Daily AI Papers — April 15, 2026

1. ClawGUI: A Unified Framework for Training, Evaluating, and Deploying GUI Agents

• Authors: Fei Tang, Zhiqiong Lu, Boxuan Zhang, Weiming Lu, Jun Xiao, Yueting Zhuang, Yongliang Shen
• arxiv: arxiv.org/abs/2604.11784
• Summary: Proposes a unified framework that addresses the full lifecycle of GUI agents — training, evaluation, and deployment — through visual interfaces rather than programmatic APIs. The system interacts with arbitrary software via taps, swipes, and keystrokes, targeting the long tail of applications that CLI-based agents cannot reach.
• Sources: HuggingFace (118 upvotes, #1), arxiv, web search
• Why trending: Massive HuggingFace engagement. GUI agents are a hot topic as the community pushes toward universal computer-use agents. The unified framework approach addresses a real bottleneck in the field.

2. KnowRL: Boosting LLM Reasoning via Reinforcement Learning with Minimal-Sufficient Knowledge Guidance

• Authors: Linhao Yu, Tianmeng Yang, Siyu Ding, Xinyu Li, Yongliang Shen, et al.
• arxiv: arxiv.org/abs/2604.12627
• Summary: Addresses reward sparsity in RLVR for hard reasoning problems by injecting minimal-sufficient knowledge guidance instead of scaling guidance by adding more tokens. Achieves better reasoning performance with more efficient training.
• Sources: HuggingFace (77 upvotes, #2), arxiv, GitHub (zjunlp/KnowRL), EmergentMind
• Why trending: RL for LLM reasoning is the dominant post-training paradigm. This paper offers a principled way to handle the reward sparsity problem that plagues harder benchmarks.

3. Rethinking On-Policy Distillation of Large Language Models: Phenomenology, Mechanism, and Recipe

• Authors: Yaxuan Li, Yuxin Zuo, Bingxiang He, et al.
• arxiv: arxiv.org/abs/2604.13016
• Summary: Provides the first systematic investigation of on-policy distillation (OPD) dynamics for LLMs, identifying two conditions that govern success/failure and offering practical recipes. Companion paper to Lightning OPD (#18).
• Sources: HuggingFace (54 upvotes, #3), arxiv, alphaXiv
• Why trending: On-policy distillation is a core post-training technique and this paper finally explains why it works (or doesn’t), giving practitioners actionable guidance.

4. Turing Test on Screen: A Benchmark for Mobile GUI Agent Humanization

• Authors: Jiachen Zhu, Lingyu Yang, Rong Shan, et al.
• arxiv: arxiv.org/abs/2604.09574
• Summary: Introduces a benchmark for evaluating whether GUI agents can pass as human users, addressing the anti-detection dimension that platforms use to block autonomous agents. Argues humanization is critical for agent survival in human-centric ecosystems.
• Sources: HuggingFace (26 upvotes, #4), arxiv
• Why trending: Directly complements #1 (ClawGUI). As GUI agents mature, the adversarial cat-and-mouse game with platforms is a pressing practical concern.

5. SPPO: Sequence-Level PPO for Long-Horizon Reasoning Tasks

• Authors: Tianyi Wang, Yixia Li, Long Li, et al.
• arxiv: arxiv.org/abs/2604.08865
• Summary: Proposes sequence-level PPO to fix the instability of temporal credit assignment over long Chain-of-Thought horizons. Standard token-level PPO struggles with long reasoning chains; SPPO assigns credit at the sequence level for more stable training.
• Sources: HuggingFace (24 upvotes, #5), arxiv
• Why trending: Long-horizon reasoning with RL is a key challenge. SPPO offers a clean solution to the credit assignment problem that has plagued token-level PPO in CoT settings.

6. Toward Autonomous Long-Horizon Engineering for ML Research (AiScientist)

• Authors: Guoxin Chen, Jie Chen, Lei Chen, et al.
• arxiv: arxiv.org/abs/2604.13018
• Summary: Introduces AiScientist, a system for autonomous ML research engineering that sustains coherent progress across task comprehension, environment setup, implementation, experimentation, and debugging over hours or days. Completed 74 experiment rounds in 23 hours.
• Sources: HuggingFace (20 upvotes, #6), arxiv, alphaXiv, Paperium, QQ News (Chinese tech press)
• Why trending: AI-for-science automation is a major theme. The 23-hour / 74-round benchmark result is eye-catching and the open-source release from RUC (Renmin University) generated Chinese tech press coverage.

7. BERT-as-a-Judge: A Robust Alternative to Lexical Methods for Efficient Reference-Based LLM Evaluation

• Authors: Hippolyte Gisserot-Boukhlef, Nicolas Boizard, Emmanuel Malherbe, et al.
• arxiv: arxiv.org/abs/2604.09497
• Summary: Proposes using fine-tuned BERT models instead of rigid lexical matching for evaluating LLM outputs against references. Shows that lexical methods conflate model capability with format compliance, while BERT-as-a-Judge provides more robust evaluation.
• Sources: HuggingFace (19 upvotes, #7), arxiv
• Why trending: LLM evaluation is a universal pain point. A lightweight BERT-based judge that avoids the cost of LLM-as-a-Judge while being more robust than string matching fills a practical gap.

8. Lyra 2.0: Explorable Generative 3D Worlds

• Authors: Tianchang Shen, Sherwin Bahmani, Kai He, et al. (NVIDIA)
• arxiv: arxiv.org/abs/2604.13036
• Summary: Generates camera-controlled videos simulating scene walkthroughs, then lifts them to explorable 3D via feed-forward reconstruction. Combines visual fidelity of video generation with the interactivity of 3D environments.
• Sources: HuggingFace (17 upvotes, #8), arxiv, alphaXiv, GitHub (nv-tlabs/lyra)
• Why trending: NVIDIA-backed project with open code. Generative 3D worlds are a frontier application combining video generation and 3D reconstruction — demos are visually stunning.

9. Towards Long-horizon Agentic Multimodal Search

• Authors: Yifan Du, Zikang Liu, Jinbiao Peng, et al.
• arxiv: arxiv.org/abs/2604.12890
• Summary: Tackles the challenge of managing heterogeneous information and high token costs in multimodal deep search agents over long horizons. Proposes methods for iteratively collecting textual and visual evidence more efficiently.
• Sources: HuggingFace (14 upvotes, #9), arxiv
• Why trending: Multimodal search agents (deep research) are a hot product category (Gemini Deep Research, Perplexity, etc.). This paper addresses the token cost problem that limits practical deployment.

10. Many-Tier Instruction Hierarchy in LLM Agents

• Authors: Jingyu Zhang, Tianjian Li, William Jurayj, et al.
• arxiv: arxiv.org/abs/2604.09443
• Summary: Formalizes the problem of conflicting instructions from multiple sources (system messages, user prompts, tool outputs, other agents) in LLM agent systems. Proposes a many-tier hierarchy where agents reliably follow the highest-privilege instruction.
• Sources: HuggingFace (12 upvotes, #10), arxiv
• Why trending: Agent safety and instruction following is a critical unsolved problem. As multi-agent systems proliferate, the “who do I listen to?” problem becomes urgent.

11. Self-Adversarial One Step Generation via Condition Shifting

• Authors: Deyuan Liu, Peng Sun, Yansen Han, et al.
• arxiv: arxiv.org/abs/2604.12322
• Summary: Achieves high-fidelity text-to-image synthesis in a single sampling step by using self-adversarial training with condition shifting. Breaks the three-way tradeoff between fidelity, inference speed, and training efficiency.
• Sources: HuggingFace (11 upvotes, #11), arxiv
• Why trending: One-step image generation is the holy grail for real-time applications. This paper avoids external discriminators, making training more practical.

12. Nemotron 3 Super: Open, Efficient MoE Hybrid Mamba-Transformer for Agentic Reasoning

• Authors: NVIDIA (Aakshita Chandiramani, Aaron Blakeman, et al.)
• arxiv: arxiv.org/abs/2604.12374
• Summary: A 120B parameter (12B active) hybrid Mamba-Attention MoE model. First to be pre-trained in NVFP4 and leverage LatentMoE. Targets agentic reasoning with strong efficiency through the Mamba-Transformer hybrid architecture.
• Sources: HuggingFace (11 upvotes, #12), arxiv, NVIDIA Developer Blog
• Why trending: NVIDIA’s open model release combining three hot trends: MoE, Mamba (state-space models), and FP4 quantization. The NVIDIA blog post amplifies reach.

13. Habitat-GS: High-Fidelity Navigation Simulator with Dynamic Gaussian Splatting

• Authors: Ziyuan Xia, Jingyi Xu, Chong Cui, et al.
• arxiv: arxiv.org/abs/2604.12626
• Summary: Upgrades Meta’s Habitat simulator with Gaussian Splatting for photorealistic rendering and dynamic human avatars. Dramatically improves visual fidelity over mesh-based rasterization for training embodied AI agents.
• Sources: HuggingFace (11 upvotes, #13), arxiv
• Why trending: Gaussian Splatting meets embodied AI simulation. Builds on Meta’s widely-used Habitat platform, making it immediately relevant to the embodied AI community.

14. Rethinking the Diffusion Model from a Langevin Perspective

• Authors: Candi Zheng, Yuan Lan
• arxiv: arxiv.org/abs/2604.10465
• Summary: Provides a unified, accessible derivation of diffusion models through the lens of Langevin dynamics. Addresses the question of how the reverse process inverts the forward process, offering a cleaner mathematical foundation than VAE/score-matching perspectives.
• Sources: HuggingFace (7 upvotes, #14), arxiv
• Why trending: Tutorial/foundational papers that simplify complex math always attract attention. Useful for the growing number of researchers entering the diffusion model space.

15. LARY: A Latent Action Representation Yielding Benchmark for Vision-to-Action Alignment

• Authors: Dujun Nie, Fengjiao Chen, Qi Lv, et al.
• arxiv: arxiv.org/abs/2604.11689
• Summary: Addresses the shortage of explicit action data for Vision-Language-Action (VLA) models by proposing a benchmark for transforming visual signals from human videos into ontology-independent latent action representations.
• Sources: HuggingFace (6 upvotes, #15), arxiv
• Why trending: VLA models for robotics are a growing area. Using human videos as a scalable data source for robot learning is an attractive direction.

16. You Only Judge Once: Multi-response Reward Modeling in a Single Forward Pass

• Authors: Yinuo Yang, Zixian Ma, Manasi Ganti, et al.
• arxiv: arxiv.org/abs/2604.10966
• Summary: A discriminative multimodal reward model that scores all candidate responses in a single forward pass, rather than requiring one pass per response. Significantly reduces inference cost for best-of-N sampling and RLHF.
• Sources: HuggingFace (5 upvotes, #16), arxiv
• Why trending: Reward model efficiency directly impacts RLHF training cost. Scoring N responses in one pass is a practical win for anyone doing best-of-N or rejection sampling.

17. Generative Refinement Networks for Visual Synthesis

• Authors: Jian Han, Jinlai Liu, Jiahuan Wang, et al.
• arxiv: arxiv.org/abs/2604.13030
• Summary: Proposes a hybrid approach that combines the complexity-awareness of autoregressive models with the quality of diffusion models. Applies non-uniform computational effort based on generation difficulty, improving efficiency over standard diffusion.
• Sources: HuggingFace (4 upvotes, #17), arxiv
• Why trending: Bridges the AR vs. diffusion debate by combining strengths of both paradigms. Addresses the inefficiency of uniform compute in diffusion models.

18. Lightning OPD: Efficient Post-Training with Offline On-Policy Distillation

• Authors: Yecheng Wu, Song Han, Hai Cai
• arxiv: arxiv.org/abs/2604.13010
• Summary: Eliminates the need for a live teacher inference server during on-policy distillation by using offline cached teacher outputs. Dramatically reduces infrastructure overhead while maintaining distillation quality.
• Sources: HuggingFace (4 upvotes, #18), arxiv
• Why trending: Companion to #3. Makes on-policy distillation practical for teams without massive GPU clusters for teacher inference. Song Han (MIT) co-author adds credibility.

19. Accelerating Speculative Decoding with Block Diffusion Draft Trees

• Authors: Liran Ringel, Yaniv Romano
• arxiv: arxiv.org/abs/2604.12989
• Summary: Uses block diffusion models as drafters for speculative decoding, generating entire draft blocks in a single forward pass. Improves upon DFlash by adding tree-based verification for higher acceptance rates.
• Sources: HuggingFace (3 upvotes, #19), arxiv
• Why trending: Speculative decoding is the main inference acceleration technique. Combining it with diffusion-based drafting is a novel and practical approach.

20. GlotOCR Bench: OCR Models Still Struggle Beyond a Handful of Unicode Scripts

• Authors: Amir Hossein Kargaran, Nafiseh Nikeghbal, Jana Diesner, et al.
• arxiv: arxiv.org/abs/2604.12978
• Summary: Introduces a comprehensive OCR benchmark spanning 100+ Unicode scripts, revealing that even state-of-the-art vision-language models fail dramatically on low-resource scripts. Highlights a major blind spot in current multimodal models.
• Sources: HuggingFace (3 upvotes, #20), arxiv
• Why trending: Exposes a concrete failure mode of frontier VLMs. The 100+ script coverage is unprecedented and the results are a wake-up call for the multilingual AI community.

Key Themes Today

1. GUI/Computer-Use Agents (#1, #4) — Building and evaluating agents that operate GUIs like humans
2. RL for LLM Reasoning (#2, #5) — Improving reward signals and credit assignment in RLVR
3. Distillation & Post-Training (#3, #18) — Understanding and optimizing knowledge distillation
4. Autonomous AI Research (#6) — AI systems that run multi-day ML experiments
5. Efficient Inference (#12, #16, #19) — MoE, reward model efficiency, speculative decoding
6. 3D Generation (#8) — Video-to-3D world generation
7. Agent Safety (#10) — Instruction hierarchy and trust in multi-agent systems

Report generated 2026-04-15 10:00 PDT. Sources: HuggingFace Daily Papers API, arxiv, web search (NVIDIA blog, GitHub, EmergentMind, alphaXiv, Chinese tech press). Reddit/HN/X direct access blocked by enterprise proxy.