Daily AI Papers — March 31, 2026

1. TAPS: Task Aware Proposal Distributions for Speculative Sampling

Authors: Mohamad Zbib, Mohamad Bazzi, Ammar Mohanna, Hasan Abed Al Kader Hammoud, Bernard Ghanem Summary: Studies how the draft model’s training distribution affects speculative decoding quality. Lightweight HASS and EAGLE-2 drafters trained on domain-specific data (MathInstruct, ShareGPT) significantly outperform generic drafters. Shows that task-aware proposal distributions can meaningfully improve speculative sampling without changing the target model. Link: arxiv.org/abs/2603.27027 Source: HuggingFace trending (#1 on Mar 31) Why trending: Speculative decoding is a key inference optimization. This paper shows a simple, actionable insight: match your drafter to your task for better acceptance rates.

2. Towards a Medical AI Scientist

Authors: Hongtao Wu, Boyun Zheng, Dingjie Song, Yu Jiang, Jianfeng Gao et al. Summary: First autonomous research framework tailored to clinical medicine. Unlike domain-agnostic AI Scientist systems, this one is grounded in medical evidence and handles specialized data modalities. End-to-end pipeline from hypothesis generation to manuscript drafting for medical research. Link: arxiv.org/abs/2603.28589 Source: HuggingFace trending (#2 on Mar 31), arxiv cs.AI Why trending: AI Scientist papers are hot, and this is the first one specialized for medicine. Jianfeng Gao (Microsoft Research) is a co-author.

3. Gen-Searcher: Reinforcing Agentic Search for Image Generation

Authors: Kaituo Feng, Manyuan Zhang, Shuang Chen, Yunlong Lin, Kaixuan Fan et al. Summary: First attempt to train a search-augmented image generation agent. Performs multi-hop reasoning and web search to collect textual knowledge needed for knowledge-intensive prompts that stump frozen image generation models. Uses RL to train the search agent. Link: arxiv.org/abs/2603.28767 Source: HuggingFace trending (#3 on Mar 31), related to Search-o1 line of work Why trending: Novel intersection of retrieval-augmented generation and image synthesis. Addresses the fundamental limitation of frozen knowledge in image models.

4. Emergent Social Intelligence Risks in Generative Multi-Agent Systems

Authors: Yue Huang, Yu Jiang, Wenjie Wang, Haomin Zhuang, Xiaonan Luo et al. Summary: Pioneer study of failure modes that emerge when multiple generative agents interact – risks that cannot be reduced to individual agent behavior. Studies negotiation, resource allocation, and planning in multi-agent systems. Identifies collective failure patterns unique to multi-agent settings. Link: arxiv.org/abs/2603.27771 Source: HuggingFace trending (#4 on Mar 31), arxiv cs.AI Why trending: Multi-agent safety is under-studied compared to single-agent alignment. Timely as agent deployments scale.

5. EpochX: Building Infrastructure for Emergent Agent Civilization

Authors: Huacan Wang, Chaofa Yuan, Xialie Zhuang, Tu Hu, Shuo Zhang et al. Summary: A credits-native marketplace infrastructure for human-agent production. Argues the binding constraint has shifted from raw AI capability to how work is delegated, verified, and rewarded at scale. Builds infrastructure for agent-to-agent and human-to-agent economic coordination. Link: arxiv.org/abs/2603.27304 Source: HuggingFace trending (#5 on Mar 31), project website (epochx.cc), blog post (epochx.cc/blog) Why trending: Bold vision paper with working infrastructure. Agent marketplaces are an emerging category.

6. GEditBench v2: A Human-Aligned Benchmark for General Image Editing

Authors: Zhangqi Jiang, Zheng Sun, Xianfang Zeng, Yufeng Yang, Xuanyang Zhang et al. Summary: Comprehensive benchmark with 1,200 real-world image editing tasks spanning six core tasks. Current benchmarks suffer from narrow task coverage and metrics that fail to capture visual consistency. GEditBench v2 provides human-aligned evaluation across generation, editing, single and multiple references. Link: arxiv.org/abs/2603.28547 Source: HuggingFace trending (#6 on Mar 31), arxiv cs.CV Why trending: Image editing benchmarks are crucial as editing models mature rapidly. Fills a clear evaluation gap.

7. Kernel-Smith: A Unified Recipe for Evolutionary Kernel Optimization

Authors: He Du, Qiming Ge, Jiakai Hu, Aijun Yang, Zheng Cai et al. Summary: Framework for high-performance GPU kernel generation combining evolutionary search agents with post-training recipes. Maintains a population of executable kernel candidates and iteratively improves them using execution feedback on compilation, correctness, and speedup. Practical tool for GPU optimization. Link: arxiv.org/abs/2603.28342 Source: HuggingFace trending (#13 on Mar 31), alphaxiv.org, AMD ROCm blog discusses related GEAK-Triton agents Why trending: GPU kernel optimization is a bottleneck for AI infrastructure teams. Evolutionary agent approach is novel and practical. Cross-platform discussion.

8. On Token’s Dilemma: Dynamic MoE with Drift-Aware Token Assignment for Continual Learning

Authors: Chongyang Zhao, Mingsong Li, Haodong Lu, Dong Gong Summary: Addresses forgetting in MoE-based continual learning of large vision-language models. Identifies “routing-drift” as the root cause: old-task tokens get misassigned as new experts are added. Proposes drift-aware token assignment to maintain routing consistency across tasks. Link: arxiv.org/abs/2603.27481 Source: HuggingFace trending (#7 on Mar 31), arxiv cs.CV Why trending: MoE architectures are everywhere now. Understanding and fixing their failure modes in continual learning is immediately relevant.

9. PRBench: End-to-end Paper Reproduction in Physics Research

Authors: Shi Qiu, Junyi Deng, Yiwei Deng, Haoran Dong, Jieyu Fu et al. Summary: Benchmark of 30 expert-curated tasks spanning 11 subfields of physics, testing whether AI agents can reliably perform end-to-end reproduction from real scientific papers. Each task requires comprehending methodology and producing verifiable results. Link: arxiv.org/abs/2603.27646 Source: HuggingFace trending (#8 on Mar 31), arxiv cs.AI Why trending: Scientific reproducibility is a key test for AI agents. Physics provides hard, verifiable ground truth.

10. Make Geometry Matter for Spatial Reasoning

Authors: Shihua Zhang, Qiuhong Shen, Shizun Wang, Tianbo Pan, Xinchao Wang Summary: Shows that naively fusing geometry tokens from 3D foundation models into VLMs via standard fine-tuning leaves geometric cues underutilized. Proposes methods to make geometry information actually matter for spatial reasoning in both static scenes and dynamic videos. Link: arxiv.org/abs/2603.26639 Source: HuggingFace trending (#9 on Mar 31), arxiv cs.CV Why trending: Spatial reasoning is a known weakness of VLMs. This paper provides a concrete diagnosis and fix.

11. MolmoPoint: Better Pointing for VLMs with Grounding Tokens

Authors: Christopher Clark, Yue Yang, Jae Sung Park, Zixian Ma, Jieyu Zhang et al. (Allen AI) Summary: Instead of generating coordinates as text (which requires learning a complex coordinate system), proposes pointing tokens that directly select visual tokens containing the target concept via cross-attention. More intuitive and efficient grounding mechanism for VLMs. Models released on HuggingFace. Link: arxiv.org/abs/2603.28069 Source: HuggingFace trending (Mar 31), Allen AI blog (allenai.org/blog/molmopoint), HF model releases (MolmoPoint-8B, MolmoPoint-Vid-4B) Why trending: Allen AI release with open models. Fundamentally simpler approach to visual grounding that outperforms coordinate-based methods.

12. On-the-fly Repulsion in the Contextual Space for Rich Diversity in Diffusion Transformers

Authors: Omer Dahary, Benaya Koren, Daniel Garibi, Daniel Cohen-Or Summary: Modern T2I diffusion models achieve great semantic alignment but converge on narrow visual solutions – “typicality bias.” Identifies a fundamental trade-off in current diversity approaches and proposes on-the-fly repulsion in contextual space to generate diverse outputs without costly optimization. Link: arxiv.org/abs/2603.28762 Source: HuggingFace trending (#10 on Mar 31), arxiv cs.CV Why trending: Addresses a real pain point in production image generation: every prompt gives the same-looking output. Clean solution.

13. MuSEAgent: A Multimodal Reasoning Agent with Stateful Experiences

Authors: Shijian Wang, Jiarui Jin, Runhao Fu, Zexuan Yan, Xingjian Wang et al. Summary: Extends research agents beyond trajectory-level retrieval by proposing a stateful experience learning paradigm. Abstracts interaction data into atomic decision patterns that can be retrieved and applied across different reasoning contexts, enhancing multimodal decision-making. Link: arxiv.org/abs/2603.27813 Source: HuggingFace trending (#12 on Mar 31), arxiv cs.AI Why trending: Agent memory and experience accumulation is a key unsolved problem. Stateful experiences is a compelling abstraction.

14. Marco DeepResearch: Unlocking Efficient Deep Research Agents via Verification-Centric Design

Authors: Bin Zhu, Qianghuai Jia, Tian Lan, Junyang Ren, Feng Gu et al. Summary: Deep research agents fail on long-horizon tasks because they lack explicit verification during data synthesis, trajectory construction, and test-time scaling. Proposes verification-centric design that catches errors early rather than compounding them through multi-step reasoning chains. Link: arxiv.org/abs/2603.28376 Source: HuggingFace trending (#15 on Mar 31), HuggingFace DeepResearch Bench Leaderboard space, blog discussion (fuchsfranklin.github.io) Why trending: Deep research is a hot capability. Verification is the key missing piece – this paper addresses it directly. Cross-platform discussion.

15. ImagenWorld: Stress-Testing Image Generation Models with Explainable Human Evaluation

Authors: Samin Mahdizadeh Sani, Max Ku, Nima Jamali, Matina Mahdizadeh Sani et al. Summary: Benchmark of 3,600 condition sets spanning six core image generation/editing tasks. Unlike existing benchmarks that give opaque scores, ImagenWorld explains failure modes. Covers generation, editing, single/multiple references with human-aligned evaluation. Link: arxiv.org/abs/2603.27862 Source: HuggingFace trending (#11 on Mar 31), arxiv cs.CV Why trending: Explainable evaluation for image generation is sorely needed. 3.6K conditions across six tasks is comprehensive.

16. HISA: Efficient Hierarchical Indexing for Fine-Grained Sparse Attention

Authors: Yufei Xu, Fanxu Meng, Fan Jiang, Yuxuan Wang, Ruijie Zhou et al. Summary: Token-level sparse attention (like DeepSeek Sparse Attention) achieves fine-grained key selection but the indexer still scans the entire prefix for every query – O(L^2) bottleneck. HISA proposes hierarchical indexing that eliminates this bottleneck while maintaining selection quality. Link: arxiv.org/abs/2603.28458 Source: HuggingFace trending (Mar 31), arxiv cs.CL Why trending: Directly extends DeepSeek’s sparse attention architecture. Addresses a concrete O(L^2) bottleneck that limits long-context scaling.

17. ResAdapt: Adaptive Resolution for Efficient Multimodal Reasoning

Authors: Huanxuan Liao, Zhongtao Jiang, Yupu Hao, Yuqiao Tan, Shizhu He et al. Summary: MLLMs scale visual understanding by increasing input fidelity, but visual token growth makes high resolution + long temporal context prohibitive. ResAdapt learns how much visual budget each frame needs, adapting resolution at the input side rather than compressing post-encoding representations. Link: arxiv.org/abs/2603.28610 Source: HuggingFace trending (#14 on Mar 31), arxiv cs.CV Why trending: Practical efficiency gain for multimodal models. Input-side adaptation is a cleaner approach than post-hoc compression.

18. DreamLite: A Lightweight On-Device Unified Model for Image Generation and Editing

Authors: Kailai Feng, Yuxiang Wei, Bo Chen, Yang Pan, Hu Ye et al. Summary: Compact 0.39B parameter diffusion model that supports both text-to-image generation and text-guided image editing on-device. Existing on-device models focus only on generation. DreamLite unifies both capabilities in a single lightweight model. Link: arxiv.org/abs/2603.28713 Source: HuggingFace trending (#16 on Mar 31), arxiv cs.CV Why trending: On-device AI is growing fast. A unified generation + editing model at 0.39B parameters is immediately practical for mobile deployment.

19. KAT-Coder-V2 Technical Report

Authors: Fengxiang Li, Han Zhang, Haoyang Huang, Jinghui Wang et al. (Kuaishou KwaiKAT Team) Summary: Agentic coding model using a “Specialize-then-Unify” paradigm: decomposes coding into five expert domains (SWE, WebCoding, Terminal, WebSearch, General), each with independent SFT and RL, before consolidation via on-policy distillation. Built KwaiEnv infrastructure sustaining tens of thousands of concurrent environments. Link: arxiv.org/abs/2603.27703 Source: HuggingFace trending (Mar 31), arxiv cs.SE Why trending: Major company (Kuaishou) release. Novel multi-domain expert approach to agentic coding. KwaiEnv infrastructure is a notable engineering contribution.

20. HandX: Scaling Bimanual Motion and Interaction Generation

Authors: Zimu Zhang, Yucheng Zhang, Xiyan Xu, Ziyin Wang, Sirui Xu et al. Summary: Synthesizing realistic hand motion and bimanual interaction remains underexplored. HandX provides a unified foundation spanning data, annotation, and evaluation for dexterous bimanual behavior – finger articulation, contact timing, and inter-hand coordination. Link: arxiv.org/abs/2603.28766 Source: HuggingFace trending (#18 on Mar 31), arxiv cs.CV Why trending: Fills a clear gap in motion generation research. Hand/finger motion is one of the hardest unsolved problems in character animation.

Honorable Mentions

Papers that didn’t make the top 20 but are worth watching:

• Story2Proposal (2603.27065) – Contract-governed multi-agent framework for scientific manuscript generation
• Density-aware Soft Context Compression (2603.25926) – Dynamic compression ratios for long-context LLMs
• SEAR (2603.26728) – Schema-based evaluation and routing for LLM gateways
• AdaptToken (2603.28696) – Entropy-based adaptive token selection for long video understanding
• Superintelligence and Law (2603.28669) – Legal analysis of AI superintelligence implications
• MOOZY (2603.27048) – Patient-first foundation model for computational pathology
• Surgical AI Comparative Study (2603.27341) – Datasets, foundation models, and barriers to Med-AGI

Methodology

Papers are ranked by: HuggingFace position > cross-platform presence > lab prestige > topic relevance.