Daily AI Papers — April 13, 2026

1. WildDet3D: Scaling Promptable 3D Detection in the Wild

• Authors: Weikai Huang, Jieyu Zhang, Sijun Li, Taoyang Jia, Jiafei Duan, Ali Farhadi, Ranjay Krishna et al.
• ArXiv: arxiv.org/abs/2604.08626
• Summary: A unified geometry-aware architecture for monocular 3D object detection that accepts text, point, and box prompts and can incorporate auxiliary depth signals at inference. Introduces the largest open 3D detection dataset (1M+ images, 13.5K categories). Achieves SOTA across Omni3D, Argoverse 2, and ScanNet benchmarks, with +20.7 AP average gain when using depth cues.
• Sources: HuggingFace (#1, 145 upvotes), Hacker News (front page), GitHub (256 stars), arXiv, alphaXiv, Allen AI project page
• Why trending: Massive community reception — highest HF upvotes of the day, HN front page, open-source from AI2. Breakthrough in open-world 3D understanding from single images.

2. FORGE: Fine-grained Multimodal Evaluation for Manufacturing Scenarios

• Authors: Xiangru Jian, Hao Xu, Wei Pang, Xinjian Zhao, Chengyu Tao et al.
• ArXiv: arxiv.org/abs/2604.07413
• Summary: A comprehensive benchmark for evaluating Multimodal LLMs in manufacturing, addressing the gap between academic evaluation and real-world industrial demands. Covers fine-grained domain semantics and autonomous execution scenarios beyond simple perception tasks.
• Sources: HuggingFace (#2, 76 upvotes), arXiv, alphaXiv
• Why trending: Strong HF engagement. Practical industrial AI evaluation is an underserved area — this fills a real gap for enterprise MLLM deployment.

3. RefineAnything: Multimodal Region-Specific Refinement for Perfect Local Details

• Authors: Dewei Zhou, You Li, Zongxin Yang, Yi Yang
• ArXiv: arxiv.org/abs/2604.06870
• Summary: Introduces region-specific image refinement as a dedicated problem: given an image and a user-specified region (scribble mask or bounding box), restore fine-grained details while keeping non-edited pixels strictly unchanged. Addresses a key limitation of modern image generation models.
• Sources: HuggingFace (#3, 34 upvotes), arXiv, alphaXiv, CatalyzeX, aimodels.fyi, GitHub (49 stars)
• Why trending: Broad cross-platform visibility. Solves a practical pain point in image editing workflows — pixel-perfect local refinement without affecting the rest of the image.

4. EXAONE 4.5 Technical Report

• Authors: Eunbi Choi, Kibong Choi, Sehyun Chun, Seokhee Hong et al. (LG AI Research)
• ArXiv: arxiv.org/abs/2604.08644
• Summary: LG AI Research’s first open-weight vision-language model, integrating a dedicated visual encoder into the EXAONE 4.0 framework for native multimodal pretraining. Trained on large-scale visual and textual data, competitive with leading open VLMs.
• Sources: HuggingFace (#4, 31 upvotes), arXiv
• Why trending: New open-weight VLM from a major industrial lab. The open-weight LLM race continues heating up with another strong entrant.

5. Matrix-Game 3.0: Real-Time and Streaming Interactive World Model with Long-Horizon Memory

• Authors: Zile Wang, Zexiang Liu, Jaixing Li, Kaichen Huang, Baixin Xu et al. (Skywork AI)
• ArXiv: arxiv.org/abs/2604.08995
• Summary: A diffusion-based interactive world model that achieves both memory-enabled long-term temporal consistency and high-resolution real-time generation. Tackles the critical bottleneck of combining streaming interaction with long-horizon memory in world models.
• Sources: HuggingFace (#5, 30 upvotes), arXiv, GitHub (Skywork AI)
• Why trending: World models are a hot research frontier. Real-time + streaming + long-horizon memory is a key unlock for interactive applications and gaming.

6. ECHO: Efficient Chest X-ray Report Generation with One-step Block Diffusion

• Authors: Lifeng Chen, Tianqi You, Hao Liu, Zhimin Bao, Jile Jiao et al.
• ArXiv: arxiv.org/abs/2604.09450
• Summary: Uses one-step block diffusion for chest X-ray report generation, dramatically reducing inference latency compared to autoregressive VLMs while maintaining quality. A promising alternative to sequential token decoding for medical AI.
• Sources: HuggingFace (#6, 14 upvotes), arXiv
• Why trending: Medical AI + diffusion-based text generation is a novel combination. Practical clinical impact — faster radiology report generation.

7. ELT: Elastic Looped Transformers for Visual Generation

• Authors: Sahil Goyal, Swayam Agrawal, Gautham Govind Anil, Prateek Jain, Sujoy Paul et al.
• ArXiv: arxiv.org/abs/2604.09168
• Summary: A parameter-efficient visual generative model using recurrent transformer architecture with weight-shared blocks. Drastically reduces parameter counts while maintaining generation quality through iterative looped computation, with elastic inference-time compute.
• Sources: HuggingFace (#7, 12 upvotes), arXiv, alphaXiv
• Why trending: Efficiency in generative models is a major research theme. Weight-sharing + elastic compute offers a practical path to deploying large generative models on constrained hardware.

8. Backdoor Attacks on Decentralised Post-Training

• Authors: Oğuzhan Ersoy, Nikolay Blagoev, Jona te Lintelo, Stefanos Koffas, Marina Krček et al.
• ArXiv: arxiv.org/abs/2604.02372
• Summary: Analyzes poisoning and backdoor attack vulnerabilities in decentralised LLM post-training that uses data and pipeline parallelism. Explores how malicious participants can compromise model integrity when training is distributed.
• Sources: HuggingFace (#8, 10 upvotes), arXiv, aimodels.fyi
• Why trending: Timely security research as decentralized training gains adoption. Critical for understanding trust boundaries in distributed LLM fine-tuning.

9. Multi-User Large Language Model Agents

• Authors: Shu Yang, Shenzhe Zhu, Hao Zhu, José Ramón Enríquez, Di Wang et al.
• ArXiv: arxiv.org/abs/2604.08567
• Summary: Addresses the gap in LLM agent design for multi-principal settings where multiple users with potentially conflicting objectives interact with the same agent. Moves beyond the single-user optimization paradigm that dominates current systems.
• Sources: HuggingFace (#9, 7 upvotes), arXiv, GitHub (5 stars)
• Why trending: Highly relevant as LLM agents become deployed in enterprise/team settings. Multi-stakeholder alignment is an underexplored but critical problem.

10. EquiformerV3: Scaling SE(3)-Equivariant Graph Attention Transformers

• Authors: Yi-Lun Liao, Alexander J. Hoffman, Sabrina C. Shen, Alexandre Duval, Sam Walton Norwood et al.
• ArXiv: arxiv.org/abs/2604.09130
• Summary: Third generation of SE(3)-equivariant graph attention transformers for 3D atomistic modeling, improving efficiency, expressivity, and physical consistency for large-scale molecular and materials science applications.
• Sources: HuggingFace (0 upvotes but listed), arXiv, GitHub (38 stars), Emergent Mind
• Why trending: Key infrastructure for computational chemistry and drug discovery. GitHub traction suggests strong niche community interest despite lower general upvotes.

11. VisionFoundry: Teaching VLMs Visual Perception with Synthetic Images

• Authors: Guanyu Zhou, Yida Yin, Wenhao Chai, Shengbang Tong, Xingyu Fu et al.
• ArXiv: arxiv.org/abs/2604.09531
• Summary: Demonstrates that targeted synthetic supervision can teach VLMs low-level visual perception skills (spatial understanding, viewpoint recognition) that natural image datasets fail to provide. A practical approach to closing the perception gap in VLMs.
• Sources: HuggingFace (#11, 5 upvotes), arXiv, alphaXiv
• Why trending: Synthetic data for VLM training is a growing paradigm. Addresses a fundamental weakness of current vision-language models.

12. Structured Causal Video Reasoning via Multi-Objective Alignment

• Authors: Zinuo Li, Yongxin Guo, Jun Liu, Jiawei Zhan, Xi Jiang et al.
• ArXiv: arxiv.org/abs/2604.04415
• Summary: Proposes structured causal representations for video reasoning that model entities, actions, and temporal relations explicitly, rather than relying on unstructured reasoning typical of current Video-LLMs.
• Sources: HuggingFace (#12, 5 upvotes), arXiv
• Why trending: Video understanding remains a frontier challenge. Causal reasoning in videos is key for real-world agent applications.

13. Large Language Models Generate Harmful Content Using a Distinct, Unified Mechanism

• Authors: Hadas Orgad, Boyi Wei, Kaden Zheng, Martin Wattenberg, Peter Henderson et al.
• ArXiv: arxiv.org/abs/2604.09544
• Summary: Reveals that LLMs use a coherent, unified internal mechanism for generating harmful content — explaining why alignment safeguards are brittle, jailbreaks succeed, and narrow fine-tuning can cause broad emergent misalignment.
• Sources: HuggingFace (#14, 2 upvotes), arXiv, Nature (related)
• Why trending: Critical safety research with implications for alignment strategy. Explains the fundamental fragility of current safety guardrails.

14. ScheMatiQ: From Research Question to Structured Data through Interactive Schema Discovery

• Authors: Shahar Levy, Eliya Habba, Reshef Mintz, Barak Raveh, Renana Keydar et al.
• ArXiv: arxiv.org/abs/2604.09237
• Summary: Uses LLM calls to automate the process of designing annotation schemas and labeling document corpora, transforming natural-language research questions into structured evidence without manual schema design.
• Sources: HuggingFace (#10, 5 upvotes), arXiv
• Why trending: Practical tool for researchers across disciplines. Automates a tedious bottleneck in data-driven research.

15. CT-1: Vision-Language-Camera Models for Camera-Controllable Video Generation

• Authors: Haoyu Zhao, Zihao Zhang, Jiaxi Gu, Haoran Chen, Qingping Zheng et al.
• ArXiv: arxiv.org/abs/2604.09201
• Summary: Transfers spatial reasoning knowledge from vision-language models to enable precise camera-controllable video generation without manual camera trajectory parameters. Bridges the gap between text-based and parameter-based camera control.
• Sources: HuggingFace (listed), arXiv, GitHub (2 stars)
• Why trending: Camera control in video generation is a practical need for content creators. Novel VL-camera model transfer approach.

16. AgentSwing: Adaptive Parallel Context Management for Long-Horizon Web Agents

• Authors: Zhaopeng Feng, Liangcai Su, Zhen Zhang, Xinyu Wang, Xiaotian Zhang et al.
• ArXiv: arxiv.org/abs/2603.27490
• Summary: Addresses the context capacity bottleneck in long-horizon web agents by routing between multiple context management strategies adaptively, rather than committing to a single fixed approach throughout a trajectory.
• Sources: HuggingFace (#13, 4 upvotes), arXiv
• Why trending: Web agents are a hot deployment target. Context management is the practical bottleneck — this tackles it directly.

17. Envisioning the Future, One Step at a Time

• Authors: Stefan Andreas Baumann, Jannik Wiese, Tommaso Martorella, Mahdi M. Kalayeh, Björn Ommer
• ArXiv: arxiv.org/abs/2604.09527
• Summary: Models future scene evolution by representing uncertainty and simulating along extended interaction chains, efficiently exploring plausible futures without dense video or latent-space prediction.
• Sources: HuggingFace (#15, 2 upvotes), arXiv
• Why trending: Scene forecasting has applications in autonomous driving and robotics. Novel approach to efficient future prediction.

18. p1: Better Prompt Optimization with Fewer Prompts

• Authors: Zhaolin Gao, Yu Wang, Bo Liu, Thorsten Joachims et al.
• ArXiv: arxiv.org/abs/2604.08801
• Summary: Studies what makes tasks amenable to prompt optimization and decomposes reward variance into components to guide more efficient prompt search, achieving better results with fewer candidate prompts.
• Sources: HuggingFace (#17, 1 upvote), arXiv
• Why trending: Prompt engineering efficiency is directly practical. Theoretical grounding for when prompt optimization works and when it doesn’t.

19. Initialisation Determines the Basin: Efficient Codebook Optimisation for Extreme LLM Quantization

• Authors: Ian W. Kennedy, Nafise Sadat Moosavi
• ArXiv: arxiv.org/abs/2604.08118
• Summary: Shows that codebook initialisation is the dominant bottleneck for extreme (2-bit) LLM quantization via additive quantization. Greedy sequential initialisation frequently fails catastrophically — better initialisation strategies unlock practical 2-bit deployment.
• Sources: HuggingFace (#19, 1 upvote), arXiv, alphaXiv, Paperium, HuggingFace model (demo)
• Why trending: Edge deployment of LLMs is a major industry need. 2-bit quantization breakthrough removes a key barrier.

20. Cross-Modal Emotion Transfer for Emotion Editing in Talking Face Video

• Authors: Chanhyuk Choi, Taesoo Kim, Donggyu Lee, Siyeol Jung, Taehwan Kim
• ArXiv: arxiv.org/abs/2604.07786
• Summary: Enables cross-modal emotion transfer for talking face generation, allowing emotion editing without label-based constraints. Improves expressiveness and flexibility for generating extended emotional expressions in synthesized face videos.
• Sources: HuggingFace (#16, 2 upvotes), arXiv, GitHub (2 stars)
• Why trending: Talking face generation is a fast-growing application area. Cross-modal emotion transfer is a creative advancement for digital content.

Report generated 2026-04-13 10:00 PDT. Sources: HuggingFace Daily Papers API, arXiv, Hacker News, Reddit, GitHub, alphaXiv, web search.