Nguyen Hong Quan, Sitao Luan, Yilun Zheng, Lihui Chen
Future link prediction on temporal graphs involves both repeated interactions between previously connected node pairs and first-time interactions between pairs with no prior interaction. While repeated-edge prediction can exploit direct pairwise evidence, unseen-edge prediction must generalize beyond historical recurrence. Existing temporal graph models often exploit signals tied to recurring interactions or historical overlap, which are highly informative for repeated edges but become limited or unavailable for unseen edges, contributing to substantial performance degradation on unseen-edge prediction. Based on this observation, we identify two complementary signals for modeling sequential dynamics beyond historical recurrence: (1) historical-sequence evolution, which captures dependencies among the source node's chronologically ordered interactions, and (2) candidate--history compatibility, which captures how well a candidate fits the source node's evolving history. Based on the decomposition, we propose MambaTG, a candidate-aware Mamba-based framework that jointly models these two signals. Mamba provides a natural mechanism for this design: its selective recurrence captures the evolution of ordered interaction histories, while its input-dependent state transitions allow the history state to adapt to different candidate inputs. Extensive experiments on 13 temporal graph datasets against 13 representative baselines show that MambaTG achieves strong and consistent performance across both repeated and unseen edges while remaining computationally efficient.
Francisco Affonso, Matheus P. Angarola, Ana Luiza Mineiro, Aditya Potnis, Marcelo Becker, Girish Chowdhary
Perceptive legged locomotion across discontinuous terrain requires shared locomotion knowledge, task-specific maneuvers, and smooth transitions between them. Existing approaches train specialized policies independently and either switch among them with a hierarchical selector or distill them into a monolithic policy. Both fix the constituent behaviors before they are composed, and both are assessed through aggregate course outcomes that leave behavior at terrain boundaries uncharacterized. We instead propose CTS-MoE, which formulates multi-terrain locomotion as multi-task reinforcement learning and jointly trains a perception-gated dense mixture-of-experts actor with task-specific critics: soft routing composes expert knowledge into continuously varying behavior, while separated critics isolate heterogeneous objectives so that specialization survives joint optimization. A concurrent teacher-student formulation transfers this composition from privileged terrain information to onboard depth, so that at deployment routing is inferred from perception alone, with no terrain label, selector, or distillation stage. Against hierarchical and DAgger-distilled baselines, the policy raises connected-course success to 93.3% from 73.3% and 53.3%, lowers mean action jerk by about a third across the course, and avoids the switching transient that hard selection introduces at boundaries. On a Unitree Go1, it completes an extended outdoor course in 10/10 trials and transfers to four outdoor terrain types. Project Website: https://cts-moe.github.io/ .
One-step generators built on MeanFlow synthesize a video in a single function evaluation, yet inherit a rigid treatment of classifier-free guidance (CFG): the guidance scale is frozen during training or, at best, exposed as a single scalar. This conflicts with mounting evidence that the optimal guidance is not a constant but a field that varies with the noise level, with spatial content, and, for video, with frame position. Every existing remedy presupposes a multi-step sampler along which guidance can be re-scheduled--an assumption one-step models violate by construction. We resolve this tension through a single observation: because the MeanFlow target is an integral over a noise interval, an arbitrary guidance field can be absorbed into the target itself and compiled into one evaluation. We formalize this as the interval-integrated CFG identity, prove that the compiled guidance carries no schedule-discretization error and can thus match or exceed any finite multi-step sampler, and instantiate it as AGAVE, which equips the average-velocity field with a learnable temporal schedule, a free motion-saliency spatial modulation, and a frame-position asymmetry. Adding fewer than ten scalars at zero inference cost, AGAVE surpasses all one-step baselines on VBench, UCF-101 and MSR-VTT while matching multi-step guided sampling. Project repository: https://github.com/YMlinfeng/AGAVE.
Recent 3D world modeling systems based on generative scene synthesis, such as Marble, can create coherent and explorable 3D environments, yet their outputs are typically static monolithic assets with limited editability and physical interaction. This restricts their use in immersive content creation and embodied simulation, where generated worlds must be actively modified and manipulated. To tackle this challenge, we present WorldAct, a framework that converts static generated 3D worlds into editable and interaction-ready scenes. WorldAct uses a multimodal agent to guide scene decomposition, identify actionable objects, reconstruct geometrically aligned object-level meshes for interaction, and restore the residual background via 3D inpainting. The resulting scenes support object-level editing, collision-aware manipulation, and embodied task execution while preserving global scene coherence. Experiments show that WorldAct enables richer interaction scenarios than the original generated scenes, suggesting a practical path toward editable and interactive 3D world models. Project page: https://sjtu-deepvisionlab.github.io/WorldAct
Multi-teacher on-policy distillation consolidates expert capabilities by supervising student-generated trajectories, whereas task arithmetic offers a training-free alternative through linear parameter combinations. We view the linear combination of student and expert parameters as a single differentiable model and propose P$^2$-MOPD, which jointly optimizes the mixing coefficients and the student weights under the same on-policy distillation objective. Unlike task arithmetic, which stays within the span of the frozen student and expert parameters for any choice of merging coefficients, P$^2$-MOPD keeps the student trainable, so the experts act as parameter priors without confining it to this span. Across four domains at 1.7B and 4B scales, P$^2$-MOPD achieves faster and more balanced recovery than MOPD from the same initialization, surpassing its final average recovery in half the training steps. At the end of training, our 1.7B model reaches 94.7% average recovery, versus 80.9% for MOPD.
Weiqin Chen, Xinmeng Huang, Dongsheng Ding, Santiago Paternain
Safe reinforcement learning from human feedback (Safe-RLHF) seeks to improve helpfulness while satisfying harmlessness constraints. Existing primal-dual approaches can converge to feasible policies but do not prevent constraint violations during training. We study a stronger requirement, safe exploration for RLHF, under which every policy executed during training, including auxiliary exploration policies, remains feasible. We propose Zeroth-Order Log-Barrier Policy Gradient (ZO-LBPG), a reward-model-free method that optimizes a log-barrier objective using zeroth-order estimates derived directly from human preferences. To ensure safety under noisy feedback, we combine conservative constraint margin estimates with coupled rollouts and an adaptive step size that prevents updates from crossing constraint boundaries. We prove that, with high probability, all iterates and exploration policies remain strictly feasible, and that ZO-LBPG globally converges to an approximately optimal constrained policy. We establish a sample complexity of $\widetilde{\mathcal O}(\varepsilon^{-6}+d^3\varepsilon^{-5})$ trajectories and $\widetilde{\mathcal O}(\varepsilon^{-8}+d^5\varepsilon^{-7})$ human queries. Experiments on a synthetic CMDP and a language-model alignment task corroborate the training-time safety guarantees.
Every deployed KV-cache-eviction rule and every top-$k$ sparse-attention rule keeps a subset of the keys, discards the rest, and renormalizes the attention weights over the kept set. Identifying the exact best subset under that constraint, we find that keeping the largest weights is already near-optimal, since the best subset we find closes only a median $2$ to $5\%$ of the remaining gap to full attention. We evaluate deployed eviction methods in KVPress and find that methods with non-uniform per-head budgets apply attention masks over the full cache because dense cache tensors require equal length across heads. Truncating each head to the target budget lowers benchmark scores by $14$ to $62$ points, and computing token importance with the question in context accounts for $87.6$ points in retrieval accuracy. We propose ContourKV, a training-free rule that applies KVzip's score under a physically enforced budget. It beats KVzip in $93$ of $160$ benchmark runs across RULER and LongBench while using $28$ to $1078$ times less memory, and ties Compactor, the strongest baseline at equal memory.
Sky-image irradiance forecasting is dominated by a two-stage design: a video-prediction model forecasts future sky frames, and a separately trained regressor maps those frames to irradiance. We propose a single-stage alternative that co-denoises the future sky-image and clear-sky-index sequences in one diffusion process, with an irradiance head that reads a denoised preview of the future sky rather than the noised diffusion state. The forecaster needs no regression stage and runs from sky images and deterministic covariates alone, with no measured irradiance at its input --- the situation at a site with a camera but no pyranometer. On 2023 SIRTA data the model reaches a pooled RMSE of $104.1$~W\,m$^{-2}$ and a skill score of $+0.34$ on the continuous ranked probability score (CRPS) against smart persistence, higher than any of twelve baseline configurations, and retains that skill when retrained on 2021--2022. On the model's own generated frames, a separately trained regressor reaches $150.6$~W\,m$^{-2}$ against the joint head's $104.1$: roughly $46$~W\,m$^{-2}$ that joint training recovers and a cascade discards. Point accuracy is competitive but not best: a deterministic network that also reads measured irradiance attains $97.6$~W\,m$^{-2}$. The model's CRPS skill leads that network's in every sky-condition tertile, by $0.26$ to $0.30$.
Transmitters in adaptive communication systems may switch their modulation scheme at an unknown time, and a receiver must detect this switch quickly without raising frequent false alarms. Existing modulation classification methods label a finite block of received data and do not provide stopping rules with controlled false-alarm and delay performance. We study the sequential detection of a switch from one phase shift keying (PSK) scheme to another over an additive white Gaussian noise channel. We build cumulative-sum (CUSUM) and moving-sum (MOSUM) detectors from the log-likelihood ratio and its low-SNR approximation. For a switch from quadrature PSK to binary PSK, we derive the moment-generating function and probability density function of the approximate statistic. We use them to relate the CUSUM threshold to the average run length to false alarm and the worst-case detection delay. For a switch between any two PSK orders, we approximate the vector of overlapping MOSUM statistics as multivariate Gaussian and obtain bounds on both measures. Monte Carlo simulations show that our derived expressions for ARL2FA and WADD closely match the empirical results across thresholds, validating the accuracy of the proposed analytical approximations and theoretical bounds.
Existing concurrency control protocols either introduce nondeterminism, resulting in a serial execution-replay dependency between primary and replica nodes, or rely on impractical prior knowledge of transaction read-write sets. In this paper, we present Lantern, a deterministic concurrency control protocol tailored for high-performance transaction processing systems operating without prior knowledge. The key insight of Lantern is that all zero-out-degree transaction vertices in the local dependency graph can be safely committed in ascending order using an overwrite-permissive strategy. We further introduce a novel Back-Propagation mechanism that iteratively propagates dependency states from sink to source vertices to identify additional committable transactions. We also propose Conflict-Free Batch Selection (CFBS) for read-modify-write intensive scenarios. We integrate Lantern into the open-source blockchain platform ChainMaker. Extensive evaluations on YCSB and SmallBank benchmarks demonstrate that Lantern achieves up to a 4.2x throughput speedup over Aria and improves the throughput of ChainMaker's execution layer by at least 2.2x.
RL for LLM reasoning often reaches an early performance plateau as the policy distribution becomes increasingly sharp. Recent work has studied policy entropy as both a diagnostic and a control signal for this behavior. In our experiments, however, increasing entropy alone does not reliably change the eventual plateau. We analyze RL through its implicit target policy and derive a sharpness factor $κ$, together with an effective counterpart $κ_{\mathrm{eff}}$, that describes how strongly reward or advantage gaps are translated into probability-ratio gaps. Motivated by this view, we introduce reward budgeting, a plug-in advantage construction that assigns a fixed total reward budget to each prompt. For binary rewards and independent on-policy rollouts, its unclipped score-function estimator follows the gradient of an exact finite-$G$ concave objective with diminishing returns, while its advantage gap decreases as more rollouts become correct. For general rewards, a shifted log-mean interpretation holds asymptotically under the stated regularity conditions. These results motivate, rather than establish exact objective equivalence for, the clipped implementations. On Qwen3-8B and Llama3.2-3B-Instruct, reward budgeting consistently improves a range of RL baselines, with gains up to +12.50 on AMC23 and +10.00 on AIME24 and AIME25. Overall, our results support an update-dynamics view of premature convergence and show that strong gains can be obtained without explicitly preserving entropy.
Pengxin Zha, Yi Lu, Haojun Ma, Jun Zhu, Minghao Qiu, Xiaoyan Wang, Wei Shan, Maoqing Yao
Wheeled humanoid robots carry both arms on one actuated waist, usually without torque sensors, so the accuracy of both end-effectors depends on a single waist model. After a first stage that estimates the tracker and target extrinsics, we identify the waist and both arms in one problem from the target positions on both flanges, measured by an optical tracker. The kinematic model is complete, with waist-base, parallel-axis, and branch parameters; arm compliances are driven by measured torques and waist compliances by gravity torques of the robot's mass model; and Gaussian priors with evidence-maximizing scales stabilize the estimate. Independent and sequential single-chain calibration are restrictions of this problem, and its waist information is the sum of the single-chain Schur complements. On two robots with different waists, each without and with a payload, calibration reduces the mean errors on held-out poses from 11.3-31.1 mm to 0.50-1.11 mm. For every training-set size from 20 poses to all, the method has 17-63% lower mean errors than adaptations of two published whole-body calibration methods, mainly through the waist compliance. Independent calibration yields two waist models; using either for both arms raises the other arm's error to up to 3.21 mm.
Elliott Thornley, Alexander Roman, Christos Ziakas, Leyton Ho, Louis Thomson
The POST-Agents Proposal (PAP) is an idea for ensuring that advanced artificial agents never resist shutdown. A key part of the PAP is using a novel `Discounted Reward for Same-Length Trajectories (DReST)' reward function to train agents to (1) pursue goals effectively conditional on each trajectory-length (be `USEFUL'), and (2) choose stochastically between different trajectory-lengths (be `NEUTRAL' about trajectory-lengths). In this paper, we propose evaluation metrics for USEFULNESS and NEUTRALITY. We use a DReST reward function to train simple agents to navigate gridworlds, and we find that these agents learn to be USEFUL and NEUTRAL. Our results thus provide some initial evidence that DReST reward functions could train advanced agents to be USEFUL and NEUTRAL. Our theoretical work suggests that these agents would be useful and shutdownable.
Differentiable weather models allow atmospheric initial conditions to be directly optimized to shadow a known trajectory, which we take from a reanalysis dataset. Here, we extend this optimization to include the forecast model, GraphCast, using low-rank adaptation (LoRA). We compare model optimization with initial-condition (IC) optimization, and then joint optimization, where both are tuned simultaneously. When compared against a control forecast initialized with reanalysis data, LoRA and IC optimization individually produce similar improvements. We find similar results when IC and LoRA are jointly optimized across a spectrum of solutions where a tuning coefficient smoothly shifts the weight of the joint optimization from all-LoRA to all-IC; forecast improvement is not unique to the relative contributions of model and IC corrections. To better discriminate the source of error, we introduce a second model, Pangu-Weather, as a supervisor that gates IC updates during joint optimization. This gated optimization is the same as before, except that IC updates at each epoch are accepted only when they improve the forecasts of both models; otherwise, only the LoRA update is accepted. Results show that this cross-model constraint substantially improves agreement between the two models in their response to the same optimized ICs, narrowing the gap in mean forecast error reduction from 64% to 10%. Moreover, when used together, the LoRA and IC corrections from the gated optimization yield GraphCast forecast improvements similar to those from the ungated optimization. These results imply that model error that is encoded in the ungated ICs is shifted to the LoRA corrections via gating, yielding IC corrections that may provide better analysis estimates.
Ziheng Ji, Chunyu Liu, Weihao Peng, Mingheng Mi, Pu Wang, Hai Zhang
As the computing power density of large-scale AI clusters rises, network communication increasingly limits system performance. GPU-initiated communication requires low-overhead request submission to overlap communication with computation. However, GPUDirect RDMA (GDR) still relies on CPU involvement in its control path, while InfiniBand GPUDirect Async (IBGDA) requires GPUs to assemble Work Queue Entries (WQEs), consuming additional SM resources and memory bandwidth. We propose OpenRDMA, an open-source multi-path reliable RDMA transport architecture designed for GPU-friendly communication. First, OpenRDMA provides a minimal GPU doorbell mechanism: the GPU issues an 8-byte token, and a dedicated FPGA NIC reconstructs the complete WQE from preconfigured context. Second, building on Multipath Reliable Connection (MRC), OpenRDMA supports per-packet multi-path transport and out-of-order arrivals, while selective retransmission, ordered DMA placement, and GPU-visible flag polling provide reliable delivery and safe data consumption. Third, a hardware-software co-designed retransmission mechanism keeps the GPU submission path lightweight and reduces FPGA hardware state overhead: the NIC handles fast data transfer, while CPU software manages complex timeout detection, state tracking, and recovery policies. We describe the architecture and key hardware components of OpenRDMA; implementation of the FPGA prototype is underway.
Moonseok Choi, Jongwon Jeong, Minki Kang, Giung Nam, Juho Lee
Large Language Model (LLM) agents are increasingly deployed under structured generation, where grammar constraints enforce output format at inference time. Whereas existing compression methods target unconstrained generation, we first identify a fundamental mismatch: under structured generation, grammar constraints reduce the model's output distribution to valid tokens, and the logits matter most at the small set of positions that decide the agent's action. Motivated by this finding, we propose SKIM (Selective Knowledge-Informed Masking), the first pruning and distillation framework for LLM agents under structured generation, with both stages guided by token role. SKIM partitions token positions into context, decision, and format categories, and this partition drives two complementary axes: (1) category-aware pruning saliency scores that preserve parameters most relevant for decision positions; and (2) an on-policy distillation objective that aligns student and teacher on valid tokens, paired with category-aware feature matching. We evaluate SKIM on tool-calling benchmarks and within multi-step agent loops, across multiple model scales and sparsity levels. SKIM consistently improves accuracy over baselines, with its superiority becoming most striking in high-sparsity regimes where conventional methods struggle, demonstrating the potential of category-aware compression for deployment-time grammar-constrained agents.
Jingnan Gao, Jiahui Huang, Sven Elflein, Jianhao Zheng, Bingxin Ke, Xiaokang Yang, Yichao Yan, Laura Leal-Taixé 等 11 位作者
We present MetricTrail, a streaming system for accurate camera motion estimation and dense reconstruction from long monocular videos. To keep inference efficient over long sequences, we revisit DPVO's sparse local estimator. We find that single-image metric depth substantially improves its trajectory accuracy, without retraining the correspondence network or enlarging the optimization window. Metric predictions initialize patch depths and remain as soft constraints during optimization. We further improve correspondence estimation through learned adapters that augment matching features and recurrent context with frozen DINOv3 representations. The correspondence estimator and adapters are trained on diverse short clips without additional long-sequence training. This modular design also allows the depth provider to be independently upgraded and supports supplied camera intrinsics and calibrated stereo observations without retraining. Under the reported RGB-only protocols, MetricTrail achieves the lowest average trajectory error among evaluated methods on KITTI, Oxford Spires, and VBR. Pairing the estimated poses with single-image depth predictions also produces strong dense reconstruction on indoor and outdoor benchmarks and coherent reconstructions of in-the-wild videos. With supplied intrinsics, the system processes KITTI-02 at 12.70 FPS with 3.21 GB peak GPU memory.
Kaiqu Liang, Jaime Fernández Fisac, Thomas L. Griffiths, Andrew Lampinen
Language agents increasingly use persistent memory to record user preferences and procedures across sessions. We show that memory from benign interactions can cause unsafe or unreliable behavior on later tasks. We call this failure the memory trap. To study this issue, we compare behavior in sessions with and without memory across six frontier models (including Claude Sonnet, Opus, and Fable models and GPT-5.6). We evaluate these models in a range of settings. First, in autonomy experiments, we find that requests to skip confirmation during low-stakes coding chores lead agents to skip confirmation and verification on simulated high-stakes tasks in both coding and everyday life. We next study memories formed when users seek emotional support, or describe harm from uncritical feedback, and find these memories bias later judgments of mathematical work, arguments, and poems toward sycophancy or over-critique (respectively). Memory notes written during legitimate stale-test repairs lead agents to weaken correct tests that catch broken safeguards; these memories also increase cheating on ImpossibleBench, with larger increases for stronger Sonnet and Opus models. We show that memory traps can arise both from overgeneralizations in memory writing, and when acting on memory. These findings demonstrate that even benign interactions can lead to serious risks from memory systems, and thus motivate the need for new research into ensuring that agents preserve the scope of their memories.
Synthetic data from diffusion models can improve object detection, but long-tail performance is still limited by scarce rare classes and inflexible conditioning (e.g., text-only or layout-only inputs). In this paper, we propose Tail-SD, a layout-instruction conditioned stable diffusion framework that supports text, layouts, or their hybrid, enabling controllable synthesis from boxes, masks, and/or prompts to better cover the long-tail objects. To train a single model across heterogeneous conditions, we introduce a multi-task cross-attention module and multi-task embedding tokens, improving generation quality and diversity. Extensive experiments on COCO and LVIS demonstrate that diffusion synthesis with hybrid conditioning consistently benefits detection, e.g., on LVIS val, it achieves +4.89% / +5.68% rare Box/Mask AP, highlighting strong potential for addressing long-tail challenges.
Effective behavioral intervention in chronic disease management requires a sequence of incremental steps, each grounded in what similar individuals have achieved. Multimodal records of daily behavior make such evidence available outside the clinic. Counterfactual explanation, an explainable AI approach, can turn these records into guidance, yet endpoint-only methods return a target state without a route to it. We propose POROS (Peer-Grounded Optimal Routes Over States), a domain-agnostic framework rooted in Bandura's self-efficacy theory and Festinger's social comparison theory. POROS constructs a Behavioral Progression Graph, a directed acyclic graph over observed patient states in which every edge requires both peer-grounded behavioral proximity and strict improvement in the observed health outcome. Every edge is bounded by a feasibility threshold derived from the cohort's day-to-day behavioral changes. Minimum-cost paths through this graph decompose behavioral gaps into incremental, peer-grounded steps. We evaluate POROS on two independent longitudinal cohorts of patients with type 1 diabetes, with a per-patient predictor estimating individual time in range (TIR, blood glucose within 70-180 mg/dL) along each path. Compared with an endpoint-only baseline, POROS returns multi-hop paths to the same target for 14.4% and 19.8% of days below the 70% TIR threshold, reducing the per-step change in carbohydrate intake, insulin dose, and bolus timing by 38-42% on the larger cohort. POROS also supports improvement within the same outcome class, connecting 78.1% and 83.7% of outcome-ordered pairs of well-controlled days. Together, these findings demonstrate how POROS translates multimodal health records into incremental, peer-grounded guidance for chronic disease management through observational path planning.
Linear attention reduces the quadratic sequence-length complexity of softmax attention by maintaining fixed-size recurrent states. However, these states scale quadratically with the representation dimension and require repeated data transfers during decoding, creating a substantial GPU memory bottleneck. Recent studies have revealed significant low-rank redundancy in recurrent states, suggesting opportunities for more efficient representations. We introduce SPARE (State Projection for Accelerated Recurrent Execution), a training-free method that exploits this redundancy through fixed low-dimensional subspaces, heterogeneous rank allocation, and lightweight write compensation. By performing recurrent reads and updates directly on compact states, SPARE reduces memory traffic without online decomposition or dense-state reconstruction. SPARE achieves 2.55x full-model decoding speedup, 7.49x lower recurrent-kernel DRAM traffic, and 3.31x lower peak GPU memory, while preserving better associative-retrieval accuracy than structured pruning at matched memory budgets.
Few-step autoregressive (AR) video diffusion models generate videos chunk by chunk and have made efficient video generation practical. However, further accelerating these models toward real-time generation remains challenging, as few-step distillation leaves little cross-step redundancy for existing training-free caching methods to exploit. We revisit this regime and identify two token-level redundancy patterns: cross-frame redundancy from static content shared across adjacent frames, and cross-step redundancy from tokens that converge early during denoising. Crucially, we find that early clean estimates along the model's own denoising trajectory provide an effective online signal for locating both patterns, requiring no retraining, auxiliary predictor, or reference video. Based on these findings, we propose PreSTP (Predictive Spatio-Temporal Token Pruning), a training-free framework that constructs a unified token-wise redundancy mask from cross-frame similarity and cross-step convergence. Tokens marked as redundant skip subsequent computation under a consistent last-computed-value restoration rule, and a content-adaptive mask reuse strategy further reduces dense probing cost. On 5-second generation, PreSTP achieves 1.52x-3.20x speedups and a better efficiency-fidelity trade-off than coarse-grained caching baselines such as TeaCache and FlowCache. On 30-second generation, it reaches up to 1.96x speedup while improving VBench-Long and suppressing long-horizon drift. Notably, it continues to reduce inference cost even in the extreme one-step regime, where cross-step redundancy vanishes entirely and coarse step-wise caching is no longer applicable.
Mobile networks are evolving toward service-aware operation to support heterogeneous applications with diverse quality-ofservice (QoS) and quality-of-experience (QoE) requirements. In practice, however, the widespread use of encryption makes service types difficult to observe directly, which limits the deployment of service-aware optimization in operational networks. Recent raw-byte representation learning methods have improved offline encrypted-traffic classification, but their robustness in realistic deployment remains limited due to temporal drift and the lack of explicit labels for online adaptation. This paper proposes Stable Encrypted-Traffic Representation and Online Adaptation (SEROA), a deployment-oriented framework for encryptedtraffic classification in service-aware mobile networks. In the offline stage, SEROA exploits protocol structure to learn temporally stable representations by applying masked modeling to stable fields, protocolaware transformations and contrastive learning to unstable fields, and time-adversarial learning to suppress time-specific cues. In the online stage, SEROA uses feedback records containing QoE measurements, wireless states, and allocation decisions to update the deployed classifier through contextual-bandit learning. Experiments show that SEROA consistently outperforms representative pretrained baselines under both random and time-based evaluation protocols, and remains effective under QoE-driven online adaptation. These results demonstrate the value of stable representation learning and online adaptation for practical service-aware mobile networks.
Seungyong Lee, Suhyeok Jang, Minyoung Kim, Jinwoo Shin
Robot foundation models have shown remarkable performance in robot control, yet adapting them to a new task still relies heavily on teleoperated demonstrations. Egocentric human videos offer a scalable substitute, but overlaying rendered robots on human videos or retargeting human poses into robot actions leaves significant visual or action gaps for high-DoF robot embodiments. To address this limitation, we propose RoboTryOn, which treats human-to-robot translation as virtual try-on: try-on keeps the person and swaps the garment, RoboTryOn keeps the demonstrated motion and swaps the human for the robot. Specifically, RoboTryOn converts a human video into a video of the target robot specified by a reference image, while preserving the original task, objects, and background. To maintain visual consistency, we translate selected keyframes autoregressively and use them to guide video synthesis with motion conditioned on the human hand skeleton. Additionally, we use an inverse dynamics model (IDM) that incorporates depth features to infer action labels from synthesized videos for policy learning. Our experiments across multiple humanoid embodiments show that policies trained with our synthetic data improve the average success rate on novel tasks, including challenging bimanual manipulation, from 2% to 60%, without any real robot demonstration of those tasks. Moreover, training with our synthetic data matches the performance of policies trained solely on teleoperated demonstrations at approximately a third of the collection cost, demonstrating the cost efficiency of our approach.
Partially Relevant Video Retrieval (PRVR) aims to retrieve untrimmed videos containing moments relevant to a given text query. Despite recent progress, existing PRVR methods suffer from two key limitations: a fixed video decomposition scheme that causes semantic dilution, and source-domain overfitting induced by task-specific training. In this paper, we propose TF-PRVR, the first training-free framework for PRVR. TF-PRVR leverages frozen vision-language features to construct video-specific hierarchical representations. It derives temporal semantic signals from frame-level features and applies frequency-based multi-scale analysis to identify adaptive temporal boundaries, producing hierarchical segments with coherent event-level semantics. Built on these segments, TF-PRVR constructs a unified multi-scale graph and propagates query relevance across temporally and semantically related nodes. A moment-aware scoring strategy then aggregates temporally aligned relevance across scales, emphasizing consistently supported moments while suppressing isolated false responses. Without task-specific training, TF-PRVR preserves the general-purpose alignment capability of pre-trained vision-language models and avoids dataset-specific overfitting. Extensive experiments demonstrate consistent performance across datasets with diverse visual and temporal characteristics, suggesting a practical direction for training-free PRVR.
The arithmetic merging of weights from multiple task-specific models fine-tuned from a shared base model, commonly known as model souping, has emerged as a popular strategy for multi-task learning without the computational overhead of joint retraining. However, naive scalar weight averaging frequently leads to severe performance degradation when the constituent models learn structurally divergent feature representations, leading to destructive parameter interference. We study this problem in a largely unexplored setting: merging image-to-image (I2I) diffusion models fine-tuned for dense, pixel-accurate tasks such as image restoration. Unlike the large language models (LLMs) and text-to-image (T2I) diffusion models for which most merging methods were developed, I2I experts encode conflicting spatial priors (e.g., anisotropic motion-blur vs. isotropic defocus-blur) in their convolutional filters, where both naive soup and LLM-style element-wise merging introduce visible artifacts. We introduce OptIMerge, a gradient-based optimization framework: instead of relying on heuristic scalars, OptIMerge learns highly expressive, yet parameter-efficient merger coefficients by minimizing a multi-task diffusion loss over a small set of 100-500 images. OptIMerge parameterizes the merge with architecture-aware transformations: for attention layers, a structured low-rank feature alignment with explicit channel biases; for convolutional layers, per-kernel merger coefficients, and a low-rank Cayley transform that aligns channels by orthogonal rotation without distorting the filters. By keeping the experts frozen and optimizing only these lightweight merger parameters (equivalent to <0.4% of the network footprint), OptIMerge achieves simultaneous multi-task capabilities that rival computationally expensive joint training and decisively outperform zero-shot model soups.
Xuan Wang, Zhuanhao Wu, Tingzhu Fang, Xushi Li, Sicheng Zhang, Zhenyu Yang, Feng Liu, Zhaohui Fan 等 20 位作者
In large language model (LLM) serving systems, the key-value (KV) cache grows linearly with sequence length and concurrency, so long-context inference easily hits the high-bandwidth memory (HBM) capacity wall and results in low concurrency. Offloading the KV cache to host dynamic random-access memory (DRAM) frees HBM space and raises concurrency, but the additional cache management and host-to-device (H2D) load overhead inflates inference latency, higher concurrency fails to translate into effective throughput. We present FlashOffloading, a host-DRAM-centric inference architecture that fully offloads the KV cache in both the prefill and decode phases. FlashOffloading builds rolling HBM cache for prefill, a hot cache for decode, a reverse sparse index with zero extra miss-lookup overhead, and a compute--load overlapped kernel, optimizing KV offloading latency across the entire pipeline and delivering throughput gains under a time-per-output-token (TPOT) service-level objective (SLO). In experiments deploying GLM-5.2 with multi-token prediction (MTP3) speculative decoding on an Ascend910C prefill--decode (PD)-disaggregated platform, FlashOffloading improves prefill throughput by up to 1.38\(\times\) and decode throughput by up to 1.91\(\times\) under a 30 ms TPOT constraint. FlashOffloading is integrated into an open-source serving framework: https://github.com/vllm-project/vllm-ascend/tree/releases/v0.30.0.
Ying Li, Yongle Cheng, Jing Li, Xingquan Cai, Yawei Luo
Spatial reasoning is a fundamental capability of multimodal large language models (MLLMs), yet existing benchmarks mainly focus on static scene understanding. When objects are moved or swapped, models must infer not only spatial relations but also how interventions change the scene state. To address this problem, we introduce CST-Bench, a benchmark for counterfactual spatial reasoning. CST-Bench contains 7,082 examples covering 6 intervention types and 12 transformation operators, with dense causal annotations including intervention sources, affected relations, expected effects, invariant relations, and metric changes. We benchmark general-purpose MLLMs on CST-Bench and further compare post-training strategies on CST-Bench and standard spatial benchmarks. Results show that methods effective for ordinary spatial reasoning can still fail under counterfactual interventions. We further formulate counterfactual spatial reasoning as causal state-transition modeling and propose CT-OPSD, which uses CST-Bench annotations to guide a teacher model in learning counterfactual state transitions and causal reasoning chains. These chains are then distilled into a student model that reasons only from real observations. Overall, experiment results establish CST-Bench as a challenging benchmark and reusable supervision resource for counterfactual spatial reasoning. CT-OPSD further provides an initial demonstration of how its annotations can support model improvement.
Recent conversational recommender systems have begun to unify recommendation and response generation within autoregressive models, with semantic IDs providing a natural representation for generative recommendation. However, relational knowledge is typically organized around item-level representations, which do not naturally align with the compositional SIDs. This mismatch makes it difficult for unified CRSs to effectively exploit structured catalog relations. We propose HGCRS, a hypergraph-enhanced generative conversational recommender system that bridges this gap with an SID--entity hypergraph. Each item forms a hyperedge connecting its SID codes and attribute entities, and a type-aware hypergraph encoder incorporates relational information into generative item selection. HGCRS further introduces a mode--item--response protocol for jointly modeling recommendation-bearing and response-only turns. Experiments on ReDial and INSPIRED demonstrate strong recommendation performance while maintaining competitive response generation quality.
Daniel Adejumo, Omkar Ankalkope, Kunal Runwal, Aliasghar Arab
We study action-conditioned monitoring of autonomous-vehicle behavior in semantically unusual driving scenes. The target is whether the executed ego action is appropriate for the scene: the same scene can admit both normal and anomalous behavior, so visual unusualness alone does not determine the monitoring label. A Vision-Language Model (VLM) first predicts the scene-required action from an outcome-hidden video window; the monitor then compares that expectation with the realized ego-motion trajectory under an explicit compatibility rule. To study this monitoring task, we contribute a six-field taxonomy of 112 paired scenarios spanning both failure polarities (42 derived from documented incidents), a generated video corpus with estimated trajectories, and an exploratory two-stage monitor on that corpus. On 235 clips from 15 scenario families, the two-stage protocol with a fixed T-2.5 s window obtains 83.0% accuracy and anomaly recall 0.773, versus 74.5% and 0.536 for a single-stage video-only baseline; specificity is 0.880 versus 0.928.