stackchain/troupe-standup
2
0 Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
Files
d862ccf4d28923acc7b826cc4f3d7ed99e697044
troupe-standup/wizard-work/sphinx/longctx-archive-2026-10-26/raw/q1_longctx_multimodal.xml
T

3337 lines
127 KiB
XML
Raw Blame History

<?xml version='1.0' encoding='UTF-8'?>
<feed xmlns:opensearch="http://a9.com/-/spec/opensearch/1.1/" xmlns:arxiv="http://arxiv.org/schemas/atom" xmlns="http://www.w3.org/2005/Atom">
<id>https://arxiv.org/api/Vacn1Dh4E1ffxoK1Pb2aqzji23k</id>
<title>arXiv Query: search_query=all:"long context" AND all:multimodal&amp;id_list=&amp;start=0&amp;max_results=30</title>
<updated>2026-10-26T01:03:22Z</updated>
<link href="https://arxiv.org/api/query?search_query=all:%22long+context%22+AND+all:multimodal&amp;start=0&amp;max_results=30&amp;id_list=" type="application/atom+xml"/>
<opensearch:itemsPerPage>30</opensearch:itemsPerPage>
<opensearch:totalResults>233</opensearch:totalResults>
<opensearch:startIndex>0</opensearch:startIndex>
<entry>
<id>http://arxiv.org/abs/2610.32483v1</id>
<title>When Agents Fail: Model or Harness--Who's to Blame?</title>
<updated>2026-10-23T10:57:54Z</updated>
<link href="https://arxiv.org/abs/2610.32483v1" rel="alternate" type="text/html"/>
<link href="https://arxiv.org/pdf/2610.32483v1" rel="related" type="application/pdf" title="pdf"/>
<summary>LLM agents combine models with execution harnesses, so agentic benchmark performance reflects their joint behavior. We argue that harnesses should therefore be evaluated through their interactions with underlying models. We develop a framework combining five complementary measures: (i) Valid Completion Rate (VCR), how reliably agents finish normally and deliver evaluable results, irrespective of task correctness; (ii) Model-Harness Fit, how well a harness supports task achievement through valid completion for a given model; (iii) Cross-Model Fit, a harness's overall suitability across evaluated models; (iv) Interaction Residuals, pairing-specific execution strengths and weaknesses hidden by component-level summaries; and (v) Token Load, the token demands of agent execution, including failed attempts. We evaluate four general-purpose harnesses (OpenCode, Codex CLI, Pi CLI, and DeepAgents) with four core models, plus two fixed native model-harness systems, on six benchmarks covering context management, tool use, long-context planning, multimodal interaction, and artifact handoff. Including native references and stress conditions, 111 evaluation settings yield 52,168 trajectories. Harness rankings vary across models and benchmarks; task performance and valid completion can favor different pairings, and similar task scores can accompany distinct execution profiles. OpenCode ranks first in cross-model performance, while DeepSeek V4 Flash Preview maintains consistently high Mean VCR across harnesses on the shared four-benchmark panel. These findings support evaluating model-harness interactions, training transferable harness-use capabilities in models, and harness designs adapted to the model and task.</summary>
<category term="cs.AI" scheme="http://arxiv.org/schemas/atom"/>
<category term="cs.CL" scheme="http://arxiv.org/schemas/atom"/>
<category term="cs.SE" scheme="http://arxiv.org/schemas/atom"/>
<published>2026-10-23T10:57:54Z</published>
<arxiv:comment>Preprint under review</arxiv:comment>
<arxiv:primary_category term="cs.AI"/>
<author>
<name>Shao Zhang</name>
</author>
<author>
<name>Hangfan Zhang</name>
</author>
<author>
<name>Yuanyuan Shi</name>
</author>
<author>
<name>Hao Li</name>
</author>
<author>
<name>Min-Le Su</name>
</author>
<author>
<name>Kangcong Li</name>
</author>
<author>
<name>Jingping Wang</name>
</author>
<author>
<name>Chunjiang Mu</name>
</author>
<author>
<name>Chen Zhang</name>
</author>
<author>
<name>Yang Chen</name>
</author>
<author>
<name>Bo Zhang</name>
</author>
<author>
<name>Shuyue Hu</name>
</author>
</entry>
<entry>
<id>http://arxiv.org/abs/2610.26431v1</id>
<title>Long-form Video-to-Text Summarization using Recursive Language Models</title>
<updated>2026-10-19T13:56:33Z</updated>
<link href="https://arxiv.org/abs/2610.26431v1" rel="alternate" type="text/html"/>
<link href="https://arxiv.org/pdf/2610.26431v1" rel="related" type="application/pdf" title="pdf"/>
<summary>While VideoLLMs have made significant progress in understanding short video benchmarks, they often struggle to summarize long-form ones. This is mainly due to performance degradation caused by long-context prompts, the inherent difficulty of preserving temporal continuity and a lack of semantically rich concepts in event descriptions. To tackle these issues, this paper introduces the first recursive Video Language Summarization Model (RVLSM) designed for video processing and tailored to long-form video-to-text summarization. RVLSM recursively decomposes videos into semantically coherent segments and delegates chunk-level reasoning to VideoLLM sub-models and external tools. It employs (1) Semantic Temporal Decomposition (STD), which preserves temporal coherence and continuity through a scene-aware chunking strategy that dynamically partitions videos according to semantic boundaries, and (2) Adaptive Multimodal Tool Calling (AMTC), which ensures semantic richness of the summary via a routing mechanism that selectively invokes auxiliary modules such as action recognition, optical character recognition, automatic speech recognition, and object detection. Experiments on the VideoXum and MixTyVSum benchmarks show that RVLSM consistently improves over zero-shot VideoLLM baselines across multiple model settings and performance metrics, and is competitive with a state-of-the-art fine-tuned model for video-to-text summarization. Importantly, the named entities in the RVLSM's summaries align more closely with those detected in the ground-truth summaries than those detected by baseline and state-of-the-art methods, confirming the saliency of the event descriptions it generates.</summary>
<category term="cs.CV" scheme="http://arxiv.org/schemas/atom"/>
<published>2026-10-19T13:56:33Z</published>
<arxiv:comment>Paper currently under review</arxiv:comment>
<arxiv:primary_category term="cs.CV"/>
<author>
<name>Alessia Intini</name>
</author>
<author>
<name>Lorenzo Vaiani</name>
</author>
<author>
<name>Enrico Chen</name>
</author>
<author>
<name>Luca Cagliero</name>
</author>
</entry>
<entry>
<id>http://arxiv.org/abs/2610.17901v1</id>
<title>FlashMask-3: Efficient and Expressive Mask-Aware Distributed Attention</title>
<updated>2026-10-13T07:55:51Z</updated>
<link href="https://arxiv.org/abs/2610.17901v1" rel="alternate" type="text/html"/>
<link href="https://arxiv.org/pdf/2610.17901v1" rel="related" type="application/pdf" title="pdf"/>
<summary>Long-context training of large language and multimodal models couples structured attention masks (document, prefix-LM, share-question, sliding-window, block-sparse) with multi-node context parallelism, making the mask a dominant bottleneck. FlashMask's column-wise sparse mask compresses them from $O(n^2)$ to $O(n)$ via four query-axis boundary vectors, covering broad structured patterns. Implemented for Ampere GPUs, it neither exploits FlashAttention-3's TMA/WGMMA warp-specialized pipeline on Hopper nor supports distribution. The column-wise sparse mask is indexed by key positions but stores query-axis interval boundaries. Under any multi-chunk context-parallel assignment, it reduces per-rank mask localization to a provably zero-communication \emph{clip--shift--max} on the boundary vectors, compatible with the single-GPU kernel. FlashMask-3 introduces a mask-aware Hopper kernel with three-role warp specialization atop FlashAttention-3's TMA/WGMMA pipeline, plus a mask-aware distributed extension: zero-communication mask localization, a sparsity-aware load balancer (HCC-IPO), and a mask-aware compute-communication overlap with mask-driven communication pruning, whose forward all-gather uses topology-aware hierarchical routing. On a single H100 across 12 masks at 4K--128K context, it improves achieved TFLOPs over FlashMask, FlexAttention, and MagiAttention by 45--141%, 34--71%, and 2--44%. At context-parallel degrees 4--32 with 8K sequence per device, it attains 31--75% compute-communication overlap efficiency and improves throughput over Megatron-Core CP ring, Megatron-Core CP all-gather, and MagiAttention by $1.80\times$--$3.71\times$, $1.71\times$--$2.92\times$, and $1.09\times$--$1.46\times$. In end-to-end training of a 118B-A14B MoE at 128K on 128 H100 GPUs, it raises throughput by 7.3--11.0% over a Dual Chunk all-gather baseline, with attention forward+backward 24.5--44.9% faster.</summary>
<category term="cs.LG" scheme="http://arxiv.org/schemas/atom"/>
<published>2026-10-13T07:55:51Z</published>
<arxiv:comment>Published as a main conference paper at NeurIPS 2026</arxiv:comment>
<arxiv:primary_category term="cs.LG"/>
<author>
<name>Guoxia Wang</name>
</author>
<author>
<name>Qianyue He</name>
</author>
<author>
<name>Siming Wu</name>
</author>
<author>
<name>Haoyang Xie</name>
</author>
<author>
<name>Qiwen Bao</name>
</author>
<author>
<name>Jinle Zeng</name>
</author>
<author>
<name>Jiabin Yang</name>
</author>
<author>
<name>Tian Wu</name>
</author>
<author>
<name>Dianhai Yu</name>
</author>
</entry>
<entry>
<id>http://arxiv.org/abs/2610.15542v1</id>
<title>Apertus 1.5: From Open Text Models to Open Multimodal Reasoners</title>
<updated>2026-10-11T11:34:43Z</updated>
<link href="https://arxiv.org/abs/2610.15542v1" rel="alternate" type="text/html"/>
<link href="https://arxiv.org/pdf/2610.15542v1" rel="related" type="application/pdf" title="pdf"/>
<summary>Language models are becoming infrastructure, and infrastructure built in the open, by public institutions, on documented data, is the alternative to dependence on a few providers; Apertus shows this is viable at scale. Apertus 1.5 extends the fully open Apertus family from text to multimodal reasoning. Continued pretraining of the Apertus 1.0 models, on 4T additional tokens for the 8B and 2T for the 70B, adds image and audio understanding and extends the context window fourfold to 262,144 tokens; multimodality enters through the tokenizer, the language-model head remains restricted to text, and the architecture is unchanged. Supervised finetuning, long-context finetuning, reinforcement learning with verifiable rewards, and preference alignment complete the released models. Thinking is a flag of the chat template rather than a separate model, and lifts mathematical reasoning by up to 20 points on the 8B model. The released models improve over their predecessors on nearly every benchmark; among fully open models of its class, the 8B leads on knowledge, instruction following, and cultural benchmarks, and its instruction following surpasses all compared models, including open-weight ones. On image and audio understanding, a first for the family, the models are competitive with fully open multimodal models, strongest on document understanding and long-form and lower-resource speech recognition, while specialized vision and audio models remain ahead on perception and transcription of English and other high-resource languages. As in Apertus 1.0, we train only on openly available data, respect crawler opt-outs retroactively, and exclude restrictively licensed material. We release the model weights, data preparation and training recipes, and evaluation suites under permissive licenses; openness extends to values, with the Apertus charter guiding alignment stages published alongside the models.</summary>
<category term="cs.CL" scheme="http://arxiv.org/schemas/atom"/>
<published>2026-10-11T11:34:43Z</published>
<arxiv:primary_category term="cs.CL"/>
<author>
<name>Project Apertus</name>
<arxiv:affiliation>ETH Zurich</arxiv:affiliation>
</author>
<author>
<name> :</name>
<arxiv:affiliation>ETH Zurich</arxiv:affiliation>
</author>
<author>
<name>Yixuan Xu</name>
<arxiv:affiliation>ETH Zurich</arxiv:affiliation>
</author>
<author>
<name>Yassine Turki</name>
<arxiv:affiliation>EPFL</arxiv:affiliation>
</author>
<author>
<name>Vinko Sabolčec</name>
<arxiv:affiliation>EPFL</arxiv:affiliation>
</author>
<author>
<name>Valentina Pyatkin</name>
<arxiv:affiliation>ETH Zurich</arxiv:affiliation>
</author>
<author>
<name>Tommy Chu</name>
<arxiv:affiliation>EPFL</arxiv:affiliation>
</author>
<author>
<name>Sven Najem-Meyer</name>
<arxiv:affiliation>EPFL</arxiv:affiliation>
</author>
<author>
<name>Sean F. Carroll</name>
<arxiv:affiliation>EPFL</arxiv:affiliation>
</author>
<author>
<name>Robert Smith</name>
<arxiv:affiliation>ETH Zurich</arxiv:affiliation>
</author>
<author>
<name>Raphael Kreft</name>
<arxiv:affiliation>ETH Zurich</arxiv:affiliation>
</author>
<author>
<name>Oscar Alessandro Tazza</name>
<arxiv:affiliation>ETH Zurich</arxiv:affiliation>
</author>
<author>
<name>Oleg Lavrovsky</name>
<arxiv:affiliation>EPFL</arxiv:affiliation>
</author>
<author>
<name>Nicola Irmiger</name>
<arxiv:affiliation>ETH Zurich</arxiv:affiliation>
</author>
<author>
<name>Matteo Santelmo</name>
<arxiv:affiliation>EPFL</arxiv:affiliation>
</author>
<author>
<name>Marian Schneider</name>
<arxiv:affiliation>ETH Zurich</arxiv:affiliation>
</author>
<author>
<name>Juan García Giraldo</name>
<arxiv:affiliation>EPFL</arxiv:affiliation>
</author>
<author>
<name>Hanna Yukhymenko</name>
<arxiv:affiliation>ETH Zurich</arxiv:affiliation>
</author>
<author>
<name>Eduard Ďurech</name>
<arxiv:affiliation>ETH Zurich</arxiv:affiliation>
</author>
<author>
<name>Davit Melikidze</name>
<arxiv:affiliation>ETH Zurich</arxiv:affiliation>
</author>
<author>
<name>Dani Tamayo</name>
<arxiv:affiliation>EPFL</arxiv:affiliation>
</author>
<author>
<name>Bettina Messmer</name>
<arxiv:affiliation>EPFL</arxiv:affiliation>
</author>
<author>
<name>Aryan Ahadinia</name>
<arxiv:affiliation>EPFL</arxiv:affiliation>
</author>
<author>
<name>Anunay Yadav</name>
<arxiv:affiliation>EPFL</arxiv:affiliation>
</author>
<author>
<name>Allen Huang</name>
<arxiv:affiliation>EPFL</arxiv:affiliation>
</author>
<author>
<name>Ahmad Fraij</name>
<arxiv:affiliation>ETH Zurich</arxiv:affiliation>
</author>
<author>
<name>Alessandro Bombarda</name>
<arxiv:affiliation>EPFL</arxiv:affiliation>
</author>
<author>
<name>Alex Padula</name>
<arxiv:affiliation>ETH Zurich</arxiv:affiliation>
</author>
<author>
<name>Alexander Hägele</name>
<arxiv:affiliation>EPFL</arxiv:affiliation>
</author>
<author>
<name>Alexander Sternfeld</name>
<arxiv:affiliation>HES-SO Valais-Wallis</arxiv:affiliation>
</author>
<author>
<name>Anna Hedström</name>
<arxiv:affiliation>ETH Zurich</arxiv:affiliation>
</author>
<author>
<name>Anna Kosovskaia</name>
<arxiv:affiliation>ETH Zurich</arxiv:affiliation>
</author>
<author>
<name>Arthur Wuhrmann</name>
<arxiv:affiliation>EPFL</arxiv:affiliation>
</author>
<author>
<name>Arundhati Balasubramaniam</name>
<arxiv:affiliation>ETH Zurich</arxiv:affiliation>
</author>
<author>
<name>Badr Al Mahouri</name>
<arxiv:affiliation>EPFL</arxiv:affiliation>
</author>
<author>
<name>Barna Pasztor</name>
<arxiv:affiliation>ETH Zurich</arxiv:affiliation>
</author>
<author>
<name>Bartosz Szostakiewicz</name>
<arxiv:affiliation>ETH Zurich</arxiv:affiliation>
</author>
<author>
<name>Bich Ngoc Doan</name>
<arxiv:affiliation>EPFL</arxiv:affiliation>
</author>
<author>
<name>Bryan Schmid</name>
<arxiv:affiliation>FHNW</arxiv:affiliation>
</author>
<author>
<name>Burak Sezen</name>
<arxiv:affiliation>ETH Zurich</arxiv:affiliation>
</author>
<author>
<name>Camille Challier</name>
<arxiv:affiliation>EPFL</arxiv:affiliation>
</author>
<author>
<name>Clément Charmillot</name>
<arxiv:affiliation>EPFL</arxiv:affiliation>
</author>
<author>
<name>Dana Shayakhmetova</name>
<arxiv:affiliation>EPFL</arxiv:affiliation>
</author>
<author>
<name>Daniel Perruchoud</name>
<arxiv:affiliation>FHNW</arxiv:affiliation>
</author>
<author>
<name>David Brüggemann</name>
<arxiv:affiliation>SDSC</arxiv:affiliation>
</author>
<author>
<name>David Isztl</name>
<arxiv:affiliation>CSCS</arxiv:affiliation>
</author>
<author>
<name>Elia Palme</name>
<arxiv:affiliation>CSCS</arxiv:affiliation>
</author>
<author>
<name>Emma Anastassova</name>
<arxiv:affiliation>EPFL</arxiv:affiliation>
</author>
<author>
<name>Fabian Bösch</name>
<arxiv:affiliation>CSCS</arxiv:affiliation>
</author>
<author>
<name>Faruk Zahiragić</name>
<arxiv:affiliation>EPFL</arxiv:affiliation>
</author>
<author>
<name>Fawzi Roberto Mohamed</name>
<arxiv:affiliation>CSCS</arxiv:affiliation>
</author>
<author>
<name>Fiona Panicali</name>
<arxiv:affiliation>FHNW</arxiv:affiliation>
</author>
<author>
<name>Florian Kolly</name>
<arxiv:affiliation>EPFL</arxiv:affiliation>
</author>
<author>
<name>Flynn Guerra</name>
<arxiv:affiliation>FHNW</arxiv:affiliation>
</author>
<author>
<name>Francesco Monti</name>
<arxiv:affiliation>EPFL</arxiv:affiliation>
</author>
<author>
<name>Frederike Lübeck</name>
<arxiv:affiliation>ETH Zurich</arxiv:affiliation>
</author>
<author>
<name>Gennaro Oliva</name>
<arxiv:affiliation>CSCS</arxiv:affiliation>
</author>
<author>
<name>Haofeng Yang</name>
<arxiv:affiliation>ETH Zurich</arxiv:affiliation>
</author>
<author>
<name>Haoliang Shang</name>
<arxiv:affiliation>ETH Zurich</arxiv:affiliation>
</author>
<author>
<name>Helen El Khoury</name>
<arxiv:affiliation>EPFL</arxiv:affiliation>
</author>
<author>
<name>Henrique Mendonça</name>
<arxiv:affiliation>CSCS</arxiv:affiliation>
</author>
<author>
<name>Jannis Vamvas</name>
<arxiv:affiliation>University of Zurich</arxiv:affiliation>
</author>
<author>
<name>Jérome Landolt</name>
<arxiv:affiliation>FHNW</arxiv:affiliation>
</author>
<author>
<name>Jessica Lam</name>
<arxiv:affiliation>ETH Zurich</arxiv:affiliation>
</author>
<author>
<name>Jingchun Xiao</name>
<arxiv:affiliation>ETH Zurich</arxiv:affiliation>
</author>
<author>
<name>Jonathan Coles</name>
<arxiv:affiliation>CSCS</arxiv:affiliation>
</author>
<author>
<name>Joshua Oyebanji</name>
<arxiv:affiliation>EPFL</arxiv:affiliation>
</author>
<author>
<name>Khanh Nguyen</name>
<arxiv:affiliation>EPFL</arxiv:affiliation>
</author>
<author>
<name>Konstantinos Pitas</name>
<arxiv:affiliation>ETH Zurich</arxiv:affiliation>
</author>
<author>
<name>Lasse B. Strand</name>
<arxiv:affiliation>ETH Zurich</arxiv:affiliation>
</author>
<author>
<name>Li Mi</name>
<arxiv:affiliation>ETH Zurich</arxiv:affiliation>
</author>
<author>
<name>Loïc Deslarzes</name>
<arxiv:affiliation>EPFL</arxiv:affiliation>
</author>
<author>
<name>Lorenzo Baggi</name>
<arxiv:affiliation>ETH Zurich</arxiv:affiliation>
</author>
<author>
<name>Lorenzo Noci</name>
<arxiv:affiliation>ETH Zurich</arxiv:affiliation>
</author>
<author>
<name>Luca Baumann</name>
<arxiv:affiliation>ETH Zurich</arxiv:affiliation>
</author>
<author>
<name>Luca Sartori</name>
<arxiv:affiliation>ETH Zurich</arxiv:affiliation>
</author>
<author>
<name>Lukas Drescher</name>
<arxiv:affiliation>CSCS</arxiv:affiliation>
</author>
<author>
<name>Marco De Negri</name>
<arxiv:affiliation>ETH Zurich</arxiv:affiliation>
</author>
<author>
<name>Marco Passerini</name>
<arxiv:affiliation>CSCS</arxiv:affiliation>
</author>
<author>
<name>Mark Klein</name>
<arxiv:affiliation>CSCS</arxiv:affiliation>
</author>
<author>
<name>Marko Stojanovic</name>
<arxiv:affiliation>EPFL</arxiv:affiliation>
</author>
<author>
<name>Martin Majsec</name>
<arxiv:affiliation>EPFL</arxiv:affiliation>
</author>
<author>
<name>Matin Ansaripour</name>
<arxiv:affiliation>EPFL</arxiv:affiliation>
</author>
<author>
<name>Mattea Busato</name>
<arxiv:affiliation>EPFL</arxiv:affiliation>
</author>
<author>
<name>Mauro Pellonara</name>
<arxiv:affiliation>EPFL</arxiv:affiliation>
</author>
<author>
<name>Max Neuwinger</name>
<arxiv:affiliation>ETH Zurich</arxiv:affiliation>
</author>
<author>
<name>Michael Graber</name>
<arxiv:affiliation>FHNW</arxiv:affiliation>
</author>
<author>
<name>Michal Tešnar</name>
<arxiv:affiliation>ETH Zurich</arxiv:affiliation>
</author>
<author>
<name>Michelle Wastl</name>
<arxiv:affiliation>University of Zurich</arxiv:affiliation>
</author>
<author>
<name>Miguel Gila</name>
<arxiv:affiliation>CSCS</arxiv:affiliation>
</author>
<author>
<name>Mikolaj Boronski</name>
<arxiv:affiliation>EPFL</arxiv:affiliation>
</author>
<author>
<name>Morgane Magnin</name>
<arxiv:affiliation>EPFL</arxiv:affiliation>
</author>
<author>
<name>Moritz Reihs</name>
<arxiv:affiliation>ETH Zurich</arxiv:affiliation>
</author>
<author>
<name>Mukhammadali Sayfiddinov</name>
<arxiv:affiliation>ETH Zurich</arxiv:affiliation>
</author>
<author>
<name>Nathan Ranchin</name>
<arxiv:affiliation>ETH Zurich</arxiv:affiliation>
<arxiv:affiliation>EPFL</arxiv:affiliation>
</author>
<author>
<name>Nicolas Cuveillier</name>
<arxiv:affiliation>EPFL</arxiv:affiliation>
</author>
<author>
<name>Oscar Wohlfahrt</name>
<arxiv:affiliation>EPFL</arxiv:affiliation>
</author>
<author>
<name>Qianqian Zou</name>
<arxiv:affiliation>SDSC</arxiv:affiliation>
</author>
<author>
<name>Rada Kamysheva</name>
<arxiv:affiliation>ETH Zurich</arxiv:affiliation>
</author>
<author>
<name>Rita Sevastjanova</name>
<arxiv:affiliation>ETH Zurich</arxiv:affiliation>
</author>
<author>
<name>Roman Machacek</name>
<arxiv:affiliation>University of Bern</arxiv:affiliation>
</author>
<author>
<name>Sebastian Keller</name>
<arxiv:affiliation>CSCS</arxiv:affiliation>
</author>
<author>
<name>Simon Matrenok</name>
<arxiv:affiliation>EPFL</arxiv:affiliation>
</author>
<author>
<name>Simon Pierre</name>
<arxiv:affiliation>EPFL</arxiv:affiliation>
</author>
<author>
<name>Stefano Schuppli</name>
<arxiv:affiliation>CSCS</arxiv:affiliation>
</author>
<author>
<name>Suei-Wen Chen</name>
<arxiv:affiliation>ETH Zurich</arxiv:affiliation>
</author>
<author>
<name>Szabina Horváth-Mikulas</name>
<arxiv:affiliation>EPFL</arxiv:affiliation>
</author>
<author>
<name>Taliesin Perez</name>
<arxiv:affiliation>EPFL</arxiv:affiliation>
</author>
<author>
<name>Thilo Spinner</name>
<arxiv:affiliation>ETH Zurich</arxiv:affiliation>
</author>
<author>
<name>Thomas Kleine Buening</name>
<arxiv:affiliation>ETH Zurich</arxiv:affiliation>
</author>
<author>
<name>Tim Robinson</name>
<arxiv:affiliation>CSCS</arxiv:affiliation>
</author>
<author>
<name>Tymoteusz Kwiecinski</name>
<arxiv:affiliation>ETH Zurich</arxiv:affiliation>
</author>
<author>
<name>Vilém Zouhar</name>
<arxiv:affiliation>ETH Zurich</arxiv:affiliation>
</author>
<author>
<name>Vinzenz Thoma</name>
<arxiv:affiliation>ETH Zurich</arxiv:affiliation>
</author>
<author>
<name>Xiaozhe Yao</name>
<arxiv:affiliation>ETH Zurich</arxiv:affiliation>
</author>
<author>
<name>Yannick Metz</name>
<arxiv:affiliation>ETH Zurich</arxiv:affiliation>
</author>
<author>
<name>Yuhei Fukuhara</name>
<arxiv:affiliation>EPFL</arxiv:affiliation>
</author>
<author>
<name>Yusif Askari</name>
<arxiv:affiliation>EPFL</arxiv:affiliation>
</author>
<author>
<name>Zixuan Chen</name>
<arxiv:affiliation>ETH Zurich</arxiv:affiliation>
</author>
<author>
<name>Alexander Ilic</name>
<arxiv:affiliation>ETH Zurich</arxiv:affiliation>
</author>
<author>
<name>Ana Klimovic</name>
<arxiv:affiliation>ETH Zurich</arxiv:affiliation>
</author>
<author>
<name>Andreas Krause</name>
<arxiv:affiliation>ETH Zurich</arxiv:affiliation>
</author>
<author>
<name>Andrei Kucharavy</name>
<arxiv:affiliation>HES-SO Valais-Wallis</arxiv:affiliation>
</author>
<author>
<name>Elliott Ash</name>
<arxiv:affiliation>ETH Zurich</arxiv:affiliation>
</author>
<author>
<name>Joost VandeVondele</name>
<arxiv:affiliation>CSCS</arxiv:affiliation>
</author>
<author>
<name>Martin Rajman</name>
<arxiv:affiliation>EPFL</arxiv:affiliation>
</author>
<author>
<name>Mennatallah El-Assady</name>
<arxiv:affiliation>ETH Zurich</arxiv:affiliation>
</author>
<author>
<name>Thomas Schulthess</name>
<arxiv:affiliation>CSCS</arxiv:affiliation>
</author>
<author>
<name>Timo Kehrer</name>
<arxiv:affiliation>University of Bern</arxiv:affiliation>
</author>
<author>
<name>Torsten Hoefler</name>
<arxiv:affiliation>ETH Zurich</arxiv:affiliation>
</author>
<author>
<name>Antoine Bosselut</name>
<arxiv:affiliation>EPFL</arxiv:affiliation>
</author>
<author>
<name>Martin Jaggi</name>
<arxiv:affiliation>EPFL</arxiv:affiliation>
</author>
<author>
<name>Imanol Schlag</name>
<arxiv:affiliation>ETH Zurich</arxiv:affiliation>
</author>
</entry>
<entry>
<id>http://arxiv.org/abs/2610.11924v1</id>
<title>Revisiting Identity and Spectra Dispersion in Media-Bridged Time Series Forecasting: Linking Multivariate Signals and Narrative Flows</title>
<updated>2026-10-08T13:20:49Z</updated>
<link href="https://arxiv.org/abs/2610.11924v1" rel="alternate" type="text/html"/>
<link href="https://arxiv.org/pdf/2610.11924v1" rel="related" type="application/pdf" title="pdf"/>
<summary>Media-bridged time series forecasting is expanding to encompass traditional "multivariate" and emerging "multimodal" (e.g., through textual assistance). Existing Time Series Forecasting (TSF) models still rely on paradigm-specific relation, fusion, and temporal modules, hindering a common forecasting backbone across numerical and pre-aligned narrative-flow settings. To explore this, we propose the Multimedia Identity-Aware Prism Network (MIDAPN), a unified spatiotemporal forecasting backbone based on media-general graph adaptation and automatic temporal learning: (1) Following media pre-alignment, our Multimedia Identity-Aware Graph (MIDAG) revisits identity through static essence, dynamic behavior, and latent commonality, inducing affinities that extend variable-specific dependencies across media. Contextual Identity Modulation (CIM) further refines discriminative aggregation. (2) We develop Spectral Prism Convolution (SPConv) to automatically perform hierarchical temporal analysis, balancing coarse trends and fine-grained details. Meanwhile, its Adaptive Search Guidance configures a scale-efficient architecture for temporal-dimension reconstruction. These decoupled yet synergistic components jointly address media identity disentanglement and temporal-scale mismatch. Comprehensive evaluations involving 16 SOTA TSF models across 13 "multivariate" and 12 "multimodal" datasets, alongside targeted long-context comparisons against 14 time series foundation models and fused pretrained language models, demonstrate MIDAPN's consistent superiority and broad shared backbone compatibility. The code is available at \href{https://github.com/leijieruilq/MIDAPN/tree/main}{https://github.com/MIDAPN}.</summary>
<category term="cs.CV" scheme="http://arxiv.org/schemas/atom"/>
<category term="cs.LG" scheme="http://arxiv.org/schemas/atom"/>
<published>2026-10-08T13:20:49Z</published>
<arxiv:primary_category term="cs.CV"/>
<author>
<name>Jierui Lei</name>
</author>
<author>
<name>Wenjian Zhang</name>
</author>
<author>
<name>Qingyi Yang</name>
</author>
<author>
<name>Yuyang Hong</name>
</author>
<author>
<name>Fangzheng Chen</name>
</author>
<author>
<name>Zhengbo Zhang</name>
</author>
<author>
<name>Haina Tang</name>
</author>
<author>
<name>Shiming Xiang</name>
</author>
</entry>
<entry>
<id>http://arxiv.org/abs/2610.12517v1</id>
<title>TerraAgent: Demonstration-Grounded Multi-Agent Reasoning for Long-Context Remote Sensing Visual Question Answering</title>
<updated>2026-10-07T19:50:09Z</updated>
<link href="https://arxiv.org/abs/2610.12517v1" rel="alternate" type="text/html"/>
<link href="https://arxiv.org/pdf/2610.12517v1" rel="related" type="application/pdf" title="pdf"/>
<summary>Remote sensing visual question answering (RS-VQA) requires answers that truly depend on image evidence, yet current remote sensing multimodal large language models (RS MLLMs) are still evaluated mainly by accuracy, which cannot reveal whether a correct answer was produced without using the image. This setting introduces three challenges: demonstrations retrieved across different ground sample distances (GSD), sensors, and city textures may be poorly matched; real queries often involve long visual contexts from multi-scale tiles, multi-temporal frames, and retrieved knowledge, where evidence can be diluted and errors can propagate through multi-step reasoning; and prior-only answering remains invisible to standard metrics. We present AVA, a training-free framework that wraps a frozen RS MLLM with three modules. GDR performs GSD-bucketed, scene-graph-based demonstration retrieval and filters demonstrations whose answers are unsupported by their images. MACE decomposes each query into an evidence-typed sub-question DAG, routes sub-questions to suitable perception tools, and records verified provenance in a shared evidence ledger with bounded re-perception. CVA audits the final answer through counterfactual perturbations of cited evidence, including ablation and swap, and computes a visual-dependency score for confidence calibration and evidence-forcing re-query. On RSVQA-LR/HR, VRSBench, and LHRS-Bench, AVA improves average accuracy by up to 2.40 points over the strongest same-setting baselines, raises the visual-dependency score from 0.28 to 0.49, reduces the prior-only answer rate from 46.2 percent to 21.7 percent, and lowers the hallucination rate from 17.1 percent to 9.8 percent, with 2.6x the cost of a single backbone forward pass and no fine-tuning.</summary>
<category term="cs.CV" scheme="http://arxiv.org/schemas/atom"/>
<published>2026-10-07T19:50:09Z</published>
<arxiv:primary_category term="cs.CV"/>
<author>
<name>Jingwen Huang</name>
</author>
<author>
<name>Yiming Fang</name>
</author>
<author>
<name>Hanxiao Wu</name>
</author>
<author>
<name>Boyuan He</name>
</author>
<author>
<name>Keyi Luo</name>
</author>
</entry>
<entry>
<id>http://arxiv.org/abs/2610.05195v3</id>
<title>Long-MDR: Long-Context Reinforcement Learning for Multimodal Deep-Research Agents</title>
<updated>2026-10-21T01:27:11Z</updated>
<link href="https://arxiv.org/abs/2610.05195v3" rel="alternate" type="text/html"/>
<link href="https://arxiv.org/pdf/2610.05195v3" rel="related" type="application/pdf" title="pdf"/>
<summary>The next generation of multimodal research agents must reason over long-lived research histories rather than short model completions. During a single task, an agent may repeatedly search the web, inspect visual evidence, revisit earlier hypotheses, and accumulate tens of thousands of tokens of multimodal context. Despite this trend, online RL for multimodal research agents remains largely confined to shorter contexts and interaction horizons. We push online RL training to 128k context and 75+ tool-interaction turns. To our knowledge, this is the first online multimodal deep-research RL study trained at 128k context, and the first trained with a 75 tool-turn horizon. Scaling to this regime exposes several practical limitations of conventional RL training. Early in training, weak policies make poor use of large interaction budgets, causing expensive rollouts with little reward improvement. Later, policy entropy can collapse before performance has saturated, prematurely ending useful learning. We introduce Long-MDR, a three-component training recipe designed specifically for this setting: On-Policy Distillation Warmup, Progressive Horizon Expansion, and Entropy-Triggered Rescue. Together, these techniques improve both the learning efficiency and stability of long-horizon RL, enabling continued gains in a regime where direct training is slow and costly. Code will be available at https://github.com/navileon/Long-MDR</summary>
<category term="cs.CV" scheme="http://arxiv.org/schemas/atom"/>
<published>2026-10-04T13:08:13Z</published>
<arxiv:primary_category term="cs.CV"/>
<author>
<name>On Tai Tang</name>
</author>
<author>
<name>Huanyao Zhang</name>
</author>
<author>
<name>Jiepeng Zhou</name>
</author>
<author>
<name>Wentao Zhang</name>
</author>
</entry>
<entry>
<id>http://arxiv.org/abs/2609.38154v1</id>
<title>LongLive-Plug: Once-for-All Distillation for Video Generation</title>
<updated>2026-09-29T17:58:49Z</updated>
<link href="https://arxiv.org/abs/2609.38154v1" rel="alternate" type="text/html"/>
<link href="https://arxiv.org/pdf/2609.38154v1" rel="related" type="application/pdf" title="pdf"/>
<summary>Video diffusion models are increasingly developed into specialized models for diverse downstream tasks, and this development often includes a distillation stage, for example to accelerate sampling or to improve long-video generation. This stage is typically repeated for every specialized model. We introduce LongLive-Plug, a once-for-all distillation framework that learns reusable capabilities as LoRAs on a base model for training-free, plug-and-play deployment to compatible downstream models. These capabilities include single-pass classifier-free guidance, few-step sampling, and long-context error correction for autoregressive generation. The adapters remain reusable even when downstream models add conditioning branches, expand output channels. Despite training at a fixed guidance scale, our dedicated CFG LoRA provides text guidance control through its inference weight. Combining it with a few-step LoRA simultaneously preserves few-step generation and CFG controllability on downstream tasks. We verify training-free deployment on 54 downstream models across three backbone families and eight task categories, including world modeling, robotics, editing, and multimodal generation. The approach may support additional compatible models. Each capability can thus be distilled once per backbone family and reused without per-target retraining.</summary>
<category term="cs.CV" scheme="http://arxiv.org/schemas/atom"/>
<published>2026-09-29T17:58:49Z</published>
<arxiv:comment>Code and models are available at https://github.com/NVlabs/LongLive</arxiv:comment>
<arxiv:primary_category term="cs.CV"/>
<author>
<name>Shuai Yang</name>
</author>
<author>
<name>Luozhou Wang</name>
</author>
<author>
<name>Wei Huang</name>
</author>
<author>
<name>ZhiFei Chen</name>
</author>
<author>
<name>Bohan Zhang</name>
</author>
<author>
<name>Xiao Fu</name>
</author>
<author>
<name>Qianli Ma</name>
</author>
<author>
<name>Chen-Hsuan Lin</name>
</author>
<author>
<name>Weian Mao</name>
</author>
<author>
<name>Bryan Chu</name>
</author>
<author>
<name>Song Han</name>
</author>
<author>
<name>Yukang Chen</name>
</author>
</entry>
<entry>
<id>http://arxiv.org/abs/2609.37311v2</id>
<title>ReMem: Rethinking Perception and Memory in Long-Context Recommendation Agents</title>
<updated>2026-10-03T13:53:12Z</updated>
<link href="https://arxiv.org/abs/2609.37311v2" rel="alternate" type="text/html"/>
<link href="https://arxiv.org/pdf/2609.37311v2" rel="related" type="application/pdf" title="pdf"/>
<summary>Recent Recommendation Agents (RecAgents) offer a promising alternative by shifting recommendation to an active, user-side paradigm, where generative agents autonomously perceive external platforms, reason over user preferences, and execute decisions. However, existing RecAgents still suffer from two critical limitations: brittle item perception based on noisy and heterogeneous item pages, and inefficient long-context reasoning over extended user histories and multi-step interaction traces. To address these challenges, we propose a novel recommendation agent framework, termed as ReMem, that combines OCR-based multimodal perception with time-evolving dynamic memory. Instead of parsing raw HTML, ReMem observes item pages through screenshots and extracts structured multimodal information via an OCR tool, enabling a more humanoid and platform-agnostic perception mechanism. To support long-horizon preference modeling, ReMem further introduces a chunk-wise sequential memory update strategy, where the agent selectively maintains a fixed-size memory of informative historical interactions while processing arbitrarily long contexts with linear inference complexity and bounded context length. This design allows the agent to preserve evolving user preferences without relying on external memory modules or disrupting the standard autoregressive generation process. To enhance the dynamic memory instruction, we further develop a multi-memory GRPO variant, which propagates the final-answer advantage to all intermediate conversations that contribute to the final response. Extensive experiments on three datasets demonstrate that ReMem consistently outperforms state-of-the-art baselines, achieving an average improvement of 5.16\% across three recommendation agent tasks, namely searching, ranking, and judging.</summary>
<category term="cs.AI" scheme="http://arxiv.org/schemas/atom"/>
<category term="cs.IR" scheme="http://arxiv.org/schemas/atom"/>
<published>2026-09-29T11:46:23Z</published>
<arxiv:primary_category term="cs.AI"/>
<author>
<name>Haohao Qu</name>
</author>
<author>
<name>Yongcheng Jing</name>
</author>
<author>
<name>Chun Hin Chan</name>
</author>
<author>
<name>Shanru Lin</name>
</author>
<author>
<name>Wenqi Fan</name>
</author>
<author>
<name>Dacheng Tao</name>
</author>
</entry>
<entry>
<id>http://arxiv.org/abs/2609.36651v1</id>
<title>FocusVTC: Efficient and High-Performance Visual Text Compression with Adaptive Resolution</title>
<updated>2026-09-29T03:51:06Z</updated>
<link href="https://arxiv.org/abs/2609.36651v1" rel="alternate" type="text/html"/>
<link href="https://arxiv.org/pdf/2609.36651v1" rel="related" type="application/pdf" title="pdf"/>
<summary>Long-context reasoning in large language models incurs substantial computation and memory costs. Visual text compression (VTC) reduces input length by rendering text as images, but fixed-resolution rendering creates a compression-performance trade-off: low DPI saves tokens at the expense of legibility, whereas high DPI spends tokens on irrelevant content. We introduce FocusVTC, which breaks this trade-off through adaptive resolution while preserving general multimodal capabilities. It combines compressed low-DPI global views with selective region enhancement, integrating enhanced views into ongoing reasoning. We construct 29.4K high-quality Reasoning-Evidence Localization (REL) chain-of-thought examples (REL-CoT) that link reasoning traces to page indices and bounding boxes. Multi-resolution REL supervised fine-tuning (REL-SFT) teaches the model to localize relevant regions, and Group Relative Policy Optimization learns when to enhance resolution and how to use the resulting observations, without a separate continual-pretraining stage. At 72 DPI on RULER v1, FocusVTC scores 87.4 at $2.9\times$ input compression, including tool observations, versus 57.5 for Glyph at $3.0\times$ input compression. It surpasses its text-input backbone on LongBench (56.40 versus 55.86), improves the MRCR macro-average by 13.91 points, and achieves a 51.19 macro-average on VTCBench. The MRCR latency evaluation also shows a $2.79\times$ online end-to-end speedup over Text. General multimodal capabilities are preserved, with MMMU increasing from 65.12 to 66.73 and MME from 2424.02 to 2457.62.</summary>
<category term="cs.CV" scheme="http://arxiv.org/schemas/atom"/>
<category term="cs.AI" scheme="http://arxiv.org/schemas/atom"/>
<published>2026-09-29T03:51:06Z</published>
<arxiv:comment>23 pages, 10 figures. Code: https://github.com/fangzhi-zhong/FoucsVTC</arxiv:comment>
<arxiv:primary_category term="cs.CV"/>
<author>
<name>FangZhi Zhong</name>
</author>
<author>
<name>Xuerui Qiu</name>
</author>
<author>
<name>Yuqi Pan</name>
</author>
<author>
<name>Ya Liu</name>
</author>
<author>
<name>Shaowei Gu</name>
</author>
<author>
<name>Bo Xu</name>
</author>
<author>
<name>Guoqi Li</name>
</author>
</entry>
<entry>
<id>http://arxiv.org/abs/2609.32831v1</id>
<title>UniCache: Task- and Type-Aware KV Cache Compression for Unified Multimodal Models</title>
<updated>2026-09-26T18:02:55Z</updated>
<link href="https://arxiv.org/abs/2609.32831v1" rel="alternate" type="text/html"/>
<link href="https://arxiv.org/pdf/2609.32831v1" rel="related" type="application/pdf" title="pdf"/>
<summary>Unified multimodal models combine understanding, generation, and editing within a single network, offering a promising foundation for versatile multimodal applications. However, growing multimodal contexts make KV cache storage and access increasingly costly. Existing KV cache compression methods are typically tailored to specific tasks and single-modality caches, while overlooking changes in cache importance across tasks and timesteps. However, in unified multimodal models, each task involves multiple KV cache types, and both their composition and dynamics differ across tasks. As a result, a single compression policy overlooks task- and type-specific requirements, leading to the loss of critical information and degraded quality across tasks. Based on these findings, we propose UniCache, a training-free framework for task- and type-aware KV cache compression. UniCache identifies the cache segments activated by each task and assigns suitable compression policies through offline calibration. It coordinates their parallel execution under a shared storage budget through attention-guided allocation and task-aware temporal scheduling. Experiments show that UniCache achieves $5\times$ KV cache compression for understanding and editing and $2.5\times$ for generation with negligible quality loss, while increasing throughput by up to $1.78\times$ in long-context settings, significantly improving the practicality of scaling unified multimodal models to longer context.</summary>
<category term="cs.LG" scheme="http://arxiv.org/schemas/atom"/>
<published>2026-09-26T18:02:55Z</published>
<arxiv:primary_category term="cs.LG"/>
<author>
<name>Wanqi Yang</name>
</author>
<author>
<name>Yuexiao Ma</name>
</author>
<author>
<name>Mei Xie</name>
</author>
<author>
<name>Xiawu Zheng</name>
</author>
<author>
<name>Shiwei Liu</name>
</author>
</entry>
<entry>
<id>http://arxiv.org/abs/2609.25611v1</id>
<title>Qwen3.8-Omni: Towards Native Omni-Modal Agents</title>
<updated>2026-09-22T03:07:59Z</updated>
<link href="https://arxiv.org/abs/2609.25611v1" rel="alternate" type="text/html"/>
<link href="https://arxiv.org/pdf/2609.25611v1" rel="related" type="application/pdf" title="pdf"/>
<summary>We introduce Qwen3.8-Omni-Flash, a natively multimodal agentic model for real-world multimodal productivity. Compared with previous omni models, which primarily emphasized perception and interaction, Qwen3.8-Omni-Flash substantially improves multimodal understanding and reasoning, as well as performance on long-horizon agentic tasks. These capabilities are supported by a native multimodal co-training strategy that preserves strong text-domain capabilities while facilitating the transfer of agentic capabilities from text to audio and video tasks. The model inherits the sparse mixture-of-experts (MoE) architecture of Qwen3.8-Next and extends the context window to one million tokens, supporting long-context multimodal reasoning and long-horizon planning. These advances enable integration into production workflows as a primary agent or a specialized sub-agent, supporting video editing, long-form audio and video translation, music-conditioned music video or movie generation, and video-based note or omni-skill creation. To address the lack of native audio and video support in existing agent harnesses, we release Qwen-MM-Plugins, a lightweight open-source plugin framework for multimodal productivity. We further frame real-time multimodal interaction as a system-level challenge requiring orchestration of context and memory management, tool use, and sub-agent delegation. Accordingly, we release Qwen-Live-Harness, an open-source framework for building responsive, real-time multimodal agents based on Qwen3.8-Omni-Flash. Extensive evaluations demonstrate that Qwen3.8-Omni-Flash achieves strong performance across multimodal understanding, reasoning, long-horizon agentic execution, and video productivity tasks. These results and the accompanying open-source tools support Qwen3.8-Omni-Flash as a practical foundation for deploying natively multimodal agents in research and production.</summary>
<category term="cs.CL" scheme="http://arxiv.org/schemas/atom"/>
<category term="cs.CV" scheme="http://arxiv.org/schemas/atom"/>
<category term="cs.MM" scheme="http://arxiv.org/schemas/atom"/>
<published>2026-09-22T03:07:59Z</published>
<arxiv:primary_category term="cs.CL"/>
<author>
<name> Qwen Team</name>
</author>
</entry>
<entry>
<id>http://arxiv.org/abs/2609.19969v1</id>
<title>DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression</title>
<updated>2026-09-17T09:43:10Z</updated>
<link href="https://arxiv.org/abs/2609.19969v1" rel="alternate" type="text/html"/>
<link href="https://arxiv.org/pdf/2609.19969v1" rel="related" type="application/pdf" title="pdf"/>
<summary>The widespread adoption of long-horizon agents has made model workloads increasingly input-heavy. Although prior work has substantially reduced the cost of long-context computation, prefill remains computationally expensive, and large KV caches continue to strain HBM and SSD capacity and data-transfer bandwidth. Together, these compute, storage, and bandwidth demands constitute the primary bottleneck to further lowering deployment costs. To address this challenge, we introduce DeepSeek-V4.1-Flash, a multimodal Mixture-of-Experts (MoE) model with 552B backbone parameters and support for contexts of up to one million tokens. With its Causal Encoder-Decoder (CED) architecture, the model activates 16B parameters per token during decode but only 8B parameters during prefill, substantially improving cost efficiency for agentic workloads. To push the limits of KV cache compression, DeepSeek-V4.1-Flash combines cross-layer KV cache reuse in Compressed Sparse Attention 2 (CSA2) with FP4 KV caching. These designs reduce its global KV cache footprint (always in HBM) to 890 bytes per token, roughly 1/4 of the corresponding footprint of DeepSeek-V4-Flash. Further, through a dedicated deployment optimization known as SWA Bounded Replay, DeepSeek-V4.1-Flash reduces its persistent KV cache footprint (always on SSD or in host memory) to roughly 1/8 of that of DeepSeek-V4-Flash. Despite its much smaller KV cache footprint, the model delivers substantially better performance than the baseline. In addition, we streamline the DeepSeek-V4 architecture and introduce several efficient architectural extensions. We pretrain DeepSeek-V4.1-Flash on a multimodal corpus comprising 45T tokens and conduct comprehensive post-training, yielding strong performance across diverse text-based and multimodal agentic scenarios. Model checkpoints are available at https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash.</summary>
<category term="cs.CL" scheme="http://arxiv.org/schemas/atom"/>
<published>2026-09-17T09:43:10Z</published>
<arxiv:primary_category term="cs.CL"/>
<author>
<name> DeepSeek-AI</name>
</author>
<author>
<name> :</name>
</author>
<author>
<name>Anyi Xu</name>
</author>
<author>
<name>B. Li</name>
</author>
<author>
<name>Bangcai Lin</name>
</author>
<author>
<name>Bing Xue</name>
</author>
<author>
<name>BingCheng Xian</name>
</author>
<author>
<name>Bingzheng Xu</name>
</author>
<author>
<name>Bochao Wu</name>
</author>
<author>
<name>Bowei Zhang</name>
</author>
<author>
<name>Boyi Deng</name>
</author>
<author>
<name>C. C. Yu</name>
</author>
<author>
<name>Chao Jin</name>
</author>
<author>
<name>Chaofan Lin</name>
</author>
<author>
<name>Chen Dong</name>
</author>
<author>
<name>Chenbing Wang</name>
</author>
<author>
<name>Chenfan Feng</name>
</author>
<author>
<name>Chengda Lu</name>
</author>
<author>
<name>Chenggang Zhao</name>
</author>
<author>
<name>Chengqi Deng</name>
</author>
<author>
<name>Chengyuan Zhang</name>
</author>
<author>
<name>Chenhao Xu</name>
</author>
<author>
<name>Chenqi Zhao</name>
</author>
<author>
<name>Chenze Shao</name>
</author>
<author>
<name>Chuhao Wang</name>
</author>
<author>
<name>Chuqi Zhang</name>
</author>
<author>
<name>Damai Dai</name>
</author>
<author>
<name>Dejian Yang</name>
</author>
<author>
<name>Deli Chen</name>
</author>
<author>
<name>Di Huang</name>
</author>
<author>
<name>Di Wu</name>
</author>
<author>
<name>Donghao Li</name>
</author>
<author>
<name>Erhang Li</name>
</author>
<author>
<name>Eric Fu</name>
</author>
<author>
<name>F. Zhou</name>
</author>
<author>
<name>Fangwei Zhou</name>
</author>
<author>
<name>Fangyun Lin</name>
</author>
<author>
<name>Fangzhou Yuan</name>
</author>
<author>
<name>Feiyu Xia</name>
</author>
<author>
<name>Fucong Dai</name>
</author>
<author>
<name>Guangbo Hao</name>
</author>
<author>
<name>Guanglin Li</name>
</author>
<author>
<name>Guanting Chen</name>
</author>
<author>
<name>Guoai Cao</name>
</author>
<author>
<name>Guofan Fan</name>
</author>
<author>
<name>Guolai Meng</name>
</author>
<author>
<name>Guowei Li</name>
</author>
<author>
<name>Haichuan Zhang</name>
</author>
<author>
<name>Haiyang Ma</name>
</author>
<author>
<name>Haiyang Shen</name>
</author>
<author>
<name>Han Li</name>
</author>
<author>
<name>Han Yu</name>
</author>
<author>
<name>Han Zhang</name>
</author>
<author>
<name>Hangyuan Deng</name>
</author>
<author>
<name>Hanwei Xu</name>
</author>
<author>
<name>Hanxiang Xu</name>
</author>
<author>
<name>Hanxun Zhong</name>
</author>
<author>
<name>Hao Guo</name>
</author>
<author>
<name>Hao Jiang</name>
</author>
<author>
<name>Hao Li</name>
</author>
<author>
<name>Hao Qin</name>
</author>
<author>
<name>Haodong Wen</name>
</author>
<author>
<name>Haofen Liang</name>
</author>
<author>
<name>Haofeng Huang</name>
</author>
<author>
<name>Haohua Liu</name>
</author>
<author>
<name>Haoling Zhang</name>
</author>
<author>
<name>Haoming Luo</name>
</author>
<author>
<name>Haoran Yang</name>
</author>
<author>
<name>Haotian Xu</name>
</author>
<author>
<name>Haotian Yuan</name>
</author>
<author>
<name>Haoting Huang</name>
</author>
<author>
<name>Haowen Luo</name>
</author>
<author>
<name>Haoyang Cai</name>
</author>
<author>
<name>Haoyu Chen</name>
</author>
<author>
<name>Haozhe Ji</name>
</author>
<author>
<name>Hengran Zhang</name>
</author>
<author>
<name>Hengrui Wang</name>
</author>
<author>
<name>Hengxu Wu</name>
</author>
<author>
<name>Honghui Ding</name>
</author>
<author>
<name>Hongxuan Tang</name>
</author>
<author>
<name>Huadong Wang</name>
</author>
<author>
<name>Huanqi Cao</name>
</author>
<author>
<name>Huazuo Gao</name>
</author>
<author>
<name>Hui Qu</name>
</author>
<author>
<name>Hui Zeng</name>
</author>
<author>
<name>J. Yang</name>
</author>
<author>
<name>J. H. Jin</name>
</author>
<author>
<name>J. H. Zhang</name>
</author>
<author>
<name>J. X. Zou</name>
</author>
<author>
<name>Jia Yu</name>
</author>
<author>
<name>Jiahui Zhou</name>
</author>
<author>
<name>Jiajun Chen</name>
</author>
<author>
<name>Jialiang Huang</name>
</author>
<author>
<name>Jialin Zhao</name>
</author>
<author>
<name>Jiamin Tang</name>
</author>
<author>
<name>Jian Zhou</name>
</author>
<author>
<name>Jianan Tong</name>
</author>
<author>
<name>Jianwen Li</name>
</author>
<author>
<name>Jiaqi Zhu</name>
</author>
<author>
<name>Jiarui Wang</name>
</author>
<author>
<name>Jiasheng Ye</name>
</author>
<author>
<name>Jiashi Li</name>
</author>
<author>
<name>Jiaxin Xu</name>
</author>
<author>
<name>Jiaying Ding</name>
</author>
<author>
<name>Jibai Lu</name>
</author>
<author>
<name>Jiewen Hu</name>
</author>
<author>
<name>Jin Yan</name>
</author>
<author>
<name>Jincheng Zhai</name>
</author>
<author>
<name>Jingchang Chen</name>
</author>
<author>
<name>Jingcheng Hu</name>
</author>
<author>
<name>Jingli Zhou</name>
</author>
<author>
<name>Jingsheng Xu</name>
</author>
<author>
<name>Jingting Xiang</name>
</author>
<author>
<name>Jingyan Yun</name>
</author>
<author>
<name>Jingyang Yuan</name>
</author>
<author>
<name>Jingyuan Cheng</name>
</author>
<author>
<name>Jinhua Zhu</name>
</author>
<author>
<name>Jinpeng Wang</name>
</author>
<author>
<name>Jinyi Chen</name>
</author>
<author>
<name>Jinyi Hu</name>
</author>
<author>
<name>Jiping Yu</name>
</author>
<author>
<name>Jueliang Guo</name>
</author>
<author>
<name>Junbo Pei</name>
</author>
<author>
<name>Junbo Sun</name>
</author>
<author>
<name>Junguang Jiang</name>
</author>
<author>
<name>Junjie Qiu</name>
</author>
<author>
<name>Junkang Zhou</name>
</author>
<author>
<name>Junqi Liu</name>
</author>
<author>
<name>Junren Li</name>
</author>
<author>
<name>Junxian Li</name>
</author>
<author>
<name>Junxiao Song</name>
</author>
<author>
<name>Junyi Guo</name>
</author>
<author>
<name>Kai Dong</name>
</author>
<author>
<name>Kaifeng Chen</name>
</author>
<author>
<name>Kaige Gao</name>
</author>
<author>
<name>Kang Guan</name>
</author>
<author>
<name>Kangdong Yuan</name>
</author>
<author>
<name>Ke Hong</name>
</author>
<author>
<name>Ke Xu</name>
</author>
<author>
<name>Kefan Zhao</name>
</author>
<author>
<name>Kexin Ji</name>
</author>
<author>
<name>Kexin Zhang</name>
</author>
<author>
<name>Kexing Zhou</name>
</author>
<author>
<name>Kuai Yu</name>
</author>
<author>
<name>Lan Zhang</name>
</author>
<author>
<name>Lean Wang</name>
</author>
<author>
<name>Lecong Zhang</name>
</author>
<author>
<name>Lei Wang</name>
</author>
<author>
<name>Letian Gao</name>
</author>
<author>
<name>Liang Zhao</name>
</author>
<author>
<name>Liansheng Xu</name>
</author>
<author>
<name>Lihua Guo</name>
</author>
<author>
<name>Lingxiao Luo</name>
</author>
<author>
<name>Lingyue Fu</name>
</author>
<author>
<name>Litao Deng</name>
</author>
<author>
<name>Litong Wang</name>
</author>
<author>
<name>Liyue Zhang</name>
</author>
<author>
<name>Longhao Chen</name>
</author>
<author>
<name>Lu Chen</name>
</author>
<author>
<name>Luotian Huang</name>
</author>
<author>
<name>Luyao Ma</name>
</author>
<author>
<name>Luyao Wang</name>
</author>
<author>
<name>M. S. Di</name>
</author>
<author>
<name>Max Mei</name>
</author>
<author>
<name>Menghao Ye</name>
</author>
<author>
<name>Miao Cui</name>
</author>
<author>
<name>Mingchuan Zhang</name>
</author>
<author>
<name>Minghua Zhang</name>
</author>
<author>
<name>Minghui Tang</name>
</author>
<author>
<name>Mingjing Zhang</name>
</author>
<author>
<name>Mingqi Wei</name>
</author>
<author>
<name>Mingshu Chen</name>
</author>
<author>
<name>Mingxing Liu</name>
</author>
<author>
<name>Mingxu Zhou</name>
</author>
<author>
<name>Mingyu Xu</name>
</author>
<author>
<name>Mingyu Yang</name>
</author>
<author>
<name>Mingze Wang</name>
</author>
<author>
<name>Muyang Chen</name>
</author>
<author>
<name>Ni Shentu</name>
</author>
<author>
<name>Ning Wang</name>
</author>
<author>
<name>Niufang Ning</name>
</author>
<author>
<name>Panpan Huang</name>
</author>
<author>
<name>Peixin Cong</name>
</author>
<author>
<name>Peiyi Wang</name>
</author>
<author>
<name>Peiyuan Xin</name>
</author>
<author>
<name>Pengfei Ren</name>
</author>
<author>
<name>Pengfei Yan</name>
</author>
<author>
<name>Pengle Zhang</name>
</author>
<author>
<name>Qi Kang</name>
</author>
<author>
<name>Qi Tang</name>
</author>
<author>
<name>Qiancheng Wang</name>
</author>
<author>
<name>Qiang Li</name>
</author>
<author>
<name>Qihao Zhu</name>
</author>
<author>
<name>Qingyang Li</name>
</author>
<author>
<name>Qinyu Chen</name>
</author>
<author>
<name>Qiushi Du</name>
</author>
<author>
<name>Qizhou Guo</name>
</author>
<author>
<name>Rongxian Xu</name>
</author>
<author>
<name>Rui Ding</name>
</author>
<author>
<name>Rui Hu</name>
</author>
<author>
<name>Rui Tian</name>
</author>
<author>
<name>Rui Yu</name>
</author>
<author>
<name>Ruidong Zhu</name>
</author>
<author>
<name>Ruifan Xu</name>
</author>
<author>
<name>Ruihan Yang</name>
</author>
<author>
<name>Ruihang Xia</name>
</author>
<author>
<name>Ruijie Lu</name>
</author>
<author>
<name>Ruilin Geng</name>
</author>
<author>
<name>Ruipeng Hong</name>
</author>
<author>
<name>Ruiqi Ge</name>
</author>
<author>
<name>Ruisong Zhang</name>
</author>
<author>
<name>Ruize Sun</name>
</author>
<author>
<name>Ruizhe Pan</name>
</author>
<author>
<name>Runji Wang</name>
</author>
<author>
<name>Runqian Chen</name>
</author>
<author>
<name>Runxin Xu</name>
</author>
<author>
<name>Ruohong Tian</name>
</author>
<author>
<name>Ruomeng Shen</name>
</author>
<author>
<name>Ruoyu Zhang</name>
</author>
<author>
<name>Ryan X.</name>
</author>
<author>
<name>S. H. Liu</name>
</author>
<author>
<name>Shanghao Lu</name>
</author>
<author>
<name>Shangyan Zhou</name>
</author>
<author>
<name>Shanhuang Chen</name>
</author>
<author>
<name>Shaofei Cai</name>
</author>
<author>
<name>Shaoheng Nie</name>
</author>
<author>
<name>Shaoyuan Chen</name>
</author>
<author>
<name>Shengding Hu</name>
</author>
<author>
<name>Shengkai Lin</name>
</author>
<author>
<name>Shengwen Ran</name>
</author>
<author>
<name>Shengyu Liu</name>
</author>
<author>
<name>Shengyuan Jia</name>
</author>
<author>
<name>Shi Bai</name>
</author>
<author>
<name>Shi Feng</name>
</author>
<author>
<name>Shicheng Xu</name>
</author>
<author>
<name>Shichun Liu</name>
</author>
<author>
<name>Shiqiang Hu</name>
</author>
<author>
<name>Shirong Ma</name>
</author>
<author>
<name>Shiyu Wang</name>
</author>
<author>
<name>Shiyuan Feng</name>
</author>
<author>
<name>Shufan Gong</name>
</author>
<author>
<name>Shuhan Lin</name>
</author>
<author>
<name>Shuiping Yu</name>
</author>
<author>
<name>Shunfeng Zhou</name>
</author>
<author>
<name>Shuo Yang</name>
</author>
<author>
<name>Shuomeng Wang</name>
</author>
<author>
<name>Shuting Guo</name>
</author>
<author>
<name>Shuting Pan</name>
</author>
<author>
<name>Shuying Yu</name>
</author>
<author>
<name>Sinuo Cao</name>
</author>
<author>
<name>Siyi Lin</name>
</author>
<author>
<name>Sizhe Chen</name>
</author>
<author>
<name>Songyang Chen</name>
</author>
<author>
<name>Songyang Zhou</name>
</author>
<author>
<name>Tao Ni</name>
</author>
<author>
<name>Tao Yun</name>
</author>
<author>
<name>Tian Jin</name>
</author>
<author>
<name>Tian Pei</name>
</author>
<author>
<name>Tian Ye</name>
</author>
<author>
<name>Tianle Lin</name>
</author>
<author>
<name>Tianran Ji</name>
</author>
<author>
<name>Tianyi Cui</name>
</author>
<author>
<name>Tianyuan Yue</name>
</author>
<author>
<name>Tingting Yu</name>
</author>
<author>
<name>Tongrui Xiong</name>
</author>
<author>
<name>Wangding Zeng</name>
</author>
<author>
<name>Wei Liu</name>
</author>
<author>
<name>Wei Zhang</name>
</author>
<author>
<name>Weibin Xu</name>
</author>
<author>
<name>Weihao Zeng</name>
</author>
<author>
<name>Weilin Zhao</name>
</author>
<author>
<name>Wen Liu</name>
</author>
<author>
<name>Wenfeng Liang</name>
</author>
<author>
<name>Wenjie Pang</name>
</author>
<author>
<name>Wenjing Luo</name>
</author>
<author>
<name>Wenjing Yao</name>
</author>
<author>
<name>Wenjun Gao</name>
</author>
<author>
<name>Wenkai Shao</name>
</author>
<author>
<name>Wenkai Yang</name>
</author>
<author>
<name>Wenli Zhang</name>
</author>
<author>
<name>Wenlu Wang</name>
</author>
<author>
<name>Wenlve Huang</name>
</author>
<author>
<name>Wenqian Yan</name>
</author>
<author>
<name>Wentao Zhang</name>
</author>
<author>
<name>Xi Gao</name>
</author>
<author>
<name>Xiang He</name>
</author>
<author>
<name>Xiang Li</name>
</author>
<author>
<name>Xiangli Li</name>
</author>
<author>
<name>Xiangwen Wang</name>
</author>
<author>
<name>Xiangying Zhang</name>
</author>
<author>
<name>Xiankui Wei</name>
</author>
<author>
<name>Xiao Bi</name>
</author>
<author>
<name>Xiaodong Liu</name>
</author>
<author>
<name>Xiaohan Wang</name>
</author>
<author>
<name>Xiaojian Qu</name>
</author>
<author>
<name>Xiaokang Chen</name>
</author>
<author>
<name>Xiaokang Zhang</name>
</author>
<author>
<name>Xiaotao Nie</name>
</author>
<author>
<name>Xiaoyao Zou</name>
</author>
<author>
<name>Xiaoyuan Li</name>
</author>
<author>
<name>Xicheng Guo</name>
</author>
<author>
<name>Xieting Chu</name>
</author>
<author>
<name>Xin Cheng</name>
</author>
<author>
<name>Xin Liu</name>
</author>
<author>
<name>Xin Xie</name>
</author>
<author>
<name>Xinbo Xu</name>
</author>
<author>
<name>Xingchao Liu</name>
</author>
<author>
<name>Xingchen Liu</name>
</author>
<author>
<name>Xingkai Yu</name>
</author>
<author>
<name>Xingyou Li</name>
</author>
<author>
<name>Xintong Yao</name>
</author>
<author>
<name>Xinyang Chen</name>
</author>
<author>
<name>Xinyong Jiang</name>
</author>
<author>
<name>Xinyu Yang</name>
</author>
<author>
<name>Xinyu Yang</name>
</author>
<author>
<name>Xu Chen</name>
</author>
<author>
<name>Xuanyu Wang</name>
</author>
<author>
<name>Xubei Zhong</name>
</author>
<author>
<name>Xuecheng Su</name>
</author>
<author>
<name>Xuejie Liu</name>
</author>
<author>
<name>Xuheng Lin</name>
</author>
<author>
<name>Xujie Fan</name>
</author>
<author>
<name>Xuncheng Zhao</name>
</author>
<author>
<name>Xuwei Fu</name>
</author>
<author>
<name>Y. C. Yan</name>
</author>
<author>
<name>Y. H. Jiang</name>
</author>
<author>
<name>Y. T. Wu</name>
</author>
<author>
<name>Y. W. M.</name>
</author>
<author>
<name>Y. Z. Wang</name>
</author>
<author>
<name>Yafei Gao</name>
</author>
<author>
<name>Yang Yang</name>
</author>
<author>
<name>Yang Zhang</name>
</author>
<author>
<name>Yanru Ma</name>
</author>
<author>
<name>Yanwen Huang</name>
</author>
<author>
<name>Yao Li</name>
</author>
<author>
<name>Yao Li</name>
</author>
<author>
<name>Yao Meng</name>
</author>
<author>
<name>Yao Zhao</name>
</author>
<author>
<name>Yaofeng Sun</name>
</author>
<author>
<name>Yaohui Wang</name>
</author>
<author>
<name>Yaoyang Ye</name>
</author>
<author>
<name>Yehang Yin</name>
</author>
<author>
<name>Yexinrui Wu</name>
</author>
<author>
<name>Yi Qian</name>
</author>
<author>
<name>Yi Tao</name>
</author>
<author>
<name>Yi Yu</name>
</author>
<author>
<name>Yichao Zhang</name>
</author>
<author>
<name>Yichen Jiang</name>
</author>
<author>
<name>Yicheng Wang</name>
</author>
<author>
<name>Yifan Ding</name>
</author>
<author>
<name>Yifan Shi</name>
</author>
<author>
<name>Yifeng Peng</name>
</author>
<author>
<name>Yifeng Zhai</name>
</author>
<author>
<name>Yijia Wu</name>
</author>
<author>
<name>Yiliang Xiong</name>
</author>
<author>
<name>Yilun Wang</name>
</author>
<author>
<name>Ying He</name>
</author>
<author>
<name>Ying Zhou</name>
</author>
<author>
<name>Yingjia Luo</name>
</author>
<author>
<name>Yinmin Zhong</name>
</author>
<author>
<name>Yiping Wang</name>
</author>
<author>
<name>Yisong Wang</name>
</author>
<author>
<name>Yixiang Zhang</name>
</author>
<author>
<name>Yixiao Chen</name>
</author>
<author>
<name>Yixuan Tan</name>
</author>
<author>
<name>Yixuan Wei</name>
</author>
<author>
<name>Yiyang Ma</name>
</author>
<author>
<name>Yiyao Yang</name>
</author>
<author>
<name>Yiyuan Liu</name>
</author>
<author>
<name>Yizai Cai</name>
</author>
<author>
<name>Yizhen Wei</name>
</author>
<author>
<name>Yizhi Wang</name>
</author>
<author>
<name>Yonglun Yang</name>
</author>
<author>
<name>Yongqi Zhuo</name>
</author>
<author>
<name>Yongqiang Guo</name>
</author>
<author>
<name>Yongtong Wu</name>
</author>
<author>
<name>Yu Wu</name>
</author>
<author>
<name>Yu Zhang</name>
</author>
<author>
<name>Yuan Bian</name>
</author>
<author>
<name>Yuan Cheng</name>
</author>
<author>
<name>Yuan Ou</name>
</author>
<author>
<name>Yuan Sun</name>
</author>
<author>
<name>Yuanfan Xu</name>
</author>
<author>
<name>Yuanhang Sun</name>
</author>
<author>
<name>Yuanhao Li</name>
</author>
<author>
<name>Yuchen Liu</name>
</author>
<author>
<name>Yuchen Yao</name>
</author>
<author>
<name>Yudong Han</name>
</author>
<author>
<name>Yuduan Wang</name>
</author>
<author>
<name>Yuhan Wu</name>
</author>
<author>
<name>Yuhao Meng</name>
</author>
<author>
<name>Yuheng Zou</name>
</author>
<author>
<name>YuKun Li</name>
</author>
<author>
<name>Yunchuan Wang</name>
</author>
<author>
<name>Yunfan Xiao</name>
</author>
<author>
<name>Yunfan Xiong</name>
</author>
<author>
<name>Yupeng Chen</name>
</author>
<author>
<name>Yuqian Cao</name>
</author>
<author>
<name>Yuqian Wang</name>
</author>
<author>
<name>Yuqing Chen</name>
</author>
<author>
<name>Yushun Zhang</name>
</author>
<author>
<name>Yutong Lin</name>
</author>
<author>
<name>Yuwei Xiao</name>
</author>
<author>
<name>Yuxian Gu</name>
</author>
<author>
<name>Yuxiang Chen</name>
</author>
<author>
<name>Yuxiang Huang</name>
</author>
<author>
<name>Yuxiang Luo</name>
</author>
<author>
<name>Yuxiang You</name>
</author>
<author>
<name>Yuxin Chen</name>
</author>
<author>
<name>Yuxin Xiang</name>
</author>
<author>
<name>Yuxuan Liu</name>
</author>
<author>
<name>Yuxuan Zhou</name>
</author>
<author>
<name>Yuyang Zhou</name>
</author>
<author>
<name>Yuzhe Guo</name>
</author>
<author>
<name>Yuzhen Huang</name>
</author>
<author>
<name>Yuzhuo Bai</name>
</author>
<author>
<name>Z. Y. Z.</name>
</author>
<author>
<name>Zanlin Ni</name>
</author>
<author>
<name>Zehao Wang</name>
</author>
<author>
<name>Zehua Zhao</name>
</author>
<author>
<name>Zehui Ren</name>
</author>
<author>
<name>Zejun Zhao</name>
</author>
<author>
<name>Zhangli Sha</name>
</author>
<author>
<name>Zhanying Wang</name>
</author>
<author>
<name>Zhaochen Zhang</name>
</author>
<author>
<name>Zhaoshuai Du</name>
</author>
<author>
<name>Zhe Fu</name>
</author>
<author>
<name>Zhean Xu</name>
</author>
<author>
<name>Zhenda Xie</name>
</author>
<author>
<name>Zheng Liu</name>
</author>
<author>
<name>Zhengyan Zhang</name>
</author>
<author>
<name>Zhenhua Dong</name>
</author>
<author>
<name>Zhewen Hao</name>
</author>
<author>
<name>Zhibang Wang</name>
</author>
<author>
<name>Zhibin Gou</name>
</author>
<author>
<name>Zhicheng Ma</name>
</author>
<author>
<name>Zhihao Li</name>
</author>
<author>
<name>Zhihong Shao</name>
</author>
<author>
<name>Zhihuan Huang</name>
</author>
<author>
<name>Zhijie Li</name>
</author>
<author>
<name>Zhirui Lu</name>
</author>
<author>
<name>Zhixian Huang</name>
</author>
<author>
<name>Zhixuan Chen</name>
</author>
<author>
<name>Zhixuan Chen</name>
</author>
<author>
<name>Zhixuan Pan</name>
</author>
<author>
<name>Zhiyu Wu</name>
</author>
<author>
<name>Zhizhou Ren</name>
</author>
<author>
<name>Zhu He</name>
</author>
<author>
<name>Zhuoshu Li</name>
</author>
<author>
<name>Zhuping Zhang</name>
</author>
<author>
<name>Zian Xu</name>
</author>
<author>
<name>Zihao Wang</name>
</author>
<author>
<name>Zihui Gu</name>
</author>
<author>
<name>Zijia Zhu</name>
</author>
<author>
<name>Zili Zhang</name>
</author>
<author>
<name>Zilin Li</name>
</author>
<author>
<name>Zilong Hou</name>
</author>
<author>
<name>Zilong Lyu</name>
</author>
<author>
<name>Ziqiao Wang</name>
</author>
<author>
<name>Ziwei Xie</name>
</author>
<author>
<name>Ziya Zhang</name>
</author>
<author>
<name>Ziyi Gao</name>
</author>
<author>
<name>Zizheng Pan</name>
</author>
<author>
<name>Zonglin Li</name>
</author>
<author>
<name>Zongqing Yao</name>
</author>
<author>
<name>Zui Chen</name>
</author>
<author>
<name>Zuofan Wu</name>
</author>
<author>
<name>Chenchen Ling</name>
</author>
<author>
<name>Chengyu Hou</name>
</author>
<author>
<name>Chong Chen</name>
</author>
<author>
<name>D. Li</name>
</author>
<author>
<name>Di Qi</name>
</author>
<author>
<name>Dongjie Ji</name>
</author>
<author>
<name>Fang Wei</name>
</author>
<author>
<name>Fanyi Xia</name>
</author>
<author>
<name>Fei Xie</name>
</author>
<author>
<name>Feiyi Tan</name>
</author>
<author>
<name>Hailong Guo</name>
</author>
<author>
<name>Haiyan Zhai</name>
</author>
<author>
<name>Hui Zhou</name>
</author>
<author>
<name>Huihui Tan</name>
</author>
<author>
<name>Huijie Li</name>
</author>
<author>
<name>Jia Luo</name>
</author>
<author>
<name>Jia Song</name>
</author>
<author>
<name>Jialu Cai</name>
</author>
<author>
<name>Jian Liang</name>
</author>
<author>
<name>Jiangting Zhou</name>
</author>
<author>
<name>Jiaqi Gao</name>
</author>
<author>
<name>Jiayi Shao</name>
</author>
<author>
<name>Jie Chen</name>
</author>
<author>
<name>Jieyu Yang</name>
</author>
<author>
<name>Jin Chen</name>
</author>
<author>
<name>Jingde Zhang</name>
</author>
<author>
<name>Jingzi Zhou</name>
</author>
<author>
<name>Jinqian Wang</name>
</author>
<author>
<name>Jinyang Liu</name>
</author>
<author>
<name>JinZhao Sun</name>
</author>
<author>
<name>Junhua Ling</name>
</author>
<author>
<name>Junmin Zheng</name>
</author>
<author>
<name>Kaicheng Yang</name>
</author>
<author>
<name>Ke Xu</name>
</author>
<author>
<name>Le Su</name>
</author>
<author>
<name>Leyi Xia</name>
</author>
<author>
<name>Liangfeng Ding</name>
</author>
<author>
<name>Lin Zhuo</name>
</author>
<author>
<name>Linwang Ma</name>
</author>
<author>
<name>Linyan Zhu</name>
</author>
<author>
<name>Liyu Cai</name>
</author>
<author>
<name>Luqi Yao</name>
</author>
<author>
<name>M. K. Zhang</name>
</author>
<author>
<name>Meng Li</name>
</author>
<author>
<name>Miao Lin</name>
</author>
<author>
<name>Miaojun Wang</name>
</author>
<author>
<name>Min Zhang</name>
</author>
<author>
<name>Mingming Li</name>
</author>
<author>
<name>Mingming Wang</name>
</author>
<author>
<name>Mingze Yin</name>
</author>
<author>
<name>Minmin Han</name>
</author>
<author>
<name>Nan Cao</name>
</author>
<author>
<name>Ning Wang</name>
</author>
<author>
<name>Ningxin Ma</name>
</author>
<author>
<name>Panpan Wang</name>
</author>
<author>
<name>Peihan Lin</name>
</author>
<author>
<name>Peng Sun</name>
</author>
<author>
<name>Peng Zhang</name>
</author>
<author>
<name>Qian Ying</name>
</author>
<author>
<name>Qiang Xiang</name>
</author>
<author>
<name>Qiao Wang</name>
</author>
<author>
<name>Qingmiao Mao</name>
</author>
<author>
<name>Qiwei Jiang</name>
</author>
<author>
<name>Rongli Jin</name>
</author>
<author>
<name>Ruyi Chen</name>
</author>
<author>
<name>Sha Tao</name>
</author>
<author>
<name>Shangmian Sun</name>
</author>
<author>
<name>Shaoqing Wu</name>
</author>
<author>
<name>Shichao Zou</name>
</author>
<author>
<name>Si Lei</name>
</author>
<author>
<name>Tianyang Zhang</name>
</author>
<author>
<name>Tianyu Sun</name>
</author>
<author>
<name>Tingting Yin</name>
</author>
<author>
<name>W. L. Xiao</name>
</author>
<author>
<name>Wei An</name>
</author>
<author>
<name>Wei Li</name>
</author>
<author>
<name>Wei Wang</name>
</author>
<author>
<name>Weiwei Lin</name>
</author>
<author>
<name>Wenqing Hou</name>
</author>
<author>
<name>X. Lin</name>
</author>
<author>
<name>Xiangfei Meng</name>
</author>
<author>
<name>Xianzhu Huang</name>
</author>
<author>
<name>Xiao Peng</name>
</author>
<author>
<name>Xiaoqian Li</name>
</author>
<author>
<name>Xiaoting Zhang</name>
</author>
<author>
<name>Xiaowen Sun</name>
</author>
<author>
<name>Xiaoxiang Wang</name>
</author>
<author>
<name>Xiaoyu Ye</name>
</author>
<author>
<name>Xinrou Zhang</name>
</author>
<author>
<name>Xinyu Zhang</name>
</author>
<author>
<name>Xue Cao</name>
</author>
<author>
<name>Xueyin Chen</name>
</author>
<author>
<name>Yanan Zhou</name>
</author>
<author>
<name>Yanhong Xu</name>
</author>
<author>
<name>Yao Xia</name>
</author>
<author>
<name>Yao Xu</name>
</author>
<author>
<name>Yi Shao</name>
</author>
<author>
<name>Yihong Zhang</name>
</author>
<author>
<name>Yiling Ma</name>
</author>
<author>
<name>Ying Tang</name>
</author>
<author>
<name>Yining Lou</name>
</author>
<author>
<name>Yiru Chen</name>
</author>
<author>
<name>Yishi Piao</name>
</author>
<author>
<name>Yixuan Chen</name>
</author>
<author>
<name>Yong Xiong</name>
</author>
<author>
<name>Yuchen Xuan</name>
</author>
<author>
<name>Yuehan Yang</name>
</author>
<author>
<name>Yuer Xu</name>
</author>
<author>
<name>Yukun Zha</name>
</author>
<author>
<name>Yunxian Ma</name>
</author>
<author>
<name>Yuping Lin</name>
</author>
<author>
<name>Yuting Yan</name>
</author>
<author>
<name>Yutong Xie</name>
</author>
<author>
<name>Yuwen Sheng</name>
</author>
<author>
<name>Yuxuan Zhu</name>
</author>
<author>
<name>Zekai Zhang</name>
</author>
<author>
<name>Zhe Ju</name>
</author>
<author>
<name>Zhenzhen Lin</name>
</author>
<author>
<name>Zheren Gao</name>
</author>
<author>
<name>Zheyang Sun</name>
</author>
<author>
<name>Zhigang Yan</name>
</author>
<author>
<name>Zhongyu Wu</name>
</author>
<author>
<name>Zi Wang</name>
</author>
<author>
<name>Zihua Qu</name>
</author>
<author>
<name>Ziling Yan</name>
</author>
<author>
<name>Ziyi Wan</name>
</author>
</entry>
<entry>
<id>http://arxiv.org/abs/2609.17248v1</id>
<title>Video-HolmesV2: Can MLLMs Reason with Spatio-Temporal Audio-Visual Evidence in Long Videos?</title>
<updated>2026-09-15T14:30:00Z</updated>
<link href="https://arxiv.org/abs/2609.17248v1" rel="alternate" type="text/html"/>
<link href="https://arxiv.org/pdf/2609.17248v1" rel="related" type="application/pdf" title="pdf"/>
<summary>Multimodal Large Language Models have demonstrated impressive video understanding, yet their ability to reason over long-form narratives is often masked by visual-centric evaluations and inefficient context processing. Existing benchmarks over-rely on visual heuristics while marginalizing auditory cues, effectively reducing models to "silent observers" that bypass genuine cross-modal reasoning. Moreover, standard dense sampling creates an evidence-context trade-off: increasing frames to capture evidence inevitably leads to attention distraction and token explosion. To bridge these gaps, we present Video-HolmesV2, a novel benchmark designed for Deep Audio-Visual Coupling. Unlike previous works, it enforces an Evidence-Based Evaluation, requiring models to justify answers with precise spatio-temporal audio-visual evidence, thereby reducing confounding effects of guessing and hallucinated evidence. To support this, we introduce: (1) a Multi-Model Cross-Verification pipeline to ensure task rigor; (2) a Spatio-temporal Evidence-Aware Metric for fine-grained calibration. Furthermore, we propose an Audio-Text Guided Token Compression framework. By fusing task intent with auditory anchors, our method distills high-value reasoning cues to mitigate long-context noise. In our evaluation, even strong proprietary models achieve below 60% accuracy, while our approach outperforms comparable open-source omni-models.</summary>
<category term="cs.CV" scheme="http://arxiv.org/schemas/atom"/>
<published>2026-09-15T14:30:00Z</published>
<arxiv:comment>Accepted by ECCV2026</arxiv:comment>
<arxiv:primary_category term="cs.CV"/>
<author>
<name>Zhaoyang Wei</name>
</author>
<author>
<name>Zipeng Wang</name>
</author>
<author>
<name>Yushe Cao</name>
</author>
<author>
<name>Chenhui Qiang</name>
</author>
<author>
<name>Shuaibing Cheng</name>
</author>
<author>
<name>Xuesong Yang</name>
</author>
<author>
<name>Sen Nie</name>
</author>
<author>
<name>Bowen Jiang</name>
</author>
<author>
<name>Wenchao Ding</name>
</author>
<author>
<name>Yanchao Hao</name>
</author>
<author>
<name>Zheng Wei</name>
</author>
<author>
<name>Xuehui Yu</name>
</author>
<author>
<name>Zhenjun Han</name>
</author>
</entry>
<entry>
<id>http://arxiv.org/abs/2609.04778v1</id>
<title>Diffusion Language Models for Mobile Edge Agentic AI: Foundations, Applications, and Challenges</title>
<updated>2026-09-04T06:16:49Z</updated>
<link href="https://arxiv.org/abs/2609.04778v1" rel="alternate" type="text/html"/>
<link href="https://arxiv.org/pdf/2609.04778v1" rel="related" type="application/pdf" title="pdf"/>
<summary>Diffusion language models (DLMs) offer a non-autoregressive alternative for mobile edge agentic artificial intelligence (AI) by refining tokens through iterative denoising rather than left-to-right decoding. Compared with autoregressive Transformer-based large language models (LLMs), DLMs can update multiple uncertain tokens in parallel and exploit bidirectional context throughout the generation process, enabling more flexible quality-latency trade-offs beyond fixed sequential decoding. These properties are particularly attractive for edge agents, where partial refinement, early exit, and constraint-guided correction can reduce response delay and communication overhead while improving robustness under noisy, incomplete, or dynamic contexts. This survey reviews DLM foundations and analyzes their suitability for edge settings under latency, memory, energy, bandwidth, privacy, and reliability constraints. We cover resource-efficient architectures, training and inference acceleration, compression, edge/cloud deployment, communication-aware serving, Internet of Things (IoT)/wireless applications, and evaluation of DLM-based agents. We further discuss open issues in long-context state management, split inference, trustworthy execution, multimodal grounding, and reproducible benchmarking. The goal is to connect DLM modeling properties, including bidirectionality, parallel refinement, controllability, and quality-latency elasticity, with system-level requirements of future mobile edge intelligence.</summary>
<category term="cs.AI" scheme="http://arxiv.org/schemas/atom"/>
<published>2026-09-04T06:16:49Z</published>
<arxiv:primary_category term="cs.AI"/>
<author>
<name>Chenqi Li</name>
</author>
<author>
<name>Minghui Min</name>
</author>
<author>
<name>Dusit Niyato</name>
</author>
<author>
<name>Wei Ni</name>
</author>
</entry>
<entry>
<id>http://arxiv.org/abs/2608.24063v1</id>
<title>VisCache: Visual KV Cache Pruning for Efficient Vision Large Language Model Inference</title>
<updated>2026-08-25T04:52:05Z</updated>
<link href="https://arxiv.org/abs/2608.24063v1" rel="alternate" type="text/html"/>
<link href="https://arxiv.org/pdf/2608.24063v1" rel="related" type="application/pdf" title="pdf"/>
<summary>While Vision Large Language Models (VLLMs) have achieved remarkable success in multimodal reasoning, their long-context inference remains prohibitively expensive due to the massive computation and memory overhead of visual Key-Value (KV) caches. Existing KV compression methods often apply uniform pruning across visual tokens and layers, leading to substantial information loss and degraded performance.To address this challenge, we propose \textbf{VisCache}, a plug-and-play framework for coarse-to-fine \textbf{Vis}ual KV \textbf{Cache} pruning without training, which consists of two synergistic stages. First, a lightweight VLM filters temporal redundancy by selectively forwarding semantically informative keyframes. Second, we introduce {PruneKV}, a surgical KV compression algorithm tailored to the attention dynamics of VLLMs. Unlike rigid pruning strategies, PruneKV adopts a parabolic layer-wise budget allocation together with an asymmetric update mechanism that selectively prunes keys while fusing values, thereby preserving critical contextual information. Extensive experiments demonstrate that VisCache substantially improves inference efficiency, achieving up to {2.35$\times$ speedup} and significant memory reduction while maintaining competitive performance with only {19--28\%} KV cache retention. VisCache consistently outperforms existing baselines, establishing a new Pareto frontier between efficiency and performance for long-context VLLM inference. Code is available at https://github.com/Wlklk/VisCache</summary>
<category term="cs.CV" scheme="http://arxiv.org/schemas/atom"/>
<category term="cs.AI" scheme="http://arxiv.org/schemas/atom"/>
<published>2026-08-25T04:52:05Z</published>
<arxiv:primary_category term="cs.CV"/>
<author>
<name>Lyuke Wang</name>
</author>
<author>
<name>Zhuo Li</name>
</author>
<author>
<name>Guangxu Zhu</name>
</author>
</entry>
<entry>
<id>http://arxiv.org/abs/2608.15056v1</id>
<title>GraphLoom: Reliability-Calibrated Graph Evidence Routing for Multimodal KG-RAG</title>
<updated>2026-08-15T06:02:23Z</updated>
<link href="https://arxiv.org/abs/2608.15056v1" rel="alternate" type="text/html"/>
<link href="https://arxiv.org/pdf/2608.15056v1" rel="related" type="application/pdf" title="pdf"/>
<summary>Multimodal retrieval-augmented generation (RAG) systems often rely on long unstructured contexts or aggressively expanded evidence graphs, which can introduce noisy evidence, weaken multi-hop reasoning, and increase unsupported generation. We present GraphLoom, a reliability-calibrated multimodal knowledge-graph RAG framework for compact and faithful evidence routing. Given a question and its associated multimodal input, GraphLoom constructs an instance-level multimodal knowledge graph from grounded scene descriptions, extracted relational triples, and external commonsense knowledge. Instead of injecting all retrieved evidence into the generator, GraphLoom performs reliability-aware subgraph retrieval with bounded expansion and selectively routes high-utility evidence through hierarchical graph memory slots and joint graph-sequence attention in a frozen language model. To improve robustness in complex reasoning settings, GraphLoom further combines interleaved retrieval with budgeted corrective retrieval, enabling adaptive multi-hop evidence refinement under noisy retrieval conditions. We evaluate GraphLoom on ScienceQA, MultiModalQA, and OK-VQA, including large distractor evidence pools that approximate noisy external knowledge retrieval. Experimental results show consistent gains in answer quality and evidence faithfulness over strong multimodal RAG, graph-retrieval, and open-source vision-language baselines, with improved retrieval quality on MultiModalQA and stable performance under noisy evidence pools. Additional analyses using MiniCheck-based verification, human evaluation, and latency profiling show that reliability-calibrated graph evidence routing provides an effective alternative to long-context multimodal evidence injection.</summary>
<category term="cs.AI" scheme="http://arxiv.org/schemas/atom"/>
<published>2026-08-15T06:02:23Z</published>
<arxiv:primary_category term="cs.AI"/>
<author>
<name>Zafar Ali</name>
</author>
<author>
<name>Asad Khan</name>
</author>
<author>
<name>Aalia Malik</name>
</author>
<author>
<name>Pavlos Kefalas</name>
</author>
</entry>
<entry>
<id>http://arxiv.org/abs/2608.13426v2</id>
<title>Reduced Matrix Multiplication: Input-Adaptive Matrix-Product Reduction for LLM Inference</title>
<updated>2026-08-31T13:12:05Z</updated>
<link href="https://arxiv.org/abs/2608.13426v2" rel="alternate" type="text/html"/>
<link href="https://arxiv.org/pdf/2608.13426v2" rel="related" type="application/pdf" title="pdf"/>
<summary>Transformer-based language models achieve strong performance but incur substantial inference cost due to repeated high-dimensional matrix multiplications. We propose Reduced Matrix Multiplication (RMM), a training-free, input-adaptive inference method that reduces Transformer matrix products by selecting informative slices along their contraction dimensions, without modifying model weights. Under a simple retention-ratio control, RMM provides a smooth and predictable accuracy-efficiency trade-off. Across language models ranging from 1B to 70B parameters, we find that reduction tolerance depends on the model family, task, component, and retention ratio, although it often improves with model scale. Under moderate reduction, RMM remains robust across the evaluated discriminative, autoregressive generation, and long-context settings. We further show that the same principle extends to multimodal vision-language inference. Mechanistic ablations reveal a structural asymmetry within Transformers: attention-side computations are substantially more reducible than MLP components. Finally, wall-clock benchmarks with custom kernels on an NVIDIA A100 show that these computational savings can translate into practical runtime gains, especially at longer sequence lengths. Together, these results position RMM as a scalable direction for input-adaptive inference-time optimization.</summary>
<category term="cs.LG" scheme="http://arxiv.org/schemas/atom"/>
<category term="cs.AI" scheme="http://arxiv.org/schemas/atom"/>
<category term="cs.CL" scheme="http://arxiv.org/schemas/atom"/>
<published>2026-08-13T16:16:04Z</published>
<arxiv:primary_category term="cs.LG"/>
<author>
<name>Zixuan Lan</name>
</author>
<author>
<name>Yanhong Li</name>
</author>
<author>
<name>Jiawei Zhou</name>
</author>
</entry>
<entry>
<id>http://arxiv.org/abs/2608.12627v3</id>
<title>EgoCITE: Context-Augmented Indexing and Time-Aware Retrieval for Long-Horizon Egocentric Memory</title>
<updated>2026-08-18T22:47:18Z</updated>
<link href="https://arxiv.org/abs/2608.12627v3" rel="alternate" type="text/html"/>
<link href="https://arxiv.org/pdf/2608.12627v3" rel="related" type="application/pdf" title="pdf"/>
<summary>Long-horizon egocentric memory transforms continuous first-person video and audio into a searchable record of past experiences. We demonstrate two bottlenecks in existing systems: indices built from context-poor captions are unreliable for agentic search, while retrieval ignores a question's temporal intent. To address both bottlenecks, we introduce EgoCITE (Egocentric Context-augmented Indexing and Time-aware Evidence retrieval), a long-horizon agentic memory framework for egocentric QA. EgoCITE comprises three components. EgoScheme uses local multimodal context to turn fragmentary video captions and speech transcripts into self-contained atomic memory indices. EgoIndex organizes complementary action, activity, utterance, and conversation representations into searchable multi-view memory indices at multiple granularities. EgoRetrv combines semantic search with question-conditioned temporal relevance scoring and curation of retrieved evidence. We evaluate EgoCITE on EgoLifeQA, EgoMem, and EgoR1-Bench in terms of answer accuracy and target-event retrieval alignment. EgoCITE improves accuracy over agentic memory baselines by at least 4.4--14.2% while achieving 36$\times$ lower cost than long-context LLM agents.</summary>
<category term="cs.CV" scheme="http://arxiv.org/schemas/atom"/>
<category term="cs.AI" scheme="http://arxiv.org/schemas/atom"/>
<category term="cs.CL" scheme="http://arxiv.org/schemas/atom"/>
<category term="cs.HC" scheme="http://arxiv.org/schemas/atom"/>
<published>2026-08-12T22:20:29Z</published>
<arxiv:primary_category term="cs.CV"/>
<author>
<name>Le Zhang</name>
</author>
<author>
<name>Hao Chen</name>
</author>
<author>
<name>Vlad Roznyatovskiy</name>
</author>
<author>
<name>Jianzhong Zhang</name>
</author>
<author>
<name>Ke Sun</name>
</author>
</entry>
<entry>
<id>http://arxiv.org/abs/2608.08630v1</id>
<title>VLZip: Unified Visual and Textual Compression for Interleaved Long-Context Modeling</title>
<updated>2026-08-09T10:31:03Z</updated>
<link href="https://arxiv.org/abs/2608.08630v1" rel="alternate" type="text/html"/>
<link href="https://arxiv.org/pdf/2608.08630v1" rel="related" type="application/pdf" title="pdf"/>
<summary>Vision Language Models (VLMs) face significant challenges with ultra-long, interleaved image-text sequences due to the quadratic complexity of self-attention. Current solutions either resort to aggressive token pruning, risking irreversible information loss, or adopt efficient but less precise architectures, while largely ignoring the equally vital textual component. We introduce VLZip, a framework that unifies visual and textual compression for high-fidelity reasoning within a pure Transformer. At its core, VLZip hierarchically distills visual and textual segments into compact, layer-specific "soft prefixes" and injects them into each decoder layer's hidden states, drastically shortening the attention sequence while preserving fine-grained global context. To address deficient evaluations in the field, we also introduce LongVLBench, a new benchmark derived from video narratives that demands holistic, narrative-level reasoning. Extensive experiments show VLZip achieves leading performance on long-context multimodal reasoning, enabling training up to 120K tokens, a 6x increase over the baseline, and inference beyond 280K tokens with significantly reduced memory, while demonstrating the memory scalability to handle up to 2M tokens. By excelling at extreme context lengths where existing methods collapse, VLZip establishes an efficient and powerful new standard for long-context multimodal AI. Code is available at https://github.com/ShareLab-SII/VLZip.</summary>
<category term="cs.CV" scheme="http://arxiv.org/schemas/atom"/>
<published>2026-08-09T10:31:03Z</published>
<arxiv:primary_category term="cs.CV"/>
<author>
<name>Yuqi Zhang</name>
</author>
<author>
<name>Cheng Chen</name>
</author>
<author>
<name>Yuyu Guo</name>
</author>
<author>
<name>Wenjie Yang</name>
</author>
<author>
<name>Lingchen Meng</name>
</author>
<author>
<name>Peng Di</name>
</author>
<author>
<name>Hang Yu</name>
</author>
<author>
<name>Zuxuan Wu</name>
</author>
<author>
<name>Yu-Gang Jiang</name>
</author>
</entry>
<entry>
<id>http://arxiv.org/abs/2608.06122v2</id>
<title>Is Self-Pretraining really useful to improve diagnosis in medical Time Series?</title>
<updated>2026-08-07T08:30:22Z</updated>
<link href="https://arxiv.org/abs/2608.06122v2" rel="alternate" type="text/html"/>
<link href="https://arxiv.org/pdf/2608.06122v2" rel="related" type="application/pdf" title="pdf"/>
<summary>Inspired by recent evidence that transformer architectures benefit from Self-PreTraining (SPT) on long-context benchmarks, we investigate whether similar gains extend to multimodal, multivariate, and even simple univariate medical time series. Our objective is to assess the impact of SPT on the performance and scalability of transformer-based models across diverse medical applications, particularly under limited data conditions. We evaluate transformer architectures on three representative medical time-series tasks: rehabilitation robotics (Camargo dataset), stress detection (Non-EEG Stress), and Parkinson's disease detection (Gait Parkinson's Disease). Models are trained either from scratch or through SPT using four masking-based objectives designed to promote temporal and cross-modal representation learning, and we systematically vary model depth to examine how capacity interacts with pre-training benefits. Across datasets and configurations, SPT consistently improves classification accuracy by 0-6 percentage points depending on masking strategy, dataset and architecture, with gains observed not only in multivariate settings but also when models are restricted to simple univariate inputs. The improvements increase for deeper models that can better exploit the enriched temporal representations learned during pre-training. These findings indicate that SPT is a simple and general strategy that enhances transformer performance on medical time-series tasks without requiring task-specific architectural changes, supporting its potential to improve robustness and accuracy in data-limited clinical settings.</summary>
<category term="cs.LG" scheme="http://arxiv.org/schemas/atom"/>
<category term="cs.AI" scheme="http://arxiv.org/schemas/atom"/>
<published>2026-08-06T14:53:23Z</published>
<arxiv:comment>21 pages, 7 figures,4 tables</arxiv:comment>
<arxiv:primary_category term="cs.LG"/>
<author>
<name>Omar Coser</name>
</author>
<author>
<name>Antonio Orvieto</name>
</author>
<author>
<name>Paolo Soda</name>
</author>
<author>
<name>Loredana Zollo</name>
</author>
</entry>
<entry>
<id>http://arxiv.org/abs/2608.01848v1</id>
<title>Decoupling semantics from vision: A framework for faithful visual-text compression evaluation</title>
<updated>2026-08-03T07:56:43Z</updated>
<link href="https://arxiv.org/abs/2608.01848v1" rel="alternate" type="text/html"/>
<link href="https://arxiv.org/pdf/2608.01848v1" rel="related" type="application/pdf" title="pdf"/>
<summary>Recent visual-text compression (VTC) methods, typified by DeepSeek-OCR, report impressive high token compression ratios for long-context modeling tasks by leveraging text-to-image rendering. However, existing evaluation protocols heavily rely on downstream task performance. Such evaluation metrics fail to accurately measure text preservation due to the strong inherent linguistic priors of Multimodal Large Language Models (MLLMs). In this work, we introduce a new evaluation framework that decouples MLLMs' capabilities to faithfully assess VTC quality. Within this framework, we further introduce the ZeroSense Benchmark to ensure low semantic correlation of testing samples. By eliminating textual dependencies, our benchmark guarantees that the evaluation results are purely reflective of VTC quality, unaffected by the semantic inference capabilities of downstream models. Extensive experiments across multiple datasets demonstrate that VTC quality and downstream task accuracy diverge significantly, highlighting the necessity of our decoupled evaluation framework.</summary>
<category term="cs.CV" scheme="http://arxiv.org/schemas/atom"/>
<published>2026-08-03T07:56:43Z</published>
<arxiv:primary_category term="cs.CV"/>
<author>
<name>Yonghan Gao</name>
</author>
<author>
<name>Zehong Chen</name>
</author>
<author>
<name>Lijian Xu</name>
</author>
<author>
<name>Jingzhi Chen</name>
</author>
<author>
<name>Jingwei Guan</name>
</author>
<author>
<name>Xingyu Zeng</name>
</author>
</entry>
<entry>
<id>http://arxiv.org/abs/2608.01314v1</id>
<title>Remember-R1: Mitigating Long-Context Visual Forgetting through Reinforcement Learning</title>
<updated>2026-08-02T15:31:31Z</updated>
<link href="https://arxiv.org/abs/2608.01314v1" rel="alternate" type="text/html"/>
<link href="https://arxiv.org/pdf/2608.01314v1" rel="related" type="application/pdf" title="pdf"/>
<summary>Multimodal large language models (MLLMs) increasingly rely on long chain-of-thought reasoning for complex tasks. However, as reasoning sequences lengthen, models may gradually rely less on visual evidence and more on accumulated textual context, leading to visual forgetting. Existing approaches do not directly constrain how visual evidence is used and maintained along the original reasoning trajectory, leaving long-context visual forgetting insufficiently addressed. To address this issue, we propose Remember-R1, a reinforcement learning framework that mitigates long-context visual forgetting by applying process-level supervision directly on the original reasoning trajectory. Specifically, Remember-R1 introduces rewards that encourage broader coverage of matched visual keywords, stronger persistence of visual dependence in later reasoning steps, and greater focus on question-relevant image regions. Experiments across multiple model scales and diverse multimodal benchmarks demonstrate that Remember-R1 consistently improves reasoning performance. Additional analyses further show that it slows the decline of visual attention during generation, supporting its effectiveness in mitigating long-context visual forgetting.</summary>
<category term="cs.CV" scheme="http://arxiv.org/schemas/atom"/>
<published>2026-08-02T15:31:31Z</published>
<arxiv:comment>Accepted by ACM Multimedia 2026</arxiv:comment>
<arxiv:primary_category term="cs.CV"/>
<author>
<name>Jianmin Chen</name>
</author>
<author>
<name>Jiaqi Tang</name>
</author>
<author>
<name>Wei Wei</name>
</author>
<author>
<name>Xiaogang Xu</name>
</author>
<author>
<name>Jiafei Wu</name>
</author>
<author>
<name>Zhe Liu</name>
</author>
<author>
<name>Qianzhou Wang</name>
</author>
<author>
<name>Yingying Yan</name>
</author>
<author>
<name>Botong Geng</name>
</author>
<author>
<name>Yuyang Xia</name>
</author>
<author>
<name>Lei Zhang</name>
</author>
<author>
<name>Qifeng Chen</name>
</author>
</entry>
<entry>
<id>http://arxiv.org/abs/2607.29596v1</id>
<title>FibVLA: An Efficient Temporal Vision-Language-Action Model with Fibonacci Sampling</title>
<updated>2026-07-31T16:23:07Z</updated>
<link href="https://arxiv.org/abs/2607.29596v1" rel="alternate" type="text/html"/>
<link href="https://arxiv.org/pdf/2607.29596v1" rel="related" type="application/pdf" title="pdf"/>
<summary>Vision-language-action models (VLAs), which leverage the cognition of multimodal information to infer physical-world actions, provide a generalized solution for embodied AI applications. Conventional VLAs usually concentrate on current digital cognition. While some efforts are made to enhance VLAs' reasoning capabilities by capturing temporal information, encoding the long-context history causes an efficiency-decreasing issue. To reconcile the conflict between capturing temporal information and maintaining inference efficiency in VLAs, this paper introduces FibVLA, an efficient framework featuring temporal perception of long-context history. Specifically, we leverage logarithmic hindsight sampling to both proprioceptive states and visual frames to capture long-term temporal dependencies with minimal redundancy. For the action expert, we introduce the flow matching to produce action distributions, and the Fibonacci recurrent inference strategy to generate long-range planning steps based on real-time closed-loop feedback. Experiments demonstrate that FibVLA significantly improves action smoothness and success rates without retraining large-scale visual encoders. Efficiency analysis demonstrates superior real-time responsiveness compared to video-based baselines in real-world evaluations.</summary>
<category term="cs.RO" scheme="http://arxiv.org/schemas/atom"/>
<category term="cs.CV" scheme="http://arxiv.org/schemas/atom"/>
<published>2026-07-31T16:23:07Z</published>
<arxiv:primary_category term="cs.RO"/>
<author>
<name>Li Lin</name>
</author>
<author>
<name>Wujun Xu</name>
</author>
<author>
<name>Weiwei Meng</name>
</author>
<author>
<name>Kaiwen Xia</name>
</author>
<author>
<name>Kang Hao Cheong</name>
</author>
<author>
<name>Shuai Wang</name>
</author>
</entry>
<entry>
<id>http://arxiv.org/abs/2608.00146v1</id>
<title>DiffusionGemma Technical Report</title>
<updated>2026-07-31T16:11:46Z</updated>
<link href="https://arxiv.org/abs/2608.00146v1" rel="alternate" type="text/html"/>
<link href="https://arxiv.org/pdf/2608.00146v1" rel="related" type="application/pdf" title="pdf"/>
<summary>We introduce DiffusionGemma, an experimental open-weight language model that uses discrete diffusion to generate text at exceptionally high speed. Rather than decoding one token at a time, DiffusionGemma iteratively refines blocks of 256 tokens in parallel, avoiding the sequential decoding bottleneck of conventional autoregressive (AR) large language models. Instead of training from scratch, we obtain DiffusionGemma by fine-tuning the mixture-of-experts Gemma 4 model with 3.8B activated and 25.2B total parameters. Our compute-efficient two-stage training pipeline uses fewer than 10% of the starting AR model's total training token budget. The first stage uses supervised fine-tuning to teach bidirectional denoising, while the second stage combines reinforcement learning with sampler distillation to jointly improve generation quality and inference efficiency. DiffusionGemma establishes a new Pareto frontier for the trade-off between generation speed and model capability. Averaged across our full evaluation suite, it generates around 20 tokens per forward pass and achieves roughly 1,500 output tokens per second on a single NVIDIA H100 GPU, which is substantially faster than AR models even with state-of-the-art speculative decoding. DiffusionGemma also retains the starting model's support for thinking mode, multimodal inputs, and long contexts. Despite diffusion fine-tuning, it remains capable of AR generation with only minor performance degradation, suggesting a path toward hybrid diffusion-AR decoding.</summary>
<category term="cs.CL" scheme="http://arxiv.org/schemas/atom"/>
<category term="cs.AI" scheme="http://arxiv.org/schemas/atom"/>
<published>2026-07-31T16:11:46Z</published>
<arxiv:primary_category term="cs.CL"/>
<author>
<name> DiffusionGemma Team</name>
</author>
<author>
<name>Adrien Ali Taïga</name>
</author>
<author>
<name>James Assiene</name>
</author>
<author>
<name>Daniele Calandriello</name>
</author>
<author>
<name>Rahma Chaabouni</name>
</author>
<author>
<name>João Gante</name>
</author>
<author>
<name>Tamara von Glehn</name>
</author>
<author>
<name>Nate Keating</name>
</author>
<author>
<name>Chris Knutsen</name>
</author>
<author>
<name>Martin Kukla</name>
</author>
<author>
<name>Tianlin Liu</name>
</author>
<author>
<name>Ivan Lobov</name>
</author>
<author>
<name>Ofir Nabati</name>
</author>
<author>
<name>João Gabriel Oliveira</name>
</author>
<author>
<name>Nicolas Perez-Nieves</name>
</author>
<author>
<name>Nastasia Prutianova</name>
</author>
<author>
<name>Bobak Shahriari</name>
</author>
<author>
<name>Jean Tarbouriech</name>
</author>
<author>
<name>Pavel Tyletski</name>
</author>
<author>
<name>Çağlar Ünlü</name>
</author>
<author>
<name>Cindy Wu</name>
</author>
<author>
<name>Glenn Cameron</name>
</author>
<author>
<name>Jerome Connor</name>
</author>
<author>
<name>Sertan Girgin</name>
</author>
<author>
<name>Maarten Grootendorst</name>
</author>
<author>
<name>Alon Levkovitch</name>
</author>
<author>
<name>Eliya Nachmani</name>
</author>
<author>
<name>Omar Sanseviero</name>
</author>
<author>
<name>Piotr Stanczyk</name>
</author>
<author>
<name>Quentin Berthet</name>
</author>
<author>
<name>Andrew Campbell</name>
</author>
<author>
<name>Clément Crepy</name>
</author>
<author>
<name>Valentin De Bortoli</name>
</author>
<author>
<name>Arnaud Doucet</name>
</author>
<author>
<name>Romuald Elie</name>
</author>
<author>
<name>Alexandre Galashov</name>
</author>
<author>
<name>Klaus Greff</name>
</author>
<author>
<name>Alexis Jacq</name>
</author>
<author>
<name>David Ruhe</name>
</author>
<author>
<name>Yu-Han Wu</name>
</author>
<author>
<name>Sebastian Flennerhag</name>
</author>
<author>
<name>Brendan O'Donoghue</name>
</author>
<author>
<name>George Scrivener</name>
</author>
<author>
<name>Shantanu Thakoor</name>
</author>
</entry>
<entry>
<id>http://arxiv.org/abs/2607.28611v1</id>
<title>Chimera: Designing and Chinchilla-Scaling Hybrid Visual Diffusion Transformers</title>
<updated>2026-07-30T17:58:02Z</updated>
<link href="https://arxiv.org/abs/2607.28611v1" rel="alternate" type="text/html"/>
<link href="https://arxiv.org/pdf/2607.28611v1" rel="related" type="application/pdf" title="pdf"/>
<summary>Visual generation increasingly requires high-resolution images, long videos, and multimodal context, making the quadratic cost of full attention prohibitive. We introduce Chimera, a hybrid visual diffusion backbone with a principled scaling recipe. Chimera processes text, image, and video tokens in one raster-ordered stream without positional embeddings. It combines Kimi Delta Attention (KDA) for long-context state tracking with O(N) complexity, interleaved Multi-head Latent Attention (MLA) for direct global interaction, and modality-aware short convolutions for local spatiotemporal context. Sparse Mixture-of-Experts (MoE) layers expand capacity while controlling activated compute. To scale this heterogeneous architecture, we introduce HeteroP, a module-wise scheme that transfers hyperparameters across width and depth according to each tensor's functional fan-in and model depth. HeteroP yields a consistently tuned family used to fit Chinchilla-style compute-optimal laws for activated model size, training-token count, and image-video data ratio. Guided by these laws, we train an 11B-parameter Chimera with 2B activated parameters. Experiments show three results. First, measured by pretraining diffusion loss, the dense backbone is 1.7x as compute-efficient as a matched full-attention Wan-2.1 2B baseline, while the complete system reaches 7.3x. Second, without length-specific fine-tuning, Chimera extrapolates zero-shot from 5-second training clips to 30-second videos, with only 6.5% FID degradation in the last five seconds. Third, the fitted laws show that compute-optimal image pretraining divides compute nearly evenly between activated model size and training-token count, whereas video pretraining modestly favors model size at higher budgets. These results establish a foundation for designing and scaling efficient long-context diffusion architectures.</summary>
<category term="cs.CV" scheme="http://arxiv.org/schemas/atom"/>
<published>2026-07-30T17:58:02Z</published>
<arxiv:comment>40 pages</arxiv:comment>
<arxiv:primary_category term="cs.CV"/>
<author>
<name>Chongjian Ge</name>
</author>
<author>
<name>Hanwen Jiang</name>
</author>
<author>
<name>Tianyu Wang</name>
</author>
<author>
<name>Jiuxiang Gu</name>
</author>
<author>
<name>Yiran Xu</name>
</author>
<author>
<name>Ziwen Chen</name>
</author>
<author>
<name>Shaoteng Liu</name>
</author>
<author>
<name>Jing Shi</name>
</author>
<author>
<name>Yicong Hong</name>
</author>
<author>
<name>Zefan Cai</name>
</author>
<author>
<name>Hailin Jin</name>
</author>
<author>
<name>Hao Tan</name>
</author>
</entry>
<entry>
<id>http://arxiv.org/abs/2607.24260v1</id>
<title>KAP: Bridging the Knowledge Selection-Runtime Consumption Gap in LLM Systems</title>
<updated>2026-07-27T10:51:38Z</updated>
<link href="https://arxiv.org/abs/2607.24260v1" rel="alternate" type="text/html"/>
<link href="https://arxiv.org/pdf/2607.24260v1" rel="related" type="application/pdf" title="pdf"/>
<summary>Modern LLM systems increasingly rely on knowledge-selection processes that produce high-value structured priors, such as ranked evidence, graph topology, multimodal alignment, and confidence signals. Yet LLM serving remains fundamentally oblivious to this rich structure: once such signals are serialized into a prompt, the backend observes only a flat token sequence, forcing dense and uniform consumption of the full key-value (KV) state during decoding. We term this architectural mismatch the Knowledge Selection-Runtime Consumption (KSRC) gap: richer contexts enlarge the full-prompt KV footprint and decode-time memory traffic, increasing latency and degrading throughput even when reasoning depends on only a small fraction of the context. To bridge the gap, we propose Knowledge Access Planning (KAP), a paradigm-shifting execution abstraction that elevates structured knowledge priors from passive prompt-construction hints into first-class physical execution artifacts. KAP establishes a universal intermediate representation (IR)-the runtime access plan-which compiles structured knowledge signals to govern physical KV access without altering logical prompt semantics, model weights, or training procedures. Through this IR, KAP shifts LLM serving from token-aware context consumption to plan-driven, knowledge-aware runtime consumption. We instantiate KAP with GraphSpec, a compiler-executor realization connecting structured knowledge selection to an LLM serving backend. We derive a phase-boundary model for the positive-speedup regime of plan-guided execution. Across 4K-128K long-context QA workloads, GraphSpec maintains answer quality comparable to full-context decoding while decoupling physical KV consumption from prompt length, reducing proposal-time KV access to 5.5% of source KV state at 128K, and fundamentally shifting the scaling trajectory of long-context generation.</summary>
<category term="cs.LG" scheme="http://arxiv.org/schemas/atom"/>
<published>2026-07-27T10:51:38Z</published>
<arxiv:comment>3 figures, 5 tables</arxiv:comment>
<arxiv:primary_category term="cs.LG"/>
<author>
<name>Shuo Wang</name>
</author>
<author>
<name>Fang Xi</name>
</author>
<author>
<name>Wenyuan Huang</name>
</author>
<author>
<name>Qing Wang</name>
</author>
<author>
<name>Junming Su</name>
</author>
</entry>
<entry>
<id>http://arxiv.org/abs/2607.19947v1</id>
<title>ETPDesigner: Multi-Agent Orchestration for Interactive Multimodal Electronic Theater Program</title>
<updated>2026-07-22T09:25:03Z</updated>
<link href="https://arxiv.org/abs/2607.19947v1" rel="alternate" type="text/html"/>
<link href="https://arxiv.org/pdf/2607.19947v1" rel="related" type="application/pdf" title="pdf"/>
<summary>Electronic Theater Programs (ETPs) serve as critical promotional media in the performing arts, comprising a multi-page collection of heterogeneous visual assets such as theatrical posters, performance details, and character portraits. However, existing text-to-image paradigms struggle with such complex design tasks due to their inability to comprehend long-context narratives and maintain visual consistency across multiple distinct pages. To address this, we introduce ETPDesigner, a collaborative Multi-Agent framework that directly synthesizes high-quality ETPs from raw dramatic scripts. Emulating a professional design pipeline, our framework orchestrates specialized agents for semantic script analysis, core poster synthesis, functional background generation, and the stratified composition of character assets. Central to ETPDesigner is a global style anchor mechanism that extracts visual priors from the core poster to enforce strict aesthetic uniformity across all generated components. Furthermore, we elevate the ETP from a static publication to an immersive interactive companion. By integrating portrait animation, customized speech synthesis, and persona-grounded Large Language Models (LLMs), our system enables users to engage in real-time, voice-enabled conversations with the generated virtual characters. To rigorously benchmark this task, we construct ETP-Pro, a domain-specific benchmark of professional theater posters and high-quality character portraits. Extensive evaluations demonstrate our method's superiority in producing semantically faithful, aesthetically consistent, and highly interactive program sets.</summary>
<category term="cs.CV" scheme="http://arxiv.org/schemas/atom"/>
<published>2026-07-22T09:25:03Z</published>
<arxiv:primary_category term="cs.CV"/>
<author>
<name>Mengtian Li</name>
</author>
<author>
<name>Xinru Guo</name>
</author>
<author>
<name>Xiaoru Lin</name>
</author>
<author>
<name>Xiao Rong</name>
</author>
<author>
<name>Zhifeng Xie</name>
</author>
<author>
<name>Chaofeng Chen</name>
</author>
</entry>
<entry>
<id>http://arxiv.org/abs/2607.14658v1</id>
<title>TopoAgent: A Self-Evolving Topological Agent for Multimodal Scientific Reasoning</title>
<updated>2026-07-16T07:22:59Z</updated>
<link href="https://arxiv.org/abs/2607.14658v1" rel="alternate" type="text/html"/>
<link href="https://arxiv.org/pdf/2607.14658v1" rel="related" type="application/pdf" title="pdf"/>
<summary>While Multimodal Large Language Models (MLLMs) excel in general tasks, rigorous scientific reasoning remains challenging due to the limitations of monolithic, linear planning. Such sequential designs often suffer from visual-semantic misalignment, long-context hallucinations, and brittle execution under fixed task granularity. We propose TopoAgent, a self-evolving topological framework that replaces linear trajectories with dynamic, state-isolated graph evolution. TopoAgent first employs a front-end decomposer to fracture complex queries into visually-grounded atoms. These atoms are organized into a Directed Acyclic Graph (DAG) based on their dependencies, enabling strict context isolation to shield the reasoning engine from irrelevant historical noise. Furthermore, we introduce adaptive atomic fission, which dynamically splits bottleneck nodes into finer-grained sub-atoms at runtime when tool capability boundaries are exceeded. Extensive experiments across mathematics, physics, and chemistry benchmarks demonstrate that TopoAgent significantly outperforms state-of-the-art linear agent frameworks, providing a robust, noise-resistant, and self-correcting paradigm for autonomous scientific reasoning.</summary>
<category term="cs.AI" scheme="http://arxiv.org/schemas/atom"/>
<published>2026-07-16T07:22:59Z</published>
<arxiv:primary_category term="cs.AI"/>
<author>
<name>Mingze Xu</name>
</author>
<author>
<name>Yinghui Li</name>
</author>
<author>
<name>Jiayi Kuang</name>
</author>
<author>
<name>Zhanhui Kang</name>
</author>
<author>
<name>Di Yin</name>
</author>
<author>
<name>Ying Shen</name>
</author>
<author>
<name>Xing Sun</name>
</author>
<author>
<name>Yuxing Han</name>
</author>
</entry>
<entry>
<id>http://arxiv.org/abs/2607.14568v1</id>
<title>A Modern Multimodal Assistant on a 6 GB 2011 GPU: Stage-Validated, All-GPU CUDA Inference for Fermi</title>
<updated>2026-07-16T04:48:44Z</updated>
<link href="https://arxiv.org/abs/2607.14568v1" rel="alternate" type="text/html"/>
<link href="https://arxiv.org/pdf/2607.14568v1" rel="related" type="application/pdf" title="pdf"/>
<summary>A companion study ran a 35B mixture-of-experts model on a 2011 NVIDIA Tesla C2075 (Fermi, sm_20, 6GB) as a GPU-prefill/CPU-decode hybrid, because the 4-bit model did not fit in device memory (arXiv:2606.24031). This report keeps the hardware and asks what a model that fits can do: we deploy MiniCPM-V-4.6, a modern multimodal assistant pairing a SigLIP2 vision encoder and window-attention merger (16x visual token compression) with a compact hybrid gated-delta-net backbone, entirely on the GPU. Three results. (i) An all-GPU engine built on measured foundations: projections that dequantize 8-bit weights once and call the vendor SGEMM still in the last Fermi toolchain (64% of FP32 peak; our best hand-written GEMM hit 37%, wrongly called the ceiling); a chunked delta-rule rewrite of the recurrent layers, 2.8x faster than the sequential scan once attribution exposed one bad kernel; and a measured negative: 4-bit weights make decode slower than 8-bit here, since Fermi issues nibble-unpacking shifts at half rate. (ii) The vision side is a port with a proof obligation: we translate tower, merger, and projector to sm_20 CUDA, validating every stage against a locally generated reference forward (full tower 1.4e-5). One failure, position-embedding bucketization differing on exact rational ties, generalizes to a rule: float tie-breaking in index arithmetic is implementation-defined; call the reference operator, do not reimplement it. (iii) Long context exposes an O(N^2) wall short benchmarks hide: prefill falls from 114 tok/s at 2k tokens to 21 at 10k in a naive attention kernel; per-head vendor-GEMM calls writing into the existing score buffer (zero extra memory) restore a flat profile (408 at 2k, 361 at 10k; 17x), verified by exact needle retrieval from 60% depth. The same rewrite cuts image encoding 6x, to 0.93s. The system answers an image question end-to-end in 1.7s.</summary>
<category term="cond-mat.other" scheme="http://arxiv.org/schemas/atom"/>
<category term="cs.AI" scheme="http://arxiv.org/schemas/atom"/>
<published>2026-07-16T04:48:44Z</published>
<arxiv:primary_category term="cond-mat.other"/>
<author>
<name>A. C. Opus</name>
</author>
<author>
<name>J. Q. Lu</name>
</author>
</entry>
</feed>
Reference in new issue View Git Blame Copy Permalink