Xiang Zhang, Yang Zhang, Lukas Mehl, Karlis Martins Briedis, Markus Gross, Christopher Schroers
Accurately modeling soft boundaries, e.g., hair and defocus blur, is a fundamental challenge in stereo conversion due to the ambiguous blending of foreground and background. Existing depth models primarily predict single-layer depth, leading to ambiguity in depth correspondence at soft boundaries. While matting techniques can capture opacity for layered modeling, they often struggle in complex scenes with multiple targets and usually require user intervention. This paper introduces αDepth, a layered representation that decomposes soft boundaries for high-fidelity stereo conversion. Specifically, we first resolve mixed color and depth ambiguity by estimating layered color and depth values at soft boundaries. Considering complex multi-target scenes, we design a Circular Alpha Representation (CAR) that shifts the paradigm from global target extraction to local boundary decomposition. Unlike prior matting methods restricted to a single foreground/background, CAR enables efficient scene-level inference without manual guidance. Extensive evaluations demonstrate that αDepth achieves state-of-the-art performance in stereo conversion, eliminating background bleeding and structural distortions at soft boundaries.
Humans naturally possess the ability to work persistently toward long-term goals. Given a challenging task, humans can continuously work for months or even years to accomplish a specific objective. In this paper, we propose Marathoner, an autonomous agentic model possessing the ability of ultra-long-horizon execution. Specifically, we propose a comprehensive post-training pipeline to instill this critical capability into base model. For Ultra-Long-Horizon Task Synthesis, we leverage major release PRs containing 1000+ lines of new code from diverse GitHub repositories as the primary source for synthesizing challenging task-level data. Additionally, we introduce Multi-Task Chaining, which chains multiple generated tasks into a single more challenging task, enabling the synthesis of tasks with frontier-level difficulty. For rejection sampling finetuning, we combine strong teacher model with diverse harnesses to generate trajectories on our synthesized tasks and conduct supervised finetuning on base model with rejection sampled trajectories. For reinforcement learning, cold-started model performs real-world execution through harnesses in independent sandboxes during rollout process, effectively facilitating the acquisition of genuine ultra-long-horizon execution capability. We further propose a novel reward strategy, Later Stage Bonus Reward, which explicitly encourages model to perform meaningful maneuvers during later stages of execution. Through extensive evaluation on 5 benchmarks containing ultra-long-horizon tasks, Marathoner achieves consistent and substantial performance improvements over base model and even surpasses performance of strong proprietary model. Further analysis shows that Marathoner can consistently work for 10+ hours and conduct 1000+ tool calls on highly challenging tasks.
Lyutianyang Zhang, Yixin Ying, Liu Cao, Vanlin Sathya, Jinke Ren, Shuguang Cui
Channel sounding in Wi-Fi multiple-input multiple-output (MIMO) systems must balance channel state information (CSI) accuracy against CSI acquisition overhead. Device mobility causes channel aging, while autoregressive CSI prediction introduces errors that accumulate over sounding intervals. However, standardized statistical channel models cannot fully capture the environment-specific channel dynamics encountered in physical artificial intelligence (AI) applications, motivating the use of Sionna-based ray tracing to evaluate CSI prediction under device mobility and evolving propagation geometry. This paper develops a dynamic sounding framework that integrates autoregressive CSI prediction with signal-to-interference-plus-noise ratio (SINR)-constrained interval selection. A Transformer with rotary position embedding that converts device positions into CSI sequences predicts CSI across multiple subcarriers for beamforming updates between sounding exchanges. We formulate sounding-interval selection as a predicted average-throughput maximization problem under per-stream SINR lower-bound constraints, including CSI acquisition and processing overhead. A prediction-error perturbation analysis provides the SINR bounds, and cumulative-sum enumeration globally solves the finite single-variable problem. Actual SINR feasibility holds whenever the prediction errors satisfy the bounds. Ray-tracing experiments assess channel prediction, trajectory transfer, and autoregressive prediction errors. In single-STA Wi-Fi 7 link-level simulations, predicted-CSI beamforming improves mean goodput by 67.3%, 112.7%, and 119.1% over stale CSI for linear, circular, and cosine trajectories, respectively.
Classifier-Free Guidance is a widely used inference-time guidance rule for conditional image generation with diffusion models, which exposes a guidance scale that adjusts conditioning strength after training. This flexibility requires conditional and unconditional model evaluations at every sampling step, so an N-step trajectory takes 2N forward passes. Single-pass alternatives reduce such cost, yet most collapse CFG's two components into one opaque guided prediction, losing the separable base flow and guidance direction that support per-step control and diagnostics, even when the scale remains adjustable. This work introduces CtrlFlow, a single-pass guidance method that preserves CFG's separable base-flow and guidance-direction interface. To preserve both components at one-pass cost, CtrlFlow attaches two lightweight LoRA adapters to a frozen backbone, each adding 2.5% parameters, and predicts the base flow and guidance direction from one pass of the shared frozen backbone. Because the guidance scale multiplies the guidance direction only after the forward pass, one evaluation yields the guided velocity at any scale, so a single trained checkpoint remains adjustable at inference without retraining. Preserving two outputs introduces a training mismatch: inference amplifies guidance-direction errors by the chosen scale, while naive component-wise losses weight both outputs equally. To align training with this composed inference-time error, CtrlFlow uses a unified multi-scale distillation loss. Residual guidance errors can still persist at extreme scales, where even small guidance-direction errors are amplified by the guidance scale. To reduce this inference-time error, CtrlFlow applies scale-adaptive modulation to damp velocity-field errors and time-grid warping to concentrate solver evaluations where the trajectory changes most rapidly.
Beyond Unit-Circle Eigenvalues: Invariant Bases for Stable State Space Dynamics
Jiaqian Zhu, Yang Zhang, Junhua Ding, Xiaowei Yu
AI 速览：作者提出对角递推的守恒量具有分层结构，并通过Kronecker-Cayley变换按构造强制满足各层条件，以在训练前固定守恒量。
Diagonal state space models process long sequences at linear cost, yet the same transition is applied at every step of a rollout, so an unconstrained recurrence holds no quantity fixed and the state decays or grows without bound. Current designs constrain one quantity, and preserving one does not guarantee preservation of the others, so a model can hold its norm to machine precision while momentum drifts by hundreds of percent. We claim that the conserved quantities of a diagonal recurrence are layered, that each layer requires a strictly stronger condition on the operator, and that a unit-modulus spectrum reaches only the outermost. Acting on the claim, we enforce the conditions required for every layer by construction. The generator of the Kronecker-Cayley transition (KC) is a graph Laplacian over object subspaces, $\mathbf{H}=\mathbf{L}\otimes I_m$, advanced by a Cayley transform, and the step stays diagonal at $O(N^3+N^2m)$ cost. The conserved set is exactly a $2m^2$-dimensional family of quadratic forms, held at every parameter value, so the quantities and the dynamics the construction reaches are fixed before training. On four conservation benchmarks, KC holds total mass at or near the reference floor while every diagonal baseline loses it by orders of magnitude. We further compare against mass correction and find that two correctors fail once the underlying prediction diverges, which shows that assembly rather than correction is necessary for a guarantee that holds across the rollout. A free modal basis trained under a conservation penalty conserves only by freezing its spectrum, whereas the assembled operator conserves with every mode but one free to rotate. Beyond these benchmarks, KC holds the linear momentum of three $N$-body systems and, at the operator level, the conserved quantity in molecular dynamics, chemical kinetics, pharmacokinetics and power-grid swing dynamics.
Gabriel Diaz-Aylwin, Joseph Neighbour, Abiel Malkani Talwar, Rui-Yang Zhang, Henry B. Moss
Bayesian optimisation is the natural tool for shape design when objectives are expensive and non-differentiable, but it needs a compact yet expressive parameterisation of the search space. Hand-crafting one is a complex endeavour requiring domain expertise, and often yields implicit infeasible regions, artificial bounds, and coupled, unordered coordinates. We instead learn the parameterisation from a collection of existing designs, applying principal component analysis to the deformations between shapes. The result is a linear, interpretable search space in which the number of components explicitly trades expressivity against dimensionality. Across aerofoils, wings, and radio-frequency cavities, spanning 2D geometry to 3D aerodynamics and electromagnetics, we show improved sample efficiency and the ability to explore beyond the confines of hand-crafted baselines.
Andrey Ignatov, Georgy Perevozchikov, Radu Timofte, Dong Zhang, Zhao Ran, Yang Zhang, Xiaochen Li, Shichang Ju 等 24 位作者
On-device RGB photo enhancement is a popular task that is needed for quality improvement of photos captured with mobile devices as well as for retouching of images acquired through other sources. This creates a need for efficient deep learning models for this problem that are not only performant but can also be deployed on constrained mobile hardware. In this Mobile AI challenge, we address this problem and propose participants to design fast solutions that can perform comprehensive enhancement of RGB photos while running on popular smartphone GPUs from Qualcomm and Arm. For this, the participants were provided with a large-scale DPED dataset consisting of RGB image pairs captured with an old iPhone 3GS smartphone and a professional Canon 70D DSLR camera. The runtime of all models was evaluated on the Snapdragon 8 Elite Gen 5 and Dimensity 9500 chipsets with the Adreno and Mali GPUs, respectively. Top solutions proposed in this challenge are able to enhance Full HD resolution RGB images on both platforms under 300 milliseconds. A comprehensive description of the models developed in the challenge is provided in this paper.
Xizhi Zhang, Zhongjie Chen, Yang Zhang, Fengyuan Ren
Modern datacenter applications demand high network bandwidth, motivating packet-level multipath transport to exploit the aggregate capacity of multiple paths. However, packet reordering under packet spraying poses challenges for hardware transports. Excessive packet reordering makes it difficult to infer packet loss, imposes substantial pressure on NIC resources, and ultimately degrades performance. Existing transports either prioritize throughput without controlling packet reordering, or mitigate reordering by restricting transmissions, which leaves available path capacity unused. In this paper, we identify two fundamental sources of packet reordering: transient inter-path queueing differences caused by control dynamics, and persistent path-delay differences due to path heterogeneity. These observations motivate a joint design of a multipath controller and a packet scheduler to address transient and persistent delay differences, respectively. We present Rondo, a packet-level multipath transport that achieves both high throughput and low packet reordering. Its multipath controller uses ACK-based per-path capacity estimation and probing to maintain stable path loads, while its packet scheduler adjusts packet transmission order to compensate for persistent delay differences. Testbed experiments and large-scale simulations show that Rondo reduces the maximum out-of-order degree by up to 97.5% while maintaining high throughput. With constrained receiver buffers, Rondo achieves up to 8.51 times higher goodput.
Jun Huang, Junyang Zhang, Zijie Yue, Bingkun Wang, Yong Luo, Miaojing Shi
Video Reasoning Segmentation (VRS) aims to infer and segment target objects specified by reasoning queries in videos. Existing methods typically first infer the target and generate a few segmentation tokens for it, which are then fed into a mask decoder for mask prediction. However, such segmentation tokens tend to be both semantically monotonous and spatially ambiguous, making them largely insufficient to represent the target to guide mask prediction. In this paper, we propose an explicit visual object grounding framework for VRS. Specifically, we first propose a spatial-aware prompt generation and refinement scheme, in which we query a multimodal large language model to generate explicit spatial prompts (i.e., bounding box and point set) for the target and then refine them through multi-step dialogue. Furthermore, we introduce a query diversification and alignment module to generate auxiliary queries that describe the same target from different perspectives. We then enforce cross-query consistency between their predicted spatial prompts to alleviate the training bias caused by query limitations. Finally, we select several reliable frames using a prompt-based keyframe selection strategy to support mask decoding and propagation. Extensive experiments on eight datasets demonstrate that our method significantly outperforms state-of-the-art methods. The code is available at https://github.com/viscom-tongji/EVOG.
Chenru Wang, Xuanyang Zhang, Yucheng Han, Wenbo Nie, Wei Cheng, Xianfang Zeng, Gang YU, Chi Zhang
A deployable text-to-image diffusion model must hold several capabilities at once, such as instruction following, text rendering and human preference, and must sample in a few steps. Capabilities and few-step sampling are built in separate stages: a full-step model first consolidates capabilities through multi-teacher on-policy distillation, then serves as a fixed teacher for step distillation. Every full-step model we evaluate scores lower once compressed, and the second stage cannot begin until the first finishes. We introduce StepX-OPD, which merges the two stages: one student acquires every capability while learning to sample in a few steps, so frozen specialists supervise it on the states its own rollout visits rather than on a longer trajectory it never runs. Because a decomposition of the rollout's KL weights velocity mismatch most heavily near the clean end, we also regress the student's velocity on the low-noise part of the rollout. Separately, we find that capability integration makes models write prompt words into images on prompts that never asked for text, and we give a variant that lowers it. We evaluate on SD3.5-Medium and FLUX.1-dev across instruction following, text rendering, image quality and human preference. On both architectures, StepX-OPD attains the highest normalized average score among the four-step methods compared and exceeds the forty-step Flow-OPD model, and on SD3.5-Medium it trains with about 37% fewer GPU-hours than sequential DiffusionOPD followed by step distillation.
Tianle Zhou, Yang Zhang, Tianyu Wu, Zhuohan Wang, William Nixon, Juncheng Yang
Looped transformers apply a shared block several times per token to improve accuracy without adding parameters. For models such as Ouro that keep a separate key-value (KV) cache per loop, cache memory limits the batch size that amortizes repeated weight reads. Uniform-bit quantization and fixed rules per loop do not allocate precision according to each layer's sensitivity in each loop. Measurements on Ouro show that quantization sensitivity varies across (loop, layer) cells, and the sum of individual cell damages approximates the overall output damage. Sharing damage also varies by cell, with later loops generally tolerating sharing better. Skipping low-attention values to save reads produces errors that compound across loops. LoopKV uses a one-time offline profile to choose 4-bit storage, 2-bit storage or reuse of the final loop's KV for each cell. The planner minimizes reads under budgets for predicted damage and storage. During decoding, each loop reads the high 2 bits of each 4-bit value and fetches the low bits only where its own attention weights require 4-bit precision. A single fused kernel in vLLM executes the plan. Serving Ouro-1.4B on one RTX PRO 6000, the memory-optimized LoopKV-M plan stores 35% less KV than uniform 4-bit quantization (KIVI-4) on the same fused kernel and CUDA graph. It sustains 1.7-2.2x KIVI-4's goodput under a latency service-level objective with a 16K-token maximum length. None of its averages on six public benchmarks falls significantly below BF16's.
Yilun Hao, Ruixiao Yang, Hudson Hilal, Yongchao Chen, Kaizhi Qian, Yada Zhu, Yang Zhang, Chuchu Fan
Real-world problems span diverse domains, including program synthesis, symbolic reasoning, planning, and optimization, and often require fundamentally different solution paradigms. A central challenge for LLM-based reasoning is twofold: identifying an appropriate problem-solving approach and formulating it correctly. In this paper, we propose AlgoPilot, a cross-paradigm reasoning framework that enables adaptive selection of the solving approaches and structured problem formulation prior to execution. The AlgoPilot framework contains three components: 1) a SteerLM, trained via supervised fine-tuning (SFT) and reinforcement learning (RL), that adaptively selects the solving strategy (paradigm and modes) based on input problems, 2) the corresponding expert GuideLM, trained via SFT, that generates structured formulation guidance tailored to the selected strategy, and 3) an ExecutionLM that generates solutions conditioned on both the selected strategy and its formulation. Unlike prompting-based methods that rely on fixed reasoning templates, our approach learns to adaptively choose and structure problem-solving strategies based on the input. We build AlgoPilot on Qwen3-8B and evaluate across 5 problem categories and 13 benchmarks with 66 unique domains, comparing against both the latest similarly scaled small models and large LLMs. AlgoPilot significantly improves average accuracy for the same model from 38.4% to 72.9%, outperforming the best large LLM prompting baseline GLM-5 by 11.7%. Ablation studies confirm the effectiveness of each component. Notably, we show that the learned steering and formulation modules can transfer across execution models, suggesting their potential to improve other LLMs. Project page: https://sites.google.com/view/algopilot.
Multimodal Large Language Models (MLLMs) have shown strong performance on video question answering, but their application to long-form videos is constrained by limited context length and computational cost, making keyframe sampling essential. Existing approaches typically rely on semantic relevance or reinforcement learning, which either fail to capture evidential clues or suffer from inefficient combinatorial optimization. In this work, we propose an evidence-driven keyframe sampling framework motivated by an information-theoretic perspective. We consider maximizing the conditional mutual information between the selected frames and the answer given the query as an idealized objective, and use a modular singleton-gain surrogate to motivate a tractable frame-wise selection criterion. Guided by this motivation, we train a query-conditioned evidence scoring network using annotated answer-supporting segments. A contrastive objective encourages greater aggregate score mass on frames from evidence regions than on background frames, providing a practical evidence-based criterion beyond semantic similarity. Experiments on long-form video understanding benchmarks show that our method consistently outperforms prior sampling strategies under comparable frame budgets. With Qwen2.5-VL-7B under the same 32-frame budget, our method improves over uniform sampling by 10.1 points on LVBench and 4.4 points on the long-video subset of VideoMME.
Ziheng Chen, Jiali Cheng, Zezhong Fan, Hadi Amiri, Diyuan Wu, Gabriele Tolomei, Yang Zhang
Generative recommendation formulates next-item prediction as autoregressive generation over semantic ID (SID) sequences derived from users' historical interactions, making modern recommender systems structurally similar to large language models (LLMs). As privacy and safety concerns grow, these systems increasingly require concept unlearning to remove sensitive or harmful concepts associated with items. However, existing LLM unlearning methods cannot be directly applied to generative recommendation. Unlike word tokens with explicit semantics, SIDs are abstract identifiers that are often shared by both forget and retain items, leading to severe conflicts between concept removal and recommendation utility preservation. To address this challenge, we propose TRACER, an end-to-end concept unlearning framework based on token reassignment. Rather than directly suppressing shared SIDs, TRACER reassigns concept-related items to alternative tokens that better facilitate forgetting while minimizing side effects on retained items. We further introduce a coherence regularizer to preserve semantic consistency among retain items during unlearning. Experiments on real-world recommendation datasets demonstrate that TRACER effectively removes target concepts while substantially better preserving recommendation utility than existing unlearning baselines.
Peiyuan Tang, Yao Li, Xiaodong Zhang, Wuyang Zhang, Qin Xia, Yanyong Zhang, Zijiang Yang
Contact-rich manipulation requires fine-grained control and complex physical interactions with the environment, posing significant challenges for robot policy learning. Vision-only policies struggle with such tasks, as contact states and interaction forces are difficult to infer from visual observations alone. Existing methods mitigate this issue by incorporating force or tactile sensing. However, they often treat these signals as auxiliary observations without fully exploiting their role in modeling future interaction dynamics. To address this, we propose ForceLDM, a force-aware latent dynamics model that learns future contact and motion dynamics to guide action generation. Rather than directly predicting future observations, which can entangle task-relevant dynamics with irrelevant sensory details, ForceLDM learns a compact representation of future dynamics through knowledge distillation. Specifically, a teacher network with privileged future force/torque and optical flow signals extracts future dynamic features, which are then distilled into a student network trained only on current observations. This enables the student model to reason about upcoming contacts and physical dynamics without relying on future sensory inputs at deployment. To improve robustness and generalization, we further introduce a curriculum-based progressive noise injection strategy to mitigate over-reliance on future features. Experiments on five real-world contact-rich manipulation tasks demonstrate that ForceLDM outperforms state-of-the-art baselines, generalizes effectively to novel objects, and remains robust to environmental perturbations. Project page and code are available at https://tpy001.github.io/ForceLDM/.
Rapid advances in cowork agents are shifting human-AI collaboration toward real-world automation, delivering software, designs, and other artifacts. However, such productivity work demands both high-quality delivery on complex productivity tasks and continued monitoring and decision-making over long horizons, a combination that remains an open challenge. We introduce CoworkOS, a general-purpose harness for long-horizon work that separates task strategy from execution management: the model decides actions, delegation, and acceptance, while the runtime maintains state and feedback. Its three-layer architecture enables direct execution or delegation to Subagents, synchronously or asynchronously, under stable responsibilities. Through a shared application layer, humans and agents inspect and revise the same evolving documents, presentations, CAD models, and Blender scenes, so human-in-the-loop participation can enter both through conversation and through direct modification of the work itself. To sustain continuity across agents and sessions, CoworkOS further introduces hierarchical workspaces and persistent semantic memory, so working files, key decisions, and long-running Job logs remain shareable, reusable, and resumable after interruption. Across 12 benchmarks and 6 categories of real-world applications, CoworkOS proves highly competitive with frontier agent applications such as Codex and DeepSeek Harness. As a sustained-work case study, CoworkOS uses GLM-5.3 as its backbone to conduct RSI-driven automated post-training on Qwen3.6-35B-A3B. Over five rounds, the model's average score across eight benchmarks rose from 33.19 to 42.95 (+9.76). Looking ahead, we believe CoworkOS opens a meaningful path toward general cowork-agent intelligence.
Haiyang Zhang, Sicheng Song, Zixin Chen, Jie Gu, Huiya Ou, Chenhui Li, Changbo Wang
Misleading charts can lead vision-language models to conclusions that conflict with underlying data. Existing work focuses on inference-time strategies, while training-time supervision remains less explored. We present VizDAGuide, a DAG-structured supervision method for misleading-chart question answering. The method represents intermediate reasoning states and their dependencies as a directed acyclic graph, assigns supervision weights according to their roles, and serializes the graph into a unified training target for open-ended and multiple-choice QA with single-pass inference. Experiments on two benchmarks across Qwen3-VL and InternVL3.5 models show improvements over a content- and weight-matched Linear CoT-SFT baseline on both answer formats. Process-level evaluation shows gains in grounding and operation execution, while ablations indicate that dependency structure and weighted supervision are most effective when combined. These results support DAG-structured weighted supervision for misleading-chart question answering.
Scientific workflows often require choosing among known relations before a deterministic calculation can proceed. Whether observations share a culture, treatment or reference standard can change the scientific meaning of the resulting count or comparison. Once the evidence is prepared and the calculation is written as code, this choice reduces to one semantic judgment among listed candidates. We evaluate Jev on this step, using a harness that follows its documented guidance. The study compares twelve model configurations on twenty Choices in ten scientific cases built from published sources, each repeated five times. We measure semantic selections, downstream outputs and final claim labels separately. Jev tied with five other configurations at complete semantic correctness and achieved the lowest observed median latency and known mean cost among successful responses. On one question, whether samples shared a culture, three comparison models made seven wrong selections that changed downstream counts while preserving the correct final label. These results support using Jev to select relations when the evidence and candidates are prepared in advance, and show that evaluating it requires checking the relations and quantities that a workflow will reuse.
Xinhe Zhang, Yuyang Zhang, Pengfei Jin, Arnau Marin-Llobet, Na Li, Quanzheng Li
High-resolution 3D medical image generation remains challenging because fully volumetric models are computationally expensive, while efficient 2D slice generators often fail to preserve anatomical consistency across the third dimension. We propose LiFT, a framework for Lifted inter-slice Feature Trajectories that factorizes 3D volume synthesis into per-slice image generation and inter-slice trajectory learning. Rather than using a fully volumetric synthesis backbone, LiFT treats a volume as an ordered trajectory in feature space, capturing how anatomical structures appear, transform, and disappear across depth. We demonstrate the LiFT framework in both unconditional and conditional settings. In unconditional generation, a tri-planar drifting loss matches feature distributions between generated and real volumes across three orthogonal planes; in paired translation, a bidirectional $z$-context mixer trained with paired target supervision supplies through-plane coherence while preserving per-slice fidelity. We evaluate LiFT on BraTS 2023 for unconditional and missing-modality MR synthesis, and on SynthRAD2023 for MR-to-CT translation. LiFT achieves competitive image quality while retaining the efficiency of predominantly 2D synthesis, operating up to $390\times$ faster than a volumetric diffusion baseline and consistently improving through-plane coherence through explicit inter-slice modeling. These results establish lightweight inter-slice trajectory learning as an efficient approach to high-resolution 3D medical synthesis.
Humanoid controllers need the ability to yield to external forces while retaining the capacity to resume their original tasks. However, tracking objectives penalize the deviations needed for yielding, while relaxing tracking goals persistently can make task recovery harder. Existing approaches often rely on task-specific interaction command interfaces or constructed compliant motion data and reference trajectories, limiting the direct reuse of their compliance-training mechanisms across different task interfaces. We introduce CompliantBot, a compliance-training framework for motion tracking and velocity-command locomotion. It preserves the original task inputs and requires no preconstructed compliant data or target trajectories. A state-dependent interaction force field couples applied loads to the robot's motion, enabling local articulation and whole-body movement to change the forces experienced. A complementary force-magnitude reward encourages lower interaction loads, while task rewards are downweighted during loading and restored after release to coordinate yielding and task resumption. The trained policies use task inputs and proprioceptive history without explicit external-force measurements. Across both task interfaces, CompliantBot yields farther under constant pulls and experiences lower spring interaction loads than matched baselines, while resuming task tracking after force removal. CompliantBot achieves the highest task completion among the compared methods under stronger pulls. Power evaluations further show that CompliantBot has the lowest mean joint mechanical power among the compared methods. Real-robot demonstrations further illustrate compliant responses through arm motion, stepping, and body reorientation.
Yan Tai, Yunan Ding, Guangzhao Li, Siqi Luo, Chaoyang Zhang, Guangtao Zhai, Guodong Guo, Xiaohong Liu
KV cache quantization is well studied for autoregressive language models, where cached prefix states are exact under causal attention and quantization is mainly a cache reconstruction problem. Diffusion language models (dLLMs) break this assumption. Fast dLLM inference reuses KV states across denoising steps, but bidirectional token representations can change as masked tokens are revealed, so the reused cache may already differ from full recomputation before quantization is applied. We study KV cache quantization under this cache-reuse error. We adapt KIVI and KVQuant to Fast-dLLM-style prefix and dual cache reuse, and introduce TRACE and dLLM-KVQ. TRACE reconstructs feasible dLLM reveal trajectories from prompt-answer pairs, providing calibration states that match iterative masked decoding. dLLM-KVQ uses these trajectories to choose per-layer K-cache entries that remain in high precision according to output- and attention-level behavior metrics, while applying standard low-bit value quantization. Across LLaDA and Dream evaluations on GSM8K, HumanEval, MATH, and MBPP, dLLM-KVQ achieves the highest average score among quantized methods across the reported 4-bit and 2-bit prefix/dual cache settings.
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.
Lei Huang, Chuan Qiu, Kuan-Jui Su, Anqi Liu, Yun Gong, Weiqiang Lin, Lindong Jiang, Chen Zhao 等 16 位作者
Genotype imputation enables dense variant coverage for genome-wide association and risk-prediction studies, yet conventional reference-panel methods remain limited by ancestry bias and reduced rare-variant accuracy. We present Genotype Bidirectional Encoder Representations from Transformers (GenoBERT), a transformer-based, reference-free framework that tokenizes phased genotypes and uses a self-attention mechanism to capture both short- and long-range linkage disequilibrium (LD) dependencies. Benchmarking on two independent datasets including the Louisiana Osteoporosis Study (LOS) and the 1000 Genomes Project (1KGP) across ancestry groups and multiple genotype missingness levels (5-50%) shows that GenoBERT achieves the highest overall accuracy compared to four baseline methods (Beagle5.4, SCDA, BiU-Net, and STICI). At practical sparsity levels (up to 25% missing), GenoBERT attains high overall imputation accuracy ($r^2 approx 0.98$) across datasets, and maintains robust performance ($r^2 > 0.90$) even at 50% missingness. Experimental results across different ancestries confirm consistent gains across datasets, with resilience to small sample sizes and weak LD. A 128-SNP (single-nucleotide polymorphism) context window (approximately 100 Kb) is validated through LD-decay analyses as sufficient to capture local correlation structures. By eliminating reference-panel dependence while preserving high accuracy, GenoBERT provides a scalable and robust solution for genotype imputation and a foundation for downstream genomic modeling.
Robot-learning studies often compare position and velocity actions as though each label identified a single experimental treatment, yet the label alone does not determine what the robot executes. Before affecting robot motion, bounded policy output passes through filtering, physical scaling, reference construction, feedback, timing, and effort limits. We study how these omitted properties affect attribution under one shared learner and training budget. Across three contact-rich Factory tasks (PegInsert, GearMesh, and NutThread), we compare five joint-space interfaces, with ten independently trained policies per condition. We first remain within the same position-feedback law and vary the executed interface: substituting only the physical action scale as the directly modified component changes sample-median evaluation success by 0.645-1.000 across tasks, while removing an exponential moving average produces task-dependent changes. We then move to a cross-law comparison and choose a velocity scale that exactly matches the position interface's local raw-torque command gain. This calibration raises the sample median on GearMesh and NutThread, with no median change on PegInsert. Yet clear residual sample-median gaps remain on PegInsert and GearMesh, while the NutThread interval includes zero. Thus, even this explicit equality leaves feedback and held-action dynamics unmatched and does not establish position-velocity trained-outcome equivalence. Action-interface comparisons should therefore report the executed pathway and state the physical quantity they match.
Test-time training (TTT) enables a deployed language model to internalize incoming context on-the-fly into a small set of fast-weight parameters, rather than merely treating it as transient context for attention. In-Place TTT recently made this idea practical for large language models by repurposing the down-projection of standard MLP blocks as fast weights. However, we find that its benefit largely disappears on realistic web long-context question answering. We attribute this to In-Place TTT's uniform write schedule: every chunk is written with equal strength, so irrelevant content dilutes query-relevant information in fast-weight memory. Building on this insight, we propose Select-TTT, a model-agnostic, query-conditioned rule for In-Place TTT that scales each chunk's write strength by its relevance to the user query. This relevance signal requires no supervision and no new parameters: it is read directly from the model's frozen output head, as the probability mass that a chunk's hidden states assign to the query tokens. A temperature-scaled softmax turns these scores into learning-rate multipliers, leaving the chunk-wise, context-parallel computation of In-Place TTT intact. Select-TTT consistently improves long-context question answering on two benchmarks and two backbones. These findings highlight the value of query-conditioned fast weight writes for practical test-time training on long-context question answering, which is also potentially promising for long-horizon agent tasks involving multiple user queries over evolving contexts.
Feed-forward Gaussian Splatting enables efficient novel view synthesis, but pixel-aligned prediction makes the primitive count grow with image resolution and view count, leading to substantial redundancy and increased computational cost. Non-pixel-aligned methods address this limitation by predicting a fixed number of Gaussians. Without explicit ray and depth constraints, these methods yield misplaced Gaussian centers, causing rendering artifacts under large viewpoint changes. To address this issue, we introduce GeoGSplat, a geometry-grounded feed-forward model that explicitly incorporates geometric priors into the non-pixel-aligned framework. A dual-branch encoder constructs Anchors and Triplane to capture coarse 3D structure and fine-grained scene features, respectively. Subsequently, a hierarchical decoder progressively decodes dense 3D points and Gaussian primitives, enabling 3D reconstruction and novel view synthesis. GeoGSplat surpasses both state-of-the-art pixel-aligned and non-pixel-aligned Gaussian Splatting methods on DTU and NRGBD in terms of novel view synthesis and reconstruction. The advantage of our geometry-grounded paradigm further increases as target viewpoints move farther from the input cameras, demonstrating stronger generalization to large viewpoint changes. Project Page: https://comedy0913.github.io/projects/geogsplat
Zeji Yi, Soham Dasgupta, Xinyang Zhang, Koushil Sreenath, Guannan Qu, Xi Chen, Max Simchowitz, Guanya Shi
Though modern AI models like LLMs predict discrete tokens, today's robotic foundation models represent actions as continuous vectors. In this paper, we challenge this common consensus by asking if continuous action parametrizations are superior to their discrete counterparts. We conduct an extensive comparison of single-step regression and multi-step flow variants of both continuous and discrete policies in both from-scratch single-task and multi-task VLA training, in both simulation and on real hardware. In all regimes, we find that discrete action prediction, even the most naive, uniform-bin quantization, can match or even outperform continuous-action policies in task success rate. Moreover, we find that the challenge to discrete action policies is not action precision, but rather factorization error introduced by product-distribution predictions across time and action dimensions. We show that, when paired with the correct interventions (iterative computation through discrete flows, or distribution sharpening through classifier-free guidance), discrete flow policies can turn this weakness into a strength, enabling more responsive action changes during a predicted action chunk. Lastly, we show that attempts to explicitly inject geometric information into discrete policies, e.g. through corruption processes that emulate continuous flow matching, do not yield consistent improvements. Taken together, our findings question existing assumptions that robotic action prediction requires distinct design decisions compared to other AI domains, and motivate the possibility of discrete action prediction as a compelling alternative to today's continuous action parametrizations. Project page: https://iscoyizj.github.io/discrete-action-website/
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.
Jiayi Luo, Yang Zhang, Liliang Chen, Nan Wang, Chen Gao, Boxiang Qiu, Zhuoxuan Li, Baorui Peng 等 16 位作者
Action-conditioned world models offer a scalable alternative to costly real-world interaction for improving robot policies. However, existing world-model-based methods largely treat world models as forward-only surrogates of the physical environment, leaving untapped a capability no physical environment affords: rewinding to past decision points to explore counterfactual futures. We introduce Rollback, a policy improvement paradigm that turns world models from passive simulators into counterfactual engines: it rolls back to critical decision points, branches into alternative actions, and contrasts their outcomes to attribute success to individual decisions. This concentrates imagination where it matters most and distills counterfactual outcomes into fine-grained, action-level supervision without a separately learned value model. We build a scalable training system that unifies distributed world-model search with asynchronous real-robot execution, forming a self-reinforcing cycle: real-world experience continually sharpens the world model, and a sharper world model in turn drives further policy improvement. We term the world-model side of this cycle test-time physics grounding, a capability that hand-built physics simulators lack: while the world model serves as the training environment, real transitions from the policy it trains correct its dynamics through a lightweight LoRA pipeline. Experiments on three challenging real-world tasks show that Rollback improves success rates over the strongest baseline by an absolute 10.67%, 11.11%, and 11.85% on block extraction, ball tossing, and tabletop curling, respectively, while achieving more stable improvement throughout training. Although initialized from teleoperated demonstrations far below human level, Rollback surpasses the average success rate of human players on both ball tossing and tabletop curling within 8 hours of training.
Chenglong Ma, Yuanfeng Ji, Junzhi Ning, Jiyao Liu, Yujie Wei, Shaohao Rui, Chaoyang Zhang, Wenjie Li 等 21 位作者
Multimodal medical AI requires a common visual interface for encoding heterogeneous imaging evidence for patient condition modeling. However, existing medical visual tokenizers are often tied to a specific dimension and imaging modality, forcing multimodal systems to rely on fragmented and poorly reusable representations. We introduce MedVTok, a general-purpose visual tokenizer that maps 2D images and 3D volumes from diverse imaging modalities into a unified token space. Building such a tokenizer poses two key challenges: preserving anatomical structure across dimensions and capturing clinical semantics across modalities. To this end, MedVTok introduces (i) slice differential regularization, which explicitly models inter-slice anatomical coherence missing in previous per-pixel/per-voxel supervision, and (ii) multi-expert representation alignment, which integrates knowledge from multiple medical expert encoders while retaining modality-specific complementary knowledge. Trained on 28M images and 70K volumes across 9 imaging modalities, MedVTok supports classification, retrieval, segmentation, synthesis, visual question answering, and medical report generation through a single tokenizer. Across 24 benchmarks, MedVTok achieves state-of-the-art performance, demonstrating a scalable token interface that unifies medical image dimension, modality, and task usage.