Inference
-
Unlocking Spatial Grounding in Flow-Matching Models via SNR Tuning
2026-10-21 11:22:58Martin Thi{\ss}en, Elke Hergenr\"otherAbstract
Despite the success of foundation models in image synthesis, achieving fine-grained spatial control remains a significant challenge, typically requiring auxiliary structural adapters or complex inference-time steering. While modern vision-language (VLM) text encoders possess the latent capacity to represent spatial coordinates, generative models often fail to leverage this information. We demonstrate that the signal-to-noise ratio (SNR) distribution is a critical determinant of spatial grounding in flow-matching models. Using a novel synthetic benchmark of geometric primitives, we isolate structural mechanics from textural complexity inherent in natural imagery and introduce a rigorous color-aware layout metric to quantify spatial fidelity. We show that training on a centered SNR distribution results in poor layout adherence (25.9%). Instead, we identify high-noise regimes as a "critical learning phase" for spatial grounding. Shifting the training focus toward these stages increases layout adherence to 80.8%, effectively reaching the 83.6% reconstruction ceiling of the underlying variational autoencoder. We find that the necessity for this shift scales with resolution and aspect ratio variance, and that the conservative noise schedules employed by models like FLUX.1 are insufficient for precise layout adherence. Finally, we show that these findings translate to real-world imagery via the COCO dataset (1.0 vs. 57.3 Average Precision), where state-of-the-art layout adherence is unlocked through distribution tuning alone, without architectural modifications.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29581
PDF
-
BiMoGen: Bidirectional Motion-Text Generation via Unified Masked Discrete Diffusion
2026-10-21 11:15:39Wanjiang Weng, Yongliang Wu, Xiaofeng Tan, Xingyu Zhu, Wenbo Zhu, Hongsong WangAbstract
Text-to-motion generation and motion-to-text captioning are two fundamental tasks in human motion modeling, both grounded in the same underlying motion-text correspondence. Existing unified approaches mostly rely on autoregressive modeling, which imposes a fixed generation order and is therefore poorly suited to the bidirectional dependencies between language and motion, allowing early prediction errors to persist as fixed context and degrade both temporal coherence and cross-modal consistency. Masked discrete diffusion, which models sequences through iterative bidirectional prediction, allows prediction from bidirectional context without a fixed generation order. We therefore propose BiMoGen (Bidirectional Motion-text Generation), a unified masked discrete diffusion framework for bidirectional motion-text modeling. To stabilize training, we design Decoupled Uni- and Cross-Modal Training, in which masked pretraining first establishes cross-modal correspondence on paired motion-text sequences, after which supervised fine-tuning specializes the model for bidirectional generation. Masked diffusion nonetheless faces a train-test mismatch, as the model is trained on clean ground-truth context yet encounters self-generated and potentially erroneous context at inference, with errors committed under heavily masked states propagating through subsequent steps. We further introduce Generation-Aware Self-Correction that exposes the model to its own predictions during training and uses correction passes at early sampling steps for further refinement of previously unmasked tokens. Extensive experiments on HumanML3D and KIT-ML demonstrate competitive performance on both tasks, validating the effectiveness of the proposed two-stage training and self-correction designs. The project page is available at this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2609.35407
PDF
-
PACE: Periodic and Channel-Aligned Encoding of Four-Bit Recurrent States
2026-10-21 10:31:52Patrick Haller, Filipe Laitenberger, Alan AkbikAbstract
Linear-attention language models keep their context in a recurrent state of fixed size, where an inference system stores one copy of this state for every sequence it decodes at the same time. Storing the state at lower precision therefore lets more sequences share an accelerator. Existing state quantizers round the state after every generated token and report that four bits lose accuracy unless the bit width is raised or the model is fine-tuned.\par We introduce PACE, a post-hoc four-bit state quantizer. PACE rounds the state only once every few generated tokens, as chunked kernels already do once per chunk during prefill, and keeps the tokens generated in between in a small buffer. Rounding less often reduces the error that accumulates in the state. Additionally, we find that the large entries of the gated-delta state gather in columns, because some channels of the update written at each token remain large across many tokens. PACE therefore shares each quantization scale within one column. It also keeps the heads that retain rounding error longest at eight bits. On four hybrid models from 4B to 35B parameters, PACE with mixed four- and eight-bit heads is competitive with eight-bit DAMP on the three larger models at about two thirds of its memory, PACE with four-bit heads throughout scores higher than Q-Mamba at equal memory, and rounding every few tokens preserves the decoding speedup of the buffer.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29545
PDF
-
RCSD: Reinforcement and Continual Learning for Domain-Specific Speculative Decoding
2026-10-21 10:16:23Mingkang Liu, Weiming Meng, Peipei Jiao, Jiawei Zhang, Tianyuan Liu, Junyan Li, Lichen PangAbstract
Speculative decoding accelerates large language model inference by employing a compact draft model to propose candidate tokens that are verified in parallel by a target model. Conventional training relies on soft-label distillation, which forces the draft model to mimic the target's full output distribution, including high-entropy regions that degrade acceptance rates and increase early rejection risk. We propose RCSD (Reinforcement and Continual Learning for Speculative Decoding), a training framework that aligns draft model optimization with the system-level objective of maximizing acceptance length. RCSD leverages naturally occurring accepted and rejected sequences during speculative inference to construct implicit preference pairs, with masked loss to avoid penalizing target-aligned tokens. Evaluated across multiple benchmarks, RCSD improves average acceptance length by 35% over HASS and achieves 1.35x faster end-to-end inference than HASS on practical workloads, without compromising generation quality.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29527
PDF
-
Beyond the Vote: When Do LLMs Add Value to Retrieval-Based Tabular Classification?
2026-10-21 10:06:29Ziquan Wang, Yuxuan Nie, Jian Ma, Yingjie LiAbstract
Retrieval-based in-context learning (ICL) enables large language models (LLMs) to predict tabular labels from similar examples, but it remains unclear whether LLM inference improves on direct voting over those same examples. We address this question by comparing four general-purpose LLMs with distance-weighted nearest-neighbor voting (Distance-KNN) using identical retrieved neighbors across 19 tabular classification datasets. With the number of demonstrations fixed, enlarging the retrieval candidate pool improves accuracy for all predictors, yet Distance-KNN achieves the highest mean accuracy at every pool size. More detailed analyses reveal that LLM gains are concentrated where retrieved neighbors exhibit low label consensus and vary across models. These gains further depend on the average feature-space distance to the test query: under low label consensus, some LLMs outperform Distance-KNN in close and intermediate neighborhoods, but all fall behind on average in the most distant group. Complementary experiments on the number of demonstrations, prompt formulation, retrieval representation, and label-balanced demonstration selection largely preserve the overall findings. These results characterize when LLM inference adds predictive value beyond distance-weighted voting in retrieval-based tabular ICL.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29520
PDF
-
Supervision-Adaptive Curriculum and Guarded Anti-Curriculum Learning for Dense Prediction
2026-10-21 10:06:22Chunming He, Dingming Zhang, Fengyang Xiao, Zhiwen Cao, Sina FarsiuAbstract
Curriculum learning orders samples by difficulty, but difficulty alone does not establish supervision reliability for dense prediction. Easy samples can expose shortcuts, hard samples can be informative, and noisy or ambiguous samples can mislead training. Building on our conference method CurriSeg, we present PaceShift, a stabilize-then-challenge paradigm combining supervision-adaptive curriculum selection with guarded spectral post-training. A learning-dynamics view motivates separating supervision reliability from temporal update value and applying spectral challenges after stabilization. The first phase estimates reliability from available supervision, augmentation consistency, model disagreement, and prediction certainty. Temporal statistics guide sample weighting, while pixel-uncertainty weights reduce the early influence of uncertain regions. The second phase ranks predictions from multiple spectral views using supervision agreement, robustness, reference consistency, and penalties. A pairwise preference objective measured against the frozen Phase-I reference, together with an anchor, refines the prediction. The framework supports full, weak, semi-supervised, and model-derived supervision. Experiments across context-entangled content segmentation (CECS), broader dense-prediction tasks, and foundation-model and cross-modal adaptations show accuracy gains. Targeted diagnostics indicate improved pooled calibration and corruption robustness in the evaluated settings. PaceShift preserves the deployed architecture and adds no inference computation. Code will be publicly released.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29519
PDF
-
BlazeVGGT: 13x Faster VGGT at 2,000 Frames via Camera Token without Training
2026-10-21 09:11:19Shujun Yao, Junhui Lin, Congying Han, Tiande Guo, Ziwen Liu, Bonan LiAbstract
Feed-forward visual geometry models, exemplified by VGGT, have demonstrated strong multi-view reconstruction capabilities through dense global attention. However, the resulting all-to-all cross-view computation scales quadratically with the number of input views, making long-sequence inference particularly expensive due to substantial frame-level redundancy among highly overlapping observations. We uncover a topology-first, geometry-later evolution in VGGT camera tokens: shallow-layer tokens reliably capture local view neighborhoods, whereas deeper-layer tokens progressively encode globally organized inter-view geometry. Motivated by this observation, we introduce BlazeVGGT, a training-free framework for structure-aware frame-level key--value sparsification. At each global-attention block, BlazeVGGT constructs a view-level similarity kernel from camera-token relations and employs greedy determinantal point process inference to select a compact set of mutually complementary views. The selected views constitute the key--value context, while all input frames are retained as queries. This asymmetric attention design preserves dense per-frame predictions while eliminating redundant cross-view computation, allowing our method to be directly integrated into pretrained VGGT as a plug-and-play module without parameter updates. Its advantages become increasingly pronounced with sequence length: at 2,000 frames, BlazeVGGT achieves approximately 13$\times$ faster inference than VGGT while reducing the Chamfer Distance from 0.41 to 0.32. Extensive experiments demonstrate a favorable accuracy--efficiency trade-off in point-cloud reconstruction and camera-pose estimation, compatibility with existing VGGT accelerators such as FastVGGT and LiteVGGT, and transferability to VGGT-based models including StreamVGGT and PAGE-4D. Codes are publicly available at this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29454
PDF
-
AdvFidelity: Unifying Preference Alignment and Few-Step Image Generation within Reinforcement Learning
2026-10-21 08:09:55Sicong Kang, Fangao Zeng, Yuhan Li, Hao Zhou, Mengfei Xu, Wenxiang Shang, Wei Li, Bingbing Ni, Pipei HuangAbstract
Practical image generation requires both human preference alignment and few-step efficiency, yet combining reinforcement learning (RL) with distillation is often multi-stage and may erode RL gains We observe that few-step RL progressively improves structural coherence, even with specialized rewards, but leaves low-level artifacts that these rewards may fail to penalize. To address this gap, we propose AdvFidelity, which integrates artifact-sensitive fidelity supervision into RL through an online adversarial reward. This enables joint reward alignment and high-fidelity few-step generation within a single RL stage, without distribution matching distillation (DMD) or other explicit distillation objectives. We train a discriminator to rank high-quality references (positives) above content-matched reconstructions (negatives) produced by the current generator. Reconstructing moderately noised references largely preserves style and scene layout while exposing low-level degradations. This encourages the discriminator to assess artifacts rather than content differences. With negatives reconstructed online, alternating updates adapt fidelity supervision to the generator's evolving failure modes. References can come from existing target-domain data or full-step model samples. Built on LeapAlign, AdvFidelity backpropagates gradients from both rewards through the few-step sampling trajectories used at inference. Across two model backbones and three task settings, AdvFidelity matches full-step LeapAlign in alignment, task performance, and overall generation quality and remains highly competitive with other evaluated full-step RL methods, using only 4 or 8 steps without CFG.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29380
PDF
-
Seeing through Denoising: Frequency-guided One-step Denoising for Infrared Small Target Detection
2026-10-21 07:51:11Xuelin Qian, Zixuan Wang, Chang Xu, Dingwen Zhang, Junwei HanAbstract
Infrared small target detection (IRSTD) is a challenging task due to the extremely small size of targets, low signal-to-noise ratios, and heavy background clutter in infrared imagery. While the progress of deep learning has advanced this field, existing methods primarily adopt discriminative paradigms that directly map images to target masks, overlooking the potential of generative modeling. In this paper, we propose DiffIRSTD, a novel generative framework that reformulates infrared small target detection as a background reconstruction problem. Instead of explicitly predicting target regions, we reinterpret small targets as noise-like signals and leverage a diffusion-based model to recover a clean, target-free background from a perturbed input image. The residual between the reconstructed background and the original image is then exploited to highlight candidate target regions, providing explicit cues for subsequent detection. To address the computational cost typically associated with diffusion sampling, we introduce a one-step denoising strategy that enables efficient background reconstruction while maintaining practical inference speed. Furthermore, we present a frequency-guided perception loss to preserve high-frequency background details and focus more on the reconstruction quality of target regions. More importantly, our DiffIRSTD operates as a plug-and-play framework, compatible with existing IRSTD models without increasing the computational complexity of detectors. Extensive experiments on IRSTD-1K, NUAA-SIRST, and NUDT-SIRST datasets demonstrate the superiority of our method, achieving state-of-the-art performance across three benchmarks.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29361
PDF
-
OpenMAP-HIE: Pathology Informed Score-based Diffusion Model for Unsupervised Neonatal Hypoxic-Ischemic Encephalopathy Lesion Segmentation
2026-10-21 07:22:50Pengyu Kan (Johns Hopkins University), Craig Jones (Johns Hopkins University, University of Alberta), Kenichi Oishi (Johns Hopkins University School of Medicine)Abstract
Accurate localization of neonatal hypoxic--ischemic encephalopathy (HIE) lesions on apparent diffusion coefficient (ADC) images is clinically important but is limited by scarce expert annotations and substantial variation in lesion appearance. We propose OpenMAP-HIE, a pathology-informed score-based diffusion model for unsupervised HIE lesion localization without voxel-wise lesion supervision. A denoising diffusion model learns the healthy neonatal ADC distribution, while a ControlNet-based conditional branch preserves patient-specific anatomy. At inference, anomaly evidence is derived directly from differences between conditional and unconditioned noise predictions at independently evaluated diffusion noise levels, avoiding dependence on a final pseudo-healthy reconstruction or sequential reverse-diffusion chain. Based on the characteristic ADC reduction associated with acute HIE, only discrepancies consistent with the expected pathological direction are retained and aggregated across noise levels using a geometric mean. Spatial mean filtering further suppresses isolated responses. The framework was evaluated on the BONBID-HIE dataset. The ADC-only method achieved a Dice score of $0.283\pm0.244$, average precision of $0.505$, Hausdorff distance of $25.45\pm15.63$ mm, mean average surface distance of $9.40\pm8.49$ mm, and normalized surface distance of $0.355\pm0.236$, providing the best overlap and surface-distance performance among the evaluated unsupervised methods and Dice performance comparable to supervised ADC-only models. Optional population-derived ADC $Z$-map refinement further increased Dice to $0.602\pm0.269$. These findings support pathology-informed score discrepancies as a useful signal for neonatal HIE lesion localization, while highlighting remaining challenges in detecting small and low-contrast lesions.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29322
PDF
-
Relational Geometry Forcing for World Action Models
2026-10-21 06:35:27Yuan Xu, Yixiang Chen, Qisen Ma, Jiabing Yang, Peiyan Li, Kai Wang, Jianhua Yang, Jianlou Si, Jun Huang, Jing Liu, Nianfeng Liu, Yan Huang, Liang WangAbstract
World Action Models (WAMs) have shown strong potential for robot manipulation by jointly learning visual dynamics and executable actions. However, their video and action objectives do not explicitly constrain the 3D geometry that is crucial for precise manipulation. Recently, geometry-aware WAMs introduce depth or 4D prediction branches, whereas representation alignment methods supervise intermediate WAM features with pretrained 3D foundation models. Yet we find that such pointwise alignment benefits the video features but degrades action learning, since dense geometric feature targets overly restrict compact action features. To address this limitation, we propose Relational Geometry Forcing (RGF), which transfers geometric priors from a frozen 3D foundation model by aligning pairwise feature relations instead of pointwise features. Specifically, RGF applies three complementary relational objectives, where intra-frame and cross-view relations transfer scene geometry and multi-view correspondence to the video features, while action-temporal relations couple action dynamics to geometric scene evolution. Applied only during training, RGF requires no explicit geometry branch or extra inference module. Extensive experiments in simulation and the real world demonstrate the effectiveness of Relational Geometry Forcing, showing consistent gains across diverse manipulation scenarios and WAM paradigms.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29282
PDF
-
PaceVLA: Adaptive Temporal Action Quantization in Vision-Language-Action Models
2026-10-21 05:55:50Hojin Jeon, Jimin Lee, Taekwan Kim, Nayoung Oh, Taeyoung Kim, Myungkyu Koo, Jinmo Koo, Joonwoo Ahn, Jinwoo ShinAbstract
Fast and reliable manipulation is crucial for practical robot deployment. Although Vision-Language-Action (VLA) models have recently shown remarkable progress as generalist robot policies, their execution often remains temporally inefficient. A key source of this inefficiency is that policies learn fine-grained temporal patterns from teleoperated demonstrations, requiring many low-level action steps. Temporal action quantization offers a simple way to accelerate execution by aggregating consecutive low-level actions into temporally coarser actions. However, applying action quantization without considering when fine-grained control is required introduces a speed-accuracy trade-off: accelerated actions may efficiently reduce redundant steps in some phases, while discarding critical motion details in precision-sensitive interactions. To address this, we propose PaceVLA, a unified framework for adaptive temporal execution in VLAs that integrates two coupled components: (1) mode-specific action decoders built on a shared action expert, supporting multiple temporal granularities and prediction horizons, and (2) a quantizability-guided mode selection scheme which decides whether coarse execution is appropriate, and a context-conditioned router that selects an execution mode at each inference step. Experiments across three multi-task simulation benchmarks and six real-world tasks show that PaceVLA improves execution efficiency while maintaining task success rate across multiple embodiments, including long-horizon and dexterous manipulation.
Abstract (translated)
URL
https://arxiv.org/abs/2610.25961
PDF
-
Learning from Fleet Experience: Collective Experience Learning Across 96 Robots
2026-10-21 05:52:57Chao Yu, Shu'ang Yu, Hao Lin, Hongzhi Zang, Yuxiang Yan, Chenxi Jiang, Zhen Guo, Chuankang Li, Yihao Chen, Xin Xu, Xu Fu, Jiakai Zhou, Liangzhi Shi, Hongyi Ren, Yiwei Liu, Jing Wu, Yuanqing Wang, Chen Wu, Bo Dai, Chunyu Yang, Hongliang Huang, Shuoyu Yue, Weiqi Yin, Yinuo Chen, Ruihua Chen, Chunyang Zhu, Han Xiao, Yanghao Li, Ruipu Wu, Zhikai Cai, Xiu Chen, Xu Tan, Si Liu, Yuchen Yan, Jialin Zhang, Wen Zeng, Yiming Zhou, Zixiao Huang, Xinlei Chen, Guanghui Ren, Quanlu Zhang, Boxun Li, Maoqing Yao, Yu WangAbstract
Robot deployment generates experience that can support policy improvement, but continued reliance on expert corrective actions makes learning difficult to scale across growing fleets. We propose learning from fleet experience: using deployed robots' own executions as a source of supervision for further policy improvement. We call this paradigm Collective Experience Learning (CEL), in which robots pool execution experience to update a shared policy that returns to the fleet. Following expert-supervised initialization, we realize CEL through failure-guided experience retrieval fine-tuning (FERT). FERT uses failure contexts to retrieve relevant segments from successful trajectories and prioritizes their recorded actions for supervised updates, without collecting additional expert demonstrations or corrective actions. Operators provide binary episode-outcome labels and environment resets. A fault-tolerant asynchronous training system (FATS) coordinates robot execution, cloud inference and policy optimization despite local interruptions. Across five real-world manipulation tasks, the resulting policies achieve success rates of 93.3-99.2%, evaluated over 120 trials per task on robots excluded from training data collection. With collection time per robot held fixed, larger fleets provide more experience and generally improve performance across the evaluated sizes. On multi-stage phone packaging, learning with 96 robots raises success from 79.2% to 98.3%. These results support CEL as a scalable approach to improving robot policy reliability through experience generated during deployment.
Abstract (translated)
URL
https://arxiv.org/abs/2610.25736
PDF
-
DisasterRoute: Evidence-Guided Dual-System Routing for Visual Disaster Question Answering
2026-10-21 05:41:44Tan Zhang, Xing Deng, Lu Zhang, Xiaofeng Zhu, Ping HuAbstract
Visual question answering from low-altitude UAV imagery must deliver timely and accurate judgments to support disaster response under limited edge computing resources. Routine questions may be answered through direct decisions, whereas more challenging questions may benefit from step-by-step reasoning. However, existing research primarily focuses on improving models' visual understanding and generative reasoning capabilities, with limited exploration of efficient direct decision-making and its adaptive coordination with generative reasoning. We develop Disaster-S1 based on a vision-language model adapted to the disaster domain. Trained for option discrimination, it outputs a probability distribution over answer options in a single multimodal forward pass for fast decision-making. We further propose DisasterRoute, a dual-system routing framework that combines prediction confidence with disaster evidence to decide whether to accept the initial answer or invoke the disaster reasoning model DisasterVL for review. By allocating inference computation on demand, the framework retains the efficiency of direct decision-making and improves overall question-answering accuracy through selective review. Experiments on DisasterBench show that Disaster-S1 achieves superior accuracy with low latency, while DisasterRoute further improves accuracy through selective review at substantially lower latency than full reasoning.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29231
PDF
-
DetM-Tracker: Deterministic and Memory-Efficient VIS for Autonomous Industrial Inspection
2026-10-21 05:09:38You-Rui Lin, Jen-Chieh Wu, Tzu-Ti Wei, Jen-Jee Chen, Yu-Chee TsengAbstract
Autonomous industrial inspection requires strictly deterministic real-time inference under tight memory and computational constraints. However, existing Video Instance Segmentation (VIS) models suffer from prohibitive scaling: dense temporal memory grows with spatial resolution and sequence length, while fixed-capacity decoders fail to exploit the inherent sparsity of industrial scenes, creating a practical memory wall on resource-constrained edge platforms. We present DetM-Tracker, a streamlined VIS framework for autonomous industrial inspection, where target objects often exhibit highly similar appearance under constrained "micro" fields of view. DetM-Tracker eliminates architectural redundancy through three synergistic components: (1) Compact Scalar Memory, which condenses high-dimensional instance features into four-scalar descriptors, reducing temporal memory complexity from $O(T \cdot N \cdot R)$ to $O(T \cdot \hat{N})$. This mechanism effectively compresses frame-level representations into a bank of active instances, where $R$ denotes resolution, $N$ and $\hat{N}$ represent global and window-specific instance counts, respectively and $T$ the temporal window size; and (2) Dynamic Instance Channels, which adaptively allocate query capacity according to the actual number of visible objects; and (3) Entropy-Gated Skip Connection, which concentrates computational budget on high-uncertainty spatial regions to refine object boundaries in a single forward pass without auxiliary refinement heads. Our approach effectively bridges the gap between modern VIS models and the practical demands of industrial edge deployment, reducing inference latency by approximately 20$\times$ and VRAM consumption by 6$\times$ compared to state-of-the-art dense-feature baselines.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29212
PDF
-
Adaptive Steering and Remasking for Safe Generation in Diffusion Language Models
2026-10-21 04:56:49Yejin Lee, Ungsik Kim, Yo-Sub HanAbstract
Diffusion Language Models(DLMs) provide a promising alternative to autoregressive language models through iterative denoising and bidirectional generation. However, their iterative generation process introduces distinct safety vulnerabilities because harmful content can emerge at arbitrary positions and persist across subsequent denoising steps. Existing defenses rely on fixed interventions or aggressive remasking, which limits adaptive control over denoising trajectories and can degrade generation quality. We propose an inference-time defense framework that combines adaptive safety steering with safety-aware remasking. Our method uses a gating direction to continuously adjust steering strength from the current denoising state and applies a steering direction to masked positions to guide subsequent predictions toward safer trajectories. Our method further employs a lightweight response detector after the first generation block to identify unsafe trajectories at an early stage. The detector triggers targeted remasking over generated content and part of the conditioning prompt, and the model regenerates the selected positions under adaptive safety steering. This design combines continuous trajectory control with explicit correction of unsafe content while requiring no modification of model parameters. Experiments on LLaDA and Dream demonstrate that our method improves robustness against diverse jailbreak attacks while preserving benign generation quality and general model capability. Our code is available at this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2605.13043
PDF
-
Two for One: Privileged Learning for Resource-Constrained Screening
Abstract
Machine learning systems are often developed under data-rich conditions but deployed when only incomplete information is available. A natural example arises in colour fundus photography (CFP), where both eyes are commonly imaged because they can provide complementary information for ocular disease assessment. Prior approaches that exploit bilateral information either fuse both eyes in a single model or use knowledge distillation to compress a dual-eye teacher into a lightweight student that still requires both images at test time. However, bilateral acquisition may not always be feasible. We thus propose a privileged learning framework in which a teacher model has access to bilateral CFP images during training and learns representations informed by information from both eyes, while the student is trained to transfer this knowledge from the bilateral teacher using only a single-eye image as input. At inference, the student operates exclusively on this single-eye input and does not require the contralateral image. Experimental results show consistent improvements in area under the receiver operating characteristic curve over conventional monocular training, with the largest gains observed for left-eye images for the detection of age-related macular degeneration (0.931 vs. 0.966) and diabetic retinopathy (0.789 vs. 0.836).%, while requiring only a single-eye image at inference. By enabling models to benefit from bilateral information during training while requiring only a single-eye image at inference, our approach may support more reliable deployment in resource-constrained clinical and screening environments.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29126
PDF
-
Agentic Video Understanding: A Survey
2026-10-21 02:49:27Xinyu Deng, Siwen Luo, Daochang LiuAbstract
As large language models (LLMs) become capable of processing increasingly diverse modalities and longer temporal contexts, an emerging line of work is moving beyond fixed video-language inference toward agentic systems that actively decide what information to inspect, retain, verify, and act upon. This survey reviews video understanding agents: systems that use video as the primary information source and solve understanding tasks through adaptive state construction and action selection. We first formalize an agent loop for video understanding, then address a central question: why do agents matter for video understanding? To answer this, we organize the literature through a challenge-to-design taxonomy, linking context bottlenecks to hierarchical evidence memory, evidence sparsity to active evidence acquisition, temporal causality to state and process tracking, and multimodal ambiguity to role-specialized coordination. We further review state space paradigms, learning paradigms, supervision signals, benchmarks, and evaluation protocols. Finally, we identify open directions toward agentic-native temporal modeling and video-native agents. Project page: this https URL
Abstract (translated)
URL
https://arxiv.org/abs/2609.31713
PDF
-
FlashNext-RL: A System Optimization Study for Long-Horizon Agentic RL of Qwen3.8-Flash-Next
2026-10-20 23:53:09Zhewei Yao, Yuxiong HeAbstract
Long-horizon agentic reinforcement learning (RL) requires both efficient rollout generation and training on long sequences. However, the memory requirements and model-specific operators of Qwen3.8-Flash-Next make this difficult with existing open-source implementations. To resolve these limitations, we introduce FlashNext-RL, a model-specific RL training optimization framework based on prime-rl, with vLLM for software engineering (SWE) rollout generation. Our main focus is BF16 training of the full 48-layer model on a single 8xB200 node, with the first two layers, the n-gram table, the sparse-attention indexer and the multi-token prediction head frozen. We optimize host offload, gradient accumulation and the kernels for mixture-of-experts, sparse attention and the gated residual stream. In the full-model development study, non-padding training throughput increases from approximately 395 tokens/s/GPU in the initial 32K-token configuration to 4,438 tokens/s/GPU in the optimized 131K-token configuration, with hardware matmul throughput (aka without considering the cost of attention components) increasing from approximately 20 to 213 TFLOPs/GPU. A controlled cumulative study on an 8-layer model further quantifies the contributions of individual optimizations. Together with three inference nodes, the system supports a live SWE RL run with a target batch of 1,024 trainable rollouts and a 131,072-token context limit. Over steps 5-15, an iteration takes 22.4 minutes on average, with bounded entropy and trainer-sampler mismatch KL. These results highlight the importance of tailoring RL training to the model's architecture, memory layout and kernel requirements, rather than relying only on general-purpose RL infrastructure.
Abstract (translated)
URL
https://arxiv.org/abs/2610.28998
PDF
-
Pair-Free Coarse-to-Real Video Generation via Real-to-Coarse Image Supervision
2026-10-20 22:09:29Gonzalo G\'omez-Nogales, Daohan Lu, Peiye Zhuang, Tinghui Zhou, Yi ZhouAbstract
Coarse-to-real (C2R) video generation converts coarse animation, simulation, or game-rendered videos into photorealistic videos while preserving scene layout, camera motion, and dynamics. Training such models requires aligned (coarse, realistic) video pairs, which are expensive to create and confine prior systems to narrow synthetic domains. We exploit a simple asymmetry: paired videos are hard to obtain, but paired real-to-coarse images are easy to construct with existing editing and 3D reconstruction tools. Using these image pairs, we train the DINO-R2C model to transfer realistic images to coarse images in DINO feature space, and apply it frame-wise to ordinary real videos to synthesize pseudo-coarse videos. The C2R# model then learns to generate each real video from its pseudo-coarse counterpart, free from needing (coarse, realistic) video pairs. At inference, the DINO-R2C model is discarded and actual coarse videos drive the C2R# model directly. Thus, our pair-free training scales with any real-video corpora rather than with paired synthetic videos. We will release the model and project webpage at this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2610.28954
PDF
