📊 Dataset #Dataset #LLM #SyntheticData #Personas

Nemotron-Personas-Korea
👤 nvidia

🎯 Task
Korean persona generation

💡 Idea
The dataset synthesizes adult Korean personas from official population statistics and public sources, covering demographics, geography, occupations, and multiple persona descriptions. It was built with NeMo Data Designer to better reflect real-world Korean population diversity.

✨ Why it's interesting
Notable as the first large-scale Korean-language persona dataset with 1M records and 7M personas.

Size: 1M records, 7M personas, 2.0 GB

Downloads: 2.0k | Likes: 46

🔗 dataset

via @Papers.Data.Code

💻 Repo #Repo #SearchEvaluation #SearchEvaluation #CitationAnalysis

Geo Citation Lab
👤 yaojingang

🎯 Task
AI search citation analysis

💡 Idea
It provides 602 prompts, raw citation/search CSVs, page crawls, 72-feature citation records, and scripts to analyze search triggering, source preferences, and citation influence across three platforms.

✨ Why it's interesting
Includes 21,143 citation records, 18,151 crawled pages, and a full reproducible analysis pipeline.

💻 Repo
⭐ yaojingang/geo-citation-lab — 130 stars (+130 3d)
Python


via @Papers.Data.Code

📄 Paper #Paper #NLP #LLMAgents #WorkflowOrchestration

AgentSPEX: An Agent SPecification and EXecution Language
👤 Pengcheng Wang, Jerry Huang, Jiarui Yao et al.

🎯 Task
LLM agent workflow specification

💡 Idea
It defines agent workflows declaratively with typed steps, branching, loops, parallelism, reusable submodules, and explicit context/state management, then runs them in a harness with tools, sandboxing, checkpointing, logging, and replay.

✨ Why it's interesting
It outperformed compared baselines on 7 benchmarks and was rated more interpretable in a user study.

💻 Repo
⭐ ScaleML/AgentSPEX — 43 stars

🔗 paper

via @Papers.Data.Code

📋 Weekly Digest | Apr 18 – Apr 25
#WeeklyDigest

📄 Papers

LLaDA2.0-Uni: Unifying Multimodal Understanding and Generation with Diffusion Large Language Model
#MultimodalLearning #ImageGeneration #DiffusionModels
Discrete diffusion LLM ⟶ unifies understanding and generation
→ Learn more...

AgentSPEX: An Agent SPecification and EXecution Language
#LLMAgents #WorkflowOrchestration #ProgramSynthesis
YAML agent workflows ⟶ beats baselines on 7 benchmarks
→ Learn more...

OneVL: One-Step Latent Reasoning and Planning with Vision-Language Explanation
#AutonomousDriving #TrajectoryPrediction #WorldModels
Latent CoT VLA ⟶ beats explicit CoT
→ Learn more...

CoInteract: Physically-Consistent Human-Object Interaction Video Synthesis via Spatially-Structured Co-Generation
#VideoGeneration #DiffusionTransformers #HumanObjectInteraction
CoInteract diffusion framework ⟶ more stable realistic HOI videos
→ Learn more...

Extending One-Step Image Generation from Class Labels to Text via Discriminative Text Representation
#TextToImage #FewStepGeneration #FlowMatching
Text-conditioned MeanFlow ⟶ 0.90 GenEval in 4 steps
→ Learn more...

💻 Repos

cosmicstack-labs/mercury-agent ⭐
#AIAgents #TelegramBots #ToolUse
TypeScript CLI Telegram agent ⟶ approval-based 24/7 tool use
→ Learn more...

yaojingang/geo-citation-lab ⭐
#SearchEvaluation #CitationAnalysis #WebCrawling
Citation analysis dataset ⟶ studies search and citation choices
→ Learn more...

intertwine/dspy-agent-skills ⭐
#AgentSkills #Dspy #Gepa
DSPy agent skills ⟶ coding workflows with tests
→ Learn more...

📊 Datasets

ParseBench
#DocumentParsing #OCR #LayoutDetection
ParseBench dataset ⟶ evaluates enterprise document parsers
→ Learn more...

Nemotron-Personas-Korea
#SyntheticData #Personas #Korean
Synthetic Korean persona dataset ⟶ model training and evaluation
→ Learn more...

➡️ Tomorrow — NLP & LLM Monthly

via @Papers.Data.Code

📊 Monthly: NLP & LLM | April 2026
#MonthlyDigest #NLPAndLLM

📄 Papers

AgentSPEX: An Agent SPecification and EXecution Language
#LLMAgents #WorkflowOrchestration #ProgramSynthesis
YAML agent workflow language ⟶ beats baselines on 7 benchmarks
→ Learn more...

💻 Repos

cosmicstack-labs/mercury-agent ⭐
#AIAgents #TelegramBots #ToolUse
TypeScript AI agent ⟶ CLI Telegram with approval actions
→ Learn more...

yzhao062/agent-style ⭐
#PromptEngineering #TechnicalWriting #AIAgents
Writing ruleset CLI ⟶ reduces style violations
→ Learn more...

📊 Datasets

Nemotron-Personas-Korea
#SyntheticData #Personas #Korean
Korean persona dataset ⟶ trains and evaluates persona models
→ Learn more...

⚡ Trends

▸ Agent tooling emphasizes explicit workflows, permissions, and human approval safeguards.
▸ Recent LLM agent systems package reusable controls for execution and output quality.
▸ Synthetic, structured persona data is expanding beyond English into localized demographics.

🧭 TL;DR

📄 AgentSPEX: An Agent SPecification and EXecution Language
Declarative agent language improves benchmark performance and interpretability with executable workflows.

⭐ cosmicstack-labs/mercury-agent ⭐
Practical permission-aware agent offers approvals, daemon mode, Telegram, and 31 tools.

💡 LLM tooling is shifting toward controllable, operationally safe agent systems.

via @Papers.Data.Code

📄 Paper #Paper #Robotics #HumanoidControl #WorldModels

UniT: Toward a Unified Physical Language for Human-to-Humanoid Policy Learning and World Modeling
👤 Boyu Chen, Yi Chen, Lu Qiu et al.

🎯 Task
Human-to-humanoid policy learning and world modeling

💡 Idea
Tri-branch visual-action-fusion tokenizer with cross-reconstruction maps heterogeneous actions into shared discrete tokens; actions predict vision and vision reconstructs actions to capture embodiment-agnostic physical intent.

✨ Why it's interesting
Achieves SOTA data efficiency, robust OOD generalization, and zero-shot task transfer on sim and real humanoids.

💻 Repo
⭐ xpeng-robotics/UniT — 37 stars

🔗 paper

via @Papers.Data.Code

📊 Dataset #Dataset #3DEditing #3DEditing #InstructionFollowing

H³D: High-quality Holistic 3D Editing Dataset
👤 ART-3D

🎯 Task
Instruction-following 3D editing

💡 Idea
~102.7K records across 10 shards: each sample has before/after 3D SLAT latents, one aligned 518×518 RGB view per side, and an edit prompt for deletion, addition, modification, scale, material, color, or global style.

✨ Why it's interesting
Paired latent+image edits across 7 types enable training and evaluation of part-level 3D editors.

Size: 102,704 records across 10 shards, ~54.5 GB total

Downloads: 441 | Likes: 10

🔗 dataset

via @Papers.Data.Code

🔥 Repo #Repo #MixtureOfExperts #MixtureOfExperts #Quantization

Tile Kernels
👤 deepseek-ai

🎯 Task
LLM GPU kernel optimization

💡 Idea
Optimized TileLang kernels for LLM ops, including top-k MoE gating/routing, FP8/FP4/E5M6 quantization, batched transpose, Engram, and Manifold HyperConnection, plus trainable torch.autograd.Function wrappers for higher-level layers.

✨ Why it's interesting
Authors say most kernels approach hardware limits for compute intensity and memory bandwidth.

💻 Repo
⭐ deepseek-ai/TileKernels — 1.2k stars (+1.1k 3d)
Python


via @Papers.Data.Code

📄 Paper #Paper #CV #NovelViewSynthesis #VideoDiffusion

Vista4D: Video Reshooting with 4D Point Clouds
👤 Kuan Heng Lin, Zhizheng Liu, Pablo Salamanca et al.

🎯 Task
Video reshooting

💡 Idea
4D-grounded point clouds with temporally persistent static pixels guide a video diffusion model, plus training on noisy reconstructed multiview data to preserve seen content and improve camera control under real-world artifacts.

✨ Why it's interesting
Best camera/3D consistency; user study wins 67.06% preservation, 68.17% camera, 77.38% fidelity.

💻 Repo
⭐ Eyeline-Labs/Vista4D — 88 stars

🔗 paper

via @Papers.Data.Code

📄 Paper #Paper #LLM #TimeSeries #VisionLanguageModels

LLaTiSA: Towards Difficulty-Stratified Time Series Reasoning from Visual Perception to Semantics
👤 Yueyang Ding, HaoPeng Zhang, Rui Dai et al.

🎯 Task
Time series reasoning

💡 Idea
Dual-view VLM input uses a time-series plot plus an index-value table for precise numerical grounding, then curriculum fine-tunes across L1-L3 reasoning levels on the 83k-sample HiTSR dataset.

✨ Why it's interesting
Best OOD results: 86.8% L1, 75.6% local L2, 97.5% global L2, 67.0% L3 accuracy.

💻 Repo
⭐ RainingNovember/LLaTiSA — 76 stars

🔗 paper

via @Papers.Data.Code

📊 Dataset #Dataset #Classification #Classification #Regression

Sleep Health & Daily Performance Dataset
👤 mohankrishnathalla

🎯 Task
Sleep health prediction

💡 Idea
100K records, 32 columns, and 3 targets spanning regression, multiclass, and binary tasks. Structured daily snapshots cover sleep metrics, behaviors, mental state, cognitive outcomes, 12 occupations, and 15 countries with no missing values.

✨ Why it's interesting
100K rows + 3 targets enable benchmarkable sleep, risk, and cognition models from beginner to expert level.

Size: 100K records, 32 columns, 14.3 MB

Downloads: 2.9k | Likes: 49

🔗 dataset

via @Papers.Data.Code

📄 Paper #Paper #LLM #AgenticRL #ToolUse

DR-Venus: Towards Frontier Edge-Scale Deep Research Agents with Only 10K Open Data
👤 Venus Team, Sunhao Dai, Yong Deng et al.

🎯 Task
Edge-scale deep research agents

💡 Idea
Cleaned and resampled long-horizon trajectories plus IGPO-based turn-level RL with information-gain rewards and format penalties train a 4B agent from ~10K open data, targeting dense credit assignment for long research runs.

✨ Why it's interesting
Beats prior agentic models under 9B on multiple deep research benchmarks.

💻 Repo
⭐ inclusionAI/DR-Venus — 50 stars
⭐ verl-project/verl — 50 stars

🔗 paper 🔗 dataset 🔗 dataset

via @Papers.Data.Code

💻 Repo #Repo #CV #FaceVerification #Webassembly

Face X
👤 facex-engine

🎯 Task
Face verification

💡 Idea
Local face embedding engine for browser, C, Go, Python, and CLI. It computes 512-d embeddings and cosine similarity, with no dependencies, optional encrypted weights, and SIMD-optimized CPU inference.

✨ Why it's interesting
Claims 3.0 ms native latency, 99.73% LFW accuracy, and 1.30x faster inference than ONNX Runtime.

💻 Repo
⭐ facex-engine/facex — 82 stars (+82 3d)
C

🔗 paper

via @Papers.Data.Code

📄 Paper #Paper #CV #TextToVideo #ReinforcementLearning

World-R1: Reinforcing 3D Constraints for Text-to-Video Generation
👤 Weijie Wang, Xiaoxuan He, Youping Gu et al.

🎯 Task
3D-consistent text-to-video generation

💡 Idea
Flow-GRPO fine-tunes a video model with rewards from 3D reconstruction, meta-view VLM scoring, trajectory alignment, and aesthetics; camera motion is injected by warping latent noise instead of adding control modules.

✨ Why it's interesting
Improves 3D consistency by 10.23 dB and 7.91 dB PSNR while preserving general video quality.

💻 Repo
⭐ microsoft/World-R1 — 197 stars

🔗 paper

via @Papers.Data.Code

📊 Dataset #Dataset #NLP #CompetitionMath #Multimodal

MathNet v0 — Olympiad Math Reasoning & Retrieval
👤 ShadenA

🎯 Task
Olympiad math reasoning and retrieval

💡 Idea
27,817 problems in v0 from 58 country/regional configs, with problem markdown, official solutions, topic paths, language, provenance, and 7,541 inline images; sourced from official booklets across 47 countries and 17 languages.

✨ Why it's interesting
30K-scale multilingual expert data enables hard reasoning, retrieval, and RAG evaluation beyond small English-only math sets.

Size: 27,817 problems, 7,541 images, 58 configs

Downloads: 9.3k | Likes: 26

🔗 dataset 🔗 paper 🔗 repo

via @Papers.Data.Code

💻 Repo #Repo #Cpp #Cpp #Gguf

Llama Cpp Deep Seek V4 Flash
👤 antirez

🎯 Task
Local LLM inference

💡 Idea
DeepSeek v4 Flash support in llama.cpp with generated GGUFs using 2-bit quantization of routed experts, targeting MacBooks with 128GB RAM; works with CPU and Metal backends.

✨ Why it's interesting
Targets 128GB MacBooks for local DSv4 inference; Metal backend is faster than CPU.

💻 Repo
⭐ antirez/llama.cpp-deepseek-v4-flash — 124 stars (+124 3d)
C++

🔗 paper 🔗 paper 🔗 paper

via @Papers.Data.Code

📄 Paper #Paper #CV #MultimodalLearning #ImageGeneration

Tuna-2: Pixel Embeddings Beat Vision Encoders for Multimodal Understanding and Generation
👤 Zhiheng Liu, Weiming Ren, Xiaoke Huang et al.

🎯 Task
Unified multimodal understanding and generation

💡 Idea
Direct patch embeddings replace VAE and representation encoders, so one transformer handles text, images, and pixel-space generation end to end. A masking-based visual feature learning scheme stabilizes training and improves pixel-space representations.

✨ Why it's interesting
At 7B, it reaches SOTA among native UMMs on understanding and stays competitive on generation.

💻 Repo
⭐ facebookresearch/tuna-2 — 139 stars

🔗 paper

via @Papers.Data.Code

📄 Paper #Paper #MultiAgentSystems #MultiAgentSystems #Reasoning

Recursive Multi-Agent Systems
👤 Xiyuan Yang, Jiaru Zou, Rui Pan et al.

🎯 Task
Multi-agent LLM reasoning

💡 Idea
Latent-state recursion across agents via lightweight RecursiveLink modules — agents pass and refine hidden states in a loop, with inner-outer training for whole-system credit assignment instead of text-based coordination.

✨ Why it's interesting
Avg +8.3% accuracy, 1.2-2.4x faster inference, and 34.6-75.6% fewer tokens vs baselines.

💻 Repo
⭐ RecursiveMAS/RecursiveMAS — 30 stars

🔗 paper

via @Papers.Data.Code

📄 Paper #Paper #AudioReasoning #AudioReasoning #Rlhf

Step-Audio-R1.5 Technical Report
👤 Yuxin Zhang, Xiangyu Tony Zhang, Daijiao Liu et al.

🎯 Task
Audio reasoning for multi-turn spoken dialogue

💡 Idea
RLHF with a rubric-guided generated reward model compares responses in multi-turn audio chats, optimizing naturalness, coherence, and instruction retention beyond label-only RLVR.

✨ Why it's interesting
77.97 avg across 8 benchmarks, +5.47 over Step-Audio-R1; 41.15 on Audio MC.

💻 Repo
⭐ stepfun-ai/Step-Audio-R1 — 647 stars

🔗 paper 🔗 dataset

via @Papers.Data.Code

💻 Repo #Repo #TestDrivenDevelopment #TestDrivenDevelopment #CodingAgents

Evan Flow
👤 evanklem

🎯 Task
AI-assisted software development workflow

💡 Idea
Single-entry workflow for Claude Code that orchestrates brainstorm → plan → execute → iterate, with vertical-slice TDD inside coding tasks, optional parallel coder/overseer subagents, and a hook blocking dangerous git commands.

✨ Why it's interesting
Keeps users in control with approval checkpoints, no auto-commits, and blocked destructive git ops.

💻 Repo
⭐ evanklem/evanflow — 356 stars (+356 3d)
Shell

🔗 paper

via @Papers.Data.Code