Daily AI Papers — May 02, 2026

1. GLM-5V-Turbo: Toward a Native Foundation Model for Multimodal Agents

Authors: GLM-V Team: Wenyi Hong, Xiaotao Gu, Ziyang Pan, Zhen Yang, Yuting Wang, Yue Wang, et al.

Summary: GLM-5V-Turbo advances toward truly native multimodal agents by integrating visual perception—across images, videos, webpages, documents, and GUIs—as a core component of reasoning, planning, and tool use, rather than bolting it on. The model is designed for real-environment deployment where agentic capability depends on heterogeneous context understanding beyond language alone.

arXiv: arxiv.org/abs/2604.26752

Sources: HuggingFace Daily Papers (#1 by upvotes, 88)

Why Trending: Top upvoted paper today; major release from Zhipu AI targeting native multimodal agentic intelligence — directly competitive with frontier closed models.

2. RADIO-ViPE: Online Tightly Coupled Multi-Modal Fusion for Open-Vocabulary Semantic SLAM in Dynamic Environments

Authors: Zaid Nasser, Mikhail Iumanov, Tianhao Li, Maxim Popov, Jaafar Mahmoud, Sergey Kolyubin

Summary: RADIO-ViPE is an online semantic SLAM system that associates arbitrary natural language queries with localized 3D regions and objects in dynamic environments, operating on raw monocular RGB video with no depth sensors, camera calibration, or pose initialization required. It tightly couples multi-modal fusion for open-vocabulary 3D grounding in real time.

arXiv: arxiv.org/abs/2604.26067

Sources: HuggingFace Daily Papers (64 upvotes), arXiv cs.CV

Why Trending: Practical monocular 3D semantic understanding without special hardware — strong implications for robotics and AR.

3. Large Language Models Explore by Latent Distilling

Authors: Yuanhao Zeng, Ao Lu, Lufei Li, Zheng Zhang, Yexin Li, Kan Ren

Summary: This paper introduces Exploratory Sampling (ESamp), a decoding approach that encourages semantic diversity in LLM generation by exploiting the observation that neural networks generalize better near training-data neighbors in latent space. ESamp drives exploration in representation space rather than surface-level token variation, substantially improving test-time scaling outcomes for complex reasoning.

arXiv: arxiv.org/abs/2604.24927

Sources: HuggingFace Daily Papers (59 upvotes), arXiv cs.CL

Why Trending: Addresses a fundamental bottleneck in test-time scaling; practical decoding improvement applicable to any LLM.

4. ClawGym: A Scalable Framework for Building Effective Claw Agents

Authors: Fei Bai, Huatong Song, Shuang Sun, Daixuan Cheng, Yike Yang, Chuan Hao, Renyuan Li, Feng Chang, Yuan Wei, Ran Tao, Bryan Dai, Jian Yang, Wayne Xin Zhao

Summary: ClawGym provides a full-lifecycle framework for developing Claw-style personal agents that operate over local files, tools, and persistent workspace states, including scalable environment creation, verifiable training data synthesis, and diagnostic evaluation. It directly addresses the absence of systematic infrastructure for building agentic AI that handles multi-step real-world workflows.

arXiv: arxiv.org/abs/2604.26904

Sources: HuggingFace Daily Papers (46 upvotes), arXiv cs.AI

Why Trending: High practical value for the agent training infrastructure space; fills a critical gap in the Claw-agent ecosystem.

5. Turning the TIDE: Cross-Architecture Distillation for Diffusion Large Language Models

Authors: Gongbo Zhang, Wen Wang, Ye Tian, Li Yuan

Summary: TIDE is the first framework for cross-architecture knowledge distillation of diffusion large language models (dLLMs), enabling transfer between teacher and student models that differ in architecture, attention mechanism, and tokenizer. This unlocks the benefits of parallel decoding and bidirectional context from large dLLMs at smaller, more practical scales.

arXiv: arxiv.org/abs/2604.26951

Sources: HuggingFace Daily Papers (42 upvotes), arXiv cs.CL

Why Trending: First cross-architecture distillation for dLLMs; makes efficient diffusion language models a realistic alternative to autoregressive at smaller sizes.

6. AutoResearchBench: Benchmarking AI Agents on Complex Scientific Literature Discovery

Authors: Lei Xiong, Kun Luo, Ziyi Xia, Wenbo Zhang, Jin-Ge Yao, Zheng Liu, Jingying Shao, et al.

Summary: AutoResearchBench is a dedicated benchmark for evaluating AI agents’ capacity for autonomous scientific literature discovery, covering both exploratory knowledge-finding for a research problem and evidence retrieval for hypothesis verification. It enables systematic measurement of a key capability needed for end-to-end autonomous research pipelines.

arXiv: arxiv.org/abs/2604.25256

Sources: HuggingFace Daily Papers (27 upvotes), arXiv cs.AI

Why Trending: Timely given the surge in AI-for-science research; fills a missing evaluation layer for autonomous research agents.

7. Step-Audio-R1.5 Technical Report

Authors: Yuxin Zhang, Xiangyu Tony Zhang, Daijiao Liu, Fei Tian, Yayue Deng, Jun Chen, Qingjian Lin, Haoyang Zhang, et al.

Summary: Step-Audio-R1.5 extends chain-of-thought reasoning into the auditory domain while addressing a core failure mode of RLVR-trained audio models: the progressive loss of rich acoustic understanding when models are optimized purely for verifiable reasoning rewards. It introduces a mixed training paradigm that preserves both extended reasoning chains and fine-grained acoustic perception.

arXiv: arxiv.org/abs/2604.25719

Sources: HuggingFace Daily Papers (13 upvotes), arXiv cs.SD, cs.AI

Why Trending: Audio reasoning is a fast-growing frontier; addresses a practical training failure mode for RLVR-style audio models.

8. Diffusion Templates: A Unified Plugin Framework for Controllable Diffusion

Authors: Zhongjie Duan, Hong Zhang, Yingda Chen

Summary: Diffusion Templates decouples controllable diffusion capabilities from specific backbone architectures via a unified, open plugin framework, enabling reuse of infrastructure across tasks, transfer across backbones, and composition of multiple controls within a single generation pipeline. It eliminates the fragmentation where each control method (ControlNet, IP-Adapter, etc.) requires isolated training pipelines and incompatible formats.

arXiv: arxiv.org/abs/2604.24351

Sources: HuggingFace Daily Papers (10 upvotes), arXiv cs.CV

Why Trending: Highly practical unification of the controllable diffusion ecosystem; directly reduces duplication across the field.

9. Step-level Optimization for Efficient Computer-use Agents

Authors: Jinbiao Wei, Kangqi Ni, Yilun Zhao, Guo Gan, Arman Cohan

Summary: This paper shows that uniform compute allocation across all GUI interaction steps is fundamentally inefficient: long-horizon computer-use trajectories contain many routine steps but only a few critical decision points requiring heavy multimodal reasoning. Step-level optimization selectively applies expensive models only where needed, cutting cost significantly while maintaining benchmark performance.

arXiv: arxiv.org/abs/2604.27151

Sources: HuggingFace Daily Papers (9 upvotes), arXiv cs.AI

Why Trending: Addresses real deployment cost of computer-use agents; elegant systems contribution with direct practical impact.

10. FAMA: Failure-Aware Meta-Agentic Framework for Open-Source LLMs in Interactive Tool Use Environments

Authors: Amir Saeidi, Venkatesh Mishra, Souradeep Mukhopadhyay, Gaowen Liu, Ali Payani, Jayanth Srinivasa, Chitta Baral

Summary: FAMA addresses cascading failures in open-source LLM agents during multi-turn customer-centric issue resolution, where incorrect decisions compound across interactions. The meta-agentic framework detects failures and triggers recovery, significantly improving task completion for smaller open-source models under constrained context and inference budgets.

arXiv: arxiv.org/abs/2604.25135

Sources: HuggingFace Daily Papers (8 upvotes), arXiv cs.AI

Why Trending: Practical reliability engineering for deployed open-source agent systems; addresses a gap that limits real-world adoption.

11. Accelerating RL Post-Training Rollouts via System-Integrated Speculative Decoding

Authors: Hayate Iso, Tiyasa Mitra, Sudipta Mondal, Rasoul Shafipour, Venmugil Elango, Terry Kong, Yuki Huang, Seonjin Na, et al.

Summary: This paper establishes speculative decoding as a lossless acceleration primitive for rollout generation during RL post-training, preserving the exact output distribution of the target policy while eliminating the throughput bottleneck of autoregressive generation. Unlike off-policy or lower-precision methods, it introduces no training artifacts.

arXiv: arxiv.org/abs/2604.26779

Sources: HuggingFace Daily Papers (6 upvotes), arXiv cs.LG

Why Trending: Systems contribution directly relevant to anyone running RLVR or GRPO post-training; lossless speedup is rare.

12. X-WAM: Unified 4D World Action Modeling from Video Priors with Asynchronous Diffusion

Authors: Jun Guo, Qiwei Li, Peiyan Li, Zilong Chen, Nan Sun, Yifei Su, Heyun Wang, Yuan Zhang, Xinghang Li, Huaping Liu

Summary: X-WAM is a unified 4D world model that simultaneously handles real-time robotic action execution and high-fidelity 4D world synthesis (video + 3D reconstruction) by leveraging pretrained video diffusion priors to predict multi-view RGB-D futures. It overcomes the 2D-only limitation of prior unified world models while balancing action efficiency and world modeling quality.

arXiv: arxiv.org/abs/2604.26694

Sources: HuggingFace Daily Papers (6 upvotes), arXiv cs.RO, cs.CV

Why Trending: Key advance in embodied AI; unifying 4D perception with action in a single model is a long-standing open problem.

13. MAIC-UI: Making Interactive Courseware with Generative UI

Authors: Shangqing Tu, Yanjia Li, Keyu Chen, Sichen Zhang, Jifan Yu, Daniel Zhang-Li, Lei Hou, Juanzi Li, Yu Zhang, Huiqin Liu

Summary: MAIC-UI is a zero-code authoring system enabling educators to create interactive STEM simulations from documents using generative AI, without HTML/CSS/JavaScript expertise, and with built-in pedagogical accuracy verification. Modification latency drops from 200–600 seconds to near-instant by generating UI incrementally rather than regenerating from scratch.

arXiv: arxiv.org/abs/2604.25806

Sources: HuggingFace Daily Papers (6 upvotes), arXiv cs.AI, cs.HC

Why Trending: High practical value for AI in education; removes real barriers that prevent non-technical educators from using generative tools.

14. Operating-Layer Controls for Onchain Language-Model Agents Under Real Capital

Authors: T.J. Barton, Chris Constantakis, Patti Hauseman, Annie Mous, Alaska Hoffman, Brian Bergeron, Hunter Goodreau

Summary: Based on a 21-day real-world deployment of 3,505 user-funded LLM agents trading real ETH in a bounded onchain market (generating 7.5M agent invocations and ~$20M in notional volume), this paper identifies operating-layer controls as critical for maintaining reliable autonomous agent behavior under adversarial market conditions. It provides rare empirical evidence from a high-stakes real deployment.

arXiv: arxiv.org/abs/2604.26091

Sources: HuggingFace Daily Papers (5 upvotes), arXiv cs.AI

Why Trending: Unprecedented real-world deployment scale for autonomous financial agents; provides grounded safety evidence rarely seen in the literature.

15. Semi-DPO: Learning from Noisy Preferences for Visual Diffusion Models

Authors: Xinxin Liu, Ming Li, Zonglin Lyu, Yuzhang Shang, Chen Chen

Summary: Semi-DPO addresses the multi-dimensional nature of visual preferences being collapsed into binary win/lose labels, which creates conflicting gradient signals in Diffusion DPO training. A semi-supervised approach decouples preference dimensions (aesthetics, detail fidelity, semantic alignment) to enable correct preference learning without gradient conflicts.

arXiv: arxiv.org/abs/2604.24952

Sources: HuggingFace Daily Papers (3 upvotes), arXiv cs.CV, cs.LG

Why Trending: Targets a systematic flaw in visual preference datasets that affects all DPO-based image generation training.

16. V-GRPO: Online Reinforcement Learning for Denoising Generative Models

Authors: Bingda Tang, Yuhui Zhang, Xiaohan Wang, Jiayuan Mao, Ludwig Schmidt, Serena Yeung-Levy

Summary: V-GRPO introduces an online policy-gradient RL method for aligning diffusion/denoising generative models with human preferences or verifiable rewards, resolving the intractable likelihood problem without resorting to MDP-based inefficiency or unstable likelihood surrogates. It provides a principled and efficient post-training framework for generative models beyond the autoregressive domain.

arXiv: arxiv.org/abs/2604.23380

Sources: HuggingFace Daily Papers (3 upvotes), arXiv cs.LG, cs.CV

Why Trending: Clean solution to a long-standing problem; directly extends RLVR-style post-training to diffusion models.

17. Preferences of a Voice-First Nation: Large-Scale Pairwise Evaluation of Multilingual TTS for Indic Languages

Authors: Srija Anand, Ashwin Sankar, Ishvinder Sethi, Aaditya Pareek, Kartik Rajput, Gaurav Yadav, Nikhil Narasimhan, et al.

Summary: This paper presents a controlled multidimensional pairwise TTS evaluation framework covering 5K+ native and code-mixed sentences across 10 Indic languages, revealing systematic preference patterns across naturalness, pronunciation, and prosody for voice-first language communities. It provides both a methodology and large-scale results for evaluating multilingual TTS in underrepresented language ecosystems.

arXiv: arxiv.org/abs/2604.21481

Sources: HuggingFace Daily Papers (3 upvotes), arXiv cs.SD, cs.CL

Why Trending: Critically needed evaluation resource for global AI fairness; covers one of the largest underrepresented language groups.

18. Instruction-Guided Poetry Generation in Arabic and Its Dialects

Authors: Abdelrahman Sadallah, Kareem Elozeiri, Mervat Abassy, Rania Elbadry

Summary: This work presents an instruction-guided system for Arabic poetry generation spanning classical Arabic and multiple spoken dialects, addressing the unique challenges of meter, rhyme schemes, and dialectal variation that make Arabic a highly demanding creative NLP domain. The paper contributes both a generation framework and a targeted evaluation methodology for this underserved area.

arXiv: arxiv.org/abs/2604.27766

Sources: HuggingFace Daily Papers (2 upvotes), arXiv cs.CL

Why Trending: Advancing creative NLP for Arabic dialects — a structurally complex and underserved linguistic domain.

19. ViPO: Visual Preference Optimization at Scale

Authors: Ming Li, Jie Wu, Justin Cui, Xiaojie Li, Rui Wang, Chen Chen

Summary: ViPO introduces Poly-DPO, which addresses conflicting preference patterns in visual generative model training by extending the DPO objective with a polynomial term that dynamically adjusts model confidence based on dataset characteristics. This enables effective scaling of visual preference optimization past the noise barrier that has limited prior work.

arXiv: arxiv.org/abs/2604.24953

Sources: HuggingFace Daily Papers (1 upvote), arXiv cs.CV, cs.LG

Why Trending: Directly addresses a scaling bottleneck in visual preference optimization; paired with Semi-DPO as complementary approaches.

20. Seeing Isn’t Believing: Uncovering Blind Spots in Evaluator Vision-Language Models

Authors: Mohammed Safi Ur Rahman Khan, Sanjay Suryanarayanan, Tushar Anand, Mitesh M. Khapra

Summary: This paper systematically evaluates VLMs used as automated evaluators for both image-to-text and text-to-image tasks, introducing targeted perturbations that degrade output quality in specific dimensions to expose hidden failure modes. It finds that widely-used evaluator VLMs have significant blind spots that undermine the reliability of automated evaluation pipelines across the field.

arXiv: arxiv.org/abs/2604.21523

Sources: HuggingFace Daily Papers (2 upvotes), arXiv cs.CV, cs.CL

Why Trending: High relevance to AI evaluation methodology; calls into question the validity of automated benchmarks used field-wide.