As generative AI systems become competent and democratized in science, business, and government, deeper insight into their failure modes now poses an acute need. The occasional volatility in their behavior, such as the propensity of transformer models to hallucinate, impedes trust and adoption of emerging AI solutions in high-stakes areas. In the present work, we establish how and when hallucinations arise in pre-trained transformer models through concept representations captured by sparse autoencoders, under scenarios with experimentally controlled uncertainty in the input space. Our systematic experiments reveal that the number of semantic concepts used by the transformer model grows as the input information becomes increasingly unstructured. In the face of growing uncertainty in the input space, the transformer model becomes prone to activate coherent yet input-insensitive semantic features, leading to hallucinated output. At its extreme, for pure-noise inputs, we identify a wide variety of robustly triggered and meaningful concepts in the intermediate activations of pre-trained transformer models, whose functional integrity we confirm through targeted steering. We also show that hallucinations in the output of a transformer model can be reliably predicted from the concept patterns embedded in transformer layer activations. This collection of insights on transformer internal processing mechanics has immediate consequences for aligning AI models with human values, AI safety, opening the attack surface for potential adversarial attacks, and providing a basis for automatic quantification of a model's hallucination risk.

Multi-modal Large Language Models (LLM) have advanced conversational abilities but struggle with providing live, interactive step-by-step guidance, a key capability for future AI assistants. Effective guidance requires not only delivering instructions but also detecting their successful execution, as well as identifying and alerting users to mistakes, all of which has to happen in real-time. This requires models that are not turn-based, but that can react asynchronously to a video stream, as well as video data showing users performing tasks including mistakes and their corrections. To this end, we introduce Qualcomm Interactive Cooking, a new benchmark and dataset built upon CaptainCook4D, which contains user mistakes during task execution. Our dataset and benchmark features densely annotated, timed instructions and feedback messages, specifically including mistake alerts precisely timestamped to their visual occurrence in the video. We evaluate state-ofthe-art multi-modal LLMs on the Qualcomm Interactive Cooking benchmark and introduce LIVEMAMBA, a streaming multi-modal LLM designed for interactive instructional guidance. This work provides the first dedicated benchmark and a strong baseline for developing and evaluating on live, situated coaching.

Graph Transformer (GT) leveraging the powerful Transformer architecture to learn graph-structured data. However, effectively representing graph information while ensuring efficiency remains challenging, as our analysis reveals that graph-scale operations still constitute the computational bottleneck in current GT designs and limit their applications to large graphs. In this work, we tackle the GT scalability issue by proposing HubGT, which is boosted by decoupled graph computation and hierarchical graph representations. HubGT represents graph information with a novel hub labeling scheme, which encompasses enriched neighborhoods for node token generation, and fast computation for distance-based positional encoding. Notably, the precomputation and training of HubGT achieve complexities linear to the number of graph edges and nodes, respectively, while the training stage completely removes graph-related computations, leading to favorable mini-batch capability and GPU utilization. Extensive experiments demonstrate that HubGT offers efficient computation and mini-batch capability over existing GT designs on large-scale datasets while achieving top-tier effectiveness. Our code is available at: https://github.com/gdmnl/HubGT.

Time series foundation models (TSFMs) demonstrate impressive zero-shot performance for time series forecasting. However, an important yet underexplored challenge is how to effectively finetune TSFMs on specific downstream tasks. While naive finetuning can yield performance gains, we argue that it falls short of fully leveraging TSFMs' capabilities, often resulting in overfitting and suboptimal performance. Given the diverse temporal patterns across sampling scales and the inherent multi-scale forecasting capabilities of TSFMs, we adopt a causal perspective to analyze finetuning process, through which we highlight the critical importance of explicitly modeling multiple scales and reveal the shortcomings of naive approaches. Focusing on encoder-based TSFMs, we propose MultiScale FineTuning (MSFT), a simple yet general framework that explicitly integrates multi-scale modeling into the finetuning process. Experimental results on three different backbones (MOIRAI, MOMENT and UNITS) demonstrate that TSFMs finetuned with MSFT not only outperform naive and typical parameter efficient finetuning methods but also surpass state-of-the-art deep learning methods. Codes are available at https://github.com/zqiao11/MSFT.

Consistency learning with feature perturbation is a widely used strategy in semisupervised medical image segmentation. However, many existing perturbation methods rely on dropout, and thus require a careful manual tuning of the dropout rate, which is a sensitive hyperparameter and often difficult to optimize and may lead to suboptimal regularization. To overcome this limitation, we propose VQ-Seg, the first approach to employ vector quantization (VQ) to discretize the feature space and introduce a novel and controllable Quantized Perturbation Module (QPM) that replaces dropout.

Video diffusion models substantially boost the productivity of artistic workflows with high-quality portrait video generative capacity. However, prevailing pipelines are primarily constrained to single-shot creation, while real-world applications urge multiple shots with identity consistency and flexible content controllability. In this work, we propose EchoShot, a native and scalable multi-shot framework for portrait customization built upon a foundation video diffusion model. To start with, we propose shot-aware position embedding mechanisms within the video diffusion transformer architecture to model inter-shot variations and establish intricate correspondence between multi-shot visual content and their textual descriptions. This simple yet effective design enables direct training on multi-shot video data without introducing additional computational overhead. To facilitate model training within multi-shot scenarios, we construct PortraitGala, a large-scale and high-fidelity human-centric video dataset featuring cross-shot identity consistency and fine-grained captions such as facial attributes, outfits, and dynamic motions. To further enhance applicability, we extend EchoShot to perform reference image-based personalized multi-shot generation and long video synthesis with infinite shot counts. Extensive evaluations demonstrate that EchoShot achieves superior identity consistency as well as attribute-level controllability in multi-shot portrait video generation. Notably, the proposed framework demonstrates potential as a foundational paradigm for general multi-shot video modeling.

Turns per Sample 1.52 Word Count Statistics Avg.

We introduce MIRAGE, a new benchmark for multimodal expert-level reasoning and decision-making in consultative interaction settings. Designed for the agriculture domain, MIRAGE captures the full complexity of expert consultations by combining natural user queries, expert-authored responses, and image-based context, offering a high-fidelity benchmark for evaluating models on grounded reasoning, clarification strategies, and long-form generation in a real-world, knowledgeintensive domain. Grounded in over 35,000 real user-expert interactions and curated through a carefully designed multi-step pipeline, MIRAGE spans diverse crop health, pest diagnosis, and crop management scenarios. The benchmark includes more than 7,000 unique biological entities, covering plant species, pests, and diseases, making it one of the most taxonomically diverse benchmarks available for vision-language models, grounded in the real world. Unlike existing benchmarks that rely on well-specified user inputs and closed-set taxonomies, MIRAGE features underspecified, context-rich scenarios with open-world settings, requiring models to infer latent knowledge gaps, handle rare entities, and either proactively guide the interaction or respond. We evaluate more than 20 closed and open-source frontier vision-language models (VLMs), using an ensemble of reasoning language models as evaluators, highlighting the significant challenges posed by MIRAGE.

Recent work has revealed that state space models (SSMs), while efficient for longsequence processing, are fundamentally limited in their ability to represent formal languages--particularly due to time-invariant and real-valued recurrence structures. In this work, we draw inspiration from adaptive and structured dynamics observed in biological neural systems and introduce the Adaptive Unitary State Space Model (AUSSM): a novel class of SSMs that leverages skew-symmetric, input-dependent recurrence to achieve unitary evolution and high expressive power. Using algebraic automata theory, we prove that AUSSM can perform modulo counting and simulate solvable group automata at precision logarithmically bounded in the input length, enabling SSMs to model a broad class of regular languages out of reach for other SSM architectures. To overcome the practical inefficiencies of adaptive recurrence, we develop a separable convolution formulation and a CUDA implementation that enables scalable parallel training. Empirically, we show that AUSSM and its hybrid variant--interleaved with Mamba--outperform prior SSMs on formal algorithmic tasks such as parity and modular arithmetic, and achieve competent performance on real-world long time-series classification benchmarks. Our results demonstrate that adaptive unitary recurrence provides a powerful and efficient inductive bias for both symbolic and continuous sequence modeling.

Moral judgment is integral to large language models' (LLMs) social reasoning. As multi-agent systems gain prominence, it becomes crucial to understand how LLMs function when collaborating compared to operating as individual agents. In human moral judgment, group deliberation leads to a Utilitarian Boost: a tendency to endorse norm violations that inflict harm but maximize benefits for the greatest number of people. We study whether a similar dynamic emerges in multi-agent LLM systems. We test six models on well-established sets of moral dilemmas across two conditions: (1) Solo, where models reason independently, and (2) Group, where they engage in multi-turn discussions in pairs or triads.