Daily AI Papers — April 12, 2026

1. Rethinking Generalization in Reasoning SFT: A Conditional Analysis on Optimization, Data, and Model Capability

• Authors: Qihan Ren, Peng Wang, Ruikun Cai, Shuai Shao
• Link: arxiv.org/abs/2604.06628
• Upvotes: 245 ⬆
• Sources: HuggingFace (#1 trending), EmergentMind
• Summary: Challenges the prevailing narrative that SFT memorizes while RL generalizes. Shows that cross-domain generalization in reasoning SFT with long chain-of-thought supervision is not absent but conditional — jointly shaped by optimization dynamics, training data, and base-model capability. Identifies that some reported failures of SFT generalization stem from confounds rather than fundamental limits.
• Why trending: Directly counters a widely-held belief in the post-training community, with implications for how labs should invest in SFT vs RL pipelines for reasoning.

2. ClawBench: Can AI Agents Complete Everyday Online Tasks?

• Authors: Yuxuan Zhang, Yubo Wang, Yipeng Zhu, Penghui Du
• Link: arxiv.org/abs/2604.08523
• Upvotes: 226 ⬆
• Sources: HuggingFace, EmergentMind, ClawBenchLabs.com
• Summary: Introduces an evaluation framework of 153 simple everyday tasks across 144 live platforms spanning 15 categories. Tests whether AI agents can automate routine online tasks like booking, inbox management, and form filling. Provides a realistic yet unsolved testbed for next-gen AI agents.
• Why trending: Practical, relatable benchmark that exposes how far agents still are from true everyday automation — a hot topic as companies race to ship agent products.

3. When Numbers Speak: Aligning Textual Numerals and Visual Instances in Text-to-Video Diffusion Models

• Authors: Zhengyang Sun, Yu Chen, Xin Zhou, Xiaofan Li
• Link: arxiv.org/abs/2604.08546
• Upvotes: 108 ⬆
• Sources: HuggingFace
• Summary: Introduces NUMINA, a training-free identify-then-guide framework that improves numerical alignment in text-to-video models. Addresses the common failure mode where video models generate the wrong number of objects specified in a prompt by selecting discriminative attention heads to derive countable latent layouts.
• Why trending: Solves a visible, frustrating failure mode in video generation that affects nearly every T2V model in production.

4. MegaStyle: Constructing Diverse and Scalable Style Dataset via Consistent Text-to-Image Style Mapping

• Authors: Junyao Gao, Sibo Liu, Jiaxing Li, Yanan Sun
• Link: arxiv.org/abs/2604.08364
• Upvotes: 90 ⬆
• Sources: HuggingFace
• Summary: Presents a novel data curation pipeline that constructs intra-style consistent, inter-style diverse, and high-quality style datasets. Leverages the consistent text-to-image style mapping capability of large generative models to generate images in the same style from style descriptions at scale.
• Why trending: Style-consistent generation is a key differentiator for creative AI tools; this enables scalable style dataset creation without manual curation.

5. LPM 1.0: Video-based Character Performance Model

• Authors: Ailing Zeng, Casper Yang, Chauncey Ge, Eddie Zhang
• Link: arxiv.org/abs/2604.07823
• Upvotes: 52 ⬆
• Sources: HuggingFace
• Summary: Tackles learning character performance (emotion, personality, intent) from video as an alternative to traditional 3D pipelines. Addresses the tension between high expressiveness, real-time inference, and long-horizon identity stability in video-based character models.
• Why trending: Character animation from video is a high-demand capability for gaming, film, and virtual production — this pushes the boundary on expressiveness vs speed tradeoffs.

6. Act Wisely: Cultivating Meta-Cognitive Tool Use in Agentic Multimodal Models

• Authors: Shilin Yan, Jintao Tong, Hongwei Xue, Xiaojun Tang
• Link: arxiv.org/abs/2604.08545
• Upvotes: 34 ⬆
• Sources: HuggingFace
• Summary: Identifies a critical meta-cognitive deficit in agentic multimodal models: they can’t decide when to use internal knowledge vs external tools, leading to blind tool invocation. Proposes methods to cultivate meta-cognitive awareness so agents use tools only when genuinely needed.
• Why trending: Directly relevant to the agent reliability problem — unnecessary tool calls waste compute and cause errors, a pain point for every agent builder.

7. OpenSpatial: A Principled Data Engine for Empowering Spatial Intelligence

• Authors: Jianhui Liu, Haoze Sun, Wenbo Li, Yanbing Zhang
• Link: arxiv.org/abs/2604.07296
• Upvotes: 31 ⬆
• Sources: HuggingFace
• Summary: Fills a critical gap by providing an open-source engine for generating high-quality spatial understanding data. Establishes design principles for robust spatial data generation systems that can fully unleash the potential of spatial intelligence research.
• Why trending: Spatial intelligence is an increasingly important frontier (robotics, AR/VR, autonomous systems), and the lack of principled data engines has been a bottleneck.

8. FIT: A Large-Scale Dataset for Fit-Aware Virtual Try-On

• Authors: Johanna Karras, Yuanhao Wang, Yingwei Li, Ira Kemelmacher-Shlizerman
• Link: arxiv.org/abs/2604.08526
• Upvotes: 18 ⬆
• Sources: HuggingFace
• Summary: Introduces a large-scale dataset that addresses a blind spot in virtual try-on: garment fit accuracy. While existing VTO methods excel at visualizing garment appearance, they ignore how different sizes look on different body types (e.g., how an XL shirt drapes differently than an S).
• Why trending: Virtual try-on is a commercially important application; fit accuracy is the missing piece that makes it actually useful for e-commerce.

9. Structured Distillation of Web Agent Capabilities Enables Generalization

• Authors: Xing Han Lù, Siva Reddy
• Link: arxiv.org/abs/2604.07776
• Upvotes: 17 ⬆
• Sources: HuggingFace
• Summary: Introduces Agent-as-Annotators, a framework that structures synthetic trajectory generation for web agents using modular LLM components (Task Designer, Annotator, Supervisor). Uses Gemini 3 Pro as teacher to distill web navigation capabilities into smaller, locally deployable models.
• Why trending: Demonstrates a practical path to deploy capable web agents locally without relying on expensive frontier API calls — key for cost-sensitive deployments.

10. Small Vision-Language Models are Smart Compressors for Long Video Understanding

• Authors: Junjie Fei, Jun Chen, Zechun Liu, Yunyang Xiong
• Link: arxiv.org/abs/2604.08120
• Upvotes: 13 ⬆
• Sources: HuggingFace
• Summary: Proposes Tempo, an efficient framework for hour-long video understanding that uses small VLMs as smart compressors. Addresses the bottleneck where dense visual streams saturate token budgets and cause lost-in-the-middle problems, replacing blind heuristics like sparse sampling.
• Why trending: Long video understanding is a practical need (meetings, lectures, surveillance) and context limits remain the key bottleneck — Tempo offers an elegant architectural solution.

11. SIM1: Physics-Aligned Simulator as Zero-Shot Data Scaler in Deformable Worlds

• Authors: Yunsong Zhou, Hangxu Liu, Xuekun Jiang, Xing Shen
• Link: arxiv.org/abs/2604.08544
• Upvotes: 13 ⬆
• Sources: HuggingFace
• Summary: Tackles the data scarcity problem for robotic manipulation of deformable objects by building a physics-aligned simulator that serves as a zero-shot data scaler. Moves beyond rigid-body sim-to-real pipelines that produce mismatched geometry and fragile soft dynamics.
• Why trending: Deformable object manipulation is one of the hardest unsolved problems in robotics — this opens a practical sim-to-real path for cloth, food, and soft materials.

12. ViVa: A Video-Generative Value Model for Robot Reinforcement Learning

• Authors: Jindi Lv, Hao Li, Jie Li, Yifei Nie
• Link: arxiv.org/abs/2604.08168
• Upvotes: 12 ⬆
• Sources: HuggingFace
• Summary: Proposes a video-generative value model for robot RL that addresses partial observability and delayed feedback in VLA-based robot manipulation. Uses video generation to assess task progress and guide policy improvement, outperforming VLM-based value models.
• Why trending: Bridges video generation and robot learning — a creative approach that leverages generative models for a traditionally discriminative task.

13. Automating Database-Native Function Code Synthesis with LLMs

• Authors: Wei Zhou, Xuanhe Zhou, Qikang He, Guoliang Li
• Link: arxiv.org/abs/2604.06231
• Upvotes: 12 ⬆
• Sources: HuggingFace
• Summary: Addresses the growing demand for database-native function synthesis by specializing LLM-based code generation for database-specific tasks. Notes that generic code generators like Claude Code are too general for database kernel functions that require deep domain knowledge.
• Why trending: Practical infrastructure problem — databases need growing function libraries, and LLMs could dramatically accelerate this previously manual process.

14. GameWorld: Towards Standardized and Verifiable Evaluation of Multimodal Game Agents

• Authors: Mingyu Ouyang, Siyuan Hu, Kevin Qinghong Lin, Hwee Tou Ng
• Link: arxiv.org/abs/2604.07429
• Upvotes: 9 ⬆
• Sources: HuggingFace
• Summary: Proposes a standardized evaluation framework for multimodal game agents that addresses latency, sparse feedback, and irreversible mistakes. Uses video games as an ideal testbed demanding fine-grained perception, long-horizon planning, and precise control.
• Why trending: Game AI is a compelling proving ground for general agent capabilities — standardized benchmarks are needed as more labs target game-playing agents.

15. Training a Student Expert via Semi-Supervised Foundation Model Distillation

• Authors: Pardis Taghavi, Tian Liu, Renjie Li
• Link: arxiv.org/abs/2604.03841
• Upvotes: 8 ⬆
• Sources: HuggingFace
• Summary: Presents a semi-supervised approach to distilling foundation models into specialized student experts. Combines labeled and unlabeled data to train efficient specialist models that retain the capabilities of large foundation models in specific domains.
• Why trending: Distillation continues to be critical for deploying AI at the edge — semi-supervised approaches reduce the labeled data requirement.

16. Lighting-grounded Video Generation with Renderer-based Agent Reasoning (LiVER)

• Authors: Ziqi Cai, Taoyu Yang, Zheng Chang, Si Li
• Link: arxiv.org/abs/2604.07966
• Upvotes: 7 ⬆
• Sources: HuggingFace
• Summary: Presents LiVER, a diffusion-based framework for controllable video generation with explicit lighting control. Disentangles layout, lighting, and camera trajectory — factors that are typically entangled in current video generation models.
• Why trending: Controllable video generation with physics-based lighting is essential for filmmaking and virtual production workflows.

17. Faithful GRPO: Improving Visual Spatial Reasoning in Multimodal Language Models via Constrained Policy Optimization

• Authors: Sai Srinivas Kancheti, Aditya Kanade, Rohit Sinha, Vineeth N Balasubramanian
• Link: arxiv.org/abs/2604.08476
• Upvotes: 5 ⬆
• Sources: HuggingFace
• Summary: Reveals that accuracy gains from RLVR in multimodal reasoning models often come at the cost of reasoning quality — generated CoT traces are frequently inconsistent with final answers. Proposes constrained policy optimization to improve faithfulness of reasoning chains.
• Why trending: Highlights a critical gap between benchmark accuracy and actual reasoning quality — important for trustworthy deployment of multimodal models.

18. The Master Key Hypothesis: Unlocking Cross-Model Capability Transfer via Linear Subspace Alignment

• Authors: Rishab Balasubramanian, Pin-Jie Lin, Rituraj Sharma
• Link: arxiv.org/abs/2604.06377
• Upvotes: 4 ⬆
• Sources: HuggingFace
• Summary: Proposes that capabilities can be transferred across different models via linear subspace alignment — a “master key” that unlocks capabilities learned by one model in another. Demonstrates cross-model capability transfer without retraining.
• Why trending: If validated, this could dramatically reduce the cost of capability development by enabling transfer between model families.

19. Phantom: Physics-Infused Video Generation via Joint Modeling of Visual and Latent Physical Dynamics

• Authors: Ying Shen, Jerry Xiong, Tianjiao Yu, Ismini Lourentzou
• Link: arxiv.org/abs/2604.08503
• Upvotes: 4 ⬆
• Sources: HuggingFace
• Summary: Argues that scaling data and model size alone doesn’t give video models understanding of physical laws. Proposes joint modeling of visual and latent physical dynamics to enforce physically consistent video generation.
• Why trending: Physics-aware generation is a growing theme — crucial for world models and simulation applications.

20. Appear2Meaning: A Cross-Cultural Benchmark for Structured Cultural Metadata Inference from Images

• Authors: Yuechen Jiang, Enze Zhang, Md Mohsinul Kabir, Qianqian Xie
• Link: arxiv.org/abs/2604.07338
• Upvotes: 4 ⬆
• Sources: HuggingFace
• Summary: Introduces a multi-category, cross-cultural benchmark for inferring structured cultural metadata (creator, origin, period) from visual input. Evaluates VLMs using an LLM-as-Judge framework measuring semantic alignment with reference annotations.
• Why trending: Cultural understanding is an underexplored dimension of VLM evaluation — important for heritage, museum, and global applications.

Report generated 2026-04-12 10:00 PDT. Primary source: HuggingFace Daily Papers API with arxiv abstract enrichment. Cross-referenced with EmergentMind, web search results.