Niclas Lietzow, Danielle Bitterman, Carsten Eickhoff, William Rudman, Michal Golovanevsky
Vision-language models must reconcile visual evidence with memorized world knowledge when the two conflict. How they resolve this conflict shapes the reliability of multimodal systems, yet prior work characterizes it behaviorally without a component-level causal account. We combine activation patching across three granularities (residual stream, attention heads, and MLP sublayers) with model-component ablation studies and mechanistic analysis. Across three VLM families, we find that visual grounding emerges by default, whereas prior grounding depends on a small set of causally necessary attention heads (2.5-4.8%) concentrated in the second half of the network. These heads enable answers from stored world knowledge (e.g., "red" for a strawberry) despite conflicting visual input. Ablating them flips predictions from knowledge-grounded to visually grounded answers in 68-96% of cases under prior-knowledge prompts, but changes only 0.8-7.5% of visually grounded predictions, establishing an asymmetric causal structure. The identified heads decompose into routing heads, which modulate information flow, and writing heads, which directly project answer tokens into the residual stream. This structure is consistent across model families and scales, revealing a sparse causal circuit underlying perception-knowledge conflict in VLMs.
Language models lose accuracy when a prompt contains non-evidential context: irrelevant statements from unrelated domains, or random off-distribution tokens, with no competing answer and no semantic overlap with the query. Such noise pervades retrieval-augmented, tool-using, and long-context deployments, yet the failure has lacked a mechanistic account. We trace the corruption to the Intrusion Pathway, which starts with MLPs enriching both query-subject and irrelevant-token representations, parroting heads carry prefix-token content to the prediction position, and relational readers largely avoid unrelated sources. Nevertheless, preserved subject enrichment and selective reading do not prevent intrusion in the broader computation. We then localize part of this interference to a sparse set of attention heads, which we call intrusion heads. Their prefix-derived writes promote distractor-related content, while their reduced support for the correct answer arises primarily from weaker non-prefix contributions rather than direct suppression by the prefix. The same heads transfer across tasks and distractor types and are largely weight-disjoint from the heads that integrate relevant contextual evidence, enabling a class-conditional intervention. Replacing the corrupted last-token residual with its clean counterpart, an oracle probe that requires a matched clean forward pass, recovers 97.6% on average of the distractor-induced drop across four downstream tasks. A deployable mask level knockout of about ten heads, roughly 1% of the model's attention heads, recovers 57% at bounded cost to helpful context.
Hao Li, Zhengkun Zhang, Gangqiang Hu, Zhen Zhang, Yude Gao, Dai Dai, Jing Liu
Reinforcement Learning with Verifiable Rewards (RLVR) is expanding from tasks with well-defined correctness signals, such as mathematics and code, toward multifaceted quality requirements specified by multi-dimensional rubrics. Since policy optimization consumes one scalar per rollout, rubric-based pipelines must map multiple criterion scores into a scalar reward. This aggregation is often treated as score scaling, but it implicitly determines how quality dimensions trade off during training. The prevailing practice, normalizing each criterion and taking a linear combination, assumes that cardinal score differences are comparable across criteria and that gains on one criterion compensate for failures on another; both assumptions are unreliable when criteria are semantically heterogeneous. We propose Reinforcement Learning with Verifiable Rubric-based Ranking (RLVR$^2$), a verifiable ranking paradigm for rubric-based RLVR. For each criterion, RLVR$^2$ converts rubric scores into criterion-specific within-group ordinal outcomes, recovers a latent utility from the resulting comparison matrix, and merges these utilities into one training signal. By retaining only within-group ordering and discarding raw score magnitudes, RLVR$^2$ avoids calibrating heterogeneous rubric scales. It further supports objective-preserving attribute adjustment: auxiliary attributes that correlate with observed rankings but are not training objectives can enter the estimation without expanding the rubric or rewarding them directly. Across three model scales and 16 benchmarks, RLVR$^2$ consistently outperforms representative rubric-based baselines, achieving the best overall performance on most benchmarks at every scale. Analysis shows it controls systematic effects tied to reasoning efficiency and response formatting while preserving the quality objective.
Large language models (LLMs) are increasingly deployed as AI agents that operate in short reasoning-action loops, interleaving model computation with external calls. Unlike traditional chat applications, these agentic workloads require inference serving systems to balance low latency, stable token emission, and throughput under multiple request arrivals from different AI agents. Recent deployments highlight a shift toward running small language models (SLMs) locally on consumer-grade GPUs, driven by privacy, compliance, and cost constraints. When heterogeneous requests overlap on a single GPU, long prefills and short decodes contend for resources, creating head-of-line blocking that destabilizes interactive performance. By analyzing agent workloads, we observe that their execution naturally separates into cold prefills, which process long system prompts, resume prefills, which append tool outputs to cached contexts, and short decodes, which are latency-critical. This mix intensifies contention compared to conventional chatbot serving. We present AgentServe, a single-GPU serving system that ensures stable multi-agent execution under such conditions by isolating prefills from decodes, applying dynamic budgeting to resume prefills, and allocating GPU resources through pre-established CUDA Green Context slots with adaptive control. Evaluation results show that AgentServe significantly improves latency stability while sustaining competitive throughput, achieving up to 2.8x TTFT improvement and 2.7x TPOT improvement over state-of-the-art baselines across different settings.
In vehicular integrated sensing and communication (ISAC) networks, each service-provider vehicle uses a limited power budget to serve multiple sensing targets and communication receivers. Maximizing the communication sum-rate subject to sensing signal-to-interference-plus-noise ratio (SINR) constraints is nonconvex, and iterative optimization is costly to rerun for every network layout. This paper proposes a graph neural network (GNN)-based power allocation algorithm, termed GNN-ISAC. Concretely, GNN-ISAC addresses the aforementioned problem by considering two distinct edge types in the graph representation, one capturing the inter-provider interference, the other representing the intra-provider power-budget sharing. GNN-ISAC is trained without labels using a sensing-penalized sum-rate loss. We also adapt the weighted mean-square error minimization (WMMSE) algorithm to this problem by introducing a sensing-power reservation step. Simulation results show that GNN-ISAC achieves communication sum-rates close to WMMSE with low sensing-link violation rates and much less computation. It also outperforms an existing heterogeneous GNN and a channel-blind baseline, and generalizes to larger networks without retraining.
Distributed Quantum Computing (DQC) offers a promising pathway to scale superconducting quantum hardware, but restricted inter- and intra-module connectivity creates a major transpilation bottleneck. Current DQC transpilation methods often assume idealized hardware connectivity, ignore subsequent intra-module routing, or suffer from substantial computational overheads that limit their applicability at scale. To address this, we introduce a scalable transpilation method tailored for the restricted connectivity of distributed superconducting quantum architectures. Specifically, our quantum interaction graph (QIG) approach couples a graph-partitioning mapping step with a fast SABRE-inspired routing heuristic. On a 48-qubit (3 QPUs $\times$ 16 qubits) architecture with all-to-all inter-core connectivity, our method reduces average aggregated costs by up to 24.72% (unstructured) and 24.01% (structured), while delivering >250$\times$ speed-ups for structured workloads compared to the state-of-the-art DMapS algorithm. On the 399-qubit (3$\times$133) IBM Flamingo, where DMapS times out on dense circuits, our approach maintains practical runtimes. Against a DQC-adapted SABRE baseline, it reduces structured circuit routing costs by up to 51.94% (all-to-all inter-core topology) and 51.96% (restricted line topology). Ultimately, this work delivers a practical transpilation method for DQC at scale.
VA-TDM: A Unified View for Improving Few-Step Diffusion Models by Velocity Adversarial Distillation
Yihong Luo, Tianyu Pang, Tianyang Hu
AI 速览：提出速度对抗蒸馏（VAD）框架并构建VA-TDM，以轨迹分布匹配和中间转移监督提升少步扩散/流模型的生成质量。
Few-step distillation offers efficient generation with diffusion and flow models, but preserving quality with only a few sampling steps remains challenging. Distribution-matching methods via score distillation provide a strong foundation, yet they rely on accurately tracking the student's changing distribution. At intermediate generation steps in few-step sampling, student-score estimation can be difficult, and commonly used estimators may describe a different distribution from the one used to update the student. Flow matching allows a more flexible approach to supervision through velocity fields constructed from the student's sampling process. Building on this perspective, we introduce Velocity Adversarial Distillation (VAD), which uses a learned critic to compare student and teacher fields and turn their disagreement into feedback for learning. VAD unifies score distillation, Fisher-based matching, and moment matching, also explaining existing heuristic MMD updates and delivering new training objectives. We further develop VA-TDM, which follows trajectory distribution matching: providing supervision for partial denoising along the student trajectory. Beyond this, VA-TDM combines transition-aware online velocity learning and dual-form VAD supervision across intermediate student transitions. Extensive results show that VA-TDM is powerful for learning few-step generators, achieving state-of-the-art four-step performance in both text-to-image and text-to-video generation across various backbones and scaling effectively to 33B MiniMax-H3. In particular, our VA-TDM-H3 notably outperforms the few-step Fast-H3 and is closer to the original many-step H3 according to human preference.
LLMs on edge devices are constrained by limited memory and bandwidth. Although 1.58-bit ternary weights greatly reduce model storage, long-context inference is increasingly bottlenecked by KV-cache traffic and attention. Existing accelerators also use separate datapaths for ternary linear layers and attention, incurring hardware and execution overhead. We present SPARTA, a sparse ternary LLM accelerator for efficient long-context edge inference. SPARTA integrates: (1) a phase-aware sparse dataflow exploiting heterogeneous sparsity across compute-bound prefill and memory-bound decode; (2) a unified compute fabric mapping ternary SpMM/SpMV and bit-serial attention onto shared processing elements to exploit zero-weight and zero-bit sparsity; and (3) an adaptive two-stage attention engine combining page-level Top-k selection to reduce K-cache accesses with token-level Top-p pruning to reduce V-cache traffic. SPARTA reduces KV traffic by 4.1-4.8x over full attention and 1.3-1.5x over a state-of-the-art sparse-attention scheme, with task-dependent quality loss. Compared with NVIDIA Jetson AGX Orin, SPARTA achieves 3.9x higher generation throughput and 8.3x higher end-to-end energy efficiency on average; versus AMD Ryzen 7 9700X, the gains are 5.0x and 81.7x. It also delivers 1.6-12.0x higher throughput, 2.0-4.2x higher energy efficiency, and 1.3-51.4x higher area efficiency than FIGLUT, TeLLMe, and MCBP.
Marcus Rolf Peter Ritt, Alexsandro Santos da Rosa Júnior, Marcos Vinicius Reballo, Cesar Augusto do Amaral, Fernando Augusto Caletti de Barros
This work studies hybrid quantum-classical neural networks for learning routing heuristics. Specifically, this paper asks whether small quantum neural networks can replace parameter-heavy modules inside a competitive attention-based routing model while maintaining solution quality. For the capacitated vehicle routing problem, encoder feed-forward replacement emerges as the most promising design: it reduces the number of model parameters by 56.6% while keeping the hybrid model close to the classical neural baseline at small and medium instance sizes, although the gap grows for larger instances. This work also compares to classical routing algorithms, which remain highly competitive and often superior on the fixed Euclidean test sets. Our results therefore do not indicate quantum advantage or solver dominance, but identify encoder feed-forward replacement as a viable hybrid-module compression strategy for neural combinatorial optimization.
Victoria Clerico, William Simon, Corey Liam Lammie, Orhun Görkem, Athanasios Vasilopoulos, Hsinyu Tsai, Jeronimo Castrillon, Abu Sebastian 等 9 位作者
LLM inference is inherently heterogeneous, with performance bottlenecks shifting throughout inference phases down to individual operators. This has motivated heterogeneous serving systems that disaggregate execution across specialized accelerators, typically through fixed partitions such as prefill/decode or attention/FFN. However, these approaches commit to a partitioning strategy in advance, limiting their ability to adapt to different workloads and model architectures while forcing emerging accelerators into predefined roles. We present HELIX, a vLLM-based software stack that discovers, compiles and dispatches workload-aware partitions across heterogeneous accelerators at the operator granularity. At engine initialization, HELIX automatically determines hardware-specific operator placements that minimize execution latency, accounting for profiled execution costs, communication overheads, device capabilities, and memory constraints. It combines dynamic programming, piecewise compilation, and heterogeneous subgraph dispatch, enabling accelerators with partial operator coverage to participate in LLM inference without predefined model partitions. HELIX provides an open-source unified framework for exploring, evaluating, and integrating heterogeneous accelerators for LLM inference, lowering the barrier to emerging hardware by exposing them as subgraph executors. In systems combining H100 and B200 GPUs with emerging accelerators, HELIX resolves LLM partitions that multi-query decoding latency by up to 3$\times$ for dense and sparse transformer models and up to 3.5$\times$ for hybrid state-space models, relative to standard attention-FFN disaggregation.
Zhehao Zhou, Tongrui Su, Shengyu Zhu, Da Li, Zeyu Liu, Jun Zhu, Xueqi Cheng
Latent Chain-of-Thought (CoT) reasoning improves efficiency by shifting intermediate computation into continuous hidden states. However, fully latent methods often struggle on precision-critical tasks such as mathematical reasoning, while existing hybrid approaches retain substantial token-generation overhead. We introduce VITAL, an importance-guided hybrid reasoning framework that selectively verbalizes informative tokens as reasoning anchors while keeping most computation latent. A lightweight gating head dynamically selects between explicit token generation and direct hidden-state feedback. VITAL constructs hybrid trajectories from explicit CoT using token- importance estimation and trains the model through curriculum-based supervised training, followed by reinforcement learning with a reward that balances correctness and conciseness. Experiments show that VITAL consistently out- performs existing fully latent reasoning methods, achieving relative average accuracy gains of 9.6% on GPT-2 and 5.1% on Llama-3.2-1B at comparable reasoning lengths. Compared with hybrid baselines, VITAL achieves higher reasoning accuracy with substantially shorter reasoning trajectories on Qwen3-1.7B.
LLM-as-a-judge is a common substitute for costly human evaluation, and aggregating several judges is the standard remedy for any single judge's bias. Judge pools mix judges of very different reliability, while the human-labeled sets that validate them hold a few hundred items or fewer. In this regime the per-item mean weights weak judges as heavily as strong ones, and confusion-table models such as Dawid-Skene must estimate a per-judge error table from too few items. We propose AGORA (Aggregation Of Reliability-weighted Assessments), an item-response-theory aggregator that places every item on one latent quality scale and describes each judge by a sensitivity parameter and ordered rating thresholds, so judge reliability is learned from all items jointly and without human labels. Across three benchmarks, AGORA matches or exceeds Dawid-Skene on large evaluation sets, and its advantage widens as the set shrinks toward a few hundred items. It remains accurate as pools grow more uneven, a label-free dispersion statistic indicates in advance when trust weighting helps, and its learned trust weights track per-judge human validity.
Serving Chain-structured Jobs with Large Memory Footprints with Application to Large Foundation Model Serving
Tingyang Sun, Ting He, I-Hong Hou
AI 速览：针对大内存足迹的链式结构作业服务，提出“服务器链组合”问题，通过块放置与缓存分配建模流水线并行下的大模型服务，并给出可扩展且具性能保证的算法。
As a current trend in Artificial Intelligence (AI), large foundation models are increasingly employed as the core of AI services. However, even after training, serving such models at scale remains a challenging task due to their heavy resource footprints, particularly in terms of GPU memory. While recent works revealed unique characteristics of systems serving foundation models that distinguish them from traditional distributed computing systems, there is still a lack of fundamental understanding of the underlying system management problems. This work aims at addressing this gap by extracting a novel problem of "server chain composition" via block placement and cache allocation for serving chainstructured jobs with large memory footprints, which models a fundamental problem in serving large foundation models through pipeline parallelism. After showing the NP-hardness of the optimal solution, the focus is turned to developing scalable algorithms with guaranteed performance under state-of-the-art load balancing. Application of the proposed solution to a distributed large language model (LLM) serving system shows significant reduction of response times compared to state-of-the-art solutions.
Modern data architectures for integrating heterogeneous data sources to discover actionable knowledge include the data warehouse, the data lake, and the data lakehouse, which combines the flexibility of the data lake with the structured data management capabilities of the data warehouse through metadata, indexes, and caching. Enterprise knowledge graphs (KGs) offer an alternative approach to such integration for the purposes of data analysis, but handling them at scale must address the Three Vs of big data: volume, velocity, and variety. Contextualized KGs (CKGs) meet part of that challenge by enriching facts with contextual information, yet many applications require only a context-specific subset of that knowledge. We propose the Knowledge Graph Lakehouse (KG Lakehouse), a cloud-native microservice architecture that adapts the data lakehouse paradigm for management of CKGs at scale. The KG Lakehouse indexes incoming source documents along configurable context dimensions and stores these documents in a data lake, from which application-specific CKGs are constructed on demand instead of maintaining a single, monolithic KG. We realize the proposed architecture as a proof of concept, with independently deployable services built from established cloud technology; context dimensions and format-specific document handling enter as configuration, leaving the architecture independent of the application domain. We evaluate the implemented system in experiments on a distributed cluster deployment.
Gino Kwun, Dhanvi Bharadwaj, Gokul Subramanian Ravi
Variational Quantum Algorithms (VQAs) potentially offer a pathway to practical quantum advantage, but their optimization is heavily hindered by barren plateaus and numerous local minima. While classically simulable Clifford circuits can warm-start VQAs to accelerate convergence, existing heuristic-based initialization methods struggle to scale within vast combinatorial search spaces. We address this bottleneck with CRiSP (a Clifford Reinforcement Learning agent for State Preparation), which casts Clifford prefix selection as a sequential decision-making problem. CRiSP utilizes Neural-Guided Monte Carlo Tree Search, driven by a Transformer-based policy trained via self-play, to insert learned Clifford gates before fixed parameterized rotations. This enables the construction of high-quality initial states entirely through polynomial-time classical stabilizer simulation. By integrating a curriculum learning strategy that progressively expands the search horizon, the agent efficiently scales to deep circuits. Evaluated on QAOA benchmarks of up to $22$ qubits and $1{,}370$ parameters, CRiSP outperforms state-of-the-art Clifford initialization methods by a mean of $3.19\times$ (max $45.02\times$) in average energy accuracy and $2.41\times$ (max $16.01\times$) in best-achieved energy accuracy. Assessments on VQE tasks further demonstrate the framework's robustness and generalizability.
We study the mixing time of the single-site Glauber dynamics for sampling proper $q$-colorings on sparse Erdős-Rényi random graphs $\mathcal{G}(n, d/n)$. We prove that, whenever $q \ge (1 + o_d(1))d$, Glauber dynamics mixes in polynomial time with high probability over $\mathcal{G}(n, d/n)$ for all sufficiently large integers $n$. This improves the previous bound $q = (α^* + o_d(1))d$ where $α^* = 1.7632...$ is the unique solution of $x = \exp(1/x)$ on $[1, \infty)$ by Efthymiou, Hayes, Štefankovič, and Vigoda (SODA, 2018) and asymptotically matches the uniqueness threshold on infinite regular trees. Our proof has three main ingredients. First we establish logarithmic bounds on the spectral and coupling independence constants of the Gibbs distribution by extending the tree analysis of Chen, Liu, Mani, and Moitra (J. ACM, 2026). Second, we extend the entropy-factorization framework for two-spin systems on random graphs, developed by Bezáková, Galanis, Goldberg, and Štefankovič (TALG, 2024), to proper colorings by introducing a buffer around high-degree vertices. Finally, we compare the resulting heat-bath block dynamics with the single-site Glauber dynamics.
Xuwei Liu, Zhimin Zhang, Xiaoguang Hu, Zhongwei Liu, Yuanhui Sun, Jun Huo, Hua Wu, Haifeng Wang
Machine learning supports materials design, atomistic simulation and experimental characterization, yet models and workflows remain fragmented. We present PaddleMaterials, an open-source framework integrating heterogeneous materials-learning tasks while preserving task-specific representations and scientific semantics. It unifies data processing, training, evaluation, pretrained-model management and workflows across six task classes: property prediction, structure generation, machine-learning interatomic potentials, electronic-structure fields, spectrum elucidation and microscopy image processing. Task interfaces record physical quantities, units, aggregation conventions and references, making cross-task composition verifiable. A shared compiler backend enables switching from dynamic-graph to compiled execution through one setting, using graph optimization, operator fusion and hardware-targeted code generation without changing model definitions. Across 65 model instances in 11 groups, compiled execution accelerated every instance in eight groups, with group median speedups of 11%-224% and group maximum speedups of 11%-336%. Case studies on LiMnO2 and toluene link structure generation or spectrum elucidation, interatomic-potential relaxation, property prediction and electron-density prediction, using same-geometry density functional theory (DFT) calculations as references. For LiMnO2, predicted formation energies for generated and relaxed structures differed from same-geometry PBE+U references by at most 7 meV per atom, while the computational cost of model steps was less than one-twentieth of a DFT single-point calculation. PaddleMaterials extends unified materials workflows beyond conventional property, potential and structure-generation tasks while supporting verifiable cross-task operations and compiled execution.
Anqi Yu, Xudong Xu, Jianzhi Lu, Yuqi Sun, Bingguo Chen, Chenxi Ma, Weimin Tan, Bo Yan
Spatial multi-omics technologies jointly profile diverse molecular modalities with spatial context, providing a comprehensive view of cellular heterogeneity and tissue organization. To integrate spatial multi-omics data and identify spatial domains, a wide range of unsupervised methods has been proposed. However, recent approaches rely on single-step fusion, directly aggregating heterogeneous modalities into a shared representation, making embeddings sensitive to modality imbalance and measurement noise that are inherent to spatial multi-omics data, as well as unstable optimization under weak supervision. Here, we propose ILR-SMO (Iterative Latent Refinement for Spatial Multi-Omics), a unified self-supervised framework that reformulates multimodal integration as a stability-aware refinement process over a shared latent representation. Instead of single-step plain fusion, ILR-SMO progressively integrates modality-specific information through sequential updates, allowing information to be injected in a controlled manner. This incremental refinement enables later updates to correct or compensate for earlier deviations, resulting in more stable and robust representation learning under heterogeneous and noisy modalities. ILR-SMO further incorporates reliability-aware modality gating for adaptive modulation of modality contributions, and employs a joint objective to enforce spatial coherence while preventing representation collapse. Extensive experiments on five spatial multi-omics benchmarks demonstrate that ILR-SMO consistently outperforms seven state-of-the-art methods and exhibits strong robustness across diverse settings.
Jules Clerc, Mohamed Sallak, Domitile Lourdeaux, Marc Ravaine
Interactive narratives aim to let players influence the unfolding of a story while preserving a coherent and satisfying narrative progression. This creates a central tension, commonly described as the narrative paradox, between player autonomy and narrative control. Experience Managers (EMs) manage the story by determining how the narrative progresses in response to player behaviour. Although this paradox is widely discussed, it is rarely modelled in a way that supports systematic analysis, simulation, and comparison across management strategies. This paper proposes a formal framework for modelling and analysing the narrative paradox as a trade-off between interdependent properties shaped by EM decisions. We represent interactive narrative behaviour through an extended state-transition system in which world states are connected by player actions and system events. From this model, we formalize key properties involved in the paradox such as the player agency, EM controllability, author intent, and robustness, and define quantitative indicators derived from simulated plans. These indicators make it possible to compare EM strategies according to their ability to balance freedom, control, and narrative satisfaction. A case study compares different EM policies, showing how simulations can reveal correlations and conflicts between properties and identify compromise zones in the design space. This work contributes a reusable modelling and analysis framework for studying the narrative paradox and evaluating heterogeneous interactive narratives on shared formal grounds, with a focus on supporting the comparison and testing of EM strategies.
FL systems are inherently subject to client heterogeneity arising from differences in hardware capabilities. We propose a realistic evaluation framework for hardware-aware federated learning methods based on lightweight emulation of client hardware. Existing evaluation approaches address this either through small-scale real-device deployments or through trace-driven and probabilistic simulations. The former are difficult to scale and reproduce, while the latter suffer from three compounding sources of uncertainty: the choice of distribution family, the choice of execution-time estimates, and the inability to capture workload-dependent behavior. Our framework reproduces heterogeneous client behavior at execution time, capturing compute capacity, memory-constrained feasibility, and their interaction within a single host system. Unlike fixed-trace or throughput-scaling approaches, the proposed method is workload-aware: the same client population can exhibit different runtime and failure behavior depending on the task under study. We evaluate the fidelity of the approach across multiple workloads and show it preserves relative device performance while accurately reflecting hardware-dependent execution constraints. Compared against direct hardware measurements and benchmark references, the emulation accurately reproduces feasibility and training performance, preserving both relative device ordering and absolute compute times. We demonstrate the importance of workload-aware compute times by evaluating different heterogeneity-management methods across multiple workloads. By coupling emulation with real-world-based device sampling, our framework enables realistic, scalable, and reproducible evaluation of federated learning systems under heterogeneous learning conditions, while providing a practical way to generate workload-specific runtime behavior for large and diverse client populations.
Electroencephalography (EEG) provides a non-invasive means for neurological and mental health assessment, yet existing EEG models are typically developed for individual tasks and are difficult to unify across heterogeneous signal configurations and assessment objectives. Large language models (LLMs) offer a common interface for multiple tasks, but adapting heterogeneous EEG signals to a shared LLM raises a key challenge: how to preserve task-specific signal characteristics while enabling effective language-side parameter sharing. To address this problem, we propose EEG-HealthLLM, a unified and parameter-efficient framework for heterogeneous EEG health assessment. We introduce an Asymmetric Hierarchical Adaptation Strategy: task-specific EEG adapters preserve heterogeneous spatio-temporal dynamics, while a unified shared LoRA is injected into the LLM backbone (e.g., Phi-3) to capture shared high-level assessment semantics across tasks. We further establish EEG-HealthBench by reorganizing five existing EEG resources into paradigms covering epileptic seizure detection, major depressive disorder assessment, anxiety risk assessment, depressive symptom risk assessment, and emotional state assessment under a unified subject-independent protocol. Extensive experiments show that EEG-HealthLLM achieves an average balanced accuracy (BAC) of 69.58%, ranking first among the evaluated methods in both average BAC and macro-F1. Compared with independent single-task LoRA adaptation, it improves average BAC from 67.70% to 69.58% while reducing the aggregate language-side adapter footprint by 80.0%. Further analyses reveal task-dependent cross-task interactions and trade-offs, supporting the proposed hierarchical adaptation strategy. The implementation is publicly available at https://github.com/Marui-88/EEG-HealthLLM.
Sparse mixtures of experts combine multiple prediction models while evaluating only a small subset for each input. Although this architecture reduces computation, how expert separation and routing sparsity jointly govern expert recovery remains an open problem. In this paper, we address this question for location Gaussian mixtures of experts with a fixed number of experts by establishing how expert separation and spatial arrangement determine recovery rates. For dense models, our results reveal that when experts form well-separated clusters, the separation geometry is governed by the size of the largest cluster; without a prescribed cluster structure, this geometry is instead driven by the total number of experts. For sparse models, the separation geometry on retained routing regions is determined by the number of active experts, while the contribution of omitted regions is controlled by their total probability. Consequently, activating fewer experts reduces the separation complexity in these bounds. Furthermore, Top-1 routing yields a particularly striking result: experts can be recovered at a rate independent of the separation quantity. Our technical approach combines polynomial interpolation and Gaussian reproducing-kernel Hilbert space (RKHS) test functions with quantitative restriction inequalities that extend pointwise comparisons to expert functions.
Riemannian Lyapunov Framework: Optimization as Closed Loop Control on Riemannian Manifolds
Yixuan Wang, Omkar Sudhir Patil, Warren E. Dixon
AI 速览：提出黎曼李雅普诺夫框架（RLF），把一阶优化统一表述为黎曼参数流形上的闭环受控动力系统，并给出收敛性分析。
We introduce the Riemannian Lyapunov Framework (RLF), a unified framework that represents first order optimization as a closed loop controlled dynamical system on a Riemannian parameter manifold. Within this framework, optimizers are specified by a quadruple $(M, g, Φ, η)$ consisting of a Riemannian metric, an internal state transition map, a direction field generator, and a lifting parameter. Classical and modern methods including SGD, Adam, AdamW, Lion and Shampoo are recovered as specific instantiations of this quadruple. The central structural object of the framework is a time indexed family of target submanifolds in the extended state space of parameters, velocities, and internal memory, which we call the Normally Attracting Invariant Manifold (NAIM) family. These submanifolds organize training dynamics into two timescales: a fast contraction of the velocity state onto the current target submanifold, followed by a slow descent along it. Using Riemannian backstepping, we construct a strict Lyapunov function on the extended state space and prove Uniform Ultimate Boundedness (UUB) of the optimization trajectory under a descent alignment assumption and bounded stochastic disturbances. Under the additional Polyak-Lojasiewicz (PL) condition, the Lyapunov function converges linearly to a bounded neighborhood of the optimum, defined as the drift of the target direction field between consecutive iterates after parallel transport. To validate the framework's generativity, we derive three concrete instantiations of the quadruple, each exercising a distinct design axis, and show on geometric diagnostics, image classification benchmarks, and language modeling that their empirical behavior matches the theoretical predictions. The framework casts optimizer design as controller synthesis, complementing heuristic perspectives by providing a constructive template with provable guarantees.
Continuous-Action World Models: Learning Cross-Embodiment Spatiotemporal Dynamics Priors
Liangyu Li, Qingwen Liu
AI 速览：提出 CoAWM 框架，通过共享潜在函数空间学习跨具身、跨执行频率的机器人动作时空动力学先验，以支持动作生成。
We introduce Continuous-Action World Models (CoAWM), a framework for learning cross-embodiment spatiotemporal dynamics priors that support robot action generation across embodiments and execution frequencies. At its core is a shared latent function space that captures the spatial and temporal structure of robot actions. We learn this space through two stages. First, function pretraining brings trajectories from different embodiments and sampling frequencies into a common representation, learning shared dynamics priors through action reconstruction and consistency across temporal sampling views. By accommodating heterogeneous robot trajectories, this stage broadens the range of data sources available for pretraining and facilitates scaling to larger and more diverse datasets. Second, we connect the pretrained function space to a world-model backbone, enabling observations, language instructions, and the current robot state to condition the prediction of a latent action function. A single decoder takes any query time and returns the corresponding end-effector action. By separating shared dynamics priors from task-conditioned action selection and temporal sampling, our approach provides a unified foundation for transferring robot experience across embodiments and control frequencies.
View- and target-specific echocardiography AI models capture useful but complementary information and may not achieve sufficient predictive performance in isolation. A learned ensemble is a natural way to combine these outputs, but heterogeneous models typically require each output to be converted into a binary decision, probability, confidence score, or fixed-length feature vector. We developed a supervised large language model (LLM)-based late-fusion module that directly aggregates view-specific predicted finding statements and predicted measurements in language space and learns a case-level fusion function. In patient-level three-fold cross-validation of 859 examinations from 646 patients, the proposed method integrated heterogeneous outputs within a single framework, performed comparably to the best-performing reference strategy evaluated, and outperformed a general-purpose LLM without task-specific training. These results suggest that a supervised LLM can combine multiple upstream models without model-specific output conversion or feature engineering that requires medical knowledge.
Francisco Martín Rico, Esther Aguado González, Francisco Miguel Moreno Olivo, Rodrigo Pérez Rodríguez, José Miguel Guerrero Hernández, Francisco J. Rodríguez Lera
Planning at run time, as in robots that re-plan while they act, needs a planner that reads the PDDL its users write (durative actions, numeric resources, time windows, continuous change), answers quickly, and delivers only validated plans. Specialised planners are strong on individual fragments of PDDL, but they differ in coverage, latency and in what their failures mean. We study whether selecting heterogeneous engines from the semantics of each task can give broad coverage of PDDL, from STRIPS to PDDL+, while remaining competitive with specialised planners and making explicit what each answer guarantees. Kairos profiles each task to decide which of five engines apply (ground heuristic search, compressed and snap-action temporal search, time steps for hybrid dynamics, lifted search) and schedules them under one budget of time and memory. Every candidate plan is checked by a validator independent of the planner's transformations, and "unsolvable" is reported only when a search whose exhaustion is a proof exhausts its space. On 1,756 tasks from nine comparative suites, Kairos solves more tasks at the full limits than the strongest evaluated planner of each suite, with the largest differences on numeric, temporal, PDDL3 and PDDL+ planning, and it also solves the most tasks at 100 ms and 1 s. Resampling over domains, the gain is above zero in six of the nine suites, while on classical planning against LAMA it is within the noise. A further 450 feature-oriented tasks and 26 one-feature probes show that Kairos handles all seventeen features probed, where no other planner evaluated handles more than eight. Ablations identify compressed temporal search and the novelty fallback as major contributors, and a comparison with VAL traces every disagreement to a PDDL+ convention or to a documented VAL defect. Planner and benchmark are open source.
Pratik Jawahar, Riccardo Maggioni, Jonas Petersen, Matteo Cozzi, Gian-Alessandro Lombardi, Marek Gajewski, Camilla Mazzoleni, Federico Martelli
Test-time compute mechanisms, such as retries, confidence-based escalation and non-local routing, are conventionally deployed as black-box additions to trained models. They typically add forward passes or halting heads, but give no structural account of what the extra compute buys or what it costs against the model's stability guarantees. We therefore propose a theoretical and empirical framework for test-time compute on a Latent Dynamics Model used as the memory. When an encoder's latent state $(q,p)$ is advanced by a damped symplectic (velocity-Verlet) step of a learned Hamiltonian, every mechanism that expands access has an exact phase-volume Jacobian, a bounded energy change and an exact latch-transport law. We split reachability failures into escape failures and reach failures. A Lorentz squeeze resolves escape failures by injecting bounded, re-absorbable energy. A gated canonical translation (a one-step channel across $C_T$) resolves reach failures at $\detJ=1$ exactly. We verify this certificate stack on a designed analytic testbed. Reaching just beyond $C_T$ with the squeeze requires energy that grows quadratically in the excess distance. The gated channel's energy cost equals the potential difference between its endpoints, which is zero on the matched channel measured here, and so does not grow with distance in this testbed. An untrained state-replacing map annihilates phase volume and erases stored charge, while the gated channel transports it. On a learned physics-structured memory, we show calibrated compute rationing and evaluate the scaling limits of non-local routing against a physics-free baseline. The main contribution is the formalization and verification of the certificate stack, not benchmark performance.
Mahmoud Elhadidy, Roshan M. D'Souza, Amirhossein Arzani
Sparse-view computed tomography (CT) reduces the number of acquired projections, but conventional sampling strategies are designed primarily for image reconstruction rather than for downstream physical inference. Here, we investigate optimal task-aware projection selection for directly predicting hemodynamics from sparse CT angiography (CTA) measurements. Using computational fluid dynamics simulations of intracranial aneurysms, we generate synthetic fan-beam CTA projections and train a Fourier neural operator (FNO) to predict two targets: the spatially resolved time-averaged wall shear stress (TAWSS) field and the spatially averaged TAWSS. The FNO was coupled with optimal input selection for simultaneous optimal sampling and training. Our goal was to investigate if agentic AI could discover optimal sampling strategies that outperform current methods. We compared fixed uniform and golden-angle sampling with three task-aware optimal selection strategies: Gumbel-Softmax, a Soft top-K method, and an agentic AI selector. The methods were compared at different view selection budgets. A detailed analysis of the AI agent was performed to understand its reasoning and performance. For spatially resolved TAWSS, Agentic and Soft top-K achieve the lowest mean errors (Agentic performs slightly better) and substantially outperform fixed sampling under extreme sparsity. For spatially averaged TAWSS, optimized selection becomes important at extremely sparse acquisitions. Analysis of the selected masks reveals multiple distinct subsets concentrated around informative regions. Moreover, masks optimized for hemodynamic prediction yield poor image reconstructions, showing that task-optimal acquisition differs fundamentally from reconstruction-optimal acquisition. This work suggests the possibility of designing sparse acquisitions directly around downstream physical quantities.
Visual rule reasoning asks whether multimodal models can recover an abstract concept from a few examples. Existing studies either infer rules from visual examples but evaluate only classification or verbalization, or supervise executable generation with rules provided as input. Consequently, the relationship between visual rule understanding and reliable rule instantiation remains insufficiently explored. We present Bongard-Forge100k, a controlled benchmark, together with RuleForge, a backbone-agnostic framework that jointly learns visual rule understanding and rule-conditioned executable generation. The benchmark contains 108,797 problems: 86,000 for training, 19,297 for testing familiar rules, and 3,500 for testing held-out rule compositions. The framework progresses through rule-understanding training, joint understanding and scene-program supervision, and refinement via execution-grounded reinforcement learning with Understanding Replay. Executing generated programs under multiple random seeds provides verifiable feedback on executability, rule satisfaction, diversity, and robustness. With UniLIP as the primary backbone, RuleForge improves multiple-choice accuracy over the base model by 20.93 percentage points on familiar rules and 28.42 percentage points on held-out rule compositions, while increasing scene-level rule satisfaction on familiar rules from 2.91% to 70.99%. Additional experiments on InternVL3.5, Qwen3-VL, and Show-o2 demonstrate the effectiveness of the proposed training data and framework across diverse multimodal architectures.
Retrieval-Augmented Generation (RAG) has become the standard paradigm for grounding Large Language Model outputs in external knowledge. Lumer et al. presented the first systematic evaluation comparing vector-based agentic RAG against hierarchical node-based reasoning systems for financial document QA across 1,200 SEC filings, finding vector-based systems achieved a 68% win rate. Concurrently, the PageIndex framework reported 98.7% accuracy on FinanceBench through purely reasoning-based retrieval. This paper extends their work by: (i) implementing and evaluating three retrieval configurations: Vector RAG, Tree Reasoning, and the proposed Adaptive Hybrid Retrieval (AHR), on a small controlled corpus of financial, legal, and medical documents; (ii) introducing a four-tier query complexity benchmark; and (iii) employing LLM-as-judge evaluation with GPT-4o-mini. Experiments reveal that Tree Reasoning achieves the highest overall score (0.900), but no single paradigm dominates across all tiers: AHR scores highest on cross-reference queries (0.850), all three methods tie on multi-section queries (0.929), and Vector RAG scores highest on the single multi-document query (0.900). Vector RAG has the lowest mean latency (1.61 s against 3.40 s for Tree Reasoning). On the first 50 publicly released FinanceBench questions, using the evidence snippets supplied with the benchmark as the retrieval corpus, Tree Reasoning scores 0.938, Hybrid AHR 0.901, and Vector RAG 0.821. Because the configurations also differ in document access, generation prompts and output length limits, these differences cannot be attributed to retrieval architecture alone. In these preliminary results the adaptive router does not outperform fixed tree reasoning overall (0.845 vs. 0.900); they motivate a larger evaluation of when query-based routing helps, rather than establishing its benefit.