Hallucination remains a persistent challenge for Multimodal Large Language Models (MLLMs), severely limiting their reliability in high-stakes applications. Existing evaluations, predominantly based on static benchmarks, suffer from narrow taxonomical coverage and rapid performance saturation, failing to reflect model robustness in evolving real-world scenarios. To bridge this gap, we present a systematic evaluation framework integrating a comprehensive benchmark with self-evolving stress testing. First, we introduce UniHall, a fine-grained dataset grounded in a unified taxonomy spanning Object, Instruction, and Knowledge dimensions. Second, to address benchmark saturation, we propose Self-Adaptive Multimodal Fuzzing (SAMF), a self-adaptive framework that employs evolutionary mutation strategies to explore the boundaries of model hallucinations. Crucially, to ensure reliable assessment of dynamic inputs, SAMF incorporates a structured metric suite driven by an ensemble of multi-modal oracles. Our extensive experiments reveal that state-of-the-art MLLMs exhibit significant performance degradation under fuzzing compared to conventional settings, exposing a dissociation between reasoning capabilities and factual grounding. Furthermore, we identify a helpfulness-hallucination trade-off, where reinforcement learning alignment inadvertently exacerbates sycophancy in instruction-following tasks. The framework, code and benchmark are available at https://github.com/LanceZPF/EvalHall.
Graph databases are becoming foundational infrastructure for connected data, and a growing body of automated testing techniques has uncovered many bugs across mature graph database systems. Each technique checks a different aspect of correctness and exposes a different class of defect. Combining them is the natural next step, but each technique is judged by the real bugs it finds, not by useless reports or tests that can never fail, and that is what decides whether a developer can afford to fix the problems. To identify the remaining gaps, we integrated test capabilities from 8 testing methods and conducted an empirical study of their behavior, during which we found two blind spots. First, 12 hours of testing produce thousands of bug reports, far more than the number of genuine defects a mature database is likely to contain. Most reports are false positives caused by dialect or pattern-matching differences, and many of the remaining reports are duplicates. Second, analysis of process-level data reveals that many equivalence-based metamorphic test cases lose their ability to detect defects because of the optimization mechanism. Based on these findings, we summarize and realize these guidelines in GraphWeave, an integrated testing platform for graph databases. This includes a multi-agent system that analyzes the flood of bug reports to produce gating rules for issue filtering and adapter rules for dialect adaptation, and a plan-aware triage routes metamorphic pairs by plan, skipping those that share a plan on the same graph state and therefore cannot detect defects. We apply GraphWeave to NeuG, a recently open-sourced industrial graph database that has not been covered by prior graph database testing work. With modest manual effort, GraphWeave uncovers 19 previously unknown bugs, all of which have since been confirmed.
Bing Cheng, Yiran Mao, Zewei Ye, Yifan Jin, Xinyi Chen, Ruifeng Tu, Shuhuai Wang, Tao Lin 等 13 位作者
General-purpose dexterous manipulation requires policies to transfer knowledge from related tasks to new tasks with only a few target demonstrations. Existing benchmarks often provide too many target demonstrations and leave the differences between source and target tasks uncontrolled. As a result, they cannot clearly distinguish knowledge transfer from task memorization. Our key insight is the Core-Gen principle. Core tasks are related tasks that share a functional interaction and provide source knowledge, while a Gen task is a novel target within the same functional family that has only a few demonstrations. This setup tests whether a policy can reuse functional interaction knowledge. Based on this principle, we propose Evo-DexBench, which controls both source-task relatedness and target-task exposure. It comprises 30 tasks in six functional families, three everyday scenes per task, and 1,248 multimodal demonstrations collected via VR teleoperation with simulated Shadow Hands equipped with 123 taxels. Experiments show that, among six representative policies, even the best-performing method achieves an average success rate of only 24.4%, highlighting the challenges existing methods face in transferring functional interaction knowledge. Further analysis shows that, across the two tasks, joint Core + Gen training improves performance by 12% and 38%, respectively, over training on Core followed by fine-tuning on Gen, indicating that joint training makes better use of Core knowledge. With matched Core data, 8 Gen demonstrations achieve performance comparable to training on 16--32 Gen demonstrations alone. On all three tested Gen targets, functionally matched Core data yields greater gains in success rate than mismatched data. We release the Evo-DexBench dataset and benchmark to support the evaluation of dexterous manipulation policies' generalization across related tasks.
Ke Cheng, Xin Xu, Zhichao Chen, Yixiao Chen, Lei Xin, Jianbo Zhao, Fanhu Zeng, Yue Liu 等 10 位作者
Induction circuits retrieve a value by matching a query to information carried from the token before that value. How this address forms, incorporates context, and supports sequential retrieval remains less clear. We use hybrid language models as a tractable testbed and identify the historical source key as a common causal interface between local preparation and global retrieval. Probes and restoration interventions connect this interface across controlled models, pretrained recurrent and sliding-window hybrids, and released language models. We organize the analysis by address context radius and sequential retrieval depth. First, shared local parameters couple predecessor Carrying to recurrent state processing; separating these demands or biasing initialization promotes address formation. Second, adapting input representations extends an established literal circuit to category matching, making semantically compatible records active competitors. Earlier context resolves this competition through the source key, with a sharp visibility boundary in a controlled single-read model and compositional relays across local layers. Third, fixed relations support memorization shortcuts, whereas sequential generalization requires learning an output-to-query bridge between retrievals. Controlled state-tracking and compositional in-context tasks demonstrate this requirement across stacked stages, looped recurrence, and intermediate-token generation. Together, these results connect circuit formation, contextual disambiguation, and sequential use through a shared retrieval interface. Code is available at https://github.com/ckpassenger/bind-match-copy.
Triangle-based splatting connects novel view synthesis with conventional graphics pipelines, but often requires excessive tessellation to preserve appearance details. We present T$^2$-Splat, which decouples topology adaptation from appearance refinement. Safe Restricted Delaunay Triangulation (Safe RDT) initializes mesh connectivity. Dynamic Topology Evolution (DTE) then adapts triangle allocation through edge collapse and vertex split operations. Residual per-face Mip-textures then refine appearance on the fixed mesh using scale-adaptive filtering. Experiments on Mip-NeRF 360, Tanks & Temples, and DTU demonstrate competitive results with compact meshes. T$^2$-Splat improves PSNR by 0.71dB over MeshSplatting on Mip-NeRF 360. Additional compression experiments demonstrate flexible control of mesh complexity, supporting face budgets as low as 10% of the original face count. Project page: https://t2-splat.github.io/.
Computer-using agents are increasingly capable of performing complex tasks in the digital world. To advance such agents, supervised learning on annotated interaction trajectories has become a common approach to improving their task-solving capabilities. However, high-quality interaction trajectories remain scarce for specialized desktop applications, where limited application knowledge, complex GUI workflows, and incomplete exploration hinder automated collection. To address these challenges, we introduce \aks{}, a knowledge-guided automated trajectory collection framework. It combines tutorial-guided exploration with a hybrid GUI--Code action space, providing application-specific knowledge and simplifying GUI workflows through direct execution of application-native commands. The framework also checkpoints and revisits pivotal states during tree exploration to explore alternative workflows, complementing linear rollouts for broader coverage. We collect 10.9K automatically validated trajectories and use them to fine-tune open-source base models. The resulting model achieves 64.1\% accuracy on a benchmark of specialized desktop workflows, improving performance by 58.2 percentage points, establishing a new state of the art. The fine-tuned model also improves performance on general desktop workflows using GUI actions only, suggesting that the collected supervision benefits general desktop tasks as well. Comparative experiments and ablations demonstrate the efficiency of our collection framework and the value of code actions in specialized workflows.
Lifan Jiang, Yibo Zhao, Senlin Cheng, Bolei Xu, Xiaobo Li, Can Ye, Dehai Zhao, Deng Cai 等 9 位作者
Digital human e-commerce video generation requires preserving product appearance, human identity, and overall video style during dynamic demonstrations. Although recent models support multiple reference images, their general conditioning schemes do not explicitly distinguish the asymmetric roles of "what is presented" and "who presents it." Consequently, generated videos may favor a single reference, resulting in product distortion, identity drift, or single-reference collapse. To address this issue, we propose DualCast, a role-addressed dual-image conditioning framework for video generation. By explicitly modeling the distinct roles of product and human conditions, DualCast reduces role ambiguity and feature competition between references. First, a dual-spatiotemporal slot injection mechanism VAE-encodes the product and human images separately and places their latent features at the front and rear boundaries of the DiT sequence, enabling sequence-level condition separation and directed injection. Second, a mask-guided region-weighted loss strengthens supervision over product, face, and product-face overlap regions, improving generation fidelity in critical areas. Third, image condition dropout enables a single model to support dual-image, product-only, and human-only generation. In addition, we construct a curated dataset of 10K e-commerce videos for model training. Experimental results demonstrate the strong performance of DualCast.
Visual Geometry Grounded Transformer (VGGT) recovers dense 3D scene structure from multi-view images in one forward pass, but quadratic cross-frame attention limits its scalability. Existing training-free accelerators exploit token, frame, or attention redundancy; we investigate how layer-wise functional roles can guide compression. Our spectral, probing, and causal analyses reveal three regimes: shallow layers lack cross-view structure, middle layers drive cross-view alignment, and deep layers are redundant for dense geometry yet their cross-frame attention remains essential for pose. RegimeVGGT applies layer-wise U-shaped compression along two axes: Saliency-Guided Banded Merging protects geometry- and edge-salient tokens, while Selectively Protected K/V Downsampling preserves cross-frame spatial coverage and the pose-critical path through a phase-shifted spatial grid, a reference-frame anchor, and uncompressed camera/register tokens. Without any training, RegimeVGGT achieves a 6.7x speedup over the memory-optimized VGGT* at matched reconstruction quality. Project page: https://tracyou07.github.io/regimevggt/
AI 速览：Möbius 将扩散生成过程直接放入与运动理解共享的多模态表示空间，在连续帧级原始运动上统一实现运动描述、文本生成运动、时序补全与指令引导编辑。
We aim to build a truly unified motion model that can understand, generate, and edit human motion within a single framework. Existing motion UMMs remain only partially unified: motion understanding is handled by a language model, while generation is often delegated to a separate diffusion module operating over learned motion latents. We introduce Möbius, which brings the entire diffusion process into the same multimodal representation space used for motion understanding. Clean motion, noisy motion, and text jointly participate in shared attention, while role-specific parameters process understanding and generation streams. Importantly, Möbius operates directly on continuous frame-level raw motion, requiring neither discrete tokenization nor a pretrained motion VAE. This simple formulation naturally supports motion captioning, text-to-motion generation, temporal completion, and instruction-guided editing within a single model. On HumanML3D, Möbius achieves state-of-the-art performance among unified motion models. Scaling to 1.7B parameters and 1,700 hours of motion data further demonstrates that our framework is both effective and scalable.
Route Anything: A Study of Model and Policy Routing for General-Purpose AI Systems
Pengfei Zhou, Yixing Ma, Hexin Wang, Weidong Tang, Zhenglin Wan, Wangbo Zhao, Yang You
AI 速览：该工作提出统一框架 Route Anything，研究在质量—成本权衡下如何为每个任务动态选择并协调专用模型、工具与策略，并构建跨模态大规模评测矩阵来检验路由的潜力。
Artificial general intelligence (AGI) requires a method that can understand, generate, reason, and act across diverse domains and modalities. Rather than relying on a single model to master every capability, a promising path toward AGI now is to orchestrate a heterogeneous ecosystem of specialized models, tools, and policies (candidates) into a unified AI system. At the heart of this vision lies a fundamental question: how can an AI system dynamically select and coordinate the most capable candidate for each task while balancing performance and computational cost? In this work, we present Route Anything, a unified framework for studying model and policy routing across general-purpose AI systems. We develop a unified theoretical foundation that characterizes optimal routing under quality--cost trade-offs and establishes how the information available to a router determines its attainable performance. To systematically investigate routing across diverse capabilities, we first construct an omni-modal trajectory matrix spanning 25 benchmarks across language understanding, image, video, and speech generation, coding and workplace agents, and robotic manipulation, comprising 374,401 scored runs from 452 candidate configurations. Building on this framework, we train and test different routers to investigate the potential of omni-modal model routing, the capabilities of learned routers, and the conditions under which routing improves efficiency. Our results reveal substantial opportunities for capability orchestration: per-input oracle routing outperforms the best fixed candidate by 5.4 points, and learned routers outperform it by 0.3 points across 16 benchmarks, by 2.1 points on video generation, and by 1.1 points on robot manipulation while using 24% less policy time. More importantly, routing raises efficiency by 8.82x on average compared with the reference model.
Smartphone agents increasingly work as teams, in which a large planner directs parallel workers across long sessions. Running such a team on the device requires 200B-class models and contexts that exceed phone DRAM many times over, yet existing tiers treat flash as storage: UFS starves per-token weight streaming, and mobile High-Bandwidth Flash (HBF) needs a redesigned SoC and cannot absorb continuous KV writes. We show that a non-volatile memory can join an unmodified LPDDR6 bus as a plain rank only if its access latency fits the DRAM timing window, a condition no NAND tier meets. We present NORmal, to our knowledge the first main-memory architecture for 3D ferroelectric NOR flash. We have fabricated and experimentally characterized back-end-of-line (BEOL) 3D FeNOR arrays that co-integrate dense 3-bit MLC and fast, high-endurance in-place-writable SLC with sub-7.5 ns sensing. NORmal turns these co-fabricated arrays into a byte-addressable, deterministic-timing load/store rank on an unmodified LPDDR6 interface through an under-array DRAM-emulation bridge, staggered bank reads with zero-cycle weight swizzling, and posted SLC writes. Across this rank and DRAM, NORmal places agent state by mutability and lifetime rather than by hotness, serving 200B-class models and persistent planner contexts in place while dynamically partitioning subchannels across agent phases. In serving-level simulation on production agent traces, NORmal is 8.9$\times$--28.0$\times$ faster than UFS offloading, within 0.4% of an unlimited-DRAM ideal, and 2.8$\times$--6.3$\times$ faster than HBF on sessions beyond 256K tokens, with 8.1$\times$--17.8$\times$ less memory energy per token.
Xinbo Zhao, Yuhang Liu, Xin Zhang, Lei Le, Zhaohan Xi, Yu Yang, Yingxue Zhang
Offline multi-task reinforcement learning (MTRL) aims to learn a unified policy from heterogeneous offline datasets that generalizes across tasks. However, existing offline MTRL methods often struggle with two fundamental issues: severe per-task data scarcity and limited cross-task generalization. In this paper, we propose Language-Model Augmented Decision Transformer (LaMA-DT), a unified sequence-modeling framework that leverages a single large language model (LLM) for both trajectory augmentation and multi-task policy learning. LaMA-DT uses natural-language task descriptions to guide task-aligned trajectory augmentation through a set of supervised generation and alignment objectives, and incorporates a scoring-and-filtering mechanism to retain high-quality augmented trajectories. The same LLM is then repurposed as a language-conditioned Decision Transformer and fine-tuned on a mixture of real and augmented data. Experiments on Meta-World and MuJoCo multi-task benchmarks demonstrate that LaMA-DT consistently outperforms competitive baselines, particularly in low-data regimes, showing that language-guided data augmentation substantially improves cross-task generalization.
Xinbo Zhao, Yuhang Liu, Xin Zhang, Lei Le, Zhaohan Xi, Yingxue Zhang
To enable effective generalization in offline meta-RL, it is essential to accurately infer the underlying task through latent task representations, which provide a compact and transferable summary of task-specific characteristics for guiding policy generation. In this paper, we propose Contrastive Task-Aware Decision Transformer (CoTA-DT), a DT-based offline meta-RL framework that learns task representations by jointly integrating four complementary components: (i) contrastive learning from offline trajectories, (ii) alignment with natural-language task descriptions, (iii) one-step supervision via a task-conditioned world model, and (iv) long-term supervision via a task-conditioned Q function. The learned task representations are used as conditioning signals for a DT-based policy, enabling task-aware meta-policy learning. To further support generalization in both zero-shot and few-shot settings, we design a language-grounded, self-adaptive strategy that enables effective and reliable adaptation to unseen tasks. Notably, in CoTA-DT, the pretrained LLM serves as a shared backbone for language encoding, world modeling, Q-function estimation, and policy learning. This unified design facilitates parameter sharing, improving both efficiency and generalization. Extensive experiments across diverse offline meta-RL benchmarks demonstrate that CoTA-DT consistently outperforms prior methods, particularly in challenging zero-shot generalization scenarios.
Deng-Ke Wu, Yifan Jin, Yida Yang, Lei Qian, Zhaoye Zhou, Hongru Zhang, Kai Ye, Shuyuan Hu 等 11 位作者
Simulation benchmarks are the standard way to compare robot policies, but each benchmark brings its own evaluator, observation layout, and action convention, and evaluation is slow: 500 RoboTwin Easy episodes with one simulator and one model on one GPU take 359.7 minutes. Profiling the official StarVLA evaluator shows that simulation occupies CPU cores while the GPU averages 14.3% utilization. RoboScope Kit wraps each model once in a Python class called a Policy, and each benchmark once in an adapter around the benchmark's own evaluator, so the benchmark's rules and scoring stay unchanged. Scheduling runs simulator workers and Policy replicas as separate pools, sizes both from measured probes, and shards work deterministically across GPU instances; runtime optimization renders only the observations an action-chunk Policy reads. Adapters cover six benchmark families. On LIBERO-Long with a StarVLA checkpoint on one 48 GiB RTX 4090, the concurrency search selects twelve simulator workers sharing three Policy replicas, which evaluate 500 episodes 4.45$\times$ faster than the official StarVLA evaluation with equivalent success rates, 2.78$\times$ faster than RLinf, and 1.52$\times$ faster than vla-eval at its tuned configuration. With CUDA MPS, which lets the replicas run their kernels concurrently, the search selects sixteen workers sharing four replicas, 6.16$\times$ faster than StarVLA, 3.85$\times$ faster than RLinf, and 2.10$\times$ faster than vla-eval. On RoboTwin Easy, four GPUs with chunk-boundary observation complete the 500 episodes in 14.4 minutes, 25.04$\times$ faster than the single-GPU baseline, with identical success outcomes and terminal steps.
Omni-modal large language models (Omni-LLMs) typically process large numbers of multimodal tokens from audio-visual inputs, incurring substantial computational costs and a large memory footprint for the key-value (KV) cache. Aggressive token pruning can alleviate this problem while maintaining competitive performance on multiple-choice tasks, which require only a single option as output. However, we observe substantial quality degradation when generating long-form, open-ended responses. We attribute this gap to query-guided selection, which can retain sufficient evidence for answer selection while discarding details needed for comprehensive generation. This motivates reducing overlapping information across tokens, termed structural redundancy, while preserving diverse content. However, the retained representations may still contain precision redundancy due to unnecessarily high numerical precision in their cached values. To address these two levels of redundancy in token representations, we introduce a training-free framework named R$^2$-Omni. First, Structure-Aware Token Consolidation (STC) consolidates redundant representations using spatiotemporal similarity in video and local continuity in audio. Then, Semantic-Sensitivity Guided Precision Allocation (SSPA) allocates value precision to the retained representations according to the estimated reduction in attention-output error under quantized keys, with query attention and audio-visual alignment providing a semantic prior. Extensive experiments on long-form generation benchmarks demonstrate an average 5.0$\times$ reduction in the memory footprint while preserving 97.98% of baseline generation quality and achieving a 1.48$\times$ end-to-end speedup, highlighting the benefits of jointly reducing structural and precision redundancy.
Weirui Ye, Yen-Jen Wang, Elvis Hsieh, Zhiyang Dou, Shibo Zhao, Pieter Abbeel, Yilun Du, Leslie Pack Kaelbling 等 9 位作者
Hierarchical control scales robot behavior by separating task-level planning from low-level control. In dynamic, high-dimensional tasks, however, a fixed planner-to-policy interface can fail when contacts, balance, or disturbances evolve faster than high-level policy updates. Execution therefore requires choosing not only which low-level policy to run, but also which policy-specific goal should be given to it. We formulate this problem as policy orchestration: runtime selection of a policy-goal pair that is likely to remain feasible while tracking the planner target. We develop EPO, an energy-based policy orchestrator for modular goal-conditioned policy families. For each policy, EPO learns a composite energy over low-level goals, where lower energy corresponds to higher feasibility and more accurate target realization. At runtime, EPO refines each policy's goal and executes the lowest-energy policy-goal pair. Because energies are policy-specific, new policies can be incorporated by learning their own energy models, without retraining the existing policies or a centralized scheduler. In simulation, EPO improves the empirical precision-robustness trade-off by refining policy-specific goals before selectively switching among policies based on energy scores. We further demonstrate sim-to-real deployment of the same frozen policies and energy models on real humanoid loco-manipulation tasks. Project website: https://yewr.github.io/epo
Agentic robots must turn a high-level goal into actions while keeping the physical conditions needed for future steps valid. A completed action does not ensure that its effects will persist: a later disturbance can invalidate support that the remaining plan still assumes. We present LOGIV, a graph engineering framework for agentic robot execution that uses an explicit task graph to connect symbolic planning, observation-based verification, and embodied action. LOGIV compiles a validated plan into condition-support and resource-ordering edges, checks their physical validity during execution, and uses dependency tracing to determine what must be restored after a failure. Repair is constrained by retained unfinished actions, and the combined continuation is symbolically validated before execution resumes. The framework operates above existing Vision-Language-Action (VLA) and World Action Model (WAM) policies without adding trainable components. Across 20 dual-arm tasks from RoboTwin V2 and BiCoord with three execution backbones, LOGIV improves mean task success without injected disturbances by 10.67 and 15.33 percentage points, respectively. Under predefined perturbations, it achieves higher task completion and condition restoration than three adapted recovery baselines for every backbone on both benchmarks. On three real-robot tasks, successful trials increase from 12/30 to 17/30. These results support explicit, continuously checked task dependencies as a practical interface for reliable agentic robot execution.
As vision language models are increasingly deployed in clinical diagnosis, understanding how they internally resolve competing visual and textual signals becomes a safety imperative. Existing mechanistic analyses remain confined to unimodal text and offer no explanation for why a single misleading sentence can override a correct image based diagnosis, or why a model commits to a confident answer despite insufficient visual evidence. We find that these two safety risks, arbitration failure where textual context overrides visual grounding and brake failure where the model commits without adequate evidence, are mediated by spatially disjoint attention head populations: arbitration heads form a mid-to-deep wideband reflecting cross-layer evidence competition, while brake heads concentrate in a narrow middle-to-late layer band that regulates evidence sufficiency and abstention behavior. To ground these observations in causal circuitry, we introduce CRAFT, which localizes each failure mode to a minimal causal head set via dual criteria and verifies necessity and sufficiency through temporal probes and Tuned Lens trajectory analysis. Excising arbitration heads sharply reduces conflict following with negligible degradation on clean inputs, while excising brake heads restores appropriate abstention under degraded visual evidence. The two interventions target spatially disjoint head sets and produce distinct corrective effects, underscoring the mechanistic separability of the failure modes. Experiments across multiple medical VQA benchmarks and VLM architectures validate both the localization and interventions, demonstrating that the identified heads causally drive each failure mode and that targeted modulation generalises without retraining. The code is available at https://github.com/zhcz328/CRAFT.
Irregular multivariate time series forecasting (IMTSF) requires capturing temporal dynamics and cross-variable dependencies from sparse, asynchronous observations. Existing methods often rely on learned temporal encoders, coupling historical representation construction with model optimisation and repeatedly processing the same observations. We propose CTLite, a lightweight forecasting framework that decouples historical summarisation from learned prediction. Specifically, a parameter-free continuous-time moment encoder summarises local levels and trends at multiple temporal scales while suppressing poorly supported slope estimates. A compact readout combines damped local extrapolation with a low-rank cross-variable correction that accounts for observation freshness, target availability, and forecast horizon. Since the historical summaries are independent of trainable parameters and forecast queries, they can be reused across optimisation steps and repeated predictions for an unchanged history without altering predictions or parameter gradients. Experiments on four real-world benchmarks show that CTLite achieves the best or second-best mean MSE and MAE among the evaluated methods with only 88-2,537 trainable parameters. Under matched GPU workloads at batch size 32, CTLite reduces inference latency by 18.6-87.3% compared with the latest models, including the cost of historical encoding. Reusing summaries further reduces total runtime by approximately 52% over 16 forecast calls on the same history compared with recomputing the encoder for every call, including the initial encoding cost. The code is at: https://github.com/SSMa913/NICLab-CTLite.
Bin Xu, Mo Zhang, Zhengfeiyang Zhang, Xinfeng Yang, Yixing Ma, Wei Wang, HeXin Wang, Pengfei Zhou 等 9 位作者
Video diffusion transformers generate high-quality videos, but the quadratic cost of 3D full attention dominates their inference. Dynamic block-sparse attention reduces this cost by computing exact attention only on key/value blocks selected from current activations. However, its fidelity degrades under aggressive sparsity: blocks are scored from coarse summaries or formed by costly global clustering, pruned blocks vanish from the softmax, and whole tiles are kept or dropped even when only a few of their keys matter. To address these limitations, we propose Accurate Block Attention (AB-Attention), a training-free method that accurately approximates dense attention at high sparsity. Its Intra-Block Clustering Selector (IBCS) estimates block importance faithfully yet cheaply by clustering queries only within each query block, without permuting tokens. Sampling with Inclusion-Probability Weighting (SIP) then samples blocks in proportion to these estimates and reweights them by inverse inclusion probability, so that one pass estimates dense attention rather than truncating it. Finally, its kernel gathers key/value blocks much smaller than a tile into full tensor-core tiles, refining selection at no extra tensor-core cost. On Wan2.1 and HunyuanVideo, AB-Attention skips 89--92\% of attention, whereas the baselines compute 1.8--3.6$\times$ as much, yet it still improves PSNR by up to 4.4~dB over the strongest baseline and achieves up to 2.23$\times$ end-to-end speedup. Our code is available at https://github.com/InfRec-Cardinal/AB-Attention.
Rigging is inherently task-dependent because the same mesh may require different skeletons and deformation behaviors across animation tasks. In practice, artists often inspect an initial rig and repeatedly edit its skeletal structure and deformation behavior to meet specific animation requirements. Existing automatic methods primarily generate a plausible rig from geometry, offering limited explicit control over the resulting skeleton and deformation behavior. In this work, we present ViP-Rig, a visual-prompted framework that supports both prompt-first rigging and result-guided editing by injecting features extracted from user-drawn or edited 2D skeletal and rigidity prompts into frozen pretrained backbones. Specifically, ViP-Rig consists of two stages, Skeleton Generation and Skinning Prediction. In the first stage, the skeletal sketch is processed by the Dense-to-Compact Visual Prompt Encoding to produce compact, fixed-length conditioning tokens. The resulting tokens are injected into a frozen pretrained autoregressive generator through gated adapters to control joint placement and branching structure while preserving the generator's geometric prior. In the second stage, the rigidity map is processed using the same visual encoding design, while the pretrained skinning backbone remains frozen. The resulting tokens are symmetrically injected into the point and joint streams to modulate point-joint compatibility and the resulting skinning weights. Experiments on Articulation-XL2.0 and zero-shot evaluation on ModelsResource show that ViP-Rig more accurately recovers target skeletons and skinning weights than geometry-conditioned baselines under prompt-guided evaluation. Qualitative results further demonstrate explicit and localized control in both prompt-first rigging and result-guided editing.
Xin Huang, Teng Ma, Zhuoran Li, Yibo Zhao, Zhaoshi Li, Qianchao Zhu, Wei Yang, Chenxi Wang 等 10 位作者
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 at session scope, 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. On a production Claude Code trace with DeepSeek-V4-Flash, FlatCake sustains a 94% cache hit rate as load increases (versus 64% for Mooncake), eliminates Mooncake's $6.5\times$--$6.7\times$ SSD read amplification, and reduces peak-load mean TTFT by $1.6\times$ while improving total token throughput by $1.6\times$. Across 8K--128K contexts on Llama-3.1-70B, FlatCake further reduces mean TTFT by up to $2.4\times$ over Mooncake and $1.8\times$ over Mooncake-GDS.
Multimodal intelligence can be viewed as human-guided co-evolution between worlds and agents through interaction. However, scaling this process faces two bottlenecks: constructing heterogeneous, verifiable environments is costly, and manual Harness iteration cannot keep pace with growing task diversity and complexity. A key insight is that tasks with different modalities and surface forms can share the same underlying capability requirements. We therefore introduce The Trinity, a recursive self-improvement framework coupling world generation with Harness evolution through three roles: Creator, Explorer, and Evolver. Drawing on verified execution trajectories, Evolver shifts the basis of task organization from surface categories to shared requirements, reorganizing the original task tree into a hierarchical capability tree. We introduce Capability-Driven Real2Sim2Real: for heterogeneous tasks at the same node, Creator abstracts their common requirements into simplified, verifiable simulated worlds for capability learning, while evaluation on original tasks measures transfer. Along this capability tree, Explorer progressively evolves reusable capability modules through selective inheritance, forming a Harness tree whose validated updates accumulate across iterations within a fixed topology. The resulting Harnesses specialize in different capabilities, motivating a Capability-Aware Harness Router that uses public task information to select compatible module configurations from the tree. Our evolved Harnesses improve mean performance by 34 points across 28 training benchmarks and 26 points across 18 unseen benchmarks, including two real-world embodied tasks.
Recovering an object's 3D structure and camera motion from a few uncalibrated views is a fundamental problem in computer vision. Classical structure from motion methods recover 3D geometry consistent with the input images but cannot recover unseen regions; native generation methods produce plausible 3D objects which are related but not necessarily faithful to the input images. Methods simply concatenating reconstruction and generation allow error to cascade across the two stages, instead of producing a 3D object consistent with both the distribution and the input. We propose GenSfM, which integrates image-based 3D reconstruction with a native 3D diffusion model. In particular, we use a multi-modal diffusion transformer to jointly denoise the shape latent and per-view pose latents under shared self-attention, grounding generated geometry in the observed views while using object-level priors to infer unobserved regions. A compact 3D-native UV-volume tokenization keeps the joint sequence small as views are added, and a pose-aware stage refines geometry and texture by sampling per-voxel image features at the recovered camera poses. On Toys4k and GSO, for $4$ uncalibrated views, GenSfM improves novel-view PSNR by $3.2$-$4.5$ dB and reduces median Chamfer distance by $2.8$-$3.8\times$ over a strong native-3D generation baseline, while recovering input-view poses over $96\%$ Acc@$5^\circ$ without any external pose-estimation backbone, post-hoc optimization, or known intrinsics. Project page: https://dotb1ue.github.io/GenSfm/.
Zhimu Zhou, Yanpeng Zhao, Qiuyu Liao, Bo Zhao, Xiaojian Ma
Visual planning represents a crucial facet of human intelligence, especially in tasks that require complex spatial reasoning and navigation. Yet, in machine learning, this inherently visual problem is often tackled through a verbal-centric lens. While recent research demonstrates the promise of fully visual approaches, they suffer from significant computational inefficiency due to the step-by-step planning-by-generation paradigm. In this work, we present EAR, an editing-as-reasoning paradigm that reformulates visual planning as a single-step image transformation. To isolate intrinsic reasoning from visual recognition, we employ abstract puzzles as probing tasks and introduce AMAZE, a procedurally generated dataset that features the classical Maze and Queen problems, covering distinct, complementary forms of visual planning. The abstract nature of AMAZE also facilitates automatic evaluation of autoregressive and diffusion-based models in terms of both pixel-wise fidelity and logical validity. We assess leading proprietary and open-source editing models. The results show that they all struggle in the zero-shot setting, finetuning on basic scales enables remarkable generalization to larger in-domain scales and out-of-domain scales and geometries. However, our best model that runs on high-end hardware fails to match the zero-shot efficiency of human solvers, highlighting a persistent gap in neural visual reasoning.
Yumeng Fu Xi'an Jiaotong University, Jiayin Zhu Xi'an Jiaotong University, Lingling Zhang Xi'an Jiaotong University, Wenjun Wu Xi'an Jiaotong University, Bo Zhao Xi'an Jiaotong University, Shaoxuan Ma, Yushun Zhang, Jun Liu Xi'an Jiaotong University
Geometry problem solving (GPS) poses significant challenges for Multimodal Large Language Models (MLLMs) in diagram comprehension, knowledge application, long-step reasoning, and auxiliary line construction. However, current benchmarks lack fine-grained evaluation for long-step problems necessitating auxiliary construction. To address these limitations, we present GeoLaux, a fine-grained annotated dataset comprising 2186 calculation and proof problems. It features long-step reasoning (with an average solution length of 6.51 steps, maximum of 24 steps) and auxiliary line construction (required in 41.8% of problems). Building on the dataset, we conduct a comprehensive five-dimensional evaluation of 23 leading MLLMs. The evaluation yields three pivotal findings: First, models perform significantly worse on long-step problems compared to short-step ones, with 18 models exhibiting a performance drop of over 50%. Second, it is crucial to enhance models’ understanding, awareness, and proficiency in auxiliary line construction, which is vital for overall geometric reasoning. Third, limited answer hints effectively improve process correctness, whereas explicit answers lead models to neglect intermediate reasoning steps. These findings position GeoLaux both to benchmark MLLMs geometry reasoning abilities and to guide their improvement. Data and code are available at https://github.com/Candice-yu/GeoLaux
尚无 AI 解读，展示原文摘要
cs.CL · cs.CV
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Jiahao Xiong, Yihe Liu, Xianming Hu, Hongbo Zhao, Nuoyi Chen, Jie Zhang, Kai Zhang East China Normal University
Low-Rank Adaptation (LoRA) for large language models (LLMs) has achieved significant success in various domains. So far, most algorithms in the LoRA-family rely on global low-rank factors spanning the entire update weight matrix (\Delta \mathbf{W}). Through careful analysis, however, we observe that the \Delta \mathbf{W} during fine-tuning typically exhibit heterogeneous subspace clusters, each corresponding to specific sub-sets of rows and columns. This structural heterogeneity suggests that global low-rank factors may not optimally capture the local variations needed for effective model adaptation. To address this limitation, we propose LoRA within Clustered Parameter Subspaces, or CPS-LoRA, which performs independent low-rank updates within clustered blocks of parameter matrices. The key idea is to group the rows/columns of the update matrix into locally coherent, and maximally uncorrelated subspaces, perform low-rank adaptations in each subspace, and iteratively update the partition and local adaptations. This allows adapting to local structures more precisely while preserving high efficiency. Theoretical analysis reveals that in case \Delta \mathbf{W} can be partitioned into subspace blocks with non-overlapping basis, CPS-LoRA have superior parameter efficiency than global adaptations. Empirical evaluations further demonstrate better rank utilization of CPS-LoRA and its consistent improvements against LoRA (and variants) by up to 3.0% in absolute accuracy in various benchmarks.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
While Large Language Models (LLMs) have increasingly assisted in historical tasks such as text processing, their capacity for professional-level historical reasoning remains underexplored. Existing benchmarks primarily assess basic knowledge breadth or lexical understanding, failing to capture the higher-order skills—such as evidentiary reasoning—that are central to historical research. To fill this gap, we introduce ProHist-Bench, a novel benchmark anchored in the Chinese Imperial Examination (Keju) system—a comprehensive microcosm of East Asian political, social, and intellectual history spanning over 1,300 years. Developed through deep interdisciplinary collaboration, ProHist-Bench features 400 challenging, expert-curated questions across eight dynasties, accompanied by 10,891 fine-grained evaluation rubrics. Through a rigorous evaluation of 18 LLMs, we reveal a significant proficiency gap: even state-of-the-art LLMs struggle with complex historical research questions. We hope ProHist-Bench will facilitate the development of domain-specific reasoning LLMs, advance computational historical research, and further uncover the untapped potential of LLMs. We release ProHist-Bench at https://github.com/inclusionAI/ABench/tree/main/ProHist-Bench.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Lirong Gao, Zewei Yu, Zhongrui Yin, Qi Zhang, Yuke Zhu Alibaba Group, Bo Zheng, Haobo Wang Zhejiang University, Junbo Zhao Zhejiang University 等 10 位作者
Tabular data is widely used in fields such as finance and healthcare. Traditional tree-based models are prevalent for tabular prediction tasks due to their ability to handle heterogeneous features. However, their heavy reliance on feature engineering limits both their generalizability and their human-readable interpretability. On the other hand, Large Language Models (LLMs) naturally provide intermediate reasoning steps, thus offering greater transparency in decision-making. Nevertheless, LLMs often fail to match the predictive performance of tree-based models on tabular data. To address these challenges, we propose a novel Logic-Graph-Enhanced LLM Reasoning (LogGER) framework that integrates the strengths of tree-based models and LLMs. Specifically, we reformulate the traditional decision tree as a human-readable logic graph, which explicitly models the causal relationships between features and targets. This logic graph is automatically constructed using LLMs based on data priors and serves as the foundation for LogGER. To fully leverage the logic graph, we further introduce a logic-graph-guided process supervision approach, which evaluates and enhances the quality of LLM’s intermediate reasoning steps using logic-graph-aided process reward. Extensive experiments demonstrate that LogGER consistently outperforms both tree-based models and state-of-the-art LLM methods on a variety of tabular prediction tasks, achieving superior accuracy and interpretability.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Zhengpeng Shi, Yanpeng Zhao, Jianqun Zhou, Yuxuan Wang Alibaba Group, Qinrong Cui, Wei Bi Independent Researcher, Song-Chun Zhu, Bo Zhao Shanghai Jiaotong University 等 9 位作者
AI models capable of comprehending humor hold real-world promise—for example, enhancing engagement in human-machine interactions. To gauge and diagnose the capacity of multimodal large language models (MLLMs) for humor understanding, we introduce v-HUB, a novel video humor understanding benchmark. v-HUB comprises a curated collection of non-verbal short videos, reflecting real-world scenarios where humor can be appreciated purely through visual cues. We pair each video clip with rich annotations to support a variety of evaluation tasks and analyses, including a novel study of environmental sound that can enhance humor. To broaden its applicability, we construct an open-ended QA task, making v-HUB readily integrable into existing video understanding task suites. We evaluate a diverse set of MLLMs, from specialized Video-LLMs to versatile OmniLLMs that can natively process audio, covering both open-source and proprietary domains. The experimental results expose the difficulties MLLMs face in comprehending humor from visual cues alone. Our findings also demonstrate that incorporating audio helps with video humor understanding, highlighting the promise of integrating richer modalities for complex video understanding tasks.
尚无 AI 解读，展示原文摘要
cs.CL · cs.CV
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表