Daily AI Papers — April 17, 2026

1. HY-World 2.0: A Multi-Modal World Model for Reconstructing, Generating, and Simulating 3D Worlds

• Authors: Team HY-World (Tencent Hunyuan), Chenjie Cao, Xuhui Zuo, et al.
• Summary: Introduces HY-World 2.0, a multi-modal world model framework that accepts text, single-view images, multi-view images, and videos to produce 3D world representations. Advances over HY-World 1.0 with unified generation, reconstruction, and simulation capabilities for creating immersive 3D environments.
• Link: arxiv.org/abs/2604.14268
• Sources found: HuggingFace (63 upvotes, #1), arxiv, Gigazine, ToolHunter, HY-World official site
• Why trending: Most upvoted paper of the day on HuggingFace. Tencent’s major release for 3D world generation with broad media coverage and practical implications for gaming/simulation.

2. RAD-2: Scaling Reinforcement Learning in a Generator-Discriminator Framework

• Authors: Hao Gao, Shaoyu Chen, Yifan Zhu, et al.
• Summary: Proposes a generator-discriminator framework for autonomous driving motion planning using diffusion-based planners with reinforcement learning. Addresses stochastic instabilities and lack of corrective negative feedback in closed-loop interactions for high-level autonomous driving.
• Link: arxiv.org/abs/2604.15308
• Sources found: HuggingFace (20 upvotes), arxiv, ICLR 2026 paper list, YouTube explainer
• Why trending: ICLR 2026 accepted paper with strong community interest in autonomous driving + RL intersection. YouTube coverage amplifying reach.

3. DR³-Eval: Towards Realistic and Reproducible Deep Research Evaluation

• Authors: Qianqian Xie, Qingheng Xiong, He Zhu, et al.
• Summary: Proposes a realistic and reproducible benchmark for evaluating Deep Research Agents (DRAs) that handle complex, long-horizon research tasks involving planning, retrieval, multimodal understanding, and report generation. Addresses challenges of dynamic web environments and ambiguous task definitions in DRA evaluation.
• Link: arxiv.org/abs/2604.14683
• Sources found: HuggingFace (20 upvotes), arxiv
• Why trending: Timely benchmark paper as deep research agents (like Gemini Deep Research, OpenAI Deep Research) are a hot product category. Fills a critical evaluation gap.

4. HiVLA: A Visual-Grounded-Centric Hierarchical Embodied Manipulation System

• Authors: Tianshuo Yang, Guanyu Chen, Yutian Chen, et al.
• Summary: Proposes a hierarchical VLA (Vision-Language-Action) system that preserves the deep reasoning capabilities of base VLMs while enabling fine-grained robotic manipulation. Resolves the fundamental trade-off between reasoning and control by separating visual grounding from action generation.
• Link: arxiv.org/abs/2604.14125
• Sources found: HuggingFace (16 upvotes), arxiv
• Why trending: Addresses a core bottleneck in embodied AI — the reasoning-vs-control trade-off in VLA models — which is a major active research area.

5. How to Fine-Tune a Reasoning Model? A Teacher-Student Cooperation Framework

• Authors: Zixian Huang, Kaichen Yang, Xu Huang, et al.
• Summary: Identifies that standard SFT with stronger-model synthetic data often fails or degrades performance for reasoning models like Qwen3-8B due to stylistic mismatches. Proposes a teacher-student cooperation framework that synthesizes student-consistent training data, maintaining reasoning quality.
• Link: arxiv.org/abs/2604.14164
• Sources found: HuggingFace (16 upvotes), arxiv
• Why trending: Directly relevant to practitioners fine-tuning open reasoning models. The finding that naive distillation hurts reasoning is counter-intuitive and practically important.

6. ASGuard: Activation-Scaling Guard to Mitigate Targeted Jailbreaking Attack

• Authors: Yein Park, Jungwoo Park, Jaewoo Kang
• Summary: Reveals that LLMs exhibit brittle refusal behaviors that can be circumvented by simple linguistic changes (e.g., tense rephrasing). Proposes ASGuard, an activation-scaling defense mechanism that addresses the underlying generalization gap in alignment methods.
• Link: arxiv.org/abs/2509.25843
• Sources found: HuggingFace (16 upvotes), arxiv
• Why trending: LLM safety/jailbreaking is a perennial hot topic. The tense-based jailbreaking insight is novel and highlights fundamental alignment weaknesses.

7. GlobalSplat: Efficient Feed-Forward 3D Gaussian Splatting via Global Scene Tokens

• Authors: Roni Itkin, Noam Issachar, Yehonatan Keypur, et al.
• Summary: Addresses the trade-off between representation compactness, reconstruction speed, and rendering fidelity in 3D Gaussian Splatting. Introduces global scene tokens for efficient feed-forward inference, improving spatial allocation of Gaussian primitives.
• Link: arxiv.org/abs/2604.15284
• Sources found: HuggingFace (13 upvotes), arxiv
• Why trending: 3D Gaussian Splatting remains one of the hottest areas in computer vision. Feed-forward approaches that eliminate per-scene optimization are highly sought after.

8. Dive into Claude Code: The Design Space of Today’s and Future AI Agent Systems

• Authors: Jiacheng Liu, Xiaohan Zhao, Xinyi Shang, et al.
• Summary: Provides a comprehensive architecture analysis of Claude Code (Anthropic’s agentic coding tool) by studying its open-source TypeScript codebase. Compares it with OpenClaw, an independent open-source AI agent system, mapping out the design space for agentic coding assistants.
• Link: arxiv.org/abs/2604.14228
• Sources found: HuggingFace (5 upvotes), arxiv, bits-bytes-nn blog
• Why trending: Meta-analysis of a widely-used AI agent tool. Highly relevant to the developer tools community as agentic coding tools become mainstream.

9. UniDoc-RL: Coarse-to-Fine Visual RAG with Hierarchical Actions and Dense Rewards

• Authors: Jun Wang, Shuo Tan, Zelong Sun, et al.
• Summary: Proposes a unified reinforcement learning framework for visual RAG that uses hierarchical actions and dense rewards. Addresses limitations of existing visual RAG systems that rely on generic retrieval signals and overlook fine-grained visual semantics for complex reasoning.
• Link: arxiv.org/abs/2604.14967
• Sources found: HuggingFace (6 upvotes), arxiv
• Why trending: RAG + vision + RL is a compelling intersection. Practical implications for document understanding and multimodal AI systems.

10. Switch-KD: Visual-Switch Knowledge Distillation for Vision-Language Models

• Authors: Haoyi Sun, Xiaoxiao Wang, Ning Mao, et al.
• Summary: Introduces a visual-switch knowledge distillation approach for compressing VLMs without increasing model size or data requirements. Makes deployment of large VLMs feasible in resource-constrained scenarios through efficient distillation.
• Link: arxiv.org/abs/2604.14629
• Sources found: HuggingFace (6 upvotes), arxiv
• Why trending: VLM compression is critical for edge deployment. Novel distillation approach addresses a practical bottleneck in deploying multimodal models.

11. TRACER: Trace-Based Adaptive Cost-Efficient Routing for LLM Classification

• Authors: Adam Rida
• Summary: Demonstrates that production LLM logs constitute a free training set for lightweight surrogates that can absorb significant traffic at near-zero marginal cost. Provides principled methods for determining when to route to the surrogate vs. the full LLM.
• Link: arxiv.org/abs/2604.14531
• Sources found: HuggingFace (5 upvotes), arxiv
• Why trending: Directly addresses LLM inference cost optimization — a top priority for production deployments. Elegant use of existing production data.

12. Boosting Visual Instruction Tuning with Self-Supervised Guidance

• Authors: Sophia Sirko-Galouchenko, Monika Wysoczanska, Andrei Bursuc, et al.
• Summary: Identifies that MLLMs underutilize visual information during instruction tuning because models take shortcuts through text. Proposes self-supervised guidance to improve fine-grained visual reasoning without additional data.
• Link: arxiv.org/abs/2604.12966
• Sources found: HuggingFace (4 upvotes), arxiv
• Why trending: Addresses a fundamental limitation in multimodal LLM training — the visual shortcut problem — with a clean self-supervised solution.

13. Representations Before Pixels (Re2Pix): Semantics-Guided Hierarchical Video Prediction

• Authors: Efstathios Karypidis, Spyros Gidaris, Nikos Komodakis, et al.
• Summary: Presents a hierarchical video prediction framework that decomposes forecasting into semantic representation prediction and representation-guided visual synthesis. Achieves both high visual fidelity and consistent scene semantics for autonomous driving scenarios.
• Link: arxiv.org/abs/2604.11707
• Sources found: HuggingFace (4 upvotes), arxiv
• Why trending: Video prediction for autonomous driving is a high-impact application. The hierarchical decomposition approach is architecturally interesting.

14. LongAct: Harnessing Intrinsic Activation Patterns for Long-Context Reinforcement Learning

• Authors: Bowen Ping, Zijun Chen, Tingfeng Hui, et al.
• Summary: Explores using intrinsic activation patterns of LLMs to guide reinforcement learning training for long-context reasoning. One of the first works to leverage model activations as RL training signals for improving long-context capabilities.
• Link: arxiv.org/abs/2604.14922
• Sources found: HuggingFace (3 upvotes), arxiv
• Why trending: Long-context reasoning is a key frontier. Novel approach of using activation patterns as training signals for RL-based improvements.

15. KV Packet: Recomputation-Free Context-Independent KV Caching for LLMs

• Authors: Chuangtao Chen, Grace Li Zhang, Xunzhao Yin, et al.
• Summary: Addresses the core limitation that standard KV caches are context-dependent, requiring recomputation when reusing cached documents in new contexts. Proposes KV Packets — recomputation-free, context-independent caching that outperforms CacheBlend, EPIC, and SAM-KV.
• Link: arxiv.org/abs/2604.13226
• Sources found: HuggingFace (3 upvotes), arxiv
• Why trending: KV cache optimization is critical for LLM serving efficiency. Context-independent caching would be a significant infrastructure improvement.

16. Don’t Retrieve, Navigate: Distilling Enterprise Knowledge into Navigable Agent Skills (Corpus2Skill)

• Authors: Yiqun Sun, Pengfei Wei, Lawrence B. Hsieh, et al.
• Summary: Challenges the standard RAG paradigm by proposing Corpus2Skill, which distills document corpora into hierarchical navigable agent skills. Enables the LLM to actively traverse knowledge rather than passively consuming search results, improving backtracking and evidence combination.
• Link: arxiv.org/abs/2604.14572
• Sources found: HuggingFace (3 upvotes), arxiv
• Why trending: “Don’t Retrieve, Navigate” is a provocative framing that challenges the dominant RAG approach. Practical implications for enterprise AI.

17. RadAgent: A Tool-Using AI Agent for Stepwise Interpretation of Chest CT

• Authors: Mélanie Roschewitz, Kenneth Styppa, Yitian Tao, et al.
• Summary: Introduces RadAgent, a VLM-based agent for CT interpretation that provides stepwise, interpretable reasoning traces for clinicians to inspect and refine. Shifts clinicians from passive observers to active participants in AI-assisted radiology.
• Link: arxiv.org/abs/2604.15231
• Sources found: HuggingFace (3 upvotes), arxiv
• Why trending: Medical AI with human-in-the-loop interpretability is a growing priority. Tool-using agents in clinical settings is a compelling application.

18. Beyond Prompts: Unconditional 3D Inversion for Out-of-Distribution Shapes

• Authors: Victoria Yue Chen, Emery Pierson, Léopold Maillard, et al.
• Summary: Demonstrates that state-of-the-art text-driven 3D generative models lose sensitivity to natural language prompts for out-of-distribution shapes. Proposes unconditional inversion methods that bypass the text-bottleneck for 3D manipulation.
• Link: arxiv.org/abs/2604.14914
• Sources found: HuggingFace (3 upvotes), arxiv
• Why trending: Exposes a fundamental limitation of text-conditioned 3D generation and offers a practical workaround.

19. Cross-Tokenizer LLM Distillation through a Byte-Level Interface

• Authors: Avyav Kumar Singh, Yen-Chen Wu, Alexandru Cioba, et al.
• Summary: Tackles cross-tokenizer distillation — transferring knowledge between LLMs with different tokenizers — via a byte-level interface. Eliminates the need for heuristic vocabulary alignment strategies used in prior approaches.
• Link: arxiv.org/abs/2604.07466
• Sources found: HuggingFace (3 upvotes), arxiv
• Why trending: Cross-tokenizer distillation is a largely unsolved problem. The byte-level interface is an elegantly simple solution with broad applicability.

20. LeapAlign: Post-Training Flow Matching Models at Any Generation Step

• Authors: Zhanhao Liang, Tao Yang, Jie Wu, et al.
• Summary: Addresses alignment of flow matching models with human preferences through reward gradient backpropagation. Solves the prohibitive memory costs and gradient explosion of long-trajectory backpropagation by building efficient two-step trajectories.
• Link: arxiv.org/abs/2604.15311
• Sources found: HuggingFace (2 upvotes), arxiv
• Why trending: Flow matching is an emerging alternative to diffusion models. Post-training alignment with human preferences is the next frontier for these generative models.

Report generated 2026-04-17 10:00 PDT. Primary source: HuggingFace Daily Papers API. Cross-referenced with arxiv, web search, Gigazine, ICLR 2026 proceedings, YouTube, and tech blogs. Reddit and Hacker News APIs were unavailable at fetch time.