World-action models jointly predict future observations and actions, yet reconstruction-oriented visual representations do not explicitly emphasize the spatial relationships and object-state changes needed for manipulation. We introduce Phoenix, which learns representations of scene structure for world-action modeling. Specifically, Phoenix Encoder enriches DINOv3's general visual semantics through dual-teacher distillation of geometric and spatiotemporal semantic knowledge. Next, CS-VAE causally aggregates and compresses Phoenix Encoder features into generative latents while preserving scene information and spatial layout. Finally, Phoenix jointly predicts these latents, complementary appearance latents, and robot actions, linking scene evolution with action generation. To study pretraining data efficiency, we curate approximately 39,000 hours of egocentric and robot manipulation data for representation learning and use the robot demonstration subset for WAM pretraining. Following benchmark-specific post-training, Phoenix achieves state-of-the-art performance across simulation and real-robot benchmarks spanning single-arm, bimanual, and dexterous manipulation. It demonstrates strong robustness to perturbations, achieving success rates of 94.8% on LIBERO-Plus and 72.4% on RoboTwin 2.0 C2R, alongside strong cross-embodiment transfer and real-world manipulation performance on Bench2Dex and RoboDojo-RealWorld, respectively. Pretraining-scale analysis shows larger performance gains from data scaling than a video-VAE-only baseline, indicating stronger scaling benefits and more effective data utilization. Moreover, Phoenix maintains high success rates when adapted with limited downstream demonstrations, demonstrating strong few-shot adaptation. Our code, pretraining data, and models are available on the project page to help the community reproduce our results.
KV cache reuse accelerates LLM serving, but existing systems manage host DRAM and NVMe SSD as a tiered hierarchy where DRAM acts as an intermediate cache. Production LLM workloads break both premises of this hierarchy. First, bimodal reuse defeats DRAM temporal filtering, since workloads are dominated either by one-shot blocks with long reuse distances or by multi-gigabyte agentic sessions that exceed DRAM capacity, putting SSD on the critical read path. Second, DRAM staging and eviction-ordered layout bottleneck SSD restoration. Holding host staging buffers for in-flight prefixes causes contention under load that aborts prefetches, while incremental eviction scatters a session's blocks across the SSD and prevents large sequential reads. We present FlatCake, a flat KV storage architecture that exposes host DRAM and NVMe SSD as peer media in a unified 64-bit address space. Exploiting the fact that KV blocks are write-once, immutable, and reused together within a prompt or session, FlatCake places each block directly onto its target medium in session-contiguous, layer-major order at creation time and never migrates it. FlatCake streams coalesced SSD reads directly into a bounded GPU window via GPUDirect Storage and decouples DRAM and SSD request queues to prevent head-of-line blocking. We implement FlatCake in SGLang and compare it against host-staged (Mooncake) and GDS-enabled (Mooncake-GDS) tiered storage, as well as SSD-direct storage (Tutti). Across three production agentic, coding, and conversational traces with DeepSeek-V4-Flash, FlatCake reduces mean TTFT near the baselines' latency knees by $1.40\times$--$2.42\times$ over the fastest baseline. On the agentic Claude Code trace, FlatCake sustains a 91.6% cache hit rate under load (versus 71.1% for all three baselines) and delivers up to $2.16\times$ higher token throughput.
Controllable real-world video generation models can condition on a rendered graybox proxy video, which fixes layout, motion and camera. Coding agents can build such videos by generating a 4D scene in Blender, yet no benchmark checks both these rendered proxy videos and the complex, multi-shot scenes behind them. We introduce VibePrevisBench, a benchmark for 4D scene generation: an agent turns a brief into an executable 4D scene and renders its proxy video. Its 242 tasks span four input kinds, including scripts, photos, continuations, and reference videos. Each delivery is scored along two evaluation dimensions. Rule-based Scene Check re-runs the code and compares the exported 4D scene with hidden ground truth facts in nine dimensions. VLM-based Screen Check asks a VLM judge whether facts that pass the rules are recognizable in the proxy video. We analyze the evaluation results and trajectories of 11 frontier coding agents and identify three key observations. First, multimodal understanding separates the agents, especially in understanding space and camera motion. Second, all agents seem to already master basic camera work and staging, and fail at challenging requests. Third, agents inspect their own renders but fail to correct their errors. Agents thus need better visual understanding of space and camera, explicit checks for challenging requests, and item-by-item verification of renders against the brief.
Grana: Compositional Reward Modeling for Long-Horizon Robot Manipulation
Zi-ang Cao, Kaiyuan Zheng, Yinzhen Xu, Haonan Duan, Jiasheng Gu, Fengyuan Hu, Nan Zhu, Kun Lei 等 17 位作者
AI 速读 基于摘要
Grana 是一种视觉-语言机器人奖励模型,通过开放词汇子任务描述识别未完成子任务并估计其连续进度,再组合子任务进度与已完成子任务信用来估计长时程任务的整体进度。
cs.RO
查看原文摘要
Robot reward modeling faces two challenges: generalizing across tasks and accurately tracking progress in long-horizon manipulation. General-purpose reward models demonstrate broad task transfer, but transfer from short-task pretraining to long-horizon tasks remains underexplored. Long-horizon progress models often rely on task-specific subtask classifiers or fixed action vocabularies, limiting the reuse of learned progress knowledge when adapting to new tasks. To address these challenges, we introduce Grana, a vision-language robot reward model that learns to identify unfinished subtasks and estimate their continuous progress using open-vocabulary subtask descriptions. We train a language-conditioned model to estimate progress in short-horizon tasks and individual subtasks of long-horizon tasks. For long-horizon tasks, we combine the selected subtask's predicted progress with credit for preceding completed subtasks to estimate whole-task progress. General progress prediction requires training the model across diverse tasks with progress supervision that accounts for imperfect execution. Elapsed-time proxies can misrepresent progress when relative subtask durations vary, or failures reverse local progress. We therefore introduce an annotation protocol that converts sparse subtask boundaries and action-quality annotations into dense progress targets, without requiring annotators to assign numerical progress scores. Pretraining with this supervision on 200 diverse, mostly short-horizon manipulation tasks enables zero-shot reward prediction on unseen tasks and improves progress prediction after adaptation to multi-minute, long-horizon tasks. Experiments demonstrate (i) generalizable reward prediction, including adaptation to five long-horizon tasks, (ii) benefits from pretraining and composing subtask progress estimates, and (iii) improved downstream policy performance.
Anticipating or Speculating? Making Futures Pay in AI Ethics
John P. Nelson
AI 速读 基于摘要
作者主张区分“预见”与“臆测”,并据此批评为医生与医疗AI判断分歧制定统一决策规则的做法。
cs.CY · cs.AI
查看原文摘要
Philosophers have recently been trying to develop uniform decision rules for instances in which physicians' judgments diverge from the recommendations of medical artificial intelligence (AI) systems. I argue that this effort is fatally flawed because the performance and appropriate use of algorithmic recommendation systems by physicians are highly context-specific, empirical medical matters rather than epistemological ones. In such instances, general decision rules can be stipulated only to the extent that they abstract away from relevant, indeed determinative, contextual circumstances. The alleged problem of physician-AI disagreement, and others like it, play into longstanding disputes in technology ethics about the viability and utility of envisioning possible futures: is this useful "anticipation" or wasteful "speculation"? I suggest that the two may be distinguished. Anticipation explores plausible futures to help inform present actions under irreducible uncertainty. Speculation abstracts away uncertainty, developing formal solutions to stipulated circumstances whose relationship to the real present is often quite unclear. Efforts to develop crisp philosophical solutions to physician-AI disagreement, among other problems, fall into the latter category. Technology ethics should remember that fruitful futures thinking must both begin from and return to the messy, uncertain present.
A language is P-selective if a polynomial-time algorithm, given two strings, picks one that belongs to the language whenever either does. Ko showed in 1983 that every P-selective language is recognized in polynomial time with quadratic advice; his argument uses $n^2+n$ bits, and for polynomial-time recognition the leading constant has stood since. We show that the right constant is $1/4$. In the selector-query model the answer is unconditional. A single computable decoder recognizes every legal cut on $n$-bit strings with $n^2/4+O(n)$ advice bits and $n$ nonadaptive comparisons. Conversely, every randomized decoder making polynomially many queries needs $n^2/4-O(n\log n)$ bits at error $1/3$, even when its queries are adaptive evaluations of individual-degree-two extensions of the selector table. The degree matters: with multilinear extensions, $n+1$ bits suffice and are optimal. Under the hypothesis that $\mathrm E$ requires exponential-size circuits with SAT gates, every P-selective language has a polynomial-time decoder with $n^2/4+O(n)$ advice bits. From this hypothesis no relativizing argument, and no argument that algebrizes with degree-two extensions, can lower the constant: one oracle satisfies the hypothesis and carries a selective language that needs $n^2/4-o(n^2)$ bits against randomized decoders with degree-two access. The cost is set by the two strong components of the selector tournament adjacent to the cut. If the smaller has $2^v$ vertices, the cost is $u(n-u)\pm O(n\log n)$ with $u=\min\{v,n/2\}$. When $\log n\ll v=o(n)$ this is $(1+o(1))nv$, attained in polynomial time with no hypothesis. The upper bounds rest on a fixed-point distribution on the vertices of a tournament. Its power moment shows that $2^{r^2-O(r)}$ different $r$-tuples cover any side, for a suitable $r$.
Diffusion probabilistic models have become a cornerstone of modern generative AI, yet the mechanisms underlying their generalization remain poorly understood. In fact, if these models were perfectly minimizing their training loss, they would just generate data belonging to their training set, i.e., memorize, as empirically found in the overparameterized regime. We revisit this view by showing that, in highly overparameterized diffusion models, generalization in natural data domains is progressively achieved during training before the onset of memorization. Our results, ranging from image to language diffusion models, systematically support the empirical law that memorization time is proportional to the dataset size, consistent with a kernel-regression bound on the time required to fit the empirical score at low noise. Generalization vs. memorization is then best understood as a competition between time scales. We show that this phenomenology is recovered in diffusion models learning a simple probabilistic context-free grammar with random rules, where generalization corresponds to the hierarchical acquisition of deeper grammar rules as training time grows, and the generalization cost of early stopping can be characterized. We summarize these results in a phase diagram. Overall, our results support that a principled early-stopping criterion - scaling with dataset size - can effectively optimize generalization while avoiding memorization, with direct implications for hyperparameter transfer and privacy-sensitive applications.
Nowadays refinery optimization utilizes sheer amounts of data, which can be handled with modern Linear Programming (LP) software, but the interpreting and applying the results remains challenging. Large petrochemical companies use massive models, with hundreds of thousands of input matrix elements. The LP solution is mathematically correct, but simplifications are made in the model, and data supply errors may occur. Therefore, further insight is needed to trust the results. The novelty of the paper consists of enhancing the LP solver with a memory, which makes possible to find potential errors in the input data, finding unusual relations in the data on different decision levels, by introducing an anomaly detection tool. Analyzing the potential "anomalies" in the input and output of LP, the subsequent LP solutions can be improved. The Anomaly Detection tool proposed comprises three methodologies. A transformed version of the popular ECOD methodology is applied. New methods are proposed to handle high-dimensional data: choosing the most informative pairs. Then, this is used alongside two 2D Anomaly Detection algorithms, revealing several business opportunities and data supply errors in the MOL refinery scheduling and planning architecture. The proposed tool learns connections in past data and runs very efficiently on present data revealing unusual connections sorted by their importance. Up to this step, optimization and Anomaly Detection happened automatically, and thereafter, decision makers may take a look at the most relevant anomalies found, and decide how to further improve the subsequent LPs. The tool also has periodic retraining built in.
When older adults trust scammers despite warnings, families face obstacles to intervention. Complainants aged 60 and older reported \$4.9 billion in losses in 2024 to the U.S. Internet Crime Complaint Center. To investigate family intervention barriers to elder scams, we collected 2M Reddit posts from May 2021 to June 2026 and 2B Baidu Tieba posts from 2007 to 2026. The final corpora contain 13,082 Reddit posts and 3,933 Baidu Tieba posts, labeled using platform-specific codebooks, manual reference annotations, and LLM coding. Through a mixed-methods analysis, we find that scams often extend beyond a single incident and are reported alongside debt problems and broken family relationships. In 68.0\% of Reddit posts recording reactions, older adults showed resistance or continued involvement; in Baidu Tieba family-confrontation posts, 41.9\% showed continued belief. Relatives identified fraud in 74.4\% of Reddit posts and narrated cases on Baidu Tieba, but without the victim's consent, relatives may lack the legal authority to intervene. In Reddit posts recording reactions, account locks or credential changes appeared in 8.9\% under resistance or continued involvement, compared with 33.3\% under acceptance. Measures such as phone or network monitoring and financial takeover were reported more often under resistance. Recovering funds is very limited in posts recording follow-up results of fraud; Baidu Tieba users reported that police classified older adults' voluntary transfers as civil disputes.
Equivariant Modeling with Polynomial Features via Minimal Homogeneous Generators
Michael Liaofan Liu, Joe Kileel
AI 速读 基于摘要
提出一个基于极小齐次生成元构造等变多项式特征的通用框架,并配套自动计算与拟合的Python包。
cs.LG · stat.ML
查看原文摘要
We present a general framework for modeling and learning equivariant functions between finite dimensional semisimple real group representations, unifying constructions developed across many application areas. Our model is a linear combination of equivariant polynomial features whose coefficients are arbitrary learnable functions of invariant polynomial features. We take as features minimal homogeneous generators for the invariant algebra and the covariant module over the invariant algebra, up to prescribed degrees. We provide a Python package that automatically computes these features and fits the resulting models to training data. The algorithm for feature computation requires explicit isotypic decompositions of the input and output representations, along with explicit Clebsch-Gordan decompositions of the tensor product of irreducible representations, and achieves speedups over alternative methods in computational algebra. Numerical experiments in scientific modeling demonstrate the full learning pipeline and yield accurate predictors with few trainable parameters, using the model both for end-to-end prediction and within layers of larger architectures.
We study continuous-review inventory models with unit Poisson demand and deterministic lead times: single sourcing with lost sales and dual sourcing with backlogging. Xin [Operations Research. 70(1):111--128 (2022)] establishes an asymptotic approximation factor of 1.79 for each model in its proportional-growth regime. Under the same assumptions, we obtain a factor of 1.248977 for the better of base-stock and constant-order policies for lost sales and a factor of 1.599528 for the better of dual-index and tailored base-surge policies for dual sourcing. The benchmark is the optimal long-run average cost over all policies using only current and past information. For lost sales, lead time and the lost-sales penalty grow proportionally. For dual sourcing, the lead-time difference grows proportionally to the express premium, the extra unit cost of faster supply, while express lead time and backorder penalty remain fixed. For sufficiently large fixed backorder penalties, the dual-sourcing factor improves to 1.248977. The improvements follow from stronger lower bounds on the optimal cost. We retain intermediate constraints on mean inventory for lost sales and use events determined only by future demand to bound backlog costs for dual sourcing.
Long-running LLM agents require memory that persists and evolves across sessions. Text-based memory retrieves and reconstructs past interactions at every query, making long-horizon performance increasingly dependent on retrieval quality and contextual reasoning as histories grow. Parametric memory encodes experience directly into model computation, but existing approaches provide limited support for cross-session memory evolution. Their coupling to a specific backbone further restricts memory reuse after model replacement. We introduce RPMem, a two-stage architecture that compiles each session into a model-independent latent memory through forward computation and selectively integrates it with retained memory via a task-trained recurrent gate. The consolidated memory is then mapped to backbone-specific low-rank adaptation (LoRA) parameters, allowing the encoding capability to transfer when the backbone is replaced. Evaluation across three long-term memory benchmarks and five diverse backbones demonstrates broad generalization with near-constant update cost and memory footprint. With Qwen3-8B on PERMA, RPMem reaches 85.52%, outperforming the strongest parametric and text-based baselines by 5.32 and 12.98 percentage points, respectively. Ablations validate the complementary roles of session compilation and cross-session consolidation, while dynamics analyses reveal that the gate acquires task-specific memory integration strategies. These results establish RPMem as a lifecycle-independent parametric memory framework that maintains evolving cross-session memory that remains reusable across backbone replacements. Our code is available at https://github.com/Quark-Medical/rpmem/tree/main.
Large-scale, multi-agent self-play trains robust driving policies effectively in simulation, but controlling these models to generate specific, safety-critical edge cases remains a key challenge. Existing goal-conditioned self-play frameworks rely primarily on static map-location goals and fail to reliably produce complex agent interactions such as aggressive tailgating or adversarial cut-offs. These agent interactions are necessary for rigorous autonomous vehicle (AV) testing. To address this, we present an extended multi-agent self-play framework that introduces agent-relative dynamic goals. By conditioning policies on this extended goal representation, our framework enables targeted, controllable scenario synthesis. We demonstrate that our approach successfully generates interactive driving behaviors that static-goal baselines cannot simulate, providing a flexible tool for long-tail scenario generation in AV testing.
Neural PDE operators are often optimized through state reconstruction, which largely constrains operator values at observed inputs but does not fully characterize their sensitivity to local perturbations. A more faithful approximation of operator behavior requires not only accurate state predictions, but also correct variations along local perturbation directions. We present TangentPDE, which uses finite differences to approximate directional derivatives and constructs task-relevant finite-step responses from the state outputs of a reference solver at sparse paired inputs. Its objective jointly constrains state fidelity, response-direction alignment, and response-magnitude calibration, thereby learning both operator values and their local variations. Because the response labels are constructed solely from state outputs at paired inputs, the method requires neither explicit derivative labels, linearized or adjoint PDE solves, nor full Jacobians, and remains applicable to nonlinear solvers without derivative interfaces. Building on this, we additionally propose a selectively executed adaptation mechanism to strengthen local operator representations under finite-difference supervision. Budget-matched and correspondence-perturbed controls isolate the effect of relational supervision from improvements attributable to additional state observations or increased data volume. Designed as a plug-and-play training framework, TangentPDE integrates with diverse neural-operator backbones. Across four PDE systems and 20 task-backbone settings, TangentPDE lowers the three-seed mean state error in 18 settings and the corresponding Rollout@20 error in 18, while intrinsic finite-step response gains are strongest on DiffReact and remain task- and backbone-dependent elsewhere.
AMR 先训练一组能力专属的 LoRA 专家,再在验证集上无梯度搜索低损失组合原型,并蒸馏出轻量路由器,为每道题选择实例级技能组合。
cs.LG
查看原文摘要
Most existing approaches enhance mathematical reasoning in large language models (LLMs) as a monolithic capability, despite the fact that different mathematical problems require different reasoning skills. Recent work has explored a more fine-grained perspective by decomposing mathematical reasoning into atomic skills and developing specialized modules for individual capabilities. However, how to effectively utilize these atomic abilities remains underexplored. To address this challenge, we propose Atomic MathReasoner (AMR), an instance-level framework for composing atomic reasoning skills. AMR first trains a library of capability-specific LoRA experts and performs gradient-free optimization over held-out validation subsets to discover a diverse set of low-loss composition prototypes. It evaluates the preference of each prototype using the token-level negative log-likelihood of reference solutions, filters ambiguous assignments through a confidence margin, and distills the resulting pseudo-labels into a lightweight router. During inference, the router selects an appropriate composition prototype for each problem, enabling problem-specific utilization of atomic skills without costly online optimization. Experiments on the AtomicMath benchmark across general and mathematics-specialized LLMs demonstrate consistent improvements. On Qwen2.5-Math-Instruct-7B, AMR increases average field accuracy from 64.4% to 68.1% at Level 1 and from 59.3% to 63.7% at Level 2, while improving average logical reasoning performance from 39.4% to 46.8%. These results demonstrate that effectively composing specialized atomic skills provides a promising approach to strengthening mathematical reasoning in LLMs.
Language models acquire their capabilities from training data, and developers withhold data during training to keep models from learning capabilities that could be misused. How long this delays a capability, and how much data is needed to acquire it later, is unknown. Here we show that the general progress a model makes on the rest of its training data substitutes for the data that was withheld. We withheld the training signal for individual skills in models of chess, Othello, molecular strings, bracket languages and web text. The delay of a withheld skill in larger models was predicted from smaller ones. A chess model never explicitly trained to respond to check ranked the legal responses almost as well as a model that was. When models were then trained on the withheld data, they acquired the skill with a small fraction of the data needed from initialization, and this fraction decreased as the models progressed. A control rule of equal difficulty was learned much more slowly. A programming language withheld from the training text of a 7-billion-parameter model was regained in the same way. Data filtering therefore delays a capability, and the data required to regain it decreases as models improve.
Large Language Models (LLMs) have shown remarkable capabilities as autonomous agents, yet existing benchmarks either focus on single-agent tasks or are confined to narrow domains, failing to capture the dynamics of multi-agent coordination and competition. In this paper, we introduce MultiAgentBench, a comprehensive benchmark designed to evaluate LLM-based multi-agent systems across diverse, interactive scenarios. Our framework measures not only task completion but also the quality of collaboration and competition using novel, milestone-based key performance indicators. Moreover, we evaluate various coordination protocols (including star, chain, tree, and graph topologies) and innovative strategies such as group discussion and cognitive planning. Notably, gpt-4o-mini reaches the average highest task score, graph structure performs the best among coordination protocols in the research scenario, and cognitive planning improves milestone achievement rates by 3%. Code and datasets are public available at https://github.com/MultiagentBench/MARBLE.
Hybrid attention models combine full and linear attention to achieve a better quality--efficiency trade-off. Converting pretrained Transformers into such models requires identifying which layers should retain full attention under a prescribed budget. We identify two off-policy mismatches in FlashMorph's layer selection: (i) continuous branch interpolation and teacher-forced hidden states differ from discrete hybrid computation; and (ii) frozen model weights fail to account for parameter adaptation during subsequent recovery. We introduce FlashMorph2, an on-policy layer-selection method that (I) evaluates sampled discrete hybrid architectures end to end using their own hidden states, and (II) jointly optimizes layer-selection scores and model weights under the recovery objective to reflect post-adaptation performance. Experiments across Qwen3 model scales and linear-attention variants show that FlashMorph2 consistently outperforms existing layer-selection methods, improving average scores over FlashMorph by $+2.42$ points on RULER, $+1.08$ on LongBench, and $+0.76$ on recall-intensive tasks, while completing layer selection in only $0.72$--$3.57$ GPU hours across 0.6B--8B models, demonstrating high efficiency and scalability.
Large-scale applications of lifelong multi-agent path finding (LMAPF) call for computationally inexpensive movement decisions under changing traffic conditions and across varied layouts. We propose a navigation policy in which a convolutional network computes a shared spatial field over the map once per timestep and a compact goal-conditioned operator derives agents' movement preferences. The complete policy is trained end-to-end using average-reward reinforcement learning with throughput as the reward, without planner demonstrations or a separate online path planner. For structured and unstructured layouts, we train one policy per class on pools of small maps and deploy each without retraining on larger, unseen layouts and agent populations. Across six layouts with 10,000 agents, our policies achieve throughput comparable to or higher than the strongest evaluated baselines. The proposed method has an average execution time below 10 ms per step, more than an order of magnitude faster than several established planners.
Global planners and path-tracking controllers are often evaluated independently, although their interaction determines how a vehicle follows a planned path. Whether simulation accurately predicts the performance of specific planner-controller combinations on Ackermann-steered hardware remains uncertain. This study evaluated three geometric controllers (Regulated Pure Pursuit, Stanley, and Vector Pursuit) paired with two global planners (Hybrid A* and NavFn) on a 1/10-scale Quanser QCar and in a matched Gazebo simulation. Both environments used shared custom planning and control algorithms, matched route geometry and goals, ground-truth pose references, and a common evaluation pipeline. Five runs per configuration and environment yielded 60 benchmark runs. Controller rankings for tracking accuracy and command smoothness transferred consistently between environments, with Spearman correlations of 0.77-0.94. However, planner-dependent effects showed substantially weaker agreement. In simulation, planner selection changed the root-mean-square (RMS) angular acceleration of the commanded velocities by no more than 0.030 rad/s^2, whereas Hybrid A* increased it by 0.141-0.843 rad/s^2 on hardware, depending on the controller. These differences coincided with planner-dependent achieved speeds that the simulation's idealised longitudinal response did not reproduce. Mean controller computation times remained below 0.5 ms on the Jetson TX2. The findings demonstrate that agreement in controller rankings does not guarantee accurate prediction of planner-dependent performance for specific controller pairings on hardware, highlighting the importance of physical validation.
Modeling systemic failure in high-dimensional environments is fundamentally limited by supervised algorithms that attempt to predict sparse, noisy or censored endpoints. To break this reliance on explicit outcome regression, we engineered the Self-Organizing Survival Manifolds (SOSM) framework, a deep learning architecture that decouples unsupervised representation learning from survival labels and treats time-to-event progression as an emergent geometric property. The system projects heterogeneous multi-omics arrays into a unified latent space governed by a pseudotime-weighted, curvature-minimizing graph Laplacian, organizing patient trajectories along low-dissipation geodesic flows. To validate the framework, we applied this architecture to a 184-patient clear cell renal cell carcinoma multi-omics cohort. Across a deterministic 5-fold cross-validation pipeline, SOSM achieved a mean Concordance Index of $0.768 \pm 0.068$, while out-of-fold risk stratification yielded significant survival divergence (log-rank $p < 0.001$) accompanied by a 14.7% increase in local manifold dissipation among high-risk patients. Furthermore, interrogating the network's encoder weights confirmed that the model autonomously recovered canonical metabolic and microenvironmental drivers of disease progression without supervised feature selection. By demonstrating that systemic failure can be mapped as a geometric phase transition, SOSM establishes a generalizable computational engine for prognostic modeling across biological, industrial and financial complex systems.
A central challenge for multimodal language models (MLLMs) in real-world tasks is to abstract visual observations into structured spatial states and reason over how those states change under actions over long horizons. Existing multimodal benchmarks often focus on static question answering, where answers can be directly extracted from images, or assume symbolic states are available and bypass the visual-to-spatial abstraction process. We present MARBLE, a diagnostic benchmark for spatially grounded multimodal reasoning and planning across three agent environments: CUBE, MAZE, and BLOX. MARBLE spans 2D and 3D spatial puzzles under geometric constraints, dynamic layouts, and spatial transformations, with difficulty systematically controlled along two axes: observation format and planning horizon. Extensive evaluations show that frontier MLLMs struggle on MARBLE, with the strongest model GPT-5.4 achieving only a 7.3% average success rate across the three environments in the hardest setting. Pairwise ablations on the open-weight models identify visual-to-spatial abstraction as the major bottleneck: when models must infer spatial states from images instead of receiving human-annotated states, success drops from 50.7% to 3.7% for Gemma-4-31B and from 87.9% to 6.5% for Kimi-K2.6. The results indicate that models often fail to construct precise spatial states from visual observations before planning, a task that is straightforward for humans. Overall, MARBLE reveals the gap in reliable visual-to-spatial abstraction for frontier MLLMs and provides a controlled testbed for measuring progress in spatially grounded multimodal reasoning and planning.
Neural networks with randomly initialized, fixed weights and trainable biases are universal approximators, yet the dynamics of bias-only training remain largely unexplored. We study these dynamics in one-hidden-layer ReLU networks through activation codes, the binary patterns of active neurons that partition the input space into convex polyhedral regions. We show that the gradients with respect to the biases depend on each input only through its activation code, and hence the training dynamics can be characterized with these codes. In particular, as long as the activation codes remain fixed, training reduces to a linear least-squares problem with the codes as features. Training can then be understood as a sequence of least-squares problems, with transitions occurring when neurons change their activation states. We show that the fraction of switching neurons vanishes as the network widens, and the training dynamics approach those of linear least squares on the initial activation codes in the infinite-width limit. Our experiments support these theoretical findings, showing that while activation switches can noticeably affect training in narrower networks, their relative frequency and the associated changes in the dynamics diminish as the width increases.
Prior work on engagement modeling in children with autism spectrum disorder (ASD) has demonstrated substantial individual variability in behavioral features, particularly in visual modalities. In this work, we extend the literature to acoustic features in naturalistic interactions, where multiple speakers and atypical interaction patterns complicate speaker-specific analysis. We investigated whether speaker diarization improved acoustic engagement modeling during month-long, in-home socially assistive robot (SAR) interventions with seven children with ASD. We trained classical machine learning models using acoustic features extracted from diarized child and parent/caregiver audio and compared their performance with models trained on undiarized audio. Diarized audio consistently yielded higher area under the receiver operating characteristic curve (AUROC) values, although these differences were not statistically significant. More importantly, the relative contributions of child and caregiver acoustic features varied across participants, indicating that the acoustic information associated with engagement was not consistent across children. These findings extend evidence of participant-level variability in ASD engagement modeling to the acoustic domain and demonstrate the feasibility of isolating speaker-specific vocalizations for studying engagement in complex, naturalistic interactions.
Self-supervised encoders for animal behavior are often trained on, or fed, hand-built movement features. We ask whether they beat those ingredients when compared fairly. A training-free representation, made of generic per-frame measurements summarized over fixed time windows and reduced with PCA, is compared with retrained and released encoders (BAMS, hBehaveMAE) under a preregistered protocol: frozen representations of equal size and one shared classifier. On CalMS21 it reaches 85.8 macro F1 against 82.4 for the best encoder, ties with four labelled videos, and leads when the annotator changes or new behaviors must be learned from a few videos; the reading holds for encoders retrained with three seeds. On the MABe22 mouse triplets it scores 83.7 against 80.3 for the released hBehaveMAE (gap +3.4, 95% interval +2.1 to +4.7). On a second species, the MABe22 fly groups, the same recipe scores 51.6 mean F1 over 50 tasks against 48.2 for BAMS, the best published method. On Shot7M2, the synthetic basketball benchmark built by hBehaveMAE's authors, it scores 68.8 against 64.7 for hBehaveMAE, the best method in that paper, though hBehaveMAE remains better on the shortest movements. Short time windows carry most of the result. Raw keypoints in each body's own frame, summarized over the same windows, already match or beat the encoders on both mouse datasets and on Shot7M2; on the fly groups they fall short, and the measurements add 5.3 points. On these benchmarks, current behavior encoders have yet to beat the hand-built features they learn from, and such baselines should be reported alongside them.
World action models (WAMs) have shown promising generalization and data efficiency in robot manipulation by coupling action generation with predictions of future visual dynamics. Translating these capabilities into reliable task execution, however, remains a challenge for existing post-training methods. Offline supervised fine-tuning provides dense supervision on expert trajectories, but its flow-matching action expert is trained only on noised demonstration actions and never on the intermediate states produced by its own sampler; task-specific adaptation can also compromise pretrained capabilities. On-policy reinforcement learning trains on the policy's own behavior, yet sparse task rewards make exploration and credit assignment difficult, while dense reward shaping still provides scalar evaluations rather than direct action targets. We introduce \textbf{HINT-WAM}, the first on-policy self-distillation framework that supervises the action expert on its own sampling trajectories using hindsight, the future video of the demonstration, as privileged information available to the WAM itself. For each demonstration observation, the student generates its action through its own denoising trajectory, and at the states along this trajectory the same model acts as a teacher conditioned additionally on this hindsight, providing a conditional action distribution as a dense training target, while the student sees only deployment-time inputs and needs no environmental rewards or simulator interaction. To transfer this supervision to a flow-matching action expert, HINT-WAM distills the teacher into the student with an energy score, a distribution-level objective that matches the teacher's action distribution from the student's own samples rather than regressing its velocity field. The project web is available in https://tuandattt.github.io/hintwam/.
People experiencing high-acuity mental health concerns are turning to AI chatbots for emotional support, but little is known about what drives this use, their experiences, and their preferences. In a sequential mixed-methods study, we inductively analyzed Reddit posts and used the findings to develop a survey. We surveyed 202 adults with recent mental health concerns who had used AI for emotional support. Most participants were already receiving clinical care and reported turning to AI, most commonly because they did not want to burden others, rather than being driven by barriers to care. Benefits were broadly shared, but participants reporting past-year suicidal ideation described more harm-facilitating AI behavior and reliance than those without. Overall, participants ranked honesty above warmth, and preference for crisis escalation declined as clinical acuity rose. These findings motivate AI support that addresses perceived costs of disclosure, evaluates benefits alongside harms, and widens feasible paths to human connection.
This paper proposes a measurement standardization framework that compresses expert-AI interactions into structured, comparable fields for prospective risk detection in deployed AI systems, without access to model internals. This concept paper defines the framework's scope, semantically and statistically, and specifies a protocol for its empirical testing. The population-level claims it is designed to support therefore belong to a staged research program rather than to results claimed here. Measurement standardization underpins three claims. The first is a reliability claim: under bounded conditions, large language models can produce reliable, standardized assessments of the evidential and policy alignment of expert-AI interactions. The second is a governance claim: alignment scores give experts an immediate signal during deployment and give institutions a basis for monitoring alignment patterns across mission types, models, and domains. The third is an outcome validation claim: once measurement standardization is established, aggregate alignment scores could be used to study associations with downstream outcomes in regulated professional settings. This introduces the possibility of an "AI epidemiology", a form of risk detection based on correlated variables instead of mechanistic analysis, inspired by epidemiological reasoning. A minimal application of the protocol to a published expert-AI corpus shows that the judge reproduces its policy and evidential alignment scores across two runs under the specified conditions. Judge reliability at scale remains to be validated in future work. The paper sets out a defined grammar of eight interaction fields, together with a statistical protocol based on paired bootstrap inference, DeLong's test for paired AUCs as a sensitivity check, a pre-specified one-sided non-inferiority margin of 0.05, and Holm-Bonferroni correction.
Long-context decoding, whether over long reasoning traces, long videos, or robot episodes, is increasingly limited by the memory of the key-value (KV) cache. Page-retrieval sparse attention makes each decoding step cheaper by attending only to the top-scoring KV pages. However, it still stores every page exactly, and adds a small digest per page for scoring, so the cache ends up slightly \emph{larger} than the full cache. We propose OVAL, which stores each page as a pair of output-aware low-rank bases that act as both the retrieval index and the storage. A key basis scores the page and rebuilds its keys when the page is selected, and a PCA basis rebuilds its values. The exact pages can then be dropped, and nothing extra is added to the cache. The key idea is that the key basis should be output-aware. A low-rank error in the keys perturbs the attention output at first order, but perturbs the page mass that retrieval ranks only at second order in expectation. OVAL therefore builds the key basis from a mix of key geometry (key PCA) and the attention-output error, with a per-head mixing weight set in closed form. At the same attended budget, OVAL is best or tied-best in eight of nine long-reasoning settings. It matches FullKV on MATH500 with 66\% of its memory, is 1.25 points above the best baseline on AIME24 with 56\%, and is on par with FullKV on LongBench~v2. On \mbox{Video-MME} and LVBench, it is within 0.32 points of exact-page retrieval at the same read, while storing 53\% of the full cache instead of 103\%. Its savings also stack with quantization and token pruning: with 4-bit bases, it is within 0.25 points of FullKV at 41\% of its memory.
Say Less, Hallucinate Less? Diagnosing and Restoring Object Coverage in Preference-Optimized Vision-Language Models
Hansang Lee
AI 速读 基于摘要
作者发现偏好优化在降低物体幻觉的同时会减少物体类别覆盖,并提出无需训练的基础模型引导恢复方法。
cs.CV · cs.AI
查看原文摘要
Preference optimization (PO) is the standard way to reduce object hallucination in large vision-language models. We find that four public PO checkpoints of LLaVA-1.5-7B also lower object coverage. They name fewer object categories, and their lower hallucination rates bring no significant gain in set-level object F1 on CHAIR. We therefore ask how much of the PO gain the base model can reproduce by stopping earlier. A positive end-of-sequence (EOS) logit bias lets the base model reach the recall of each PO model, and this early stopping alone reproduces 58 to 86 percent of the CHAIRi gain of three checkpoints on two disjoint COCO sets. Writing longer does not bring the lost categories back. Under greedy decoding, a negative EOS bias recovers only 2 to 9 percent of the lost CHAIR recall of these checkpoints, and 94 to 96 percent of the added object mentions repeat categories already named. On CHAIR, the base caption of the same image still names about half of the missed object categories. We therefore propose base-guided object-category restoration, which appends high-confidence base-only categories to the PO caption without training. With thresholds fixed on a development set, it brings three checkpoints within one point of the base recall on a held-out COCO set with a lower distinct-object hallucination rate than the base model. Against the recall-matched base stopping curve, its gain holds for RLAIF-V near the base recall and for RLAIF-V and CHAIR-DPO at an intermediate recall. We will release the code of the recall-controlled evaluation.