accuracy, close to a text‑based baseline, while on ANLI it matches RoBERTa‑large on rounds R2 and R3 but lags on R1, illustrating a trade‑off between interpretability and performance.
The paper introduces MASA, a method for Wild Test-Time Adaptation that uses a frozen multimodal large language model to provide structured semantic anchors, thereby avoiding the self-referential loop common in existing WTTA techniques. MASA selects a small, diverse set of reliability-ranked anchors, encodes their descriptions, propagates them to nearby test samples, and stores this visual‑semantic information in an online prototype memory. The stored descriptors enable lightweight adaptation of normalization parameters, and MASA is evaluated on the WTTA ImageNet‑C benchmark with ResNet and ViT backbones under limited‑batch, mixed‑domain, and imbalanced‑label‑shift scenarios.
The paper introduces ORDER, a task‑conditioned retrieval‑augmented generation framework that dynamically adapts both indexing and retrieval strategies to each incoming query. It first clusters questions to learn cluster‑specific chunking, metadata filtering, and reranking settings, then routes queries to the appropriate pre‑built index via nearest‑centroid assignment. Additionally, a supervised query router predicts relevant collections and a Uniform Multi‑source Sampler distributes the retrieval budget evenly across selected sources, yielding superior performance on heterogeneous historical archives compared to existing RAG systems.
The paper introduces a schema‑adaptive action‑conditioned Joint‑Embedding Predictive Architecture (SAAC‑JEPA) for cross‑machine CNC transfer when only a subset of sensors overlap between source and target machines. Experiments show that pretraining does not improve source‑only forecasting, but a carefully selected action‑conditioned JEPA model achieves a zero‑shot RMSE of 0.546 on the target, outperforming persistence but falling short of certain baseline models. Ablation studies reveal that adding RevIN improves RMSE but harms calibration, and limited post‑lock adaptation can further reduce error.
OPD‑Aha is a privileged on‑policy distillation method that improves multimodal reasoning by reconstructing the distillation target from the teacher’s isolated visual preference instead of relying on fragile teacher‑student discrepancies. It suppresses continuations that contradict the image, encouraging students to interrupt flawed reasoning with reflection tokens such as "wait" and "actually." This approach leads to consistent improvements across fine‑grained perception and complex multimodal reasoning benchmarks.
RoleBreak is an open benchmark designed to evaluate long‑horizon role‑playing robustness in spoken dialogue systems. It includes 310 character‑based and user‑centered roles, 6,688 human‑verified dialogue turns, and 11,743 fine‑grained evaluation criteria, with 1,856 turns specifically targeting expressive vocal emotion. The benchmark stresses role consistency, interaction quality, safety, and affect over extended conversations, and the authors evaluated nine system configurations across full‑duplex, omni‑modal, and cascaded ASR–LLM–TTS paradigms.
The paper surveys continuous‑time (CT) machine learning, a framework for modeling temporal dynamics as continuous processes, especially useful when data are sampled irregularly or over long horizons. It introduces a unified taxonomy that groups major CT methods by their underlying mathematical formulations and shows how different architectural choices—such as vector‑field parameterization, stochasticity, memory mechanisms, and discretization—relate these families. The survey compares training algorithms, optimization strategies, failure modes, computational complexity, and benchmarks, reviews supporting software ecosystems, and outlines open challenges and future research directions.
RegRet is a large multimodal model framework that improves region-level retrieval by adding a Region‑Aware Encoder and a multi‑stage training pipeline featuring localized captioning and regional contrastive learning. It also introduces the REGMB benchmark, containing 225k contrastive pairs across four multimodal retrieval tasks. Experiments show RegRet surpasses strong baselines in zero‑shot settings and gains over 20% improvement on REGMB and public benchmarks while maintaining global retrieval performance.