Attention
-
QK-Edit: Revisiting Attention-based Injection in MM-DiT for Image and Video Editing
2026-10-14 05:32:56Tiancheng Shen, Zilong Huang, Xiangtai Li, Zhijie Lin, Jiyang Liu, Yitong Wang, Jiashi Feng, Ming-Hsuan Yang, Jun Hao LiewAbstract
Multimodal Diffusion Transformers (MM-DiTs) have recently emerged as a powerful framework for unified text-vision synthesis, surpassing traditional U-Net architectures in generative tasks. One key innovation lies in its Multimodal Self-Attention (MM-SA) interaction, where image and text tokens are concatenated and processed via self-attention. However, this mechanism poses significant challenges for editing, rendering conventional U-Net-based attention manipulation methods ineffective. To address this limitation, we propose QK-Edit, a training-free framework that exploits the unique attention dynamics of MM-DiTs for precise text-guided image and video editing. By introducing a novel query-key manipulation strategy, our method isolates and adjusts critical attention components to achieve an optimal balance between prompt fidelity and structural consistency. This enables seamless edits across various tasks, including object addition, object removal, object replacement, changing background, changing material, changing color, and style transformation. Notably, it can be easily implemented with feature replacement in inference. QK-Edit demonstrates superior editing performance on state-of-the-art models, such as FLUX and HunyuanVideo, effectively bridging the gap between generative power and editable flexibility in MM-DiTs, and paving the way for scalable multimodal content creation.
Abstract (translated)
URL
https://arxiv.org/abs/2610.17829
PDF
-
PN-Mix: Pair-Norm Guided Mixed-Precision Key Cache Quantization for Efficient Long-Output LLM Inference
2026-10-14 05:29:08Gil Jae Song, Sungjoo YooAbstract
Long-output inference in large language models suffers from severe KV cache memory and bandwidth bottlenecks. While KV cache quantization alleviates this issue, low-bit quantization (e.g., 2-bit) severely degrades model accuracy. Recent mixed-precision key-cache quantization methods mitigate this degradation by allocating higher precision to important key channels, but estimate channel saliency using only key statistics. Through attention logit perturbation analysis at the RoPE channel-pair level, we show that quantization errors in channel pairs with large query--key pair-norm products contribute significantly to logit variation. Based on this insight, we propose PN-Mix, a mixed-precision key-cache quantization method for RoPE-based decoder-only LLMs, which allocates higher precision to salient channel pairs identified once during the prefill stage and reused throughout decoding. Across standard reasoning, code-generation, and AIME benchmarks, PN-Mix with 2.26-bit average precision achieves comparable or better average accuracy than prior mixed-precision methods using 2.32--2.39 bits. Furthermore, in our Pareto sweep, PN-Mix degrades gracefully down to 1.5 bits, whereas the compared baselines exhibit sharper performance drops below 2.25 bits.
Abstract (translated)
URL
https://arxiv.org/abs/2610.19605
PDF
-
PA1-OPD: One-Step Prefix-Augmented On-Policy Distillation for Language Model Pretraining
2026-10-14 05:23:00Zhenhe Wu, Jian Yang, Yaping Jin, Xinlai Kang, Jie Hu, Hanting ChenAbstract
Knowledge distillation (KD) improves language-model pretraining by matching teacher distributions at corpus prefixes, but provides no direct supervision at contexts induced by student predictions. We introduce PA1-OPD, which augments corpus-path distillation with teacher feedback after one student-sampled transition per selected prefix. This expands the supervised states while retaining next-token prediction and corpus-path KD. Data-anchored roots support parallel sampling, and differentiable tree attention shares prefix computation without changing inference. In two continued-pretraining settings, full coverage improves held-out cross-entropy over KD across C4, PG-19, and Pile at every evaluated checkpoint. In Stage2, selecting the 10% of roots with highest student surprisal adds 48.85% to KD's Macro cross-entropy gain over next-token training. Gains remain when branch supervision excludes successor states already on the corpus path, supporting the value of feedback beyond observed contexts. Additional depth or breadth improves aggregate results at unchanged per-term supervision weight, with benefits varying by corpus. These findings support local student-induced supervision as a complement to pretraining distillation.
Abstract (translated)
URL
https://arxiv.org/abs/2610.19601
PDF
-
RoboSpan: Learning Modular Coordination Policies for High-DoF Whole-Body Mobile Manipulation with Teleoperation
2026-10-14 05:10:48Keqiang Ren, Xuecheng Ye, Hanyu Liu, Zhihua Wang, Tao Jian, Yunsheng Ma, Jiayi Wen, Xinran Li, Zhigong SongAbstract
Whole-body mobile manipulation suffers from two fundamental challenges: inherent motion heterogeneity across locomotion, arm, and hand modules, and progressive error accumulation over long-horizon multi-stage tasks. Existing teleoperation platforms also lack integrated dexterous tactile sensing to support coordinated whole-body decision-making. This paper presents RoboSpan, a high-DoF whole-body teleoperation and demonstration platform, and the Hierarchical Modular Framework (HMF) for learning modular coordination policies. HMF functionally decouples the 29-DoF action space into fine hand manipulation, arm pose adjustment, and mobile-base spatial configuration branches, and introduces the Coordinated Modular Embodied Transformer (COMET) with inter-modular cross-attention to enable dynamic information exchange and action coordination. Vision-language model driven hierarchical task decomposition structurally suppresses temporal error propagation, while a full-palm 2D tactile representation encodes distributed contact signals for cross-module tactile empowerment. Experiments on a physical 29-DoF robot demonstrate that the proposed method achieves over 20\% higher task success rate than mainstream baselines across a range of long-horizon whole-body tasks, markedly improving stability and robustness in contact-rich manipulation.
Abstract (translated)
URL
https://arxiv.org/abs/2610.19594
PDF
-
Scaling Vision Transformers Without Training: Prune-and-Merge Token Compression
2026-10-14 04:42:32Vatsal, Soumil Malhotra, Priyanshu Chahal, Sanjeev KumarAbstract
Vision Transformers achieve strong performance but incur high computational costs from quadratic token interactions. Training-free token reduction methods improve efficiency, but their robustness remains underexplored. We propose Prune-and Merge ViT, a training-free framework for off-the-shelf models. It combines class attention-based pruning with similarity-based merging to reduce redundancy while preserving semantic information. To evaluate robustness we introduce two metrics. Relative Accuracy Drop (RAD) measures relative accuracy degradation under transformations, and Prediction Flip Rate (PFR) quantifies prediction consistency. On pretrained ViT and DeiT, Prune-and-Merge reduces computation competitively while maintaining robustness and better accuracy than existing token reduction methods. Our analysis suggests that hybrid token reduction does not introduce hallucination-like behavior and preserves prediction stability for common image transformations.
Abstract (translated)
URL
https://arxiv.org/abs/2610.18039
PDF
-
Speaking Longer Than Learned: Training-Free Long-Form Generation for Full-Stream Text-to-Speech
Abstract
Recent full-stream text-to-speech (TTS) models enable low-latency personalized synthesis as input text arrives. However, sustaining generation beyond the lengths seen during training without additional training remains largely unexplored. We systematically analyze VoXtream2, which extends generation by periodically discarding past context while retaining the acoustic prompt. We find that this strategy does not prevent degradation: pronunciation accuracy deteriorates near the end of the training range due to the range of positions in which the model operates rather than the loss of past context, while speaker similarity gradually declines throughout generation. Based on these findings, we propose two training-free modifications that reduce the attention window and strengthen attention to the acoustic prompt. Experiments show that these modifications preserve pronunciation accuracy and speaker similarity throughout long-form generation while retaining the naturalness of the original model. Ablations further show complementary effects of the two modifications.
Abstract (translated)
URL
https://arxiv.org/abs/2610.19553
PDF
-
Cross-Scenario Industrial Defect Detection via Domain-Robust Training and Metric-Aligned Selection
2026-10-14 03:57:52Dikai Li, Weijia Wang, Yuan-Gen WangAbstract
Automated defect detection of industrial boards must remain highly reliable under cross-scenario variations in lighting, surface contaminants, and production line conditions. This makes cross-scenario defect detection a domain generalization problem, wherein models are required to detect, localize, and classify defects across unseen scenarios while minimizing false alarms on clean boards. For the IDA 2026 Track 1 challenge, we propose a cross-scenario defect detection method that includes four complementary modules: cross-scenario robust training, small-defect-aware learning, score calibration and inference threshold selection, and metric-aligned checkpoint selection. These four modules respectively introduce cross-scenario robust normalization and DSU-style statistic perturbation with progressive random convolution, NWD and parameter-free spatial attention, the VFL learning objective, and a custom SWAD variant driven by the Track 1-oriented proxy. Experimental results demonstrate that our model achieves a Track 1 score of 0.564 with a screening score of 0.696 on the test dataset on a single RTX 4090 GPU. The code is available at this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2610.19540
PDF
-
Towards Efficient Simultaneous Inverse Text Normalization with Pretrained Text-to-Text Language Model and Read-Tag-Write Policy
2026-10-14 03:13:23Kiet Anh Hoang, Khanh Le, Bao Nguyen, Linh Pham, Dung Vo, Thai Tran, Mai Nguyen, Tri Nguyen, Vu LeAbstract
Streaming Inverse Text Normalization (ITN) is vital for converting spoken-form outputs from streaming Automatic Speech Recognition into formatted written text. Existing streaming ITN methods rely on hybrid systems combining neural tagging with expert-crafted finite-state transducer rules, limiting scalability across domains and languages. While end-to-end models offer superior scalability by learning directly from data, standard encoder-decoder architectures are inherently non-streaming due to global attention. In this paper, we propose an efficient streaming end-to-end ITN system, adapting a pretrained text-to-text model to leverage its robust linguistic knowledge. To enable streaming, we introduce architectural adaptations, a specialized training strategy, a Read-Tag-Write decoding policy, and inference optimizations. Experiments on Vietnamese datasets show accuracy comparable to non-streaming baselines, outperforming hybrid methods while satisfying real-time latency requirements.
Abstract (translated)
URL
https://arxiv.org/abs/2610.19492
PDF
-
WinARDet: Compressed Window Auto-Regression for Long-Horizon Vision-Language Grounding
2026-10-14 02:09:36Chenxu Wang, Hangyi Kuang, Bo-Wen Yin, Yuxuan Li, Quan-Sheng Zeng, Yang Zhang, Xu Guo, Xiang Li, Qibin HouAbstract
MLLM-based detectors formulate vision-language grounding as autoregressive coordinate generation, yet in dense scenes, they often collapse into a box repetition loop, repeatedly emitting identical boxes until the decoding budget is exhausted. We trace this failure to historical-answer attention takeover: Under full-history autoregression, repeated coordinate tokens in the growing answer history attract disproportionate attention, crowding out the attention sinks needed to attend to new visual evidence. % To break this loop, in this paper, we propose \textbf{Compressed Window Auto-Regression} (WinAR). This new mechanism partitions a long answer sequence into consecutive generation windows and restricts full token-level autoregression to the current window only while compressing each previous window into a fixed set of latent tokens that condition subsequent decoding. This design reduces the repeated exposure to raw historical coordinates that triggers the attention takeover, yet preserves cross-window dependencies essential for coherent multi-object generation. On RSDense, which contains 200-2,000 objects per image, our model WinARDet improves RexOmni F1-score (IoU=0.5) from 32.16 to 39.36 over the baseline, while reducing the redundancy rate from 35.48% to 13.00%. WinARDet also generalizes across decoding strategies and output formats, including parallel box decoding and polygon generation, providing a practical paradigm for long-horizon vision-language grounding across scene densities.
Abstract (translated)
URL
https://arxiv.org/abs/2610.19419
PDF
-
Do Syntax Heads Influence Model Behavior?
2026-10-13 21:47:26William Timkey, Tal LinzenAbstract
Transformer language models develop attention heads that specialize in syntactic relations, for example, attending from verbs to their subjects. Do these heads play a causal role in the model's syntactic predictions? Evidence for such a role is mixed to negative: even when we force these heads to attend to a distractor noun other than the subject, the model's syntactic predictions are still more consistent with the subject than with the distractor. Interventions on specialized syntax heads had a significantly stronger effect on syntactic predictions than interventions on random heads, but only in one of the two measures we propose.
Abstract (translated)
URL
https://arxiv.org/abs/2610.19231
PDF
-
Riding the Fixed-Point Curve: Image Classification in the Continuous Thought Machine
2026-10-13 21:36:28Thomas HannaganAbstract
The Continuous Thought Machine (CTM) is a recurrent network that attends to inputs and makes predictions based on the synchronization between its neurons. On tasks spanning many domains, it displays emergent thought processes that have not been explained. We report a number of findings about how the CTM classifies images, focusing on ImageNet. Each thinking step is an attracting but strongly non-normal map exhibiting balanced amplification (Murphy and Miller, 2009); while attraction is already present at CTM initialization, training highly promotes non-normality. By Banach's theorem, the neural state has a locally unique attracting fixed point, which moves at every step, and the thought trajectory is the curve these fixed points trace, followed one tick behind. Most of that curve's motion is shared across images, and visible as a previously reported traveling wave of activity across neurons. The attention schedule along the thought is largely generic: shuffling queries across images at every tick costs only 7 percent of classification accuracy, while image-independent increments predict the next query with R^2 = 0.98 (held out). Our results suggest that the CTM has found a single strategy to process all ImageNet images.
Abstract (translated)
URL
https://arxiv.org/abs/2610.19220
PDF
-
Titan-Llama: Retrofitting Pre-trained LLMs with External Neural Memory Modules
2026-10-13 21:23:58Sarah Pan, Ali BehrouzAbstract
Transformers are efficient to pre-train at scale because attention permits parallel computation across sequence positions, but long-context inference requires increasingly large KV caches. Efficient recurrent alternatives reduce the asymptotic complexity of inference, but their sequential dependence limits token-wise parallelism during training, making them harder to scale. Further, the current paradigm of extending long context tasks creates competing notions of retaining information precisely and keeping long-context inference efficient. To address this gap, we introduce Titan-Llama, a post-training method for retrofitting attention-based language models with a hierarchical memory architecture that updates parameters at test-time. Titan-Llama applies a blockwise attention mask and preserves full-fidelity attention over recent tokens while compressing historical context into recurrent, fixed-size neural memory modules (NMMs). Our formulation allows pre-trained attention layers to be replaced without training a new language model from scratch. On Llama-3.2-1B-Instruct, Titan-Llama raises accuracy over segmented attention windows from 23.87\% to 40.51\% on QuALITY and from 14.50\% to 40.75\% on LongHealth. Inference benchmarks show more than 20\% faster prefill, 37\% faster decode at medium context lengths, and 24\% lower steady-state decode memory usage in the evaluated configurations. Together, these results demonstrate that incorporating memory architecture mechanisms that utilize test-time training effectively modulates long-context behavior of pre-trained models, offering a practical route to adapting pre-trained transformers for efficient long-context inference. Our code is released on Github \href{this https URL}{GitHub}
Abstract (translated)
URL
https://arxiv.org/abs/2610.19210
PDF
-
Unlocking Structured Pruning for Reasoning Language Models via Attention-Preserving MLP Trimming
2026-10-13 19:55:33Woosang Lim, Haneul Yoo, Vincent-Daniel Yun, Sungjoo YooAbstract
Structured pruning reduces LLM inference costs by removing model components to produce smaller dense models. However, existing methods can cause reasoning performance to collapse on long-form reasoning tasks such as AIME. We propose \textit{LayerTrim}, a training-free structured pruning method that identifies redundant decoder layers and selectively trims their MLP features while preserving attention and decoder depth. LayerTrim is the first training-free structured pruning method designed to preserve long-form reasoning without gradients or recovery training. Across three reasoning language models and five benchmarks, LayerTrim outperforms existing structured pruning methods in most evaluated settings, improving average accuracy by up to 21.4 percentage points over RESP. On \texttt{Qwen3-14B}, LayerTrim achieves 1.24$\times$ faster prefill and 19.7\% lower peak GPU memory than the dense model, while better preserving reasoning ability. It also achieves the shortest average generation length among evaluated pruned models, while retaining the highest average accuracy. These results demonstrate that LayerTrim improves reasoning efficiency through both structural compression and reduced token generation.
Abstract (translated)
URL
https://arxiv.org/abs/2610.19121
PDF
-
MonoMIP: Multi-Height Inverse Perspective Mapping for Monocular 3D Object Detection
2026-10-13 19:06:51Emir Ulurak, Furkan K{\i}ra\c{c}Abstract
Monocular 3D object detection methods increasingly rely on depth-guided transformers, but they operate entirely in the perspective image plane where spatial relationships are distorted by foreshortening. We propose MonoMIP, a dual-decoder framework that adds complementary 3D geometric reasoning through Bird's-Eye View (BEV) features constructed via Multi-Height Inverse Perspective Mapping. We perform IPM at K elevation planes spanning ground to rooftop, producing an ordered stack of height-conditioned metric BEV features that reduces above-ground misprojection and exposes depth-dependent appearance variation across planes. A BEV decoder with multi-scale deformable attention aggregates this multi-height evidence per query, while a depth-guided 3D decoder provides complementary perspective-space predictions. Without extra training data, MonoMIP attains 25.84% Moderate AP3D|R40 on KITTI val and 21.66% on the official test set for Car, outperforming recent published monocular methods on both the KITTI and Waymo benchmarks.
Abstract (translated)
URL
https://arxiv.org/abs/2610.19071
PDF
-
Joint Subspace Sparse Attention for Efficient Long-Context Decoding
Abstract
Sparse attention reduces the KV-cache traffic that dominates long-context LLM decoding, but practical methods largely operate at coarse granularity. Token-level selection offers finer control, yet presents two challenges: low-dimensional scoring may lose query--key structure needed for accurate ranking, while gathering scattered selected tokens can offset the resulting latency benefit. We introduce Joint Subspace Sparse Attention (JSSA) to address both challenges. For selection quality, JSSA learns a shared low-dimensional query--key subspace by minimizing token-score preservation error. We upper-bound the resulting non-convex objective with a tractable marginal objective that admits a closed-form spectral solution; direct refinement of the exact objective provides no measurable improvement in the tested settings. For execution efficiency, we develop a gather-free token-level indexed-attention kernel that supports distinct per-KV-head selections under the standard grouped-query attention (GQA) KV-cache layout. Across standard GQA, hybrid sliding-window/global attention, and Multi-head Latent Attention (MLA) models, JSSA consistently remains close to full-KV accuracy while avoiding the substantial generation-length inflation and architecture-dependent failures observed with competing methods. On an NVIDIA B200 GPU, JSSA achieves up to $6.43\times$ the decoding throughput of full-KV attention at 128K context. JSSA achieves reasoning accuracy comparable to model-native DSA without model training and reaches 96--99\% of its decoding throughput even at 700B scale large models.
Abstract (translated)
URL
https://arxiv.org/abs/2610.19026
PDF
-
Tev: Scaling Long-Horizon Agentic Environments with Dense Progress Rewards
2026-10-13 17:59:51Wenbo Hu, Shraman Pramanick, Krishna Kumar Singh, Nanyun Peng, Kai-Wei Chang, Yong Jae LeeAbstract
Long-horizon agents have drawn growing attention for their ability to work autonomously -- from hours to days -- on economically valuable tasks, a capability essential to automated discovery and recursive self-improvement (RSI). Yet open-source models still struggle on long-horizon tasks, largely for lack of both training data and a working recipe, and scaling both is hard: long, complex tasks are difficult to curate and verify, and their rewards are sparse. We address both. First, we introduce Tev-Gym, the first long-horizon agent training dataset with verifiable, dense, progressive rewards -- 1,170 verified task environments across 7 domains. It is built on a latent-world-first approach that first constructs a latent world: a complete and correct hidden specification of the task domain, workflow, and expected state after each step. Each task is then progressively rendered from this specification into an agent-facing environment, enabling controllable complexity and a natural progress metric. Second, we turn this metric into a training recipe: via progress-potential reward shaping (PPRS), we convert progress into a dense per-step reward that improves training efficiency and makes otherwise-intractable tasks trainable. Training with this recipe yields Tev-27B, which achieves substantial gains on the Agents' Last Exam computing and math sciences domains (ALE-CM), matching GPT-6 Astra and Claude Opus 5 at Pass@5. Despite training only on ALE-related data, it generalizes to the Terminal-Bench (TB) 3.0 and 4.0 subsets, surpassing GLM 5.2 and Gemini 3.7 Flash, and reaches Claude Opus 4.8-level performance on middle-horizon tasks such as TB 2.1.
Abstract (translated)
URL
https://arxiv.org/abs/2610.18868
PDF
-
RadarGSOcc: Doppler-Aware 3D Semantic Occupancy Prediction Using 4D Radar Tensor with Geometric-Semantic Feature Extraction
2026-10-13 16:40:59Christofel Rio Goenawan, Dong-Hee Paek, Seung-Hyun KongAbstract
3D semantic occupancy prediction is crucial for autonomous driving, but camera-based methods degrade in low-light conditions and LiDAR-based methods are sensitive to adverse weather. 4D radar is more robust, yet its sparsity and the high cost of radar tensor processing limit its use. We propose Radar Geometric-Semantic Occupancy Prediction (RadarGSOcc), an efficient framework that directly processes 4D radar tensors by separately extracting geometric and semantic features, including Doppler information, with a deformable attention-based geometric refinement module as its key component. We also introduce the K-Radar 3DSemOcc benchmark, the first large-scale benchmark for 3D semantic occupancy prediction using 4D radar tensors, built on the K-Radar dataset with refined LiDAR-based annotations. RadarGSOcc predicts 15 classes within a 51.2m range and achieves 38.2% IoU, 7.6% mIoU, and 25.5% weighted mIoU, outperforming RadarOcc by 17% relative (25.5 vs. 21.9) in weighted mIoU while using fewer parameters and lower computation. It substantially improves truck prediction while trading off car accuracy, outperforms camera-based approaches, and narrows the gap with LiDAR-based methods. To assess robustness without per-frame ground truth in adverse weather, we further propose a novel proxy metric, Environment-Based KL Divergence (EBKL), under which 4D radar proves a reliable and efficient modality for 3D semantic occupancy prediction. The dataset and code will be publicly available.
Abstract (translated)
URL
https://arxiv.org/abs/2610.18686
PDF
-
From Relation-Specific Heads to Shared Interventions: Contextual Entrainment in Language Models
Abstract
Large language models rely on prompt context to perform in-context learning, follow formats, and adapt to local evidence. This ability, however, also raises a robustness concern: models may reuse irrelevant tokens from earlier context when answering a later query. Prior work identifies this behavior as contextual entrainment and finds attention heads whose ablation reduces it, arguing that these heads are largely task- or relation-specific. But are the mechanisms behind contextual entrainment truly isolated by relation, or do relation-conditioned searches recover reusable components shared across relations? Using factual-distractor relation-completion prompts across 19 lexical-semantic subtypes, we identify attention heads whose ablation reduces entrainment and test how the resulting head sets transfer recovery ability across exact relations, semantic families, and mixed relation distributions. Same-family transfer exceeds different-family transfer, family residuals weaken once globally shared heads are removed, and subtype-specific searches repeatedly rediscover many global heads. Clean-completion and context-use probes further suggest that this shared intervention can reduce harmful copying while preserving ordinary relation completion and simple in-context abilities. Our findings provide a more precise perspective on contextual entrainment and its mitigation: harmful context following is not best addressed by searching for isolated relation-specific circuits, but by using heterogeneous attack prompts to reveal reusable intervention targets whose effects are shaped by semantic relation structure.
Abstract (translated)
URL
https://arxiv.org/abs/2610.18674
PDF
-
Kwai Summary Attention V2 Technical Report
2026-10-13 16:18:44OneRec Team, Chenglong Chu, Guohong Mu, Guowang Zhang, Hao Peng, Hui Wang, Jian Liang, Jiangxia Cao, Jiao Ou, Jiaxin Deng, Lu Ren, Muhao Wei, Ruiming Tang, Tianle Xu, Wenwu Ou, Xingmei Wang, Zhixin Ling, Zhiyuan Liang, Ziqi Wang, Chaoyi Ma, Haixiao Gao, Kuo Cai, Renpu Liu, Ruochen Yang, Wenlong Xu, Xinqi Jin, Yan Sun, Yifan Wang, Yingzhi He, Yiyu Wang, Yusen Huo, Zhaojie Liu, Zihan ZhouAbstract
Summary-token architectures such as Kwai Summary Attention (KSA) compress distant context into learnable Summary token per chunk, and hybrid variants interleave these layers with full attention to retain exact access to the original text. In such hybrid KSA-Full architecture, the interleaving inserted summary tokens are served as a streaming flow role across different layers. Although effective, however, it also introduce some bottlenecks: (1) the different layers summary compression chunk must be same, therefore hard to reach a flexible KV Cache compression, e.g., different layer set different summary compression chunk window. (2) in the full-attention layers, the summary token introduced additional computation cost, but such layer ignore them and still pay quadratic prefilling, therefore unfriendly to the computation side. Motivated by them, we developed the KSA V2, which not only achieves a more flexible KV cache memory compression but also reduces the computation cost in Full attention layer. Specifically, we found that the summary tokens can serve two roles: the KV memory and sparse indexer.
Abstract (translated)
URL
https://arxiv.org/abs/2610.18654
PDF
-
AMemOps: Evaluating Memory Updates in Long Audio Streams
2026-10-13 16:16:19Peiwei Chang, Li Zhang, Mark D. Plumbley, Philip J. B. Jackson, Wenwu WangAbstract
Audio memory must recognise recurring sources across acoustic changes while keeping their attributes up to date. Overall accuracy does not reveal how these demands interact. We introduce Audio Memory Operations (AMemOps), a benchmark for latest-attribute retrieval, preservation of unchanged attributes and absent-source rejection in long audio streams. Paired controls vary recordings, training identities and earlier events to examine their effects on updating. Training on other recordings of the test instruments increases a reference delta-rule memory's update advantage over matched attention by 13.4 percentage points. At retrieval delays of 120\,s, matched histories separate benefits of consistent repetition from costs of conflicting earlier values, while additive-memory controls identify benefits of erasure.
Abstract (translated)
URL
https://arxiv.org/abs/2610.18652
PDF
