Meta tags:
keywords=alphaxiv, arxiv, forum, discussion, explore, trending papers;
description=Discuss, discover, and read arXiv papers. Explore trending papers, see recent activity and discussions, and follow authors of arXiv papers on alphaXiv.;

Headings (most frequently used words):

ask, or, search, anything, alphaxiv, events, explore, sign, in, blog, feedback, upgrade, to, pro, dark, mode,

Text of the page (most frequently used words):
and (61), the (28), computer (26), blog (21), 2026 (21), audio (20), bookmark (20), science (20), view (20), may (20), for (18), models (17), agents (16), that (15), language (14), agentic (11), world (11), model (10), performance (9), learning (9), vision (9), action (9), diffusion (8), with (8), large (7), system (7), artificial (7), intelligence (7), developed (7), researchers (7), pattern (7), recognition (7), this (7), training (7), efficient (6), while (6), llm (6), framework (6), task (6), reinforcement (6), achieving (6), tokens (6), agent (6), video (6), visual (6), latent (6), attention (5), mechanisms (5), introduces (5), memory (5), achieved (5), zhang (5), methods (5), wang (5), outperforming (5), across (5), tasks (5), policy (5), from (5), significantly (5), frameworks (5), reasoning (5), search (5), into (4), state (4), benchmarks (4), designed (4), scale (4), enables (4), real (4), optimization (4), multi (4), human (4), integrates (4), computation (4), like (4), inference (4), token (4), generation (4), benchmark (4), university (4), explore (4), are (3), parameter (3), motion (3), hybrid (3), transformer (3), architecture (3), art (3), physical (3), chen (3), mit (3), complex (3), distilled (3), between (3), approach (3), costs (3), distillation (3), paradigm (3), quality (3), baselines (3), research (3), embodied (3), new (3), architectures (3), single (3), one (3), oriented (3), within (3), tool (3), method (3), liu (3), rope (3), long (3), fidelity (3), demonstrating (3), exploration (3), events (3), accuracy (3), grep (3), self (3), sign (2), more (2), matching (2), your (2), mem (2), online (2), mechanism (2), information (2), via (2), low (2), overall (2), compared (2), average (2), score (2), percentage (2), holomotion (2), horizon (2), robotics (2), humanoid (2), approximately (2), previous (2), time (2), control (2), technical (2), report (2), engram (2), based (2), including (2), dynamic (2), improving (2), machine (2), diffusionopd (2), effectively (2), capabilities (2), existing (2), both (2), embedded (2), flows (2), elf (2), competitive (2), fewer (2), than (2), yiyang (2), investigated (2), register (2), transformers (2), dits (2), pixel (2), space (2), registers (2), improves (2), scores (2), wams (2), represent (2), understanding (2), environment (2), their (2), comprehensive (2), structured (2), architectural (2), data (2), shi (2), frontier (2), autonomously (2), neural (2), optimizing (2), lin (2), aira (2), flash (2), grpo (2), step (2), reducing (2), alignment (2), dexjoco (2), toolkit (2), dexterous (2), manipulation (2), mujoco (2), actions (2), robustness (2), generalization (2), zhao (2), atlas (2), them (2), leading (2), physbrain (2), higher (2), success (2), rates (2), improvements (2), positions (2), contexts (2), study (2), specific (2), dila (2), disentangled (2), high (2), sana (2), open (2), minute (2), achieves (2), larger (2), meituan (2), you (2), futuresim (2), evaluate (2), adaptive (2), arvindh (2), arun (2), how (2), vggt (2), feed (2), zhengxi (2), pro (2), acm (2), alphaxiv (2), there, papers, filters, moment, 332, 115, augmenting, frozen, backbones, compressing, historical, fixed, size, associative, dynamically, steers, rank, corrections, highest, among, boosting, qwen3, instruct, backbone, points, intensive, incurring, minimal, overhead, junxian, jingdi, lei, soujanya, poria, 205, 446, foundation, zero, shot, whole, body, tracking, integrating, corpus, sparse, mixture, experts, reduction, mean, per, keypoint, position, error, robots, kaihui, maiyue, csail, sustained, coherent, progress, transferring, knowledge, sequential, generated, solutions, diverse, discovered, novel, algorithms, programming, cloud, multicast, below, mohammad, alizadeh, kimia, noorbakhsh, pantea, karimi, coherence, persistence, 285, work, two, stage, text, image, min, max, normalized, 929, evaluated, metrics, efficiency, kaixun, jiang, junqiu, quanhao, unified, perspective, 441, 132, 281, establishes, continuous, utilizing, flow, generative, translation, summarization, requiring, times, steps, linlu, qiu, keya, 102, yandex, role, finding, not, exhibit, patch, outliers, adding, achieve, serving, norm, sinks, global, carriers, resulting, cleaner, internal, feature, maps, compact, dual, stream, leveraging, specialization, enhanced, fid, various, scales, resolutions, ilya, drobyshevskiy, ilia, sudakov, nikita, starodubcev, matter, 893, 159, emerging, unifying, predictive, enable, anticipate, changes, under, interventions, survey, provides, formal, definition, taxonomy, systematic, analysis, ecosystems, evaluation, protocols, outlining, critical, challenges, future, directions, zhaoyang, junhao, siyin, next, fair, meta, airaformers, airahybrids, established, such, llama, nemotron, these, also, engineering, scripts, despoina, magka, chien, alberto, pepe, discovery, compose, design, 151, stable, aligning, full, trajectory, computational, robust, preference, addressing, core, instability, issues, zeyue, xue, siming, xiaoxuan, 135, introduced, simulator, featuring, assess, fine, grained, coordination, use, bimanual, includes, cost, demonstration, evaluates, modern, imitation, policies, revealing, current, limitations, xiangyu, weizhi, hanwen, 444, discrete, functional, autoregressive, enabling, execute, operations, standard, loop, allows, perform, interleaved, efficiently, interpretably, prior, approaches, improved, wemath, blink, xinyan, rain, ziyu, guo, word, enough, 291, commonsense, derived, egocentric, enhancing, physically, grounded, vlm, vla, simulations, robot, demonstrated, notable, point, gain, object, grasping, xiaopeng, bin, shijie, lian, 190, illinois, urbana, champaign, collaborators, demonstrates, rotary, positional, embeddings, inherently, struggle, reliably, distinguish, provable, degradation, identifies, failure, modes, emerge, context, length, increases, undermining, locality, bias, consistent, relevance, minyang, tian, phillip, harris, yufeng, distinguishes, neither, nor, provably, 271, resolves, trade, off, abstraction, disentanglement, produces, semantically, transferable, generates, transfer, planning, yufan, muyang, lyu, tianqiu, 556, 250, 111, nvidia, source, capable, generating, 720p, videos, precise, dof, camera, comparable, industrial, strong, second, yuyang, haozhe, haoyi, zhu, modeling, linear, 199, technology, china, unfamiliar, environments, measuring, capability, metric, called, checkpoint, coverage, ecc, which, interleaves, focused, employs, then, act, adaptability, environmental, perturbations, yuxin, wentao, ziang, look, before, leap, autonomous, 832, 100, replays, timestamped, news, forecasting, revealed, significant, room, test, adaptation, gpt, top, but, many, weight, negative, brier, skill, nikhil, chandak, shashwat, goel, replaying, 865, pricewaterhousecoopers, empirically, effectiveness, retrieval, strategies, focusing, lexical, semantic, vector, interact, different, output, delivery, reveals, often, when, results, delivered, inline, heavily, influenced, orchestration, layer, elias, lumer, akhil, kasturi, sahil, sen, all, need, harnesses, reshape, 094, 137, investigates, scalability, forward, reconstruction, static, scenes, power, law, datasets, innovations, reduce, gpu, boost, speed, shangzhan, minghao, jianyuan, 842, 161, zhejiang, tsinghua, sdar, robustly, opsd, turn, utilizes, level, gating, selectively, apply, teacher, guidance, improvement, alfworld, other, internalizing, skills, without, needing, zhuowen, han, zhiyuan, yao, personalize, likes, hot, cais, watch, recordings, alt, enter, style, smart, what, most, popular, math, ask, anything, hiring, dark, mode, upgrade, feedback,

Text of the page (random words):
explore alphaxiv alphaxiv explore sign in blog feedback upgrade to pro dark mode we re hiring ask or search anything what are the most popular benchmarks for math reasoning smart search style alt enter to search events watch recordings acm cais 2026 05 29 acm hot likes pro personalize your feed self distilled agentic reinforcement learning 14 may 2026 zhengxi lu zhengxi lu zhiyuan yao zhuowen han researchers at zhejiang university meituan and tsinghua university developed self distilled agentic reinforcement learning sdar a method that robustly integrates on policy self distillation opsd with reinforcement learning for multi turn large language model agents this approach utilizes a token level gating mechanism to selectively apply teacher guidance achieving up to a 9 4 improvement in success rates on alfworld and outperforming other hybrid methods while effectively internalizing skills without needing them at inference view blog agentic frameworks agents computer science 161 bookmark audio 4 2 842 vggt ω ω ω 14 may 2026 jianyuan wang minghao chen shangzhan zhang vggt ω investigates the scalability of feed forward 3d reconstruction achieving state of the art accuracy across static and dynamic scenes and demonstrating power law like performance improvements with larger models and datasets the framework introduces architectural and training innovations that significantly reduce gpu memory by 70 and boost inference speed by 20 25 view blog attention mechanisms computer science computer vision and pattern recognition 137 bookmark audio 2 094 is grep all you need how agent harnesses reshape agentic search 14 may 2026 sahil sen akhil kasturi elias lumer researchers at pricewaterhousecoopers empirically investigated the effectiveness of retrieval strategies for llm agents focusing on how lexical grep and semantic vector search interact with different agent architectures and tool output delivery methods the study reveals that grep often achieves higher accuracy when tool results are delivered inline and that overall agent performance is heavily influenced by the agent s specific orchestration layer view blog agentic frameworks agents computer science 68 bookmark audio 865 futuresim replaying world events to evaluate adaptive agents 14 may 2026 arvindh arun shashwat goel nikhil chandak arvindh arun researchers developed futuresim an environment that replays real world events from timestamped news to evaluate ai agents long horizon adaptive forecasting capabilities the benchmark revealed significant room for improving test time adaptation in frontier models with gpt 5 5 achieving 25 top 1 accuracy but many open weight models demonstrating negative brier skill scores view blog computer science artificial intelligence computation and language 100 bookmark audio 1 832 look before you leap autonomous exploration for llm agents 15 may 2026 ziang ye wentao shi yuxin liu researchers from the university of science and technology of china and meituan developed a training framework that enables large language model llm agents to autonomously explore unfamiliar environments measuring this capability with a new metric called exploration checkpoint coverage ecc their method which interleaves task oriented and exploration focused reinforcement learning and employs an explore then act paradigm significantly improves agent adaptability generalization and robustness to environmental perturbations view blog computer science artificial intelligence computation and language 11 bookmark audio 199 sana wm efficient minute scale world modeling with hybrid linear diffusion transformer 14 may 2026 haoyi zhu haozhe liu yuyang zhao nvidia researchers developed sana wm a 2 6b parameter open source world model capable of generating high fidelity one minute 720p videos with precise 6 dof camera control the system achieves visual quality comparable to larger industrial baselines while significantly reducing data training and inference costs demonstrating strong performance on a new 60 second world model benchmark view blog computer science computer vision and pattern recognition 111 bookmark audio 5 250 1 556 dila disentangled latent action world models 15 may 2026 tianqiu zhang muyang lyu yufan zhang dila a disentangled latent action world model resolves the trade off in latent action models between action abstraction and video generation fidelity by co optimizing disentanglement and latent action learning within a single framework the model produces semantically structured transferable latent actions and generates high fidelity video outperforming baselines in action transfer and visual planning tasks view blog computer science artificial intelligence computer vision and pattern recognition 19 bookmark audio 271 rope distinguishes neither positions nor tokens in long contexts provably 15 may 2026 yufeng du phillip harris minyang tian research from the university of illinois at urbana champaign and collaborators demonstrates that rotary positional embeddings rope inherently struggle to reliably distinguish between positions and tokens in long contexts leading to provable performance degradation in large language models the study identifies specific failure modes that emerge as context length increases undermining rope s locality bias and consistent token relevance view blog attention mechanisms computer science artificial intelligence 12 bookmark audio 190 physbrain 1 0 technical report 14 may 2026 shijie lian bin yu xiaopeng lin physbrain 1 0 integrates physical commonsense understanding derived from human egocentric video into embodied ai enhancing physically grounded reasoning and achieving higher success rates across vlm benchmarks vla simulations and real world robot tasks the system demonstrated notable improvements including a 16 2 percentage point gain in real world single object grasping view blog agents computer science artificial intelligence 22 bookmark audio 291 atlas agentic or latent visual reasoning one word is enough for both 14 may 2026 ziyu guo rain liu xinyan chen atlas is a visual reasoning framework that integrates discrete functional tokens into autoregressive vision language models enabling them to represent and execute visual operations within a standard generation loop this method allows models to perform complex interleaved visual reasoning more efficiently and interpretably than prior approaches leading to improved performance on benchmarks like wemath and blink view blog agentic frameworks agents computer science 49 bookmark audio 444 dexjoco a benchmark and toolkit for task oriented dexterous manipulation on mujoco 15 may 2026 hanwen wang weizhi zhao xiangyu wang dexjoco is introduced as a comprehensive benchmark and toolkit for task oriented dexterous manipulation within the mujoco simulator featuring 11 complex tasks designed to assess fine grained coordination tool use and bimanual actions it includes a low cost human demonstration system and evaluates modern imitation learning policies revealing current limitations in robustness and task generalization view blog computer science robotics 14 bookmark audio 39 135 flash grpo efficient alignment for video diffusion via one step policy optimization 15 may 2026 xiaoxuan he siming fu zeyue xue flash grpo introduces an efficient and stable single step policy optimization framework for aligning video diffusion models achieving full trajectory performance while significantly reducing computational costs it enables robust preference alignment for large scale video models by addressing core instability issues in reinforcement learning based training view blog computer science computer vision and pattern recognition 13 bookmark audio 22 151 agentic discovery of neural architectures aira compose and aira design 15 may 2026 alberto pepe chien yu lin despoina magka fair at meta developed llm agent frameworks that autonomously designed new neural architectures airaformers and airahybrids outperforming established human designed models such as llama 3 2 and nemotron 2 at the 1b parameter scale these agents also achieved competitive performance in engineering efficient attention mechanisms and optimizing language model training scripts view blog agentic frameworks agents computer science 10 bookmark audio 94 world action models the next frontier in embodied ai 12 may 2026 siyin wang junhao shi zhaoyang fu world action models wams represent an emerging paradigm in embodied ai unifying predictive physical understanding with action generation to enable agents to anticipate environment changes under their interventions this comprehensive survey provides a formal definition a structured architectural taxonomy and a systematic analysis of data ecosystems and evaluation protocols for wams outlining critical challenges and future research directions view blog computer science computation and language computer vision and pattern recognition 159 bookmark audio 1 893 registers matter for pixel space diffusion transformers 15 may 2026 nikita starodubcev ilia sudakov ilya drobyshevskiy researchers from yandex research investigated the role of register tokens in diffusion transformers dits finding that while pixel space dits do not exhibit vision transformer like patch token outliers adding registers significantly improves generation quality register tokens achieve this by serving as norm sinks and global information carriers resulting in cleaner internal feature maps a compact dual stream architecture leveraging this token specialization enhanced fid scores across various model scales and resolutions view blog computer science computer vision and pattern recognition 10 bookmark audio 102 elf embedded language flows 11 may 2026 yiyang lu mit keya hu linlu qiu yiyang lu embedded language flows elf establishes a continuous diffusion language model paradigm utilizing flow matching achieving competitive generative quality in tasks like machine translation and summarization while requiring approximately 10 times fewer training tokens and fewer inference steps than existing baselines view blog computer science artificial intelligence computation and language 281 bookmark audio 132 4 441 diffusionopd a unified perspective of on policy distillation in diffusion models 14 may 2026 quanhao li junqiu yu kaixun jiang this work introduces diffusionopd a two stage on policy distillation framework that effectively integrates multi task capabilities into text to image diffusion models the approach achieved an average min max normalized score of 0 929 across evaluated metrics outperforming existing multi task reinforcement learning methods in both performance and training efficiency view blog computer science computer vision and pattern recognition machine learning 37 bookmark audio 285 improving coherence and persistence in agentic ai for system optimization 19 may 2026 mit pantea karimi kimia noorbakhsh mohammad alizadeh researchers at mit csail developed engram an agentic ai architecture that enables sustained coherent progress in complex system optimization by transferring distilled knowledge between sequential llm agents engram generated solutions outperforming previous llm based methods across diverse tasks and discovered novel algorithms including a dynamic programming approach for multi cloud multicast that achieved costs below the human state of the art view blog 2 bookmark audio 26 holomotion 1 technical report 14 may 2026 maiyue chen kaihui wang bo zhang holomotion 1 developed by horizon robotics introduces a humanoid motion foundation model designed for zero shot whole body motion tracking integrating a large scale hybrid motion corpus and a sparse mixture of experts transformer architecture the system achieved approximately a 40 reduction in mean per keypoint position error compared to previous state of the art methods and enables real time control on physical humanoid robots view blog agents computer science artificial intelligence 22 bookmark audio 446 205 δ δ δ mem efficient online memory for large language models 12 may 2026 soujanya poria jingdi lei di zhang junxian li δ mem introduces an efficient online memory mechanism for large language models augmenting frozen backbones by compressing historical information into a fixed size associative state that dynamically steers attention via low rank corrections the system achieved the highest overall performance among compared memory mechanisms boosting the qwen3 4b instruct backbone s average score by 4 87 percentage points to 51 66 on memory intensive benchmarks while incurring minimal parameter overhead view blog agents attention mechanisms computer science 115 bookmark audio 25 1 332 there are no more papers matching your filters at the moment sign in