Daily AI Papers — April 24, 2026

1. LLaTiSA: Towards Difficulty-Stratified Time Series Reasoning from Visual Perception to Semantics

• Authors: Yueyang Ding, HaoPeng Zhang, Rui Dai, Yi Wang, Tianyu Zong, et al. (7 total)
• arXiv: arxiv.org/abs/2604.17295
• Sources: HuggingFace Daily Papers (75 upvotes — top of the day), arXiv
• Why trending: Introduces HiTSR — an 83k-sample hierarchical time-series reasoning dataset with verified CoT trajectories — and a 4-level taxonomy formalizing TSR. The companion VLM-based model fuses visualized patterns with precision-calibrated tables and shows strong OOD generalization, addressing a long-standing fragmentation in time-series LLM evaluation.

2. WorldMark: A Unified Benchmark Suite for Interactive Video World Models

• Authors: Xiaojie Xu, Zhengyuan Lin, Kang He, Yukang Feng, Xiaofeng Mao, et al. (8 total)
• arXiv: arxiv.org/abs/2604.21686
• Sources: HuggingFace Daily Papers (29 upvotes), arXiv
• Why trending: First unified benchmark to put Genie, YUME, HY-World, Matrix-Game and others on apples-to-apples scenes via a shared WASD action vocabulary. Ships a 500-case test suite plus a public “World Model Arena” (warena.ai) for live side-by-side battles — directly attacks the comparability crisis in world-model research.

3. UniT: Toward a Unified Physical Language for Human-to-Humanoid Policy Learning and World Modeling

• Authors: Boyu Chen, Yi Chen, Lu Qiu, Jerry Bai, Yuying Ge, et al. (6 total)
• arXiv: arxiv.org/abs/2604.19734
• Sources: HuggingFace Daily Papers (25 upvotes), arXiv
• Why trending: Unified Latent Action Tokenizer that bridges human egocentric data and humanoid kinematics through a tri-branch cross-reconstruction (vision↔action) trick. Demonstrates zero-shot task transfer on humanoid sim and real hardware — a meaningful step toward scaling humanoid foundation models past the data bottleneck.

4. StyleID: A Perception-Aware Dataset and Metric for Stylization-Agnostic Facial Identity Recognition

• Authors: Kwan Yun, Changmin Lee, Ayeong Jeong, Youngseo Kim, Seungmi Lee, et al. (6 total)
• arXiv: arxiv.org/abs/2604.21689
• Sources: HuggingFace Daily Papers (18 upvotes), arXiv
• Why trending: Releases StyleBench-H (human verification judgments) and StyleBench-S (psychometric 2AFC supervision) for measuring facial identity across diffusion- and flow-matching-based stylization. Calibrating semantic encoders against human perception yields large robustness gains on artist-drawn portraits — a tooling gap many stylization papers have ignored.

5. Seeing Fast and Slow: Learning the Flow of Time in Videos

• Authors: Yen-Siang Wu, Rundong Luo, Jingsen Zhu, Tao Tu, Ali Farhadi, et al. (9 total)
• arXiv: arxiv.org/abs/2604.21931
• Sources: HuggingFace Daily Papers (13 upvotes), arXiv
• Why trending: Treats time as a first-class learnable visual concept: self-supervised playback-speed estimation curates the largest in-the-wild slow-motion dataset to date, then enables speed-conditioned generation and temporal super-resolution. Opens a fresh axis for video world models that reason about how events unfold.

6. COSPLAY: Co-Evolving LLM Decision and Skill Bank Agents for Long-Horizon Tasks

• Authors: Xiyang Wu, Zongxia Li, Guangyao Shi, Alexander Duffy, Tyler Marques, et al. (8 total)
• arXiv: arxiv.org/abs/2604.20987
• Sources: HuggingFace Daily Papers (10 upvotes), arXiv
• Why trending: Pairs an LLM decision agent with a learnable skill bank curated by a second skill-management agent — both co-evolve from unlabeled rollouts. An 8B base model beats four frontier-LLM baselines by >25% average reward across six game environments, a strong signal for skill-reuse architectures.

7. TingIS: Real-time Risk Event Discovery from Noisy Customer Incidents at Enterprise Scale

• Authors: Jun Wang, Ziyin Zhang, Rui Wang, Hang Yu, Peng Di, et al. (6 total)
• arXiv: arxiv.org/abs/2604.21889
• Sources: HuggingFace Daily Papers (9 upvotes), arXiv
• Why trending: Production-scale system handling 300k messages/day with P90 alert latency of 3.5 minutes and 95% high-priority discovery. Combines LLM-driven event linking, cascaded routing, and behavioral noise reduction — practical, deployed work that resonates with infra/SRE crowds.

8. Hybrid Policy Distillation for LLMs

• Authors: Wenhong Zhu, Ruobing Xie, Rui Wang, Pengfei Liu
• arXiv: arxiv.org/abs/2604.20244
• Sources: HuggingFace Daily Papers (9 upvotes), arXiv
• Why trending: Reframes LLM knowledge distillation as a unified token-level reweighted log-likelihood and proposes HPD, blending forward+reverse KL with off-policy data and lightweight on-policy sampling. Improves stability and final quality across math reasoning, dialogue, and code — a clean theory-driven contribution that practitioners can drop in.

9. VLAA-GUI: Knowing When to Stop, Recover, and Search — A Modular Framework for GUI Automation

• Authors: Qijun Han, Haoqin Tu, Zijun Wang, Haoyue Dai, Yiyang Zhou, et al. (14 total)
• arXiv: arxiv.org/abs/2604.21375
• Sources: HuggingFace Daily Papers (6 upvotes), arXiv
• Why trending: Three-component framework (Completeness Verifier, Loop Breaker, on-demand Search Agent) hits 77.5% on OSWorld and 61.0% on WindowsAgentArena — and three of five backbones (Opus 4.5/4.6, Gemini 3.1 Pro) surpass human performance on OSWorld in a single pass. Big numbers in a hot benchmark space.

10. Context Unrolling in Omni Models

• Authors: Ceyuan Yang, Zhijie Lin, Yang Zhao, Fei Xiao, Hao He, et al. (19 total)
• arXiv: arxiv.org/abs/2604.21921
• Sources: HuggingFace Daily Papers (5 upvotes), arXiv
• Why trending: Unified multimodal model natively trained on text/image/video/3D/hidden-rep, with explicit “Context Unrolling” reasoning across modal representations before predicting. Strong scores on multimodal generation + understanding plus in-context generation across modalities.

11. EditCrafter: Tuning-free High-Resolution Image Editing via Pretrained Diffusion Model

• Authors: Kunho Kim, Sumin Seo, Yongjun Cho, Hyungjin Chung
• arXiv: arxiv.org/abs/2604.10268
• Sources: HuggingFace Daily Papers (5 upvotes), arXiv
• Why trending: Edits images at resolutions far above the T2I model’s training resolution without any fine-tuning, using tiled inversion plus a noise-damped manifold-constrained CFG (NDCFG++). A pragmatic recipe for the “I have a 4K photo” problem most editing pipelines ignore.

12. UniGenDet: A Unified Generative-Discriminative Framework for Co-Evolutionary Image Generation and Generated Image Detection

• Authors: Yanran Zhang, Wenzhao Zheng, Yifei Li, Bingyao Yu, Yu Zheng, et al. (8 total)
• arXiv: arxiv.org/abs/2604.21904
• Sources: HuggingFace Daily Papers (3 upvotes), arXiv
• Why trending: Couples a generator and a generated-image detector via symbiotic multimodal self-attention so each improves the other (more interpretable detection ↔ higher-fidelity generation). Argues against the architectural divide between gen and det — relevant to the deepfake / authenticity arms race.

13. WebGen-R1: Incentivizing LLMs to Generate Functional and Aesthetic Websites with Reinforcement Learning

• Authors: Juyong Jiang, Chenglin Cai, Chansung Park, Jiasi Shen, Sunghun Kim, et al. (7 total)
• arXiv: arxiv.org/abs/2604.20398
• Sources: HuggingFace Daily Papers (3 upvotes), arXiv
• Why trending: End-to-end RL framework with a cascaded multimodal reward (structure + execution + aesthetics) that turns a 7B base into a multi-page website generator beating 72B open-source baselines and rivaling DeepSeek-R1 (671B) on functional success while exceeding it on rendering and aesthetics.

14. Trust but Verify: Introducing DAVinCI — A Framework for Dual Attribution and Verification in Claim Inference for Language Models

• Authors: Vipula Rawte, Ryan Rossi, Franck Dernoncourt, Nedim Lipka
• arXiv: arxiv.org/abs/2604.21193
• Sources: HuggingFace Daily Papers (2 upvotes), arXiv
• Why trending: Two-stage framework: attribute generated claims to internal components and external sources, then verify via entailment + confidence calibration. 5-20% gains on FEVER / CLIMATE-FEVER over verification-only baselines, with a modular pluggable implementation released.

15. Vista4D: Video Reshooting with 4D Point Clouds

• Authors: Kuan Heng Lin, Zhizheng Liu, Pablo Salamanca, Yash Kant, Ryan Burgert, et al. (12 total)
• arXiv: arxiv.org/abs/2604.21915
• Sources: HuggingFace Daily Papers (2 upvotes), arXiv
• Why trending: Re-synthesizes a video from a different camera trajectory by grounding both the input video and the target camera in a 4D point cloud — with static-pixel segmentation and multiview dynamic training to survive depth-estimation artifacts. Strong baseline for dynamic-scene expansion and 4D recomposition.

16. Coevolving Representations in Joint Image-Feature Diffusion (CoReDi)

• Authors: Theodoros Kouzelis, Spyros Gidaris, Nikos Komodakis
• arXiv: arxiv.org/abs/2604.17492
• Sources: HuggingFace Daily Papers (2 upvotes), arXiv
• Why trending: Lets the semantic representation space evolve during diffusion training via a lightweight learnable projection (with stop-gradient + normalization to prevent collapse). Works in both VAE-latent and pixel-space diffusion, yielding faster convergence and higher quality than fixed-rep counterparts.

17. Encoder-Free Human Motion Understanding via Structured Motion Descriptions (SMD)

• Authors: Yao Zhang, Zhuchenyang Liu, Thomas Ploetz, Yu Xiao
• arXiv: arxiv.org/abs/2604.21668
• Sources: HuggingFace Daily Papers (1 upvote), arXiv
• Why trending: Skips the motion encoder entirely — converts joint sequences into rule-based natural language describing joint angles, body-part motion, and global trajectory. Hits 66.7% BABEL-QA / 90.1% HuMMan-QA / R@1 0.584 on HumanML3D and ports across 8 LLMs with only LoRA — a refreshingly simple alternative to learned alignment.

18. Explainable Disentangled Representation Learning for Generalizable Authorship Attribution in the Era of Generative AI

• Authors: Hieu Man, Van-Cuong Pham, Nghia Trung Ngo, Franck Dernoncourt, Thien Huu Nguyen
• arXiv: arxiv.org/abs/2604.21300
• Sources: HuggingFace Daily Papers (1 upvote), arXiv
• Why trending: EAVAE separates style from content via supervised contrastive pretraining + dual-encoder VAE, with a discriminator that not only judges style/content matches but generates natural-language explanations for its decisions. SOTA on Amazon Reviews / PAN21 / HRS plus strong few-shot AI-text detection on M4.

19. Test-Time Adaptation for EEG Foundation Models: A Systematic Study under Real-World Distribution Shifts

• Authors: Gabriel Jason Lee, Jathurshan Pradeepkumar, Jimeng Sun
• arXiv: arxiv.org/abs/2604.16926
• Sources: HuggingFace Daily Papers (1 upvote), arXiv
• Why trending: Introduces NeuroAdapt-Bench and finds that gradient-based TTA methods degrade EEG foundation models under realistic distribution shifts, while optimization-free methods are stable — a useful negative result that should reset expectations for clinical EEG deployment.

20. PersonalAI: A Systematic Comparison of Knowledge Graph Storage and Retrieval Approaches for Personalized LLM Agents

• Authors: Mikhail Menschikov, Dmitry Evseev, Victoria Dochkina, Ruslan Kostoev, Ilia Perepechkin, et al. (8 total)
• arXiv: arxiv.org/abs/2506.17001
• Sources: HuggingFace Daily Papers, arXiv
• Why trending: External-memory framework built on AriGraph with a hybrid design supporting standard edges and two hyper-edge types for richer temporal/semantic representation. Compares A*, WaterCircles, beam search and hybrid retrieval across TriviaQA / HotpotQA / DiaASQ — a useful systematic baseline for the personalized-agent memory crowd.